Artificial intelligence-based exploration data anomaly detection method and device

Abstract

The invention provides an exploration data anomaly detection method and device based on artificial intelligence, relates to the technical field of data processing, and thoroughly gets rid of the constraint of a fixed frequency band on an unsteady logging signal by enabling a frequency band range to be dynamically adjusted along with data complexity through adaptive frequency band division driven by information entropy. Furthermore, high-frequency dynamic features are reserved through normalized logging values, global information is supplemented through window statistics, cooperative information of associated channels is effectively utilized in combination with cross-channel correlation, feature fusion is performed by using frequency band energy vectors, and high-frequency abnormal features are still kept in a remarkable identification degree after fusion. And the suppression by low-frequency backgrounds or statistical characteristics is avoided. In conclusion, the model can accurately distinguish the easily confused exceptions, and the exception recognition precision is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and in particular to an exploration data anomaly detection method and device based on artificial intelligence. BACKGROUND

[0002] In the field of geophysical exploration, well logging data, as an important means to understand the underground geological structure and identify abnormal layer, is widely used in oil and gas exploration, mineral resources exploration and geological disaster monitoring, and its analysis results have a key influence on exploration decision. At present, in the process of processing well logging data, the existing technology usually adopts a method of pre-setting fixed frequency band division to analyze the frequency distribution of data, and relies on a single energy feature as the core basis for anomaly judgment. However, well logging curves generally have complex internal characteristics, such as significant non-stationary characteristics, different geological events often correspond to different frequency distributions, for example, faults usually show high-frequency abrupt signal, while lithology gradually changes and shows low-frequency gradual change characteristics, resulting in dynamic changes of data frequency characteristics with geological events. In addition, the energy of effective signal and noise is easy to overlap, increasing the difficulty of feature extraction. Moreover, there are short-period high-frequency characteristics and long-period low-frequency characteristics in the data, showing the characteristics of coexistence of high and low frequency characteristics.

[0003] The fixed frequency band division method of the existing technology is difficult to adapt to the non-stationary characteristics of well logging data changing with geological events, and cannot accurately capture the real frequency distribution corresponding to different geological events. At the same time, the judgment mode of single energy feature has obvious limitations, which not only makes it difficult to effectively distinguish between effective signal and noise, but also misses the multi-dimensional key information required for anomaly recognition, resulting in the inability to fully depict abnormal characteristics in the scene of coexistence of high and low frequency characteristics. These problems directly affect the accuracy of abnormal layer identification, making it difficult for the existing technology to meet the actual needs of accurate identification of abnormal layer and reliable early warning of high-risk anomalies in the field of geophysical exploration. SUMMARY

[0004] Therefore, the embodiments of the present application provide an exploration data anomaly detection method and device based on artificial intelligence, which can accurately distinguish confusing anomalies and improve the accuracy of anomaly recognition.

[0005] In a first aspect, an embodiment of the present application provides an exploration data anomaly detection method based on artificial intelligence, which comprises: obtaining a formation physical response of a target exploration layer section, recording depth coordinates, instrument states and environmental parameters of the formation physical response, and generating original exploration data of the target exploration layer section; determining an adaptive frequency band energy ratio of the original exploration data according to an information entropy of the original exploration data; determining a frequency band energy vector contained in the original exploration data based on the adaptive frequency band energy ratio; the frequency band energy vector comprises a low-frequency vector, a medium-frequency vector and a high-frequency vector; constructing a basic feature matrix of the original exploration data based on a normalized logging value and a window statistic of the original exploration data; performing cross-channel feature fusion on the basic feature matrix by using the frequency band energy vector according to a cross-channel correlation of the basic feature matrix, to generate a fusion feature of the original exploration data; inputting the fusion feature as a to-be-tested feature into a pre-constructed artificial intelligence model, performing anomaly recognition on the to-be-tested feature by using the artificial intelligence model, and determining an anomaly recognition result corresponding to the target exploration layer section; and determining an abnormality category to which the target exploration layer section belongs based on the anomaly recognition result.

[0006] In combination with the first aspect, an embodiment of the present application provides a first implementation manner of the first aspect, wherein the step of determining the abnormality category to which the target exploration layer section belongs based on the anomaly recognition result comprises: determining an abnormality confidence degree corresponding to the anomaly recognition result of each to-be-tested region of the target exploration layer section by using a sliding window mode; and determining the abnormality category to which the target exploration layer section belongs based on a cumulative result of the abnormality confidence degrees.

[0007] In combination with the first aspect, an embodiment of the present application provides a second implementation manner of the first aspect, wherein the step of determining the adaptive frequency band energy ratio of the original exploration data according to the information entropy of the original exploration data comprises: performing wavelet packet decomposition on the original exploration data, to divide the original exploration data into multiple layers of data; each layer of data comprises multiple frequency band nodes; calculating an information entropy of each frequency band node, determining a set of optimal sub-tree nodes from a decomposition tree to which the frequency band node belongs based on the information entropy, calculating a node energy of the current frequency band node according to a wavelet packet coefficient of each frequency band node in the set of optimal sub-tree nodes, 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] With reference to the first aspect, the third implementation of the first aspect is provided, and the step of performing cross-channel feature fusion on the basis feature matrix according to the frequency band energy vector and the cross-channel correlation of the basis feature matrix to generate the fusion feature of the original exploration data comprises: calculating a cross-channel correlation coefficient matrix of the basis feature matrix, and constructing a diagonal matrix of the frequency band energy vector; performing channel weighting on the cross-channel correlation coefficient matrix by using the diagonal matrix to generate a cross-channel feature fusion matrix; and performing selective filtering on the frequency domain feature of the original exploration data, and superimposing the original exploration data after the selective filtering and the cross-channel feature fusion matrix to generate the fusion feature of the original exploration data, which is the to-be-detected feature after the feature fusion.

[0009] With reference to the first aspect, the fourth implementation of the first aspect is provided, and the step of inputting the fusion feature into a pre-constructed artificial intelligence model as the to-be-detected feature, performing abnormality recognition on the to-be-detected feature by using the artificial intelligence model, and determining an abnormality recognition result corresponding to the target exploration layer section comprises: performing adaptive feature selection on the to-be-detected feature based on a preset feature channel attention mechanism to determine an attention output feature matrix of the to-be-detected feature; performing short-period feature extraction and long-period feature extraction on the attention output feature matrix respectively, and performing gated residual connection on the features extracted by the short-period feature extraction and the long-period feature extraction to generate a double-path residual output feature matrix; performing frequency band modulation processing on the double-path residual output feature matrix based on the frequency band energy vector to generate a frequency domain modulation feature matrix; performing attention gate probability conversion on the frequency domain modulation feature matrix to determine an abnormality probability distribution corresponding to the to-be-detected feature; and determining an abnormality recognition result corresponding to the to-be-detected feature according to the abnormality probability distribution.

[0010] With reference to the first aspect, the fifth implementation of the first aspect is provided, and the step of performing adaptive feature selection on the to-be-detected feature based on a preset feature channel attention mechanism to determine an attention output feature matrix of the to-be-detected feature comprises: inputting the to-be-detected feature into a single-hidden layer perceptron, performing linear activation processing on the to-be-detected feature based on a weight matrix of a hidden layer of the single-hidden layer perceptron; performing feature channel normalization on the to-be-detected feature after the linear activation processing to generate an attention weight matrix corresponding to the to-be-detected feature; and performing adaptive weighting on the to-be-detected feature by using the attention weight matrix to generate the attention output feature matrix corresponding to the to-be-detected feature.

[0011] In combination with the first aspect, an embodiment of the present invention provides a sixth implementation of the first aspect, wherein the step of performing short-period feature extraction on the attention output feature matrix includes: applying a depthwise separable convolution operator to perform short-period feature extraction on the attention output feature matrix, and outputting the 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 step size to perform feature extraction; the step of performing long-period feature extraction on the attention output feature matrix includes: performing standard convolution operation and maximum pooling operation on the attention output feature matrix in sequence, performing feature extraction on the attention output feature matrix, and outputting the long-period feature output matrix of the attention output feature matrix.

[0012] In combination with the first aspect, an embodiment of the present invention provides a seventh implementation method of the first aspect, wherein the method also includes: obtaining a pre-constructed training sample set; using a small batch gradient descent strategy to input the training sample set into the artificial intelligence model for model training; calculating the training gradient through a pre-constructed joint loss function, and using a preset adaptive moment estimation optimizer to dynamically adjust the learning rate of the artificial intelligence model to update the network parameters of the artificial intelligence model.

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

[0014] In a second aspect, the embodiments of the present application also provide an artificial intelligence-based exploration data anomaly detection device, which comprises: a data acquisition module, configured to acquire a formation physical response of a target exploration layer section, record depth coordinates, instrument states and environmental parameters of the formation physical response, and generate original exploration data of the target exploration layer section; a data processing module, configured to determine an adaptive frequency band energy ratio of the original exploration data according to information entropy of the original exploration data; determine a frequency band energy vector contained in the original exploration data based on the adaptive frequency band energy ratio; the frequency band energy vector comprises a low-frequency vector, a medium-frequency vector and a high-frequency vector; a feature fusion module, configured to construct a basic feature matrix for the original exploration data based on normalized logging values and window statistics of the original exploration data; perform cross-channel feature fusion on the basic feature matrix by using the frequency band energy vector to generate fusion features of the original exploration data according to cross-channel correlation of the basic feature matrix; an execution module, configured to input the fusion features as to-be-detected features into a pre-constructed artificial intelligence model, perform anomaly identification on the to-be-detected features by using the artificial intelligence model, and determine an anomaly identification result corresponding to the target exploration layer section; and an output module, configured to determine an anomaly category to which the target exploration layer section belongs based on the anomaly identification result.

[0015] The embodiments of the present application provide an artificial intelligence-based exploration data anomaly detection method and device. Through adaptive frequency band division driven by information entropy, the frequency band range is dynamically adjusted according to the data complexity, and the fixed frequency band is completely freed from the constraint of non-steady-state logging signals. Further, the high-frequency dynamic characteristics are retained by using normalized logging values, the global information is supplemented by using window statistics, the cross-channel correlation is used to effectively utilize the cooperative information of the associated channels, and the frequency band energy vector is used for feature fusion, so that the high-frequency abnormal features still maintain significant recognition after fusion and are not suppressed by low-frequency background or statistical characteristics. In summary, the model can accurately distinguish between confusing anomalies and improve the anomaly identification accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to more clearly illustrate the specific embodiments of the present application or the technical solutions in the prior art, the drawings needed in the specific embodiments or prior art description will be briefly introduced as follows. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0017] Figure 1 A flowchart of an artificial intelligence-based exploration data anomaly detection method provided by the embodiments of the present application is provided. Figure 2 A flowchart of another artificial intelligence-based exploration data anomaly detection method provided by the embodiments of the present application is provided. Figure 3A schematic diagram of the decomposition effect of a frequency band decomposition method provided by an embodiment of the present invention; Figure 4 A schematic diagram comparing the anomaly detection performance of different methods provided by an embodiment of the present invention; Figure 5 A schematic diagram showing a performance comparison of different feature fusion methods provided by an embodiment of the present invention; Figure 6 A schematic diagram of frequency domain energy vector distribution of different geological events provided by an embodiment of the present invention; Figure 7 A schematic diagram of the structure of an artificial intelligence-based exploration data anomaly detection device provided by an embodiment of the present invention; Figure 8 A schematic structural diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0018] The following will clearly and completely describe the technical solutions of the present invention in conjunction with the embodiments. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0019] The embodiments of the present invention provide an artificial intelligence-based exploration data anomaly detection method and device, which can ensure that abnormal information is not concealed, and can take into account key cross-channel, time-frequency domain, and geology-engineering correlations, ensure that features are not fragmented and information is not isolated, and achieve a high capture rate, low misjudgment rate, and accurate classification rate of geological anomalies.

[0020] For ease of understanding, the following is first described about an artificial intelligence-based exploration data anomaly detection method provided by an embodiment of the present invention. Figure 1 , the method comprises the following steps: Step S102 : obtaining the formation physical response of the target exploration layer, recording the depth coordinates, instrument status and environmental parameters of the formation physical response, and generating original exploration data of the target exploration layer.

[0021] The present invention uses a multi-channel logging instrument system to collect original exploration data, specifically including five types of logging curves: acoustic wave transit time, gamma ray, resistivity, neutron porosity and density. The data acquisition depth covers the target exploration layer, such as 0-1500 meters, and continuously records the physical response of the formation at a fixed sampling interval, such as 0.1 meter / point, to form an initial logging data set. The data acquisition process synchronously records the depth coordinates, instrument status and environmental parameters to ensure the temporal and spatial consistency of the data.

[0022] Step S104, according to the information entropy of the original exploration data, an adaptive frequency band energy ratio of the original exploration data is determined; and based on the adaptive frequency band energy ratio, a frequency band energy vector contained by the original exploration data is determined.

[0023] The frequency band energy vector contains a low frequency vector, a medium frequency vector and a high frequency vector. The conventional wavelet decomposition or Fourier method adopts a fixed frequency band, which cannot adapt to non-stationary signals, resulting in insufficient detection accuracy of transient anomalies such as faults. The embodiment of the present application screens sensitive frequency bands of the original exploration data based on information entropy, and further calculates the energy ratio of each frequency band; based on the energy ratio, the contribution of different frequency bands to the total energy of the signal is determined to highlight the importance of high-entropy frequency bands, so that the determined frequency band energy vector not only reflects the energy distribution of each frequency band, but also highlights high-uncertainty regions based on information entropy, effectively capturing transient anomalies such as faults.

[0024] Step S106, based on the normalized logging values and window statistics of the original exploration data, a basic feature matrix of the original exploration data is constructed; and based on the cross-channel correlation of the basic feature matrix, cross-channel feature fusion is performed on the basic feature matrix by using the frequency band energy vector to generate a fusion feature of the original exploration data.

[0025] The basic feature matrix is constructed by the normalized logging values and the window statistics, which can preserve the original feature information of the data to a certain extent, including high-frequency transient features. The window statistics can capture the statistical characteristics of the data in different time windows, and the normalization processing makes different logging values comparable, avoiding the situation that some features are ignored due to the difference in value size, thereby reducing the possibility that high-frequency transient features are covered by statistical features. The cross-channel feature fusion is performed on the basic feature matrix by using the frequency band energy vector, which can fully exploit the correlation information between channels. The frequency band energy vector reflects the energy distribution of different frequency bands in each channel, and through the frequency band energy vector, it can be determined which channels have strong correlation in a specific frequency band, and then the correlation information is fused, so that the fused features can more comprehensively reflect the true situation of the data, improving the representation ability of complex information in the logging data. The embodiment of the present application not only considers the features of the data in a single channel, but also comprehensively utilizes the correlation information across channels, so that the generated fusion features are more accurate and reliable. In practical applications, it can more accurately identify the changes of geological structures, abnormal information of formation parameters and the like, providing stronger support for geological exploration and resource evaluation.

[0026] Step S108, the fusion feature is input into a pre-constructed artificial intelligence model as a to-be-measured feature, and the artificial intelligence model is used to perform anomaly recognition on the to-be-measured feature to determine an anomaly recognition result corresponding to the target exploration layer section.

[0027] Step S110, based on the abnormality recognition result, determine the abnormality category to which the target exploration layer section belongs.

[0028] The above data processing process converts the original exploration data into model recognizable feature correlation, provides multi-dimensional basis for lithology gradual change and fluid-bearing layer determination, and can also avoid the model preferentially learning large value features (such as high resistance layers) due to feature magnitude differences, and ignoring small value but key abnormal signals (such as small drift caused by instrument failure). The fusion feature artificial intelligence model provides a complete information view that contains both local transients and global statistics, and integrates both single-channel features and cross-channel coupling.

[0029] Finally, the probability distribution vector of the abnormality category is output, which includes four types of abnormalities, i.e., fault, fluid-bearing layer, lithology gradual change, and instrument failure. The embodiment of the present application provides clear category guidance under the premise of preserving uncertainty information, and the cooperative decision of model recognition and expert artificial verification provides flexible support, significantly improving the accuracy and reliability of abnormal interpretation of the target exploration layer section.

[0030] To sum up, the embodiment of the present application reflects the data complexity based on the information entropy of the original exploration data of the target exploration layer section, performs adaptive frequency band division, dynamically matches the data fluctuation characteristics, and provides accurate frequency band range. Further, the high-frequency dynamic characteristics are preserved by normalizing the logging values, the global information is supplemented by window statistics, the correlation information of associated channels is effectively utilized by combining cross-channel correlation, and the feature fusion is performed by using the frequency band energy vector, so that the high-frequency abnormal features still maintain significant recognition after fusion and are not suppressed by low-frequency background or statistical characteristics. The model can accurately distinguish between confusing abnormalities and improve the abnormality recognition accuracy.

[0031] On the basis of the above embodiment, the embodiment of the present application further provides another artificial intelligence-based exploration data anomaly detection method, Figure 2 The flowchart of another artificial intelligence-based exploration data anomaly detection method provided by the embodiment of the present application is shown, and the details are described with reference to Figure 2 The method comprises the following steps: Step S202, obtaining the formation physical response of the target exploration layer section, recording the depth coordinates, instrument state and environmental parameters of the formation physical response, and generating the original exploration data of the target exploration layer section.

[0032] Step S204, determining the adaptive frequency band energy ratio of the original exploration data according to the information entropy of the original exploration data.

[0033] Original frequency domain analysis methods do not differentiate and process the specific frequency bands corresponding to different geological events. For example, faults typically correspond to high-frequency signals, while lithologic gradients correspond to low-frequency signals. Conventional processing uses a fixed frequency band partitioning strategy, which is difficult to adapt to the non-steady-state characteristics of logging signals, resulting in missed detection of transient anomaly information. The present invention dynamically partitions wavelet packet subtrees based on information entropy and extracts adaptive frequency band energy ratios. In specific implementation, embodiments of the present invention use wavelet packet decomposition to partition raw exploration data into frequency bands. This method possesses unique features such as dynamic frequency band partitioning and integration of geological significance, overcoming the limitations of conventional fixed frequency band analysis and providing adaptive, highly targeted frequency domain energy information. Furthermore, by determining information entropy based on the statistical characteristics of the wavelet packet coefficients of the subnodes, the complexity / uncertainty of the coefficient distribution of each subnode after wavelet packet decomposition is measured, enabling precise screening of geologically significant signals and suppressing noise interference. After further determining the optimal subtree node set, the corresponding adaptive frequency band energy ratio is determined. This can combine the physical energy of the subnodes with their geological significance to determine the importance of the frequency band for anomaly detection. Specifically, embodiments of the present invention include the following steps: 1) Perform wavelet packet decomposition on the original exploration data and divide it into multiple layers of data, where each layer of data contains multiple frequency band nodes.

[0034] Apply the wavelet packet decomposition algorithm to the original logging signal to perform J-layer decomposition, and obtain the wavelet packet coefficient vector of each sub-node to store the time-frequency details of the signal in a specific frequency band. Layer decomposition The wavelet packet coefficients of the child nodes are , which contains the detailed information of the signal in a specific frequency band, and defines It is Tier The coefficient vector of child nodes stores the time-frequency coefficients of the frequency band. represents the layer index of wavelet packet decomposition, ; Indicates the maximum number of layers of wavelet packet decomposition, for example, when the signal length is greater than When, usually set , ensuring moderate frequency band resolution; Represents the child node index, each layer of decomposition produces child nodes, The corresponding lowest level is child nodes, .

[0035] It should be noted that a child node is a frequency band node in the wavelet packet decomposition tree, which contains a set of coefficient vectors representing the detailed information of the signal in a specific frequency band, not a single value.

[0036] 2) Calculate the information entropy of each band node and determine the optimal subtree node set from the decomposition tree to which the band node belongs based on the information entropy.

[0037] Calculate the information entropy of each child node and measure the complexity of its wavelet packet coefficient distribution. Then, select the optimal subtree node set from all possible subtrees by maximizing the entropy value and determine the optimal subtree containing the key frequency band node, which is expressed as:

[0038] Where, Indicates the The information entropy of a child node measures the complexity or uncertainty of the coefficient distribution of the node; Indicates the The first The probability of an interval is obtained through histogram statistics. Specifically, The wavelet packet coefficient values ​​of the child nodes are divided into interval, calculate the The frequency of the coefficients in the interval ,but ; b is the interval index of the histogram, , corresponding to the binning of coefficient values; For logarithmic functions, the default base is a natural constant; Indicates the number of intervals divided by the histogram, that is, the number of bins. The number of histogram intervals affects the accuracy of entropy calculation. When the coefficient value range is large, it is usually set , when computing resources are limited, it is usually set . It means to find the subtree that maximizes the objective function value in the set of all possible subtrees of the wavelet packet decomposition tree; Represents a subtree in the wavelet packet decomposition tree, which is the A subset of child nodes; Represents the set of nodes contained in the optimal subtree.

[0039] 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.

[0040] 3) Calculate the node energy of the current frequency band node based on the wavelet packet coefficient of each frequency band node in the optimal subtree node set.

[0041] 4) Based on the node energy and information entropy, the adaptive frequency band energy ratio of the original exploration data for each frequency band node is determined.

[0042] For each child node in the optimal subtree, the square sum of the modulus of its wavelet packet coefficients is calculated as the energy, and then the normalized adaptive energy ratio is calculated in combination with the information entropy of the node to measure the importance of the frequency band for anomaly detection, which is expressed as:

[0043] Where, Indicates the first The energy of a child node is calculated as the sum of the squares of the moduli of all wavelet packet coefficients of the node; Represents the sequence index of the wavelet packet coefficients in the node, ; Indicates the The length of the wavelet packet coefficient of the child node is determined by the original signal length And the maximum number of layers J of wavelet packet decomposition is determined by the formula: ,like, , hour, ; Indicates the Layer decomposition The first child node wavelet packet coefficient values; Indicates the square of the absolute value, which means the modulo square operation for vector calculations; Indicates the The adaptive energy ratio of the child nodes. The larger the value, the more important the frequency band is for anomaly detection. Indicates the first The energy of each child node; The set of nodes contained in the optimal subtree The node index in .

[0044] It should be noted that geological anomalies such as faults often present low entropy values ​​in specific frequency bands and high signal purity. The project uses the inverse logarithm of entropy as a weight amplifier to nonlinearly enhance the energy ratio of low-entropy frequency bands, which can amplify the contribution of transient events such as faults, while suppressing high-entropy noise frequency bands, such as the background noise mixed in lithologic gradient zones. By combining information entropy, the energy ratio carries both physical intensity and geological significance, significantly improving the anomaly sensitivity without increasing computational complexity.

[0045] Step S206 : determining the frequency band energy vector included in the original exploration data based on the adaptive frequency band energy ratio.

[0046] Among them, the frequency domain characteristics of geological events can be based on the prior (such as faults corresponding to high frequencies, lithology gradients corresponding to low frequencies), and the central frequency of the child node can be used to calculate the frequency domain characteristics of geological events. For the sake of basis, all the sub-nodes in the optimal sub-tree M are divided into three non-overlapping frequency band sets, and further, the adaptive frequency band energy ratios of all the sub-nodes in the three divided low / middle / high frequency sets are summed up, and the three summation results are taken as the three dimensions of a vector, and finally a frequency band energy vector is formed.

[0047] In a specific implementation, represents a low-frequency sub-node set, and the center frequency satisfies . represents a middle-frequency sub-node set, and the center frequency satisfies . represents a high-frequency sub-node set, and the center frequency satisfies . represents the center frequency of the frequency band represented by the sub-node; represents a low-frequency and middle-frequency division threshold, such as . represents a middle-frequency and high-frequency division threshold, such as . represents a vector construction operation, which combines three scalars into a three-dimensional vector.

[0048] Then, the adaptive energy ratios of each set are aggregated to form a frequency band energy vector containing three dimensions of low frequency, middle frequency, and high frequency, which is represented as:

[0049] In the formula, represents a frequency band energy vector containing three dimensions of aggregated energy ratios of low frequency, middle frequency, and high frequency. It should be noted that in the process of generating the frequency band energy vector based on the node set dynamically selected by the optimal sub-tree, based on the frequency band distribution characteristics of geological events, such as faults corresponding to high frequencies and lithology gradual changes corresponding to low frequencies, the adaptive classification is performed, and the original energy is used as the aggregation object, so that the vector fuses the physical energy of the frequency band and the geological importance weight. When the energy of a certain frequency band is high but the entropy value is also high, such as instrument noise, the value of the adaptive energy ratio of the frequency band is suppressed to avoid interference with the vector representation. The double filtering mechanism makes the frequency band energy vector more purely represent the frequency band energy distribution with geological significance.

[0050] In one embodiment, the embodiment of the present application also analyzes the difference in the effect of adaptive frequency band decomposition and conventional fixed frequency band decomposition in processing the logging signal. Figure 3 A decomposition effect diagram of a frequency band decomposition method is shown in Figure 3The four subgraphs include: the original logging signal display depth 0-1000m range comprehensive signal characteristics, about 450m mark the fault position and 700m lithology gradual change starting point; the conventional fixed frequency band decomposition subgraph shows that the low frequency, medium frequency and high frequency components in the fault area respond weakly, and false high frequency fluctuations are generated in the lithology gradual change area; the technical decomposition subgraph of the application significantly enhances the high frequency response of the fault area, while maintaining a smooth low frequency feature in the lithology gradual change area; the high frequency component contrast subgraph directly proves that the application produces a more sharp and more significant high frequency peak near the fault position of 450m, and maintains a lower background noise in the non-anomaly area, the experimental results show that the dynamic frequency band division mechanism based on information entropy of the application can adaptively adjust the frequency band for different geological events, and significantly improve the transient anomaly detection capability.

[0051] Step S208, based on the normalized logging value and window statistics of the original exploration data, the original exploration data is constructed into a basic feature matrix.

[0052] The basic feature matrix is constructed by integrating the normalized logging value and the window statistics basic feature, which contains multiple feature dimensions, and the basic feature matrix is defined as , the dimension is , including normalized logging value, window statistics and other basic information. Wherein represents the total number of samples, that is, the total number of logging data sampling points, each sampling point corresponds to a sample, and the window is used to indicate the depth neighborhood range with a single sampling point of the original exploration data as the core and containing a fixed number of adjacent sampling points around it. It should be noted that the basic feature matrix contains 15 feature dimensions, that is, the total number of integrated basic features is 15, which are divided into two categories of normalized logging value and window statistics, wherein the normalized logging value includes normalized acoustic travel time, normalized gamma ray, normalized resistivity, normalized neutron porosity and normalized density; the window statistics includes window mean, window standard deviation, window skewness, window kurtosis, window minimum value, window maximum value, window energy, window first order gradient mean, window second order gradient mean and window autocorrelation coefficient.

[0053] Step S210, according to the cross-channel correlation of the basic feature matrix, the cross-channel feature fusion of the basic feature matrix is carried out by using the frequency band energy vector, and the fusion feature of the original exploration data is generated.

[0054] The logging data contains dynamic numerical value, static statistics and cross-channel correlation feature, the conventional feature splicing method ignores the space-time coupling relationship between the features, resulting in that the high frequency transient feature is covered by the statistical feature, and the cross-channel correlation information cannot be effectively utilized. The application adopts a feature interaction matrix to realize dynamic weighted fusion, and determines the fusion feature by constraining the weight distribution through frequency band energy.

[0055] In a specific implementation, the fusion features are determined by the following steps: 1) Calculate the cross-channel correlation coefficient matrix of the base feature matrix, and construct the diagonal matrix of the frequency band energy vector.

[0056] The cross-channel correlation coefficient matrix contains the correlation coefficients calculated between different logging channels. The Pearson correlation coefficient between different logging channels is calculated, such as the correlation coefficient between the acoustic travel time and the gamma ray channel in the sliding window, to form the matrix elements. The matrix dimension is ; represents the number of channel combinations, that is, the number of different logging channel pairs participating in the correlation calculation, which depends on the number of logging channels , such as , corresponding to the acoustic travel time, gamma ray, resistivity, neutron porosity, and density channels, the number of channel combinations is obtained by calculating all possible channel pairs .

[0057] The diagonal weight matrix is used to weight the cross-channel correlation coefficient matrix. The diagonal elements are determined by the frequency energy, which is represented as ; wherein represents the diagonal weight matrix, is a natural constant; represents the frequency band energy vector, which contains the aggregate energy ratio of the low, medium, and high frequency dimensions; represents the operation of constructing a diagonal matrix from a vector; represents the weight adjustment factor, which is used to control the influence strength of the frequency energy on the weight. The typical value is , which is greater than 1, indicating that the value is too large, which can easily lead to excessive weight bias towards high frequency. When the value is less than 0.1, it indicates that the value is too small, which can easily lead to the weakening of dynamic weighting. is the frequency band index, which is taken from the energy information of different frequency bands, including low, medium, and high frequency; represents the sum of the frequency energy vector, that is, the sum of the low, medium, and high frequency energy vectors.

[0058] 2) Use the diagonal matrix to weight the cross-channel correlation coefficient matrix to generate the cross-channel feature fusion matrix.

[0059] Apply the weight matrix determined by the frequency energy vector to the cross-channel correlation coefficient matrix, and then concatenate it with the base feature matrix to form the cross-channel feature fusion matrix, which is represented as:

[0060] In the formula, denotes a cross-channel feature fusion matrix, which is a concatenation of the base feature matrix and the weighted cross-channel correlation coefficient matrix; denotes a cross-channel correlation coefficient matrix; denotes a diagonal weight matrix; denotes a concatenation operation; denotes a Hadamard product, i.e., an element-wise multiplication of matrices.

[0061] 3) Selective filtering is performed on the frequency domain features of the original exploration data, and the original exploration data after selective filtering is superimposed with the cross-channel feature fusion matrix to generate a fusion feature of the original exploration data, i.e., a feature to be measured after feature fusion.

[0062] In a specific implementation, the gradient feature matrix is converted to the frequency domain, selective filtering is performed by applying a frequency band mask vector, and then inverse Fourier transform is performed to return to the time domain, which is superimposed with the cross-channel feature fusion matrix to form a fusion feature matrix, denoted as:

[0063] In the formula, denotes a fusion feature matrix, which contains a base feature, a cross-channel weighted feature, and a reconstructed gradient feature, and is used to enhance high-frequency transient information; denotes a gradient feature matrix, which contains first-order gradient information of the original logging data, the first-order gradient information of the original logging data is a rate of change of logging values with depth, and is calculated as a difference value between adjacent sampling points. The gradient feature matrix stores gradient values of all sampling points. denotes a Fourier transform operator, which converts a time domain signal to a frequency domain; denotes an inverse Fourier transform operator, which converts a frequency domain signal back to a time domain; denotes a frequency band mask vector, which is used for selective filtering of the gradient feature in the frequency domain, and is defined as ; denotes a weight coefficient of a low-frequency component, which controls the contribution of low-frequency gradient information in reconstruction, and is set based on geological priori, such as , to enhance low-frequency signals of gradual lithology changes; denotes a weight coefficient of a medium-frequency component, which controls the contribution of medium-frequency gradient information in reconstruction, such as a default weight of a medium-frequency, which is set as ; denotes a weight coefficient of a high-frequency component, which controls the contribution of high-frequency gradient information in reconstruction, such as , to enhance high-frequency signals of faults.

[0064] By defining a frequency band mask vector , frequency band selective enhancement is realized, and a weight coefficient The low-frequency gradient related to the lithology gradual change is set with a weight coefficient Meanwhile, The item represents the gradient feature after inverse Fourier transform, only the geological effective frequency band information is reserved, and then the cross-channel feature fusion matrix is fused The superposition can realize the balance of "high-frequency transient enhancement" and "low-frequency trend fidelity", and solve the problem that the real anomaly is submerged by high-frequency noise in the conventional method.

[0065] Further, Figure 4 An abnormality detection performance comparison diagram of different methods is shown, which is used for comparing the detection accuracy of feature fusion and different data preprocessing methods. In an embodiment, different method abnormality detection performance comparison can be performed, the detection accuracy of four kinds of processing technologies on four kinds of geological anomalies is quantitatively compared through a heat map, and then the feature fusion method is compared. The horizontal axis is the abnormal type (fault, fluid-containing layer, lithology gradual change, instrument fault), the vertical axis is the processing method (fixed frequency band wavelet, Fourier transform, conventional wavelet packet, the present application), and the color scale from blue to red represents the accuracy from low to high. The most notable feature is that the present application technology is in the row, which presents a full red / yellow tone, which is better than the conventional wavelet packet in the conventional method, indicating that the frequency energy constraint mechanism and the feature fusion strategy of the present application technology effectively enhance the sensitivity of high-frequency transient anomalies, while retaining the integrity of low-frequency gradual change characteristics.

[0066] Further, the embodiment of the present application also provides a performance comparison diagram of different feature fusion methods, which is referred to Figure 5 The influence of different feature fusion methods on the model training effect can be evaluated by comparing and analyzing the feature fusion effect. The performance changes of three methods, conventional feature splicing, static weighted fusion and the present application adaptive fusion, in the training process are compared. The figure contains two subgraphs: the upper graph is the accuracy curve, and the lower graph is the F1 score curve; the horizontal coordinate is the training round (unitless), and the vertical coordinate is the index score (unitless). Each method is represented by a broken line with different marks, and the training fluctuation range is displayed with a translucent error band. From the accuracy curve, it can be seen that the present application method rises rapidly at the beginning of training, and stabilizes at a high level of more than 0.9 after about 30 rounds, and the error band is the narrowest. The static weighted fusion reaches about 0.9 after 40 rounds, and the conventional feature splicing (blue square line) is similar, and the F1 score curve presents the same trend: the present application method finally stabilizes at 09, which is significantly higher than other methods, proving that the present application effectively solves the problem that high-frequency transient features are covered by statistical features through dynamic weighting of feature interaction matrix and selective filtering of frequency band mask. The narrow error band shows that the present application method is more stable and faster in convergence.

[0067] Step S212, based on the preset feature channel attention mechanism, adaptive feature selection is performed on the to-be-tested feature to determine an attention output feature matrix of the to-be-tested feature.

[0068] The sensitivity of different logging channels to geological anomalies is significantly different, the artificial feature screening method has strong subjectivity and cannot adapt to sample specificity, resulting in that key abnormal response features are weakened or ignored. The present application constructs a lightweight attention mechanism to realize sample-level feature channel adaptive optimization, and in specific implementation, the following steps are included: 1) The to-be-tested feature is input into a single hidden layer perceptron, and linear activation processing is performed on the to-be-tested feature based on a weight matrix of the hidden layer of the single hidden layer perceptron.

[0069] 2) The to-be-tested feature subjected to linear activation processing is subjected to feature channel normalization to generate an attention weight matrix corresponding to the to-be-tested feature.

[0070] In one embodiment, the linear activation processing is performed by using a ReLU activation function, which represents a rectified linear unit activation function. In the embodiment of the present application, the fusion feature matrix (i.e., the to-be-tested feature) is input into the single hidden layer perceptron, and the rectified linear unit activation function is used for processing; then, the softmax function is used to generate the attention weight matrix, and the feature channel weight of each sample is normalized, which is represented as:

[0071] In the formula, which represents a rectified linear unit activation function; which represents a weight matrix from the input layer to the hidden layer, and the dimension is . which is the number of hidden units, and is usually valued at . which is the number of feature channels of the fusion feature matrix . Since the basic feature matrix contains 15 feature dimensions, then . which represents a bias vector of the hidden layer, and the dimension is . which represents a weight matrix from the hidden layer to the output layer, and the dimension is . which represents a bias vector of the output layer, and the dimension is . which represents a Softmax normalization exponential function, the item represents normalization according to the sample dimension, and ensures that the sum of all channel weights of each sample is 1. which represents an attention weight matrix, and the dimension is the same as the attention input feature matrix . The element in the i-th row and the j-th column of the attention weight matrix represents the attention weight of the j-th feature channel of the i-th sample. ​ represents the importance weight of the i-th feature channel of the j-th sample, satisfying is the feature channel index, .

[0072] 3) using the attention weight matrix to adaptively weight the to-be-tested feature, to generate an attention output feature matrix corresponding to the to-be-tested feature.

[0073] In the specific implementation, the attention weight matrix is multiplied element by element with the fusion feature matrix (i.e., the to-be-tested feature) to realize adaptive weighting of the feature channel, and an attention output feature matrix is output, denoted as:

[0074] In the formula, represents the attention output feature matrix, and the dimension is It should be noted that the feature selection module is a lightweight single hidden layer perceptron structure, including an input layer, a hidden layer and an output layer. It should be noted that the lightweight single hidden layer perceptron structure realizes sample-level feature channel adaptation, generates sample-specific weights by calculating the attention weight matrix , and defines the importance weight vector of the i-th sample as which dynamically allocates channel importance according to its fusion feature, such as automatically increasing the neutron porosity channel weight of the fluid-containing layer, so that the model can enhance the attention of the transient feature channel of the high-frequency dominant sample according to the implicit information of the frequency band energy vector . The dynamic mechanism enables the model to autonomously learn the association rules of "channel- abnormal type" without explicitly inputting the geological label. Further, with reference to

[0075] , the embodiment of the present application also provides a frequency band energy vector distribution diagram of different geological events, which is used for analyzing the frequency band energy vector distribution and displaying the distribution law of different geological events in the low frequency-mid frequency-high frequency energy space in a three-dimensional space. The X axis represents the low frequency energy (0-1 dimensionless ratio), the Y axis represents the mid frequency energy, and the Z axis represents the high frequency energy. The fault event (red point cloud) is concentrated in the region with high frequency energy greater than 0.5, forming obvious aggregation; the fluid-containing layer (blue point cloud) is mainly distributed in the region with high mid frequency energy; the lithology gradient (green point cloud) is concentrated in the region with high low frequency energy; and the instrument noise (purple point cloud) is randomly scattered in the whole space. Experimental results show that the frequency band energy vector constructed based on the frequency band distribution characteristics of geological events (fault corresponding to high frequency, lithology gradient corresponding to low frequency) has significant representation ability. Figure 6

[0076] ​​​​In step S214, short-period feature extraction and long-period feature extraction are respectively performed on the attention output feature matrix, and the features extracted by the short-period feature extraction and the long-period feature extraction are connected by a gated residual connection to generate a double-channel residual output feature matrix.

[0077] The geological anomaly pattern exists in both the short-period instantaneous value and the long-period trend, and a conventional single volume kernel structure cannot simultaneously consider feature extraction in different time scales, which is prone to cause loss of short-time mutation or long-range trend features. The present application adopts a parallel convolution channel and an adaptive residual connection mechanism to fuse short-period transient features and long-period trend features.

[0078] The short-period feature extraction is as follows: A depth separable convolution operator is applied to the attention output feature matrix, a small convolution kernel time size and a unit step size are used to extract high-frequency transient features, and a short-period feature output matrix is output. The depth separable convolution operator is a lightweight convolution operation for extracting high-frequency transient features from the attention output feature matrix. The depth separable convolution operator separates spatial (or temporal) convolution and channel convolution, reduces the computational complexity while ensuring the accuracy of high-frequency feature extraction, and adapts to the detection requirements of high-frequency transient anomalies such as faults in well logging data. Specifically, the depth separable convolution operator is represented as:

[0079] In the formula, represents the depth separable convolution operator, wherein represents that the convolution kernel time size is 3, represents that the time step size is 1, which realizes lightweight and is suitable for capturing local patterns; is the time size of the convolution kernel; is the step size of the convolution operation; is the short-period feature output matrix, representing high-frequency transient features.

[0080] The long-period feature extraction is as follows: The attention output feature matrix is first subjected to a standard convolution operation to expand the receptive field, and then subjected to a maximum pooling operation to downsample, extract low-frequency trend features, and output a long-period feature output matrix, represented as:

[0081] In the formula, represents the standard convolution operator, wherein represents that the convolution kernel time size is 7, represents that the time step size is 2, which is used to expand the receptive field and capture long-period patterns; represents the maximum pooling operator, wherein represents that the pooling window time size is 5, which further samples and extracts significant features by reducing the resolution to extract coarse-grained features. represents a long-period characteristic output matrix, representing a low-frequency trend characteristic. represents an operator combination symbol, represents a maximum pooling operation after a standard convolution operation; Further, the short-period characteristic output matrix and the long-period characteristic output matrix are added, processed by a rectified linear unit activation function, and then added to the attention output feature matrix filtered by the gated residual function to form a double-path residual output feature matrix, represented as:

[0082] In the formula, represents a gating weight matrix, which is a trainable parameter, used to learn a gating signal; represents a Sigmoid activation function, which compresses the input to the interval (0, 1); represents an element-wise multiplication; represents a gated residual function, The term is used to adaptively gate the attention output feature matrix to retain important information; The double-path residual output feature matrix is obtained.

[0083] It should be noted that the structure of the double-path residual time domain convolution module is a double-path parallel structure. The short-period path uses deep separable convolution to capture transient characteristics, and the long-period path uses standard convolution and pooling to capture trend characteristics. Then, the double-path outputs are fused by a gated residual connection to retain key information.

[0084] It should be further noted that the short-period characteristic output matrix and the long-period characteristic output matrix are added and processed by a ReLU activation function, and then added to the attention output feature matrix filtered by the gated residual function to form a residual connection, which is the double-path residual output feature matrix .

[0085] It should be further noted that The term represents a soft selection of the original feature using a learnable gate, retaining only the important components for the current task, and combining the double-path convolution output to achieve triple feature fusion, i.e., the short-period characteristic output matrix , the long-period characteristic output matrix , and the key original feature filtered by the gate , which is suitable for sampling points with a large depth interval. When the convolution path loses details due to down-sampling, the gating mechanism can supplement the high-frequency abnormal signals in the original feature.

[0086] Step S216 , 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.

[0087] The probability of geological anomalies is affected by the energy distribution of frequency bands, and different channels contribute differently to the discriminant nature of anomalies. Conventional fully connected layers that directly map probabilities weaken this critical information. This paper employs a frequency-domain modulation layer and an attention gating mechanism, combined with an optimal subtree node set, to generate the final predicted probability distribution vector.

[0088] In a specific implementation, the frequency domain modulation fully connected layer of the embodiment of the present invention performs the following steps: The Hadamard product is calculated between the dual-path residual output feature matrix and the projected frequency band energy vector, and then input into the fully connected layer, and the ELU activation function is applied to output the frequency domain modulation feature matrix, which is expressed as:

[0089] Where, Represents the frequency domain modulation feature matrix, with dimension ; is the ELU activation function, used to enhance nonlinearity; is the frequency domain modulation weight matrix, which is a trainable parameter with a dimension of ; is the frequency domain modulation bias vector, which is a trainable parameter with a dimension of ; Output feature matrix for the dual-path residual The number of feature channels; is the frequency band energy projection operator, The three-dimensional frequency band energy vector Mapped to the dual-path residual output feature matrix The same dimension, the mapping method is expressed as : is the frequency band modulation matrix, with dimension , the row vector of the d-th dimension of the frequency band modulation matrix is , control the The sensitivity of the dimensional feature to low-frequency / mid-frequency / high-frequency energy; For the The sensitivity weight of each feature channel to low-frequency energy is a trainable parameter; For the The sensitivity weight of each feature channel to the intermediate frequency energy is a trainable parameter; For the The sensitivity weight of each feature channel to high-frequency energy is a trainable parameter; Copy along the timeline dimensional vector to samples; Represents the frequency band energy vector, which includes the aggregate energy ratio of low frequency, medium frequency and high frequency dimensions; for The transpose of .

[0090] It should be noted that the frequency band energy vector Output after projection by frequency band energy projection operator dimensional vector, but the dual-path residual output feature matrix The dimension is ,need dimensional matrix performs Hadamard product operation by replicating along the time axis dimensional vector to samples, generate dimensional matrix, so that each sample shares the same frequency band energy weight, ensuring that the frequency band energy vector can modulate all sample features and enhance the model's utilization of frequency domain priors.

[0091] It should also be noted that through the trainable frequency band modulation matrix Learn the sensitivity of feature channels to each frequency band, and at the same time, The term causes each characteristic channel to be weightedly modulated by the energy of the corresponding frequency band. When the high-frequency energy is significant, it indicates a fault, and the model automatically enhances the response strength of the high-frequency sensitive channel. The "frequency domain-feature domain" coupling mechanism enables the subsequent fully connected layer to more accurately utilize the frequency domain prior of geological events.

[0092] Step S218: Perform attention gate probability conversion on the frequency domain modulation feature matrix to determine the abnormal probability distribution corresponding to the feature to be tested.

[0093] Step S220: determining the abnormality recognition result corresponding to the feature to be tested according to the abnormality probability distribution.

[0094] The extended attention weight matrix is ​​applied to the frequency domain modulation feature matrix for weighted summation, and then the original probability distribution vector is generated through the Softmax function, which is expressed as:

[0095] Where, For the 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 failures, and lithologic gradients. represents the Softmax normalized exponential function, Item representations are normalized along the category dimension. To expand the attention matrix, a single hidden layer perceptron network is preset, the activation function used is the ReLU activation function, and the input feature is the frequency domain modulation feature matrix , the output feature is the expanded attention matrix , the i-th row of the expanded attention matrix The column elements are , indicating the Sample No. The importance weight of each feature channel. A high weight indicates that the channel is more sensitive to abnormality discrimination of the current sample. is the frequency domain modulation feature matrix The i-th row Column elements are feature elements processed by the frequency domain modulation fully connected layer.

[0096] Furthermore, the embodiment of the present invention also calculates the adaptive temperature coefficient based on the average information entropy of the optimal subtree node set, and then adjusts the original probability distribution to generate a predicted probability distribution vector, which is expressed as:

[0097] Where, For the The predicted probability distribution vector of samples; For the The original probability distribution vector of samples; is the number of categories; is the category index, ; For the Adaptive temperature coefficient of samples, , the calculation method is expressed as ; Indicates the The set of nodes contained in the optimal subtree of the sample The average information entropy on ; is the entropy sensitive factor, such as, , controlling the average information entropy to the adaptive temperature coefficient The intensity of the impact.

[0098] It should be noted that the node set contained in the optimal subtree Dynamically adjust the adaptive temperature coefficient based on the average information entropy , so that the high entropy samples with complex signals correspond to larger adaptive temperature coefficients With smooth probability distribution, low entropy samples with pure signal reduce the adaptive temperature coefficient By sharpening the prediction, the problem of misjudgment in geologically ambiguous areas is solved. For example, in the lithologic gradient zone with high average information entropy, the model outputs a relatively flat probability distribution to avoid overconfident predictions of the "fault / normal" category, while in the clear fault with low average information entropy, it outputs a high-confidence abnormal probability.

[0099] Step S222, based on the abnormality recognition result, determine the abnormality category to which the target exploration layer belongs.

[0100] In combination with the above steps, the embodiment of the present application also determines the abnormality confidence corresponding to the abnormality recognition result of the target exploration layer in each to-be-measured region in a sliding window manner. Based on the cumulative result of the abnormality confidence, the abnormality category to which the target exploration layer belongs is determined. Through the above steps S212-S220, the to-be-measured features are identified, and a probability distribution vector of multiple abnormality categories about the target exploration layer is output, such as the following four categories of abnormalities: fault, fluid-containing layer, lithology gradual change, and instrument failure. These categories can be determined by data labeling on the training sample set, and in an implementation, the data labeling can be realized by manual labeling. Among them, the geological abnormality position can be identified according to the drilling core data, seismic profile and formation dip logging data, and the four categories of labels are labeled: fault, corresponding to the high-frequency mutation signal in the depth interval; fluid-containing layer, fluid response in a specific lithology combination; lithology gradual change, slowly changing zone of physical property parameters; instrument failure, non-geological noise interference. In specific implementation, the logging profile is processed continuously in a sliding window manner (window length 512 sampling points, step length 64 points), and the abnormality confidence in the interval [0, 1] is output for each depth point. When the confidence of a specific abnormality category continuously exceeds the depth threshold value, such as the fault confidence being greater than 0.7 and the continuous depth being greater than 0.5 meters, the abnormal event marker is automatically triggered, and the final abnormality category is determined.

[0101] In summary, the embodiment of the present application proposes an exploration data abnormality detection method based on artificial intelligence, which has the following innovations compared with the prior art: 1. The present application selects the optimal sub-tree by maximizing the entropy value, adaptively determines the frequency band, and effectively enhances the discrimination of high-frequency faults and low-frequency lithology gradual changes. 2. The existing method only uses energy size as a feature, which cannot distinguish meaningful high-energy signals from noise energy, resulting in false abnormalities or missed detection. The present application combines energy and information entropy, nonlinearly amplifies low-entropy high-frequency signals such as transient events such as faults, while suppressing high-entropy noise such as instrument interference, and improves the sensitivity of abnormality detection. 3. Conventional methods are mostly feature splicing or static weighting, ignoring cross-channel correlation and spatio-temporal coupling relationship, and high-frequency transient features are easily masked by statistical features. The present application fuses basic features, cross-channel correlation and frequency domain enhanced gradient features to construct a dynamic weighted interaction matrix, taking into account high-frequency transients and low-frequency trends, and avoiding the masking of single features. 4. Conventional convolutional neural network models mostly rely on single convolution kernel or fixed receptive field, short and long period feature extraction is unbalanced, and abnormal information is easily lost. The present application uses sample-level channel attention to realize channel adaptive weighting, and double-path convolution to extract short-period transient features and long-period trend features at the same time, and a gating residual to retain key original signals.

[0102] Further, the embodiment of the present application also proposes a model training method for the artificial intelligence model. The training sample set constructed in advance is obtained, and the training sample set is input into the artificial intelligence model by using the mini-batch gradient descent strategy to perform model training. In specific implementation, a fixed number of samples such as 256 depth points are input in each batch. In the forward propagation stage, the wavelet packet decomposition with the frequency energy constraint, the multi-source feature adaptive fusion, the channel attention feature selection, the double-path residual convolution, and the joint probability estimation are sequentially performed to generate a prediction probability distribution. Further, in the backward propagation stage, the gradient is calculated based on the joint loss function, and the learning rate is dynamically adjusted by the adaptive moment estimation optimizer to update the network parameters. The adaptive moment estimation optimizer is an adaptive optimizer combining the momentum gradient descent and the adaptive learning rate two classical optimization ideas. In the training process, the performance of the validation set is monitored in real time. When the validation set loss does not decrease or the abnormal classification F1 score fluctuation is less than 0.5% in 10 consecutive training rounds, the early stopping mechanism is triggered to save the optimal model parameters. Based on the gradient calculated by the joint loss function, the update step (learning rate) of each network parameter is dynamically adjusted to ensure that the model can quickly converge to the optimal parameters in the training process and avoid overfitting or underfitting caused by the fixed learning rate, which adapts to the complex training requirements of the multi-module (channel attention, double-path convolution, frequency domain modulation) and multi-parameter in the well logging data anomaly detection model.

[0103] The logging anomaly has the depth continuity feature, and the misjudgment cost of different anomaly types is significantly different. The standard cross-entropy loss function ignores the spatial continuity constraint and cost sensitivity, which easily leads to isolated point misjudgment and high-cost anomaly missing. The present application fuses the depth continuity constraint and dynamic cost weight to construct the loss function, and combines the weighted cross-entropy loss, the depth continuity penalty term and the regularization term to form the joint loss function. In specific implementation, the construction steps of the joint loss function of the embodiment of the present application are as follows: a- According to the JS divergence of the prediction probability distribution of the adjacent samples of the training sample set and the absolute value of the depth difference of the adjacent samples, the depth continuity penalty term of the training sample set is constructed.

[0104] The depth continuity penalty term is constructed: the JS divergence of the prediction probability distribution of the adjacent samples is calculated, the depth attenuation factor is applied in combination with the absolute value of the depth difference to form the depth continuity penalty term, which encourages the consistency of the prediction in the depth direction, and is expressed as:

[0105] In the formula, The depth continuity penalty term is used to punish the case that the prediction probability distribution difference of the adjacent depth points is too large, and encourages the model to have continuity in the depth; The number of samples in the batch training is represented by N; The prediction probability distribution of the model to the i-th sample in the training sample set is represented by 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 failures, and lithologic gradients. Indicates the model The predicted probability distribution vector of samples; Represents JS divergence, which is used to measure the difference between two probability distributions; Indicates the The depth value of each sample; Indicates the The depth value of the samples, Characterizes the absolute value of the depth difference between adjacent samples; Represents the depth attenuation coefficient, which controls the rate at which the continuity constraint weakens as the depth difference increases. The larger the value, the more significant the impact of the depth difference. For example, It should be noted that through The term realizes adaptive penalty strength. When the depth difference between adjacent samples is small, such as , then the probability distribution is forced to be similar, and when the depth difference is large, such as , distribution jumps are allowed. This strategy conforms to the law of geological layer changes, maintains prediction consistency within the homogeneous layer, and tolerates mutations at the layer interface.

[0106] b- Obtain the misjudgment cost matrix corresponding to the training sample set, and calculate the cost-sensitive weight corresponding to the training sample set based on the misjudgment cost matrix.

[0107] According to the true label and the predicted label, the misclassification cost value is obtained from the cost matrix, and the cost-sensitive weight is calculated in combination with the frequency of the category samples, which is expressed as:

[0108] Where, Indicates the The cost-sensitive weight of each sample is used to weight the cross entropy loss term; Indicates that according to The true labels of the samples Hedi The predicted labels of samples Take the corresponding misjudgment cost from the cost matrix, which is The cost matrix, whose row index corresponds to the true category label and the column index corresponds to the predicted category label, is used to define the cost of misjudgment between different categories. For example, the cost matrix is ​​defined as , where the elements If the value of is 0.3, it means the cost of mispredicting a sample whose true category is the second category as the third category, and the diagonal elements indicate that the prediction is correct and there is no cost; represents the total sample size; y i represents the true label of the i-th sample, is a scalar value, and takes values 1, 2, 3, or 4; y i represents the true label of the i-th sample, is a scalar value, and takes values 1, 2, 3, or 4; y i represents the true label of the i-th sample, is a scalar value, and takes values 1, 2, 3, or 4; y i represents the true label of the i-th sample, is a scalar value, and takes values 1, 2, 3, or 4; y i represents the true label of the i-th sample, is a scalar value, and takes values 1, 2, 3, or 4; y i represents the true label of the i-th sample, is a scalar value, and takes values 1, 2, 3, or 4; y i represents the true label of the i-th sample, is a scalar value, and takes values 1, 2, 3, or 4; y i represents the true label of the i-th sample, is a scalar value, and takes values 1, 2, 3, or 4; y i represents the true label of the i-th sample, is a scalar value, and takes values 1, 2, 3, or 4; y i represents the true label of the i-th sample, is a scalar value, and takes values 1, 2, 3, or 4;

[0109] c-Utilize the cost-sensitive weight to weight the cross-entropy loss of the training sample set, combine with the depth continuity penalty term, and construct the joint loss function of the training sample set.

[0110] Combine the weighted cross-entropy loss, the depth continuity penalty term, and the L2 regularization term to form the joint loss function as the model optimization objective, denoted as:

[0111] In the formula, L represents the joint loss function; y i represents the one-hot encoding vector of the true label of the i-th sample; y i represents the one-hot encoding vector of the true label of the i-th sample; y i represents the one-hot encoding vector of the true label of the i-th sample; y i represents the one-hot encoding vector of the true label of the i-th sample; y i represents the one-hot encoding vector of the true label of the i-th sample; y i represents the one-hot encoding vector of the true label of the i-th sample; y i represents the one-hot encoding vector of the true label of the i-th sample; y i represents the one-hot encoding vector of the true label of the i-th sample; y i represents the one-hot encoding vector of the true label of the i-th sample; y i represents the one-hot encoding vector of the true label of the i-th sample; y i represents the one-hot encoding vector of the true label of the i-th sample; y i represents the one-hot encoding vector of the true label of the i-th sample; y i represents the one-hot encoding vector of the true label of the i-th sample; y i represents the one-hot encoding vector of the true label of the i-th sample; expected average thickness of a typical geologic anomaly and sampling interval of the well logging data binding, so that the constraint strength is adaptive to the data sampling density, when enhancing the continuity constraint, and vice versa, thereby endowing the model with geological scale perception ability, so that it maintains optimal performance on well logging data of different sampling densities.

[0112] In summary, the embodiment of the present application constructs a joint loss function based on the entropy-based adaptive temperature calibration probability distribution, combined with the depth continuity penalty and the cost-sensitive weight, which can effectively reduce misjudgment and high-cost anomaly omission.

[0113] Further, on the basis of the above-mentioned embodiment, the embodiment of the present application also provides an exploration data anomaly detection device based on artificial intelligence, Figure 7 shows the structure schematic diagram of the exploration data anomaly detection device based on artificial intelligence provided by the embodiment of the present application, referring to Figure 7 The device comprises: a data acquisition module 100, which is used to acquire the formation physical response of a target exploration layer section, record the depth coordinates, instrument state and environmental parameters of the formation physical response, and generate original exploration data of the target exploration layer section; a data processing module 200, which is used to determine the adaptive frequency band energy ratio of the original exploration data according to 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; the frequency band energy vector comprises a low-frequency vector, a medium-frequency vector and a high-frequency vector; a feature fusion module 300, which is used to construct a basic feature matrix for the original exploration data based on the normalized logging value and window statistics of the original exploration data; perform cross-channel feature fusion on the basic feature matrix by using the frequency band energy vector according to the cross-channel correlation of the basic feature matrix, to generate a fusion feature of the original exploration data; an execution module 400, which is used to input the fusion feature as a to-be-tested feature into a pre-constructed artificial intelligence model, perform anomaly recognition on the to-be-tested feature by using the artificial intelligence model, and determine an anomaly recognition result corresponding to the target exploration layer section; and an output module 500, which is used to determine an anomaly category to which the target exploration layer section belongs based on the anomaly recognition result. The exploration data anomaly detection device based on artificial intelligence provided by the embodiment of the present application has the same technical features as the above-mentioned method embodiment, so it can also solve the same technical problems and achieve the same technical effects.

[0114] The embodiment of the present application also provides an electronic device, which comprises a memory, a processor and a computer program stored in the memory and capable of running on the processor, and the processor implements the above-mentioned Figures 1-2The steps of the method shown. The embodiments of the present application also provide a computer readable storage medium, the computer readable storage medium stores a computer program, the computer program is executed by a processor to perform the steps of the method shown above Figures 1-2 The steps of the method shown. The embodiments of the present application also provide a structural diagram of an electronic device, as shown in Figure 8 The structural diagram of the electronic device is shown, wherein the electronic device comprises a processor 81 and a memory 80, the memory 80 stores computer executable instructions capable of being executed by the processor 81, and the processor 81 executes the computer executable instructions to implement the steps of the method shown above Figures 1-2 Figure 8 In the embodiment shown, the electronic device further comprises a bus 82 and a communication interface 83, wherein the processor 81, the communication interface 83 and the memory 80 are connected through the bus 82. The memory 80 can include a high-speed random access memory (RAM) and can also include a non-volatile memory such as at least one disk memory. The communication connection between the system network element and at least one other network element is realized through at least one communication interface 83 (which can be wired or wireless), and the Internet, a wide area network, a local area network, a metropolitan area network, etc. can be used. The bus 82 can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc., and can also be an AMBA (Advanced Microcontroller Bus Architecture) bus, wherein AMBA defines three buses including an APB (Advanced Peripheral Bus) bus, an AHB (Advanced High-performance Bus) bus and an AXI (Advanced eXtensible Interface) bus. The bus 82 can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 8 In the embodiment shown, only one bidirectional arrow is used to represent the bus, but it does not mean that there is only one bus or only one type of bus.

[0115] ​The processor 81 can be an integrated circuit chip with signal processing capability. In implementation, each step of the above method can be completed by integrated logic circuit of hardware in the processor 81 or by instructions in the form of software. The processor 81 described above can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor. The steps of the method disclosed in combination with the embodiments of the present application can be directly embodied as a hardware decoding processor for execution, or a combination of hardware and software modules in the decoding processor for execution. The software module can be located in a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, or other mature storage medium in the art. The storage medium is located in the storage memory, and the processor 81 reads the information in the storage memory, and combines the hardware to complete the foregoing Figures 1-2 any of the methods described.

[0116] The computer program product of the method and device for detecting exploration data anomalies based on artificial intelligence provided by the embodiments of the present application includes a computer readable storage medium storing program codes, the instructions included in the program codes can be used to execute the method described in the foregoing method embodiments, and the specific implementation can be referred to the method embodiments, which will not be described here. Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the system described above can refer to the corresponding process in the foregoing method embodiments, which will not be described here. In addition, in the description of the embodiments of the present application, unless otherwise explicitly specified and limited, the terms "mounting", "connection", "connecting" should be understood in a broad sense, for example, can be fixedly connected, can also be detachably connected, or integrally connected; can be mechanically connected, can also be electrically connected; can be directly connected, can also be indirectly connected through an intermediate medium, and can be the communication between two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0117] If the functions are realized in the form of software function units and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application or the parts of the present application that essentially contribute to the prior art or the parts of the technical solutions can be embodied in the form of software products. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes. In the description of the present application, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", and the like indicate the orientation or positional relationship shown in the drawings, and are only used for the purpose of describing the present application and simplifying the description, and therefore cannot be understood as indicating or implying that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as limiting the present application. In addition, the terms "first", "second", "third" are only for descriptive purposes, and cannot be understood as indicating or implying relative importance. Finally, it should be noted that: the above embodiments are only specific embodiments of the present application, used to illustrate the technical solutions of the present application, and are not limiting, the protection scope of the present application is not limited thereto, although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand: any person skilled in the art within the technical range disclosed by the present application, can still modify or easily think of changes to the technical solutions recorded in the foregoing embodiments, or make equivalent replacements to part of the technical features; and these modifications, changes or replacements do not make the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. An artificial intelligence-based exploration data anomaly detection method, characterized in that: The method comprises: Acquiring a formation physical response of a target exploration interval, recording the depth coordinates, instrument status, and environmental parameters of the formation physical response, and generating original exploration data of the target exploration interval; Determining an adaptive frequency band energy ratio of the original exploration data according to the information entropy of the original exploration data; determining a frequency band energy vector included in the original exploration data based on the adaptive frequency band energy ratio; the frequency band energy vector includes a low-frequency vector, a medium-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 fusion features of the original exploration data; Inputting the fused features as features to be measured into a pre-built artificial intelligence model, performing anomaly identification on the features to be measured by the artificial intelligence model, and determining an anomaly identification result corresponding to the target exploration layer segment; Based on the anomaly identification result, the anomaly category to which the target exploration layer section belongs is determined.

2. The method according to claim 1, characterized in that The step of determining the anomaly category to which the target exploration layer section belongs based on the anomaly identification result includes: Determine the anomaly confidence corresponding to the anomaly identification results of the target exploration layer in each area to be tested by a sliding window method; Based on the accumulated results of the anomaly confidences, the anomaly category to which the target exploration layer 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 according to the information entropy of the original exploration data comprises: Performing wavelet packet decomposition on the original exploration data to divide the original exploration data into multiple layers of data; each layer of data includes multiple frequency band nodes; Calculating the information entropy of each of the frequency band nodes, and determining an 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 according to the wavelet packet coefficient of each frequency band node in the optimal subtree node set; An adaptive frequency band energy ratio of the original exploration data for each frequency band node is determined based on the node energy and the information entropy.

4. The method according to claim 1, wherein The step of performing cross-channel feature fusion on the basic feature matrix according to the frequency band energy vector and the cross-channel correlation of the basic feature matrix to generate the fusion feature of the original exploration data includes: Calculating a cross-channel correlation coefficient matrix of the basic feature matrix, and constructing a diagonal matrix of the frequency band energy vectors; Performing channel weighting on the cross-channel correlation coefficient matrix using the diagonal matrix to generate a cross-channel feature fusion matrix; The frequency domain features of the original exploration data are selectively filtered, and the original exploration data after selective filtering are superimposed with the cross-channel feature fusion matrix to generate fusion features of the original exploration data and features to be measured after feature fusion.

5. The method according to claim 1, wherein The steps of inputting the fused features as the features to be measured into a pre-built artificial intelligence model, performing anomaly identification on the features to be measured by the artificial intelligence model, and determining an anomaly identification result corresponding to the target exploration layer segment include: Based on a preset feature channel attention mechanism, adaptive feature selection is performed on the feature to be measured to determine the attention output feature matrix of the feature to be measured; Performing short-term feature extraction and long-term feature extraction on the attention output feature matrix respectively, and performing gated residual connection on the features of the short-term feature extraction and the long-term feature extraction 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 gate probability conversion on the frequency domain modulation feature matrix to determine the abnormal probability distribution corresponding to the feature to be tested; According to the abnormal probability distribution, an abnormality recognition result corresponding to the feature to be tested is determined.

6. The method according to claim 5, characterized in that The step of performing adaptive feature selection on the feature to be measured based on a preset feature channel attention mechanism and determining an attention output feature matrix of the feature to be measured includes: Inputting the feature to be measured into a single hidden layer perceptron, and performing linear activation processing on the feature to be measured based on a weight matrix of the single hidden layer perceptron with respect to the hidden layer; Performing feature channel normalization on the feature to be tested after linear activation processing to generate an attention weight matrix corresponding to the feature to be tested; The attention weight matrix is ​​used to adaptively weight the features to be measured to generate an attention output feature matrix corresponding to the features to be measured.

7. The method according to claim 5, characterized in that The step of performing short-term feature extraction on the attention output feature matrix comprises: Applying a depthwise separable convolution operator to perform short-term feature extraction on the attention output feature matrix, and outputting a short-term feature output matrix of the attention output feature matrix; the depthwise separable convolution operator uses a small convolution kernel time size and a unit step size for feature extraction; The step of performing long-term feature extraction on the attention output feature matrix comprises: After performing standard convolution operations and maximum pooling operations on the attention output feature matrix in sequence, feature extraction is performed on the attention output feature matrix, and a long-period feature output matrix of the attention output feature matrix is ​​output.

8. The method according to claim 5, characterized in that The method further comprises: Obtain a pre-built training sample set; Using a small batch gradient descent strategy, the training sample set is input into the artificial intelligence model to perform model training; The training gradient is calculated by a pre-constructed joint loss function, and the learning rate of the artificial intelligence model is dynamically adjusted using a preset adaptive moment estimation optimizer to update the network parameters of the artificial intelligence model.

9. The method according to claim 8, characterized in that The method further comprises: Constructing a depth continuity penalty term of the training sample set according to the JS divergence of the predicted probability distribution of adjacent samples of the training sample set and the absolute value of the depth difference between adjacent samples; Obtaining a misjudgment cost matrix corresponding to the training sample set, and calculating a cost-sensitive weight corresponding to the training sample set based on the misjudgment cost matrix; The cross entropy loss corresponding to the training sample set is weighted by using the cost-sensitive weight, and a joint loss function of the training sample set is constructed in combination with the depth continuity penalty term.

10. An artificial intelligence-based exploration data anomaly detection device, characterized in that: The device comprises: A data acquisition module is used to obtain the formation physical response of the target exploration layer, record the depth coordinates, instrument status and environmental parameters of the formation physical response, and generate original exploration data of the target exploration layer; a data processing module, configured to determine an adaptive frequency band energy ratio of the original exploration data according to the information entropy of the original exploration data; and determine frequency band energy vectors included in the original exploration data based on the adaptive frequency band energy ratio; the frequency band energy vectors including a low-frequency vector, a medium-frequency vector, and a high-frequency vector; a feature fusion module for constructing a basic feature matrix for the original exploration data based on the normalized logging values ​​and window statistics of the original exploration data; and performing 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 fused features of the original exploration data; an execution module, configured to input the fused features as features to be measured into a pre-built artificial intelligence model, perform anomaly identification on the features to be measured through the artificial intelligence model, and determine an anomaly identification result corresponding to the target exploration layer segment; An output module is used to determine the anomaly category to which the target exploration layer segment belongs based on the anomaly identification result.

Images