Salt rock expansion point identification method and electronic equipment

Abstract

The invention provides a salt rock expansion point identification method and electronic equipment, and relates to the technical field of geological engineering monitoring. The method comprises the following steps: carrying out sliding sampling on acoustic emission data in an initial search interval according to a first preset window length to obtain a plurality of initial window data segments; determining an initial expansion point position according to the plurality of initial window data segments; determining a target search interval according to the initial capacity expansion point position; wherein the target search interval is smaller than the initial search interval; performing sliding sampling on the acoustic emission data in the target search interval with a second preset window length to obtain a plurality of target window data segments; wherein the second preset window length is smaller than the first preset window length; and determining a target expansion point position according to the plurality of target window data segments. Acoustic emission data is adopted as a data source and can be used for real-time monitoring; and meanwhile, a two-section identification method of rough search and fine search is adopted, so that accurate identification of the expansion point can be realized.

Description

Technical Field

[0001] This invention relates to the field of geological engineering monitoring technology, and in particular to a method and electronic equipment for identifying salt rock expansion points. Background Technology

[0002] Utilizing salt rock formations for energy storage is a priority development direction for large-scale energy reserves. During the long-term injection and production cycle of underground salt cavern reservoirs, the stress state of the surrounding rock changes due to periodically fluctuating operating pressures. When the stress exceeds the expansion initiation threshold, grain boundary slip and dislocation proliferation induce microcrack initiation, causing the volumetric strain to change from compression to expansion—a phenomenon known as expansion. Expansion not only increases the permeability of the surrounding rock, creating potential gas leakage channels and reducing the reservoir's sealing performance, but its continued development also alters the stress field distribution of the surrounding rock structure, affecting the reservoir's stability. Therefore, accurately identifying and dynamically monitoring salt rock expansion points is crucial for ensuring the long-term safety and extending the service life of underground reservoirs.

[0003] In existing technologies, volumetric strain analysis is typically used to determine the expansion area by identifying the inflection point of the volumetric strain curve. However, this method has significant limitations: first, stress-strain data is difficult to collect directly at the underground reservoir site, making real-time monitoring impossible; second, a single volumetric strain measurement is insufficient to describe the complex characteristics of salt rock under multi-scale, nonlinear loading, and it lacks adaptive processing capabilities for outliers, missing data, and non-stationary fluctuations, easily leading to identification delays, misjudgments, or omissions. Therefore, this method cannot meet the requirements for real-time and accurate monitoring of salt rock expansion points. Summary of the Invention

[0004] This invention provides a method and electronic device for identifying expansion points in salt rock, in order to solve the problem that existing volumetric strain analysis cannot meet the requirements for real-time and accurate monitoring of expansion points in salt rock.

[0005] In a first aspect, embodiments of the present invention provide a method for identifying salt rock expansion points, including: The acoustic emission data within the initial search interval is sampled using a first preset window length to obtain multiple initial window data segments. The initial expansion point location is determined based on multiple initial window data segments; The target search interval is determined based on the initial expansion point location; the target search interval is smaller than the initial search interval. The acoustic emission data within the target search interval is sampled using a second preset window length to obtain multiple target window data segments; wherein, the second preset window length is less than the first preset window length. The location of the target expansion point is determined based on multiple target window data segments.

[0006] In a second aspect, embodiments of the present invention provide an electronic device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the salt rock expansion point identification method as described in the first aspect or any possible implementation of the first aspect.

[0007] This invention provides a method and electronic device for identifying salt rock expansion points. The method includes: sliding sampling of acoustic emission data within an initial search interval using a first preset window length to obtain multiple initial window data segments; determining the location of an initial expansion point based on the multiple initial window data segments; determining a target search interval based on the location of the initial expansion point, wherein the target search interval is smaller than the initial search interval; sliding sampling of acoustic emission data within the target search interval using a second preset window length to obtain multiple target window data segments, wherein the second preset window length is smaller than the first preset window length; and determining the location of a target expansion point based on the multiple target window data segments. This invention uses acoustic emission data as the data source, which can be conveniently and non-destructively collected on-site, effectively solving the problem of on-site data acquisition and meeting the needs of real-time monitoring of expansion points. Furthermore, since acoustic emission data can characterize the internal structural evolution and elastic energy release characteristics of salt rock during loading, reflecting changes in physical properties such as microcrack initiation, expansion, and damage accumulation, it provides a reliable data foundation for the analysis of salt rock expansion points, thereby improving the accuracy of expansion point identification results. At the same time, the two-stage identification method of coarse search and fine search can achieve accurate identification of expansion points. Attached Figure Description

[0008] Figure 1 This is a flowchart illustrating the implementation of a method for identifying salt rock expansion points provided in an embodiment of the present invention; Figure 2 This is a comparison diagram of the target expansion point location and the actual expansion point location provided in the embodiments of the present invention; Figure 3 yes Figure 2 Enlarged view of point A in the middle; Figure 4 This is a schematic diagram of the structure of the salt rock expansion point identification device provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of the electronic device provided in an embodiment of the present invention. Detailed Implementation

[0009] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0010] See Figure 1 The document illustrates a flowchart of a method for identifying salt rock expansion points according to an embodiment of the present invention, which is described in detail below: The above-mentioned method for identifying salt rock expansion points includes: S101: Slide sampling is performed on the acoustic emission data within the initial search interval using the first preset window length to obtain multiple initial window data segments; The initial search interval is the starting range for the analysis of acoustic emission data. It can be determined based on the query time, and the time period corresponding to the query time is the initial search interval.

[0011] It should be noted that under compression conditions, salt rock materials undergo stages such as compaction, elastic deformation, stable crack propagation, and expansion and post-peak failure. The expansion point is the critical stress state (physical state threshold) where the volumetric strain transforms from compression to expansion, marking the transition of cracks from stable propagation to unstable opening, and the damage variable enters an accelerated evolution stage.

[0012] Acoustic emission data is obtained in real time by collecting acoustic emission signals during the loading process of salt rock using acoustic emission monitoring equipment. The acquisition process does not damage the material structure and is considered non-destructive testing. Acoustic emission data includes ring count, energy, amplitude, average frequency, signal strength, and peak frequency. Acoustic emission refers to the physical phenomenon of the instantaneous release of elastic energy caused by irreversible damage evolution processes such as the initiation, expansion, closure, and interparticle slippage of internal microcracks in materials like salt rock during loading. Essentially, it reflects the dynamic response process of damage variables evolution within the salt rock and can be used to determine the degree of damage development and deformation of the salt rock. This invention, through the acquisition of acoustic emission data from salt rock, obtains acoustic emission characteristic parameters characterizing the expansion changes in salt rock, providing a data foundation for subsequent analysis and identification of expansion points in the salt rock.

[0013] In one possible implementation, prior to S101, the method may further include: S106: Preprocess the raw acoustic emission data to obtain acoustic emission data.

[0014] For example, preprocessing may include outlier detection and handling, adaptive smoothing, and standardization.

[0015] I. Outlier Detection and Handling: To reduce the impact of anomalous disturbances on feature statistical distribution and subsequent recognition accuracy, a statistical distribution-driven hybrid strategy is adopted for outlier identification, including methods based on interquartile range (IQR) and 3D model. σ The method combines the percentile method and the criterion method to determine outliers, and corrects outliers by removing or replacing them with the neighborhood mean.

[0016] The interquartile range method is specifically as follows: The interquartile range (ICM) is calculated using the following formula:

[0017]

[0018]

[0019] in, For the original acoustic emission data, the first line, number The value of the column, For the first The interquartile range of a column is used to measure the median distribution range of the data. It is the 25th quantile. It is the 75th percentile.

[0020] like ,or Then determine This is an outlier.

[0021] 3 σ The specific law is as follows: Calculate the original acoustic emission data. The mean and standard deviation of the column are calculated using the following formulas:

[0022]

[0023]

[0024] in, For the first The mean of the column, For the first The total amount of data in the column; For the first The standard deviation of the column, express How many standard deviations is it from the mean?

[0025] like Then determine This is an outlier.

[0026] The percentile method is as follows: like or Then determine This is an outlier.

[0027] in, For the first The lower percentile value of the column, For the first The highest percentile value of the column; here Take 0.01, Take 0.99.

[0028] Outlier handling: The detected outliers can be replaced using the neighborhood substitution method, as shown in the following formula:

[0029] in, for The corresponding replacement value.

[0030] II. Adaptive Smoothing Processing Adaptive smoothing is achieved by selecting either Savitzky-Golay or mean filtering methods based on the characteristic volatility.

[0031] For each column of the raw acoustic emission data, based on its volatility It adaptively selects different smoothing methods.

[0032]

[0033] in, For the first The volatility of the column.

[0034] if > If the fluctuation is considered large, the Savitzky-Golay (SG) filter should be used first; otherwise, the mean filter should be used. The calculation formula is as follows:

[0035] in, For the smoothed data, The threshold value can be set to 0.5. The number of samples on both sides of the center point of the filter window. The filter coefficients are obtained by least squares fitting. For relative position index, This represents the width of the filtering window.

[0036] III. Standardization Processing To unify the dimensions and scales of different features, laying the foundation for subsequent feature extraction, the calculation formula is as follows:

[0037] in, The standardized value. For the first The mean of the column after adaptive smoothing. For the first The standard deviation of the column after adaptive smoothing.

[0038] Standardization unifies the dimensions and scales of different features, eliminating the potential adverse effects of differences in feature units on subsequent data processing. For example, features such as ring counts and energy in acoustic emission data have vastly different dimensions. Without standardization, features with large dimensions might overly dominate the model's learning process during training, leading to insufficient learning of other important features. After standardization, all features are on the same order of magnitude, laying a solid foundation for subsequent feature fusion and extraction, and accurate model training.

[0039] S102: Determine the location of the initial expansion point based on multiple initial window data segments; This application uses a larger window for coarse search to determine the approximate location of the expansion point, i.e., the initial expansion point location.

[0040] In one possible implementation, S102 may include: S1021: Extract the features of each initial window data segment to obtain multiple initial feature vectors; For any initial window of data, multimodal features of acoustic emission data can be extracted from time domain, frequency domain, statistics, shape and dynamic time warping (DTW) to form a multimodal feature vector.

[0041] Multimodal features can include: basic statistical features, time-domain features, frequency-domain features, shape features, and DTW distance features.

[0042] I. Basic Statistical Characteristics

[0043]

[0044]

[0045]

[0046] in, For the first The mean of samples within the window of the column. Indicates the first in the window One sample, For the first The standard deviation of the sample within the window of the column. Indicates the first Ranked in The maximum value of a window. For the first Column skewness.

[0047] II. Temporal Characteristics The time-domain features include the autocorrelation features, difference features, and trend fitting features of each sequence in the sliding window.

[0048] (1) Autocorrelation characteristics

[0049] in, For lag The sample autocorrelation coefficient, Indicates the lag step size.

[0050] (2) Difference characteristics

[0051]

[0052]

[0053]

[0054] in, It is a first-order difference. It is a second-order difference. Difference mean This represents the difference standard deviation.

[0055] (3) Trend fitting characteristics

[0056] Among them, each coefficient is derived from Obtain; The curvature coefficient, The coefficient of the linear term, This is a constant term.

[0057] III. Frequency Domain Characteristics Information such as signal energy distribution, main frequency components, and spectrum is extracted to reveal the changes in energy and frequency structure during the transition from the stable phase to the expansion phase.

[0058] Perform a Fast Fourier Transform on each column of data, using the following formula:

[0059] in, Indicates the first Fourier transform.

[0060] IV. Shape Characteristics Shape features characterize the trend, direction change and curvature of the signal curve, and can effectively reflect the morphological differences of acoustic emission data before and after expansion.

[0061] Shape feature vector Represented as:

[0062] Curvature is expressed as:

[0063] Convexity is expressed as:

[0064] in, For curvature, For convexity, This represents a very small quantity (here, 1x10). -10 Avoid division by zero; These represent the number of peaks, the number of valleys, and the average peak height, respectively.

[0065] V. DTW Distance Characteristics The dynamic time warping distance of signals between different time periods is calculated to characterize the dynamic consistency of multi-channel signals within a window. The calculation formula is as follows:

[0066] in, This is the minimum cumulative distance calculated by the dynamic time warping algorithm. For two sequences in the first... The value of the column, Indicates in The index offset for sliding upwards.

[0067] This application extracts multimodal features from acoustic emission data from multiple dimensions, including time domain, frequency domain, statistics, shape, and dynamic time warping, which greatly enriches the feature information and effectively solves the problem that traditional methods cannot describe the complex features of salt rock under multi-scale, nonlinear loads using single volume strain.

[0068] S1022: Input each initial feature vector into the expansion point recognition model to obtain the detection probability of each initial window data segment containing expansion points; The expansion point identification model is a pre-trained machine learning model or deep learning model. The training data must cover window data segments containing expansion points and window data segments not containing expansion points. The model learns the feature differences between the two types of data segments to make a probability judgment on the input feature vector. The output detection probability ranges from [0,1]. The closer the probability value is to 1, the higher the probability that the initial window data segment contains expansion points, and vice versa.

[0069] S1023: Determine the initial expansion point location based on each detection probability.

[0070] By combining multiple detection probabilities, the approximate location of the expansion point is determined, thus achieving a coarse search.

[0071] Furthermore, multiple different algorithms can be used to calculate the expansion point location to ensure the diversity of candidate locations and reduce the decision bias of a single algorithm.

[0072] In one possible implementation, S1023 includes: 1. Based on the various detection probabilities, the location of the first candidate expansion point is determined using the maximum probability change method; The maximum probability change method is used for calculation. The core logic is to identify the location where the detection probability changes by the largest magnitude and use this location as the first candidate expansion point. The maximum probability change can be measured by indicators such as the absolute value of the difference between the detection probabilities of adjacent windows and the rate of change, and the location with the most significant change is selected as the candidate.

[0073] Specifically, the detection probabilities are sorted according to the time order of the initial window data segments, and the difference between two adjacent detection probabilities is calculated. The midpoint of the initial window data segment corresponding to the previous detection probability between the two detection probabilities with the largest difference is taken as the first candidate expansion point position.

[0074] 2. Based on the various detection probabilities, the location of the second candidate expansion point is determined using the high-probability median method; The high-probability median method is used for calculation. The core logic is to select the initial window data segments with a detection probability higher than the preset probability threshold, and take the median value of the corresponding position of these high-probability data segments as the position of the second candidate expansion point.

[0075] The preset probability threshold can be set according to the scenario requirements and is used to filter out window data segments that are more likely to contain expansion points. For example, the preset probability threshold can be 0.7.

[0076] 3. Based on the detection probabilities, the location of the third candidate expansion point is determined using the probability-weighted average method; The probability-weighted average method is used for calculation. The core logic is to treat each initial window data segment as a sample point, use the detection probability of that window as the weight, and calculate a weighted average of all sample points. The result is then used as the position of the third candidate expansion point. This method can take into account the differences in detection probabilities across different windows, improving the rationality of the candidate positions.

[0077]

[0078] in, This is the index value of the initial window data segment where the third candidate expansion point is located. For the first The detection probability of an initial window of data segments.

[0079] 4. Calculate the differences between the locations of the first candidate expansion point, the second candidate expansion point, and the third candidate expansion point, respectively; 5. If none of the differences are greater than the preset difference, the average of the first candidate expansion point position, the second candidate expansion point position, and the third candidate expansion point position will be used as the initial expansion point position. 6. If there is at least one difference greater than the preset difference, the comprehensive change index of each initial window data segment is determined according to the first formula; the midpoint of the initial window data segment with the largest comprehensive change index is taken as the initial expansion point position. The first formula may include:

[0080] in, For the first The overall degree of change index of each initial window data segment. and The first The mean of the previous initial window data segment and the mean of the next initial window data segment. and The first The standard deviation of the previous initial window data segment and the standard deviation of the next initial window data segment. and The first The linear fitting slope of the previous initial window data segment and the linear fitting slope of the next initial window data segment.

[0081] If the difference between the three expansion point positions is within the allowable deviation range, the model is considered stable, and the average of the three results is selected as the initial expansion point position. If the results of the three probability strategies are dispersed, it indicates that there is noise interference or multi-peak response in the acoustic emission data. In this case, a data-driven detection method is introduced to determine the initial expansion point position. This application uses three different probability strategies to calculate candidate positions, combined with consistency verification and a fallback strategy, to achieve complementary fusion of probability information and signal morphology information. This effectively avoids misjudgment caused by a single strategy under strong noise or complex signal conditions, and improves the stability and reliability of expansion point identification.

[0082] In one possible implementation, S1021 may include: 1. For any initial window data segment, extract the multimodal features of the initial window data segment to form a multimodal feature vector; determine whether the dimension of the multimodal feature vector exceeds the preset dimension; if so, determine the discriminant value corresponding to each feature in the multimodal feature vector, and filter the multimodal feature vector based on each discriminant value to obtain the initial feature vector corresponding to the initial window data segment; if not, use the multimodal feature vector as the initial feature vector corresponding to the initial window data segment. The multimodal features include: basic statistical features, time-domain features, frequency-domain features, shape features, and DTW distance features.

[0083] In one possible implementation, the discriminative power corresponding to each feature in the multimodal feature vector is determined, and the multimodal feature vector is filtered based on each discriminative power to obtain the initial feature vector corresponding to the initial window data segment, including: (1) Calculate the inter-class variance and intra-class variance of each feature in the multimodal feature vector; (2) For any feature, the discriminative power of the feature is obtained based on the ratio of the inter-class variance of the feature to the intra-class variance of the feature. (3) Sort the discrimination of each feature in descending order, and select the features corresponding to the first preset number of discrimination values ​​to form the initial feature vector corresponding to the initial window data segment.

[0084] Based on the above methods, the multimodal features of the initial window data segment are extracted. For example, there are 90 basic features, 48 ​​time-domain features, 36 frequency-domain features, 30 shape features, and 3 DTW features.

[0085] Because the multimodal characteristics of acoustic emission data from salt rock encompass multiple dimensions such as time domain, frequency domain, statistics, shape, and DTW distance, they easily form high-dimensional feature vectors. Some of these features have overlapping information, lack effective distinguishing value for expansion states, and may even contain noise information from on-site acquisition. Therefore, when the feature dimension is greater than a preset dimension (e.g., 100), this application uses discriminative filtering to eliminate redundant and invalid features with low discriminative power. This reduces the interference of worthless information on subsequent feature vector input, model training, and inference processes, avoids expansion point identification bias caused by redundant features, and allows the subsequent expansion point identification model to focus on core features with strong characterization capabilities for salt rock expansion behavior, thereby improving the overall anti-interference capability of the identification process.

[0086] This application uses the ratio of between-class variance to within-class variance as a quantitative indicator of feature discriminative power. Between-class variance accurately reflects the difference in the same feature between expanded and non-expanded salt rock data, while within-class variance reflects the dispersion of a feature within the same category. The ratio of the two can objectively and quantitatively measure the ability of a single feature to distinguish between expanded and non-expanded salt rock states, avoiding the subjectivity and experience of manual feature selection. By ranking and selecting the most discriminative features based on discriminative power, the application simplifies dimensions while preserving the core multimodal features with strong discriminative power for expanded and non-expanded salt rock states. This ensures that the initial feature vector after selection can still comprehensively and accurately represent the feature changes in salt rock acoustic emission data before and after expansion, improving the accuracy of the final expansion point identification location.

[0087] Specifically, the discriminative power of each feature is determined by... The statistical measure is calculated using the following formula:

[0088]

[0089]

[0090] in, It is a feature The inter-class variance It is a feature The within-class variance. Indicates the number of categories. Indicates the first Number of class samples Representation of features In the Mean of class feature The global mean.

[0091] according to To sort values ​​by size and retain highly distinctive features, two filtering methods can be used: Fixed threshold method: Set a threshold Value threshold (e.g.) ≥3), keep all Features with values ​​greater than or equal to the threshold are removed. Features with values ​​less than a threshold.

[0092] Proportional screening method: by Sort the values ​​from largest to smallest and retain the top K% of features (e.g., retain the top 20%, meaning the number of features after filtering is 20% of the original dimension), or directly specify the number of features to retain (e.g., retain the top 50).

[0093] When the dimension of the multimodal feature vector is too high, the above process can accurately select key features that contribute highly to the identification of expansion points, eliminate redundant features, reduce the interference of invalid information on subsequent model calculations, and at the same time reduce computational overhead and improve the overall process efficiency. When the dimension does not exceed the preset value, the multimodal feature vector is directly used to avoid feature loss caused by over-selection and to take into account feature integrity.

[0094] S103: Determine the target search interval based on the initial expansion point location; wherein the target search interval is smaller than the initial search interval; The neighborhood near the initial expansion point is defined as the target search interval. For example, taking the same number of data points forward and backward from the initial expansion point as the center, the target search interval is formed, thus narrowing the search range and improving detection accuracy.

[0095] S104: Slide sampling is performed on the acoustic emission data within the target search interval using a second preset window length to obtain multiple target window data segments; wherein, the second preset window length is less than the first preset window length; S105: Determine the location of the target expansion point based on multiple target window data segments.

[0096] Using the same sliding sampling method as in coarse search, multiple target window data segments are obtained, thus achieving fine search.

[0097] Because salt rock exhibits significant viscoplasticity and creep characteristics, its acoustic emission response is related to local non-coordinated deformation processes such as grain boundary slip, dislocation movement, and shear slip. Therefore, a single acoustic emission parameter (such as ring count, energy, amplitude, etc.) is difficult to directly and accurately define the expansion point, as there is no one-to-one correspondence between these parameters and the damage stage, and they may exhibit similar response characteristics at different stress stages. Furthermore, the expansion of salt rock is essentially a nonlinear, multi-scale, and multi-mechanism coupled process. This invention constructs a multimodal feature vector by integrating multi-dimensional features such as acoustic emission ring count, energy, and frequency, and utilizes an expansion point identification model to mine the relationships between different parameters, thereby effectively identifying the critical expansion state and improving the stability and reliability of the identification.

[0098] In one possible implementation, S105 may include: S1051: Extract the features of each target window data segment to obtain multiple target feature vectors; S1052: Input each target feature vector into the expansion point recognition model to obtain multiple candidate expansion point positions; S1053: Determine the target confidence level for each candidate expansion point location, and take the candidate expansion point location with the highest target confidence level as the target expansion point location.

[0099] The target feature vector is extracted using the same method as S102. However, unlike S102, the target feature vector is input into the expansion point recognition model, and the expansion point position with the highest confidence is selected as the target expansion point position.

[0100] In one possible implementation, S1053 includes: 1. For any candidate expansion point location, determine the model confidence and data consistency confidence of that candidate expansion point location, and calculate the mean of the model confidence and data consistency confidence, which is used as the target confidence of that candidate expansion point location.

[0101] This application uses a dual-reset reliability fusion evaluation, which, compared to S102, emphasizes both accuracy and reliability, thus improving the reliability and accuracy of fine positioning results.

[0102] In one possible implementation, determining the model confidence and data consistency confidence of the candidate expansion point location may include: (1) The detection probability of the candidate expansion point location is used as the model confidence of the candidate expansion point location; (2) Determine the data consistency confidence level of the candidate expansion point location according to the second formula; The second formula may include:

[0103] in, Assess the data consistency confidence level for this candidate expansion point location. This is the average value of the data segment in the target window following the location of the candidate expansion point. The average data of the target window segment preceding the candidate expansion point location; The distance is Euclidean.

[0104] This application employs a two-stage search range reduction and window adaptation to achieve hierarchical optimization of computational efficiency. The initial stage uses a large window and wide-range coarse positioning to quickly eliminate invalid data intervals without expansion characteristics, eliminating the need for detailed analysis of all acoustic emission data and reducing unnecessary computational overhead. Subsequent stages use a small window for detailed analysis only on the narrowed target search interval, improving accuracy while avoiding the computational burden caused by detailed window analysis of all data. This application can significantly reduce data processing costs and improve identification efficiency while ensuring positioning accuracy, adapting to the real-time monitoring needs of salt rock engineering sites. This method can guide the optimization of reservoir injection and production strategies, improve reservoir stability and sealing, thereby extending the service life of the reservoir; it can also be applied to the long-term monitoring and safety assessment of underground salt cavern reservoirs, providing technical support for the implementation of large-scale underground energy reserve projects.

[0105] In one possible implementation, the expansion point identification model can be a weighted fusion model based on random forest, extreme gradient boosting, and lightweight gradient boosting machine.

[0106] Specifically, the expansion point recognition model may include: a feature input layer, a base model parallel training layer, a gradient weight optimization layer, and a probability fusion output layer; The input to the feature input layer is the feature vector extracted from the current window data segment, and the output is the standardized feature matrix. The current window data segment can be the initial window data segment mentioned above, the target window data segment, or the training window data segment used in the training process. The base model parallel training layer includes three independent branches, each of which takes a standardized feature matrix as input and outputs three expansion point detection probability vectors. There is no gradient interaction between the independent branches. The independent branches include: random forest feature decision branch, extreme gradient boosting gradient iteration branch, and lightweight gradient boosting machine histogram optimization branch. The gradient weight optimization layer takes three expansion point detection probability vectors as input, calculates the initial value of the fusion probability based on the initial weight vector corresponding to the three expansion point detection probability vectors, calculates the loss function based on the initial value of the fusion probability, calculates the partial derivative of the initial weight vector based on the loss function, updates the initial weight vector based on the partial derivative through the gradient descent algorithm, and outputs the dynamically optimized weight vector. The input to the probabilistic fusion output layer is three expansion point detection probability vectors and a dynamically optimized weight vector. The three expansion point detection probability vectors are weighted and fused based on the dynamically optimized weight vector to obtain the detection probability that the current window data segment is identified as containing expansion points.

[0107] I. Feature Input Layer The feature input layer receives feature vectors and performs data normalization and dimension matching, providing a unified input format for the base model of the parallel training layer. Specifically, this includes operations such as feature dimension verification, missing value completion, and batch standardization.

[0108] II. Parallel Training Layer of Base Model There is no gradient interaction between the three independent branches of the base model's parallel training layer; gradient backpropagation only occurs at the probabilistic fusion output layer. These independent branches include: a random forest feature decision branch, an extreme gradient boosting gradient iteration branch, and a lightweight gradient boosting machine histogram optimization branch.

[0109] The random forest feature decision branch captures the nonlinear correlation of acoustic emission features through multi-decision tree ensemble, resists overfitting, and extracts key patterns of high-dimensional features. Internally, it consists of K CART decision trees (K is determined by cross-validation, preferably 100-200 trees). Each tree randomly selects N features for splitting, which includes Gini coefficient splitting of decision tree nodes, out-of-bag (OOB) validation, and multi-tree probability voting (taking the average of the predicted probabilities of K trees).

[0110] The extreme gradient boosting gradient iteration branch corrects model errors through the gradient boosting framework, accurately capturing the characteristic abrupt changes in the acoustic emission data of salt rocks from the stationary to the expansion stage. Internally, it consists of T rounds of gradient boosting tree iteration (T is preferably 50~100 rounds), with each round generating a regression tree. It includes a regularization term (L1+L2) to suppress overfitting, specifically including: Taylor second-order expansion to optimize the loss function, column sampling feature selection, and gradient descent to update the tree node weights.

[0111] The lightweight gradient booster histogram optimization branch achieves efficient gradient boosting based on histogram optimization, improving the processing speed of large-scale acoustic emission data while ensuring accuracy, and is suitable for real-time monitoring in engineering sites. Internally, it adopts a gradient boosting tree with a leaf node growth strategy, and the feature values ​​are discretized into H histogram intervals (H is preferably 255). The splitting feature is selected according to mutual information, including histogram difference optimization, single-sided sampling to reduce the sample size, and mutual information feature splitting.

[0112] The parallel training layers of the aforementioned base models can capture the patterns of different types of features: Random Forest (RF) excels at mining nonlinear correlations between statistical and shape features; Extreme Gradient Boosting (XGB) and Light Gradient Boosting Machine (LGB) gradient boosting structures can accurately capture the dynamic changes of time-domain and frequency-domain features (such as the energy of acoustic emission signals and the abrupt changes in the dominant frequency before and after expansion). By fully mining multimodal features, the shortcomings of traditional volumetric strain analysis, which relies on single-feature description and cannot reflect the complex mechanical behavior of salt rock, are solved, making the basis for expansion point identification more comprehensive and more in line with the physical essence of salt rock expansion.

[0113] III. Gradient Weight Optimization Layer The gradient weight optimization layer takes as input the three expansion point detection probability vectors output by the parallel training layer of the base models. Based on the initial weight vectors corresponding to these three expansion point detection probability vectors, it calculates initial fusion probabilities and a loss function. Based on the loss function, it calculates the partial derivatives of the initial weight vectors and updates them using a gradient descent algorithm. The output is the dynamically optimized weight vectors corresponding to the three expansion point detection probability vectors. The gradient weight optimization layer dynamically adjusts the weights based on the prediction errors of each base model, achieving weight optimization through gradient backpropagation, replacing static fusion with fixed weights.

[0114] For example, suppose the detection probability vectors of the three base models are... , , ; Probability label for actual expansion point detection Initial weights The core operations of this layer include: The initial value of the fusion probability is calculated using the following formula: ; The loss function is calculated using the following formula:

[0115] The loss function mentioned above is the cross-entropy loss, which is suitable for binary classification. Take the partial derivatives of the loss function with respect to each weight. Update the weights using gradient descent: ( (The learning rate is preferably 0.01). Weight normalization: Ensure that the weight sum is 1.

[0116] Output the dynamically optimized weight vector .

[0117] IV. Probability Fusion Output Layer The input to the probabilistic fusion output layer is the three expansion point detection probability vectors output by the parallel training layer of the base models and the dynamically optimized weight vector output by the gradient weight optimization layer. Based on the optimized weight vector, the three expansion point detection probability vectors are weighted and fused to output the expansion point detection probability for the current acoustic emission window data segment. The probabilistic fusion output layer completes the weighted fusion of the probabilities of the three base models based on the optimized weights, outputting the final single-sample expansion point detection probability.

[0118] It should be noted that the gradient direction propagation path of the above expansion point identification model is as follows: loss function (cross entropy) → probability fusion output layer → gradient weight optimization layer (weight gradient update) → base model parallel training layer (each base model independently backpropagates gradient: XGB / LGB updates tree nodes, RF updates decision tree split threshold) → feature input layer (feature importance ranking, reverse selection of key acoustic emission features).

[0119] This expansion point identification model integrates the core advantages of three base models: the overfitting and high-dimensional feature processing capabilities of random forests, which can effectively handle the high-dimensional characteristics of multimodal acoustic emission features (time domain / frequency domain / statistics, etc.) of salt rock, avoiding identification bias caused by feature redundancy; the second-order gradient optimization and regularization design of extreme gradient boosting, which can accurately capture the abrupt change in acoustic emission features of salt rock from compression to expansion, solving the problem that traditional methods cannot describe complex features under multi-scale and nonlinear loads; and the histogram optimization and efficient iteration capabilities of lightweight gradient boosting machines, which significantly reduce the computational cost of model training and inference while ensuring identification accuracy. Furthermore, by adopting a gradient backpropagation-driven dynamic weight optimization strategy, the weights are adjusted in real time during training based on the prediction errors of each base model: when the noise of the salt rock acoustic emission data is low and the feature patterns are obvious, the weights are tilted towards the XGB with higher fitting accuracy; when the data has strong noise and outliers, the weights are tilted towards the more robust random forest; and when large-scale acoustic emission data is collected on-site and real-time inference is required, the weights are tilted towards the more computationally efficient LGB. This dynamic weight design enables the model to adapt to different environments, solving the engineering challenges of non-stationary fluctuations, numerous outliers, and dynamic changes in data volume in the acoustic emission data of underground salt rock storage sites. It also avoids the problems of misjudgment and omission by a single model under complex data conditions.

[0120] In practical applications, the aforementioned expansion point identification model, based on acoustic emission data, utilizes multi-base model fusion and dynamic weight optimization to learn the common characteristics and unique patterns of salt rock expansion under different geological conditions. It captures weak acoustic emission signals before expansion and accurately identifies the true expansion point. The model's strong generalization ability allows it to be applied to underground salt cavern storage facilities in different regions, providing reliable technical support for the safe operation of large-scale energy storage projects.

[0121] For the expansion point identification model based on a weighted fusion ensemble model of Random Forest, Extreme Gradient Boosting (XGB), and Lightweight Gradient Boosting (LGB), a sample set needs to be constructed during training. When the distribution of sample numbers in different categories is unbalanced, corresponding weights need to be set for each category. Three base models—Random Forest (RF), Extreme Gradient Boosting (XGB), and Lightweight Gradient Boosting (LGB)—are trained. Each base model undergoes supervised learning based on weights, outputting the detection probability of each sample including expansion points.

[0122] The formula for calculating the detection probability is:

[0123]

[0124] in, For the weights of each base model, For the sample Detection probabilities of each base model For the expansion point identification model, the sample The detection probability.

[0125] For example, the weights of the three base models—Random Forest (RF), Extreme Gradient Boosting Tree (XGB), and Lightweight Gradient Boosting Machine (LGB)—can be 0.4, 0.35, and 0.25, respectively. During training, dynamic weight allocation can be adaptively performed, allowing the better-performing model to have a larger weight in the prediction results. This enables more efficient use of the prediction information from each model, improving the accuracy and reliability of the model's prediction of expansion points.

[0126] It should be noted that the extraction and feature selection of sample features in the training of the expansion point recognition model are the same as the feature extraction and feature selection methods in the detection stage mentioned above, and will not be repeated here.

[0127] The above method will be described in detail below with reference to specific embodiments.

[0128] Location of target expansion point Location of actual expansion point The difference between the sampling points is defined as the error index:

[0129] Based on the error range, the recognition results are divided into three accuracy levels: Level I (High): The target expansion point location is basically consistent with the actual expansion point location, with minimal error; Level II (Intermediate): The locations of the target expansion points may vary slightly, but the overall trend remains consistent. Level III (Low): The location of the target expansion point was significantly off, resulting in poor prediction results.

[0130] Nine samples were identified using the salt rock expansion point identification method provided in this application. The identification results are shown in Table 1.

[0131] Table 1. Comparison of Target Expansion Point Locations and Actual Expansion Point Locations

[0132] As shown in Table 1, the target expansion point locations obtained from the eight samples are basically consistent with the actual expansion point locations, with prediction errors within 5 sampling points. The identification results all achieve Level I accuracy, indicating that the salt rock expansion point identification method of this application can accurately determine the expansion point location and has high identification accuracy. Among them, the identification result of sample D-3 is Level II accuracy; its target expansion point location deviates slightly from the actual expansion point location, but overall, it still maintains good consistency. Furthermore, taking sample D-9 as an example (e.g....) Figure 2 and Figure 3 As shown in the figure, the dashed lines indicate the locations of the expansion points; among them, Figure 2 In the diagram, the Y-axis represents the amplitude after Z-score normalization and is dimensionless; the X-axis represents the sampling time (in seconds). The predicted expansion point location largely coincides with the actual location. In summary, the salt rock expansion point identification method provided by this invention achieves stable identification results under different sample conditions and exhibits good robustness.

[0133] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0134] The following are device embodiments of the present invention. For details not described in detail, please refer to the corresponding method embodiments described above.

[0135] Figure 4 A schematic diagram of the structure of the salt rock expansion point identification device provided in an embodiment of the present invention is shown. For ease of explanation, only the parts related to the embodiment of the present invention are shown, and are described in detail below: like Figure 4 As shown, the salt rock expansion point identification device includes: The coarse search module 21 is used to slide sample the acoustic emission data within the initial search interval with a first preset window length to obtain multiple initial window data segments; The initial expansion point determination module 22 is used to determine the location of the initial expansion point based on multiple initial window data segments; The search range narrowing module 23 is used to determine the target search interval based on the initial expansion point position; wherein the target search interval is smaller than the initial search interval. The fine search module 24 is used to perform sliding sampling of acoustic emission data within the target search interval with a second preset window length to obtain multiple target window data segments; wherein, the second preset window length is less than the first preset window length; The target expansion point output module 25 is used to determine the location of the target expansion point based on multiple target window data segments.

[0136] In one possible implementation, the initial expansion point determination module 22 may include: The first feature vector extraction unit is used to extract the features of each initial window data segment to obtain multiple initial feature vectors. The first identification unit is used to input each initial feature vector into the expansion point identification model to obtain the detection probability of each initial window data segment being identified as containing an expansion point. The first expansion point output unit is used to determine the initial expansion point position based on each detection probability.

[0137] In one possible implementation, the first expansion point output unit can be specifically used for: 1. Based on the various detection probabilities, the location of the first candidate expansion point is determined using the maximum probability change method; 2. Based on the various detection probabilities, the location of the second candidate expansion point is determined using the high-probability median method; 3. Based on the detection probabilities, the location of the third candidate expansion point is determined using the probability-weighted average method; 4. Calculate the differences between the locations of the first candidate expansion point, the second candidate expansion point, and the third candidate expansion point, respectively; 5. If none of the differences are greater than the preset difference, the average of the first candidate expansion point position, the second candidate expansion point position, and the third candidate expansion point position will be used as the initial expansion point position. 6. If at least one difference is greater than the preset difference, the comprehensive change index of each initial window data segment shall be determined according to the first formula. The midpoint of the initial window data segment with the largest comprehensive change index is taken as the initial expansion point position. The first formula may include:

[0138] in, For the first The overall degree of change index of each initial window data segment. and The first The mean of the previous initial window data segment and the mean of the next initial window data segment. and The first The standard deviation of the previous initial window data segment and the standard deviation of the next initial window data segment. and The first The linear fitting slope of the previous initial window data segment and the linear fitting slope of the next initial window data segment.

[0139] In one possible implementation, the first feature vector extraction unit can be specifically used for: 1. For any initial window data segment, extract the multimodal features of the initial window data segment to form a multimodal feature vector; determine whether the dimension of the multimodal feature vector exceeds the preset dimension; if so, determine the discriminant value corresponding to each feature in the multimodal feature vector, and filter the multimodal feature vector based on each discriminant value to obtain the initial feature vector corresponding to the initial window data segment; if not, use the multimodal feature vector as the initial feature vector corresponding to the initial window data segment. The multimodal features include: basic statistical features, time-domain features, frequency-domain features, shape features, and DTW distance features.

[0140] In one possible implementation, the discriminative power corresponding to each feature in the multimodal feature vector is determined, and the multimodal feature vector is filtered based on each discriminative power to obtain the initial feature vector corresponding to the initial window data segment, including: (1) Calculate the inter-class variance and intra-class variance of each feature in the multimodal feature vector; (2) For any feature, the discriminative power of the feature is obtained based on the ratio of the inter-class variance of the feature to the intra-class variance of the feature. (3) Sort the discrimination of each feature in descending order, and select the features corresponding to the first preset number of discrimination values ​​to form the initial feature vector corresponding to the initial window data segment.

[0141] In one possible implementation, the target expansion point output module 25 may include: The second feature vector extraction unit is used to extract the features of each target window data segment to obtain multiple target feature vectors. The second recognition unit is used to input each target feature vector into the expansion point recognition model to obtain multiple candidate expansion point positions. The second expansion point output unit is used to determine the target confidence level of each candidate expansion point location, and to take the candidate expansion point location with the highest target confidence level as the target expansion point location.

[0142] In one possible implementation, the second expansion point output unit can be specifically used for: 1. For any candidate expansion point location, determine the model confidence and data consistency confidence of that candidate expansion point location, and calculate the mean of the model confidence and data consistency confidence, which is used as the target confidence of that candidate expansion point location.

[0143] In one possible implementation, determining the model confidence and data consistency confidence of the candidate expansion point location may include: (1) The detection probability of the candidate expansion point location is used as the model confidence of the candidate expansion point location; (2) Determine the data consistency confidence level of the candidate expansion point location according to the second formula; The second formula may include:

[0144] in, Assess the data consistency confidence level for this candidate expansion point location. This is the average value of the data segment in the target window following the location of the candidate expansion point. The average data of the target window segment preceding the candidate expansion point location; The distance is Euclidean.

[0145] In one possible implementation, the expansion point recognition model includes: a feature input layer, a base model parallel training layer, a gradient weight optimization layer, and a probability fusion output layer; The input to the feature input layer is the feature vector extracted from the current window data segment, and the output is the standardized feature matrix; where the current window data segment includes: the initial window data segment or the target window data segment; The base model parallel training layer includes three independent branches, each of which takes a standardized feature matrix as input and outputs three expansion point detection probability vectors. There is no gradient interaction between the independent branches. The independent branches include: random forest feature decision branch, extreme gradient boosting gradient iteration branch, and lightweight gradient boosting machine histogram optimization branch. The gradient weight optimization layer takes three expansion point detection probability vectors as input, calculates the initial value of the fusion probability based on the initial weight vector corresponding to the three expansion point detection probability vectors, calculates the loss function based on the initial value of the fusion probability, calculates the partial derivative of the initial weight vector based on the loss function, updates the initial weight vector based on the partial derivative through the gradient descent algorithm, and outputs the dynamically optimized weight vector. The input to the probabilistic fusion output layer is three expansion point detection probability vectors and a dynamically optimized weight vector. The three expansion point detection probability vectors are weighted and fused based on the dynamically optimized weight vector to obtain the detection probability that the current window data segment is identified as containing expansion points.

[0146] Figure 5 This is a schematic diagram of the electronic device 3 provided in an embodiment of the present invention. Figure 5 As shown, the electronic device 3 of this embodiment includes a processor 30 and a memory 31. The memory 31 stores a computer program 32. When the processor 30 executes the computer program 32, it implements the steps in the various method embodiments described above. Alternatively, when the processor 30 executes the computer program 32, it implements the functions of each module / unit in the various device embodiments described above.

[0147] For example, computer program 32 may be divided into one or more modules / units, which are stored in memory 31 and executed by processor 30 to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of computer program 32 in electronic device 3.

[0148] Electronic device 3 may include, but is not limited to, processor 30 and memory 31. Those skilled in the art will understand that... Figure 5 This is merely an example of electronic device 3 and does not constitute a limitation on electronic device 3. It may include more or fewer components than shown, or combine certain components, or different components. For example, electronic device 3 may also include input / output devices, network access devices, buses, etc.

[0149] The processor 30 can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor.

[0150] The memory 31 can be an internal storage unit of the electronic device 3, such as a hard disk or memory of the electronic device 3. The memory 31 can also be an external storage device of the electronic device 3, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the electronic device 3. Furthermore, the memory 31 can include both internal and external storage units of the electronic device 3. The memory 31 is used to store the computer program 32 and other programs and data required by the electronic device 3. The memory 31 can also be used to temporarily store data that has been output or will be output.

[0151] For the sake of simplicity and clarity, only the above-described functional modules / units are used as examples. In practical applications, the functions described above can be assigned to different functional modules / units as needed. These modules / units can be implemented in hardware, software, or a combination of both.

[0152] This invention also provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the methods described in the above-described method embodiments.

[0153] This invention also provides a computer program product, including a computer program. When the computer program is executed by a processor, it implements the methods described in the above-described method embodiments.

[0154] Computer programs include computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. Computer-readable media can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0155] In the above embodiments, the descriptions of each embodiment have their own emphasis. Parts not detailed or described in a particular embodiment can be referred to in the relevant descriptions of other embodiments. Unless otherwise specified or in conflict with logic, the terminology and / or descriptions between different embodiments are consistent and can be referenced interchangeably. Technical features in different embodiments can be combined to form new embodiments based on their inherent logical relationships.

[0156] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for identifying salt rock expansion points, characterized in that, include: The acoustic emission data within the initial search interval is sampled using a first preset window length to obtain multiple initial window data segments. The initial expansion point location is determined based on the multiple initial window data segments; Based on the initial expansion point location, a target search interval is determined; wherein, the target search interval is smaller than the initial search interval; The acoustic emission data within the target search interval is sampled using a second preset window length to obtain multiple target window data segments; wherein, the second preset window length is less than the first preset window length; The location of the target expansion point is determined based on the multiple target window data segments.

2. The method for identifying salt rock expansion points according to claim 1, characterized in that, The step of determining the initial expansion point location based on the plurality of initial window data segments includes: Features are extracted from each initial window data segment to obtain multiple initial feature vectors; Each initial feature vector is input into the expansion point recognition model to obtain the detection probability of each initial window data segment containing expansion points. The initial expansion point location is determined based on each detection probability.

3. The method for identifying salt rock expansion points according to claim 2, characterized in that, Determining the initial expansion point location based on each detection probability includes: Based on the various detection probabilities, the location of the first candidate expansion point is determined using the maximum probability change method; Based on the various detection probabilities, the high-probability median method is used to determine the location of the second candidate expansion point; The location of the third candidate expansion point is determined by using a probability-weighted average method based on the various detection probabilities. Calculate the differences between the positions of the first candidate expansion point, the second candidate expansion point, and the third candidate expansion point, respectively; If none of the differences are greater than the preset difference, then the average of the first candidate expansion point position, the second candidate expansion point position, and the third candidate expansion point position is taken as the initial expansion point position. If there is at least one difference greater than the preset difference, then the comprehensive change index of each initial window data segment is determined according to the first formula; the midpoint of the initial window data segment with the largest comprehensive change index is taken as the initial expansion point position. The first formula includes: in, For the first The overall degree of change index of each initial window data segment. and The first The mean of the previous initial window data segment and the mean of the next initial window data segment. and The first The standard deviation of the previous initial window data segment and the standard deviation of the next initial window data segment. and The first The linear fitting slope of the previous initial window data segment and the linear fitting slope of the next initial window data segment.

4. The method for identifying salt rock expansion points according to claim 2, characterized in that, The process of extracting features from each initial window data segment yields multiple initial feature vectors, including: For any initial window data segment, extract the multimodal features of the initial window data segment to form a multimodal feature vector; determine whether the dimension of the multimodal feature vector exceeds a preset dimension; if so, determine the discriminative power of each feature in the multimodal feature vector, and filter the multimodal feature vector based on each discriminative power to obtain the initial feature vector corresponding to the initial window data segment; if not, use the multimodal feature vector as the initial feature vector corresponding to the initial window data segment. The multimodal features include: basic statistical features, time-domain features, frequency-domain features, shape features, and DTW distance features.

5. The method for identifying salt rock expansion points according to claim 4, characterized in that, The step of determining the discriminative power corresponding to each feature in the multimodal feature vector, and filtering the multimodal feature vector based on each discriminative power to obtain the initial feature vector corresponding to the initial window data segment, includes: Calculate the inter-class variance and intra-class variance of each feature in the multimodal feature vector; For any given feature, the discriminative power of that feature is obtained by the ratio of its inter-class variance to its intra-class variance. The discriminative power of each feature is sorted in descending order, and the features corresponding to the first preset number of discriminative powers are selected to form the initial feature vector corresponding to the initial window data segment.

6. The method for identifying salt rock expansion points according to claim 1, characterized in that, Determining the location of the target expansion point based on the multiple target window data segments includes: Features of each target window data segment are extracted separately to obtain multiple target feature vectors; Each target feature vector is input into the expansion point recognition model to obtain multiple candidate expansion point locations. The target confidence level of each candidate expansion point location is determined, and the candidate expansion point location with the highest target confidence level is taken as the target expansion point location.

7. The method for identifying salt rock expansion points according to claim 6, characterized in that, The determination of the target confidence level for each candidate expansion point location includes: For any candidate expansion point location, determine the model confidence and data consistency confidence of that candidate expansion point location, and calculate the mean of the model confidence and the data consistency confidence as the target confidence of that candidate expansion point location.

8. The method for identifying salt rock expansion points according to claim 7, characterized in that, The model confidence and data consistency confidence for determining the location of the candidate expansion point include: The detection probability of the candidate expansion point location is used as the model confidence of the candidate expansion point location; The data consistency confidence level of the candidate expansion point location is determined according to the second formula. The second formula includes: in, Assess the data consistency confidence level for this candidate expansion point location. This is the average value of the data segment in the target window following the location of the candidate expansion point. The average data of the target window segment preceding the candidate expansion point location; The distance is Euclidean.

9. The method for identifying salt rock expansion points according to any one of claims 2 to 8, characterized in that, The expansion point identification model includes: a feature input layer, a base model parallel training layer, a gradient weight optimization layer, and a probability fusion output layer; The input to the feature input layer is a feature vector extracted from the current window data segment, and the output is a standardized feature matrix; wherein, the current window data segment includes: the initial window data segment or the target window data segment; The base model parallel training layer includes three independent branches, each branch taking the standardized feature matrix as input and outputting three expansion point detection probability vectors; there is no gradient interaction between the independent branches, and the independent branches include: a random forest feature decision branch, an extreme gradient boosting gradient iteration branch, and a lightweight gradient boosting machine histogram optimization branch; The input to the gradient weight optimization layer is the detection probability vector of the three expansion points. Based on the initial weight vector corresponding to the detection probability vector of the three expansion points, the initial value of the fusion probability is calculated, and the loss function is calculated based on the initial value of the fusion probability. Based on the loss function, the partial derivative of the initial weight vector is calculated, and the initial weight vector is updated by the gradient descent algorithm based on the partial derivative, and the dynamically optimized weight vector is output. The input to the probability fusion output layer is the three expansion point detection probability vectors and the dynamically optimized weight vector. The three expansion point detection probability vectors are weighted and fused based on the dynamically optimized weight vector to obtain the detection probability that the current window data segment contains expansion points.

10. An electronic device, characterized in that, The method includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the salt rock expansion point identification method as described in any one of claims 1 to 9.

Images