Wavelet time-frequency analysis and deep learning fused oil layer dessert identification method and device, and medium

Abstract

The invention provides an oil layer dessert identification method and device fusing wavelet time-frequency analysis and deep learning and a medium, and belongs to the field of oil-gas exploration and development. Comprising the following steps that logging data are collected, measurement data of logging curves changing along with the depth and an oil testing and production testing result data set of corresponding layers are established, standardized preprocessing is conducted on original logging data, and a cross plot of different logging curves is drawn; carrying out sensitivity analysis on the logging curve by utilizing the cross plot, carrying out wavelet decomposition after carrying out normalization processing on the sensitive logging curve, and determining an optimal wavelet function and an optimal decomposition scale; a BiLSTM-FCNN oil layer sweet spot recognition model is constructed and trained, sensitive logging curve data is processed by adopting the determined optimal wavelet function and the optimal decomposition scale and then input to the trained model, and an oil layer sweet spot recognition result is output; according to the method, the problem of separation of noise and effective features in original data signals is solved, and the generalization ability of the model under complex geological conditions is improved.

Description

Technical Field

[0001] This application relates to the field of oil and gas exploration and development technology, and in particular to a method, device and medium for identifying sweet spots in oil layers that integrates wavelet time-frequency analysis and deep learning. Background Technology

[0002] With the deepening of oil and gas exploration and development, the identification of sweet spots in oil reservoirs under complex geological conditions has become a key issue for the efficient development of oilfields. Sweet spots refer to oil and gas-rich sections in the reservoir with excellent physical properties; their accurate identification directly affects drilling success rates and development benefits. Taking the Lower Wuerhe Formation reservoir in the Ma2 well area as an example, its oil layers have a large span, uneven thickness distribution, and relatively thin individual layers, making systematic evaluation difficult using traditional methods. Therefore, developing an intelligent identification technology that can adapt to strong heterogeneity and complex vertical correlations is of great significance for reducing exploration costs and improving recovery rates.

[0003] Traditional reservoir identification methods, such as cross-plotting, rely heavily on expert experience and well logging analysis. Their interpretations are often influenced by subjective experience and are difficult to standardize. Furthermore, the complexity of geological conditions and reservoir heterogeneity lead to a non-linear relationship between well logging response and reservoir characteristics, which traditional linear statistical models struggle to accurately represent. Moreover, existing methods are mostly based on single-point data mapping, failing to consider the vertical correlation of well logging sequences, resulting in depth-series deviations in identification results that do not conform to actual geological patterns.

[0004] To address the shortcomings of traditional methods, deep learning-based intelligent identification technology for oil reservoir sweet spots has overcome the bottleneck of parameterization difficulties in traditional physical models, providing an efficient and reliable technical means for the development of complex oil reservoirs. However, existing single deep learning models (such as RNN and LSTM) can process sequential data, but they have problems such as ignoring the multi-scale features of the data (such as high-frequency noise interference and low-frequency trend signals), not making full use of the bidirectional sequence dependence of logging data (the influence of upper and lower formations), and insufficient sensitivity to local features such as thin layer boundaries and abrupt change points. Summary of the Invention

[0005] To address the aforementioned technical issues, this application proposes a method, apparatus, and medium for identifying oil reservoir sweet spots by integrating wavelet time-frequency analysis and deep learning. It utilizes wavelet multi-scale decomposition of logging data to extract time-frequency domain energy features, and captures bidirectional sequence dependencies using BiLSTM and classifies multi-source features by fusing FCNN.

[0006] The technical solution adopted in this application is: a method for identifying oil-based dessert layers that integrates wavelet time-frequency analysis and deep learning, comprising the following steps: Step 1: Collect a series of physical parameters continuously measured and recorded along the wellbore by different logging instruments during the oil and gas field logging process as logging data, forming a logging curve sequence that varies with depth. Correlate the measurement data of the logging curves with depth with the reservoir or non-reservoir results obtained after oil testing and production at the corresponding formation to form a structured dataset. Standardize the logging data. Step 2: Plot cross-plots of different logging curves and use these plots to perform sensitivity analysis on the logging curves. Normalize the sensitive logging curves, and then perform wavelet transforms on the normalized sensitive curves using different wavelet functions at the same decomposition scale and the same wavelet function at different decomposition scales to determine the optimal wavelet function and optimal decomposition scale. Step 3: Construct a BiLSTM-FCNN oil layer dessert recognition model; Step 4: Train the BiLSTM-FCNN oil layer sweet spot recognition model, and process the sensitive logging curve data using the determined optimal wavelet function and optimal decomposition scale, then input the processed data into the trained BiLSTM-FCNN oil layer sweet spot recognition model to output the oil layer sweet spot recognition result.

[0007] Furthermore, the logging curves in step 1 include one or more of the following: spontaneous potential (SP) curve, undisturbed formation resistivity (RT) curve, intrusive zone resistivity (RI) curve, flushed zone resistivity (RXO) curve, sonic transit time (AC) curve, natural gamma ray (GR) curve, lithological density (DEN) curve, and neutron porosity (CNL) curve.

[0008] Furthermore, in step 2, the standardized sensitive logging curves are subjected to wavelet transforms with different wavelet functions at the same decomposition scale and wavelet transforms with the same wavelet function at different decomposition scales. The results are compared and analyzed with the oil testing and production results to determine the wavelet function and decomposition scale that are most effective in identifying sweet spots in oil reservoirs.

[0009] Furthermore, the sensitive logging curve is decomposed into high-frequency, mid-frequency, and low-frequency components using the optimal wavelet function and decomposition scale, and frequency domain features are extracted. The frequency domain features are then input into the BiLSTM-FCNN oil layer sweet spot recognition model to extract time domain features.

[0010] Furthermore, the construction steps of the BiLSTM-FCNN oil layer dessert recognition model are as follows: 3.1: Constructing the BiLSTM sequence processing module: The high-frequency, mid-frequency, and low-frequency logging sequences processed by wavelet were input into three BiLSTM models respectively. 3.2: Constructing the FCNN feature integration and processing module: The forward and backward information extracted from the three BiLSTM models are concatenated. After independently modeling the high-frequency, mid-frequency, and low-frequency components, the results of the three are further fused and used as input to the FCNN feature integration processing module.

[0011] Furthermore, the logging sequences of high-frequency, medium-frequency, and low-frequency components are processed bidirectionally using the BiLSTM model to extract features of the logging data from both directions. The forward and backward outputs of the BiLSTM model are concatenated into a vector, which is then input into the FCNN feature integration module for further processing.

[0012] Furthermore, the FCNN feature integration processing module is used to integrate and nonlinearly map the feature information extracted by the BiLSTM sequence processing module, mapping the output value of FCNN to between 0 and 1, and outputting the identification probability of oil-bearing reservoirs.

[0013] Furthermore, the Adam optimizer was used to train the BiLSTM-FCNN oil layer dessert recognition model.

[0014] A computer device includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method.

[0015] A computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implement the steps of the method.

[0016] The advantages of this application compared to the prior art are as follows: Compared with existing oil layer sweetness identification methods, the advantages of this application are mainly reflected in the following aspects: 1. Wavelet multi-scale decomposition of logging data to extract time-frequency domain energy features for learning frequency domain information. Well logging wavelet transform enables depth-frequency domain conversion, displaying one-dimensional depth curve data in two-dimensional space, reflecting two-dimensional spectral signals, and providing information on the variation of various frequency intensities with depth. The frequency domain often contains more intuitive information; learning this frequency domain information can achieve higher recognition accuracy.

[0017] 2. BiLSTM captures bidirectional sequence dependencies, while FCNN integrates multi-source feature classification for learning time-domain information. When analyzing well logging data, BiLSTM-FCNN enables each depth sampling point to capture data features containing information about the upper and lower formations simultaneously. This overcomes the limitation of traditional LSTM models, which can only process unidirectional time-series data, effectively uncovering the correlations between sequence data and accurately extracting the sequence features of oil-bearing reservoirs and non-reservoir areas from the well logging curves. The combination of BiLSTM and FCNN, through the nonlinear mapping of FCNN, further improves the model's ability to represent complex nonlinear relationships, solving the highly nonlinear problem in oil sweet spot identification and effectively suppressing the influence of interference factors. Attached Figure Description

[0018] The following description, in conjunction with the accompanying drawings, further illustrates this application: Figure 1 A schematic diagram of the BiLSTM structure is provided for the embodiments of this application; Figure 2 This is a structural diagram of the BiLSTM-FCNN oil layer sweet spot recognition model provided in the embodiments of this application; Figure 3 Histograms and cross plots of resistivity and acoustic transit time of the Ma2239 test section provided in this application embodiment; Figure 4 The flowchart illustrates the oil layer dessert recognition method that integrates wavelet time-frequency analysis and deep learning, as provided in the embodiments of this application. Detailed Implementation

[0019] like Figures 1 to 4 As shown, this application provides a method for identifying oil layer sweet spots by integrating wavelet time-frequency analysis and deep learning. It aims to realize the depth-frequency domain conversion of well logging curves through wavelet transform, extract information on the variation of various frequency intensities with depth, and combine bidirectional long short-term memory neural network (BiLSTM) and fully connected neural network (FCNN) to fully mine the feature information of the upper and lower formations in the well logging data, so as to improve the identification accuracy of oil layer sweet spots.

[0020] The main implementation steps of this method include: Step 1: Establish a dataset of measurement data showing the variation of well logging curves with depth and the corresponding oil testing and production results for each layer; Step 2: Standardize the raw logging data and perform sensitivity analysis on the logging curves using cross plots. Normalize the sensitive logging curves to determine the optimal wavelet function and optimal decomposition scale. Step 3: Construct a BiLSTM-FCNN oil layer sweet spot recognition model. Use BiLSTM to perform bidirectional processing on the preprocessed data to extract the features of the logging data from both the front and rear directions. The forward and backward outputs of the BiLSTM are concatenated into a vector, which is then fed into the FCNN network for further processing. The FCNN is responsible for integrating and nonlinearly mapping the feature information extracted by the bidirectional LSTM to obtain the identification probability of the oil layer sweet spot; the model will output whether the depth location is an oil-bearing reservoir (output is 1) or a non-reservoir (output is 0).

[0021] Step 4: Train the BiLSTM-FCNN oil layer sweet spot recognition model, and process the sensitive logging curve data using the determined optimal wavelet function and optimal decomposition scale, then input the processed data into the trained BiLSTM-FCNN oil layer sweet spot recognition model to output the oil layer sweet spot recognition result.

[0022] The following is a further explanation of this application based on the Lower Urho Formation reservoir in the Ma2 well area of ​​the Mabei Oilfield.

[0023] The data used in this embodiment are actual logging data from 12 wells in the Lower Wuerhe Formation reservoir of the Ma2 well area in the Mabei Oilfield. One well was used for training, and the remaining 11 wells were used for testing. These 11 test wells are located in different areas of the entire study area, with depths of 160.875 m, 125.375 m, 79 m, 155.75 m, 91 m, 109.5 m, 100 m, 53.75 m, 31.875 m, 48.25 m, and 32.625 m, respectively. Through feature extraction and sensitivity factor analysis of the logging data, a total of 8 logging curves were found to affect the identification of sweet spots in the reservoir: spontaneous potential (SP) curve, original formation resistivity (RT) curve, intrusive zone resistivity (RI) curve, flushed zone resistivity (RXO) curve, sonic transit time (AC) curve, natural gamma ray (GR) curve, lithological density (DEN) curve, and neutron porosity (CNL) curve.

[0024] The proposed method for identifying oil reservoir sweet spots based on well logging wavelet analysis and BiLSTM-FCNN first performs standardized preprocessing on well logging data from the Lower Wuerhe Formation reservoir in the Ma2 well area. Then, sensitivity analysis of the well logging curves is performed using cross-plots. Wavelet transform is used to decompose the sensitive well logging curves into high-frequency, mid-frequency, and low-frequency components, extracting the frequency domain features. Subsequently, the data is input into BiLSTM-FCNN to extract time domain features, deeply mining the temporal features and nonlinear relationships in the well logging data, ultimately obtaining the oil-bearing reservoir identification results. The specific steps are as follows.

[0025] Step 1: Establish a dataset of well logging curves varying with depth and corresponding oil testing and production results. Using real well logging data from 12 wells in the Lower Wuerhe Formation reservoir of the Ma2 well area in the Mabei Oilfield, standardize the data and plot cross-sections of different well logging curves, such as... Figure 4As shown, the uncirculated formation resistivity (RT) curve was determined to be the most sensitive logging curve for identifying oil sweet spots. Subsequently, the uncirculated formation resistivity curve data were normalized to remove dimensional influences and retain data characteristics.

[0026] Step 2: Perform wavelet multi-scale preprocessing on the undisturbed formation resistivity curve data. The standardized undisturbed formation resistivity curves are subjected to wavelet transforms with different wavelet functions at the same decomposition scale and wavelet transforms with the same wavelet function at different decomposition scales. Comparative analysis with oil testing and production results is conducted to determine the Daubechies wavelet basis (db4) and 3-layer decomposition as the wavelet function and decomposition scale that best identify oil sweet spots. The output of the discrete transformation of the input using the db4 wavelet is used as the input to the high-, medium-, and low-frequency BiLSTM-FCNN hybrid neural network models.

[0027] Step 3: Build and train the model.

[0028] 3.1: Constructing the BiLSTM sequence processing module: The high-frequency, mid-frequency, and low-frequency logging sequences, processed by wavelet processing, were input into three BiLSTM models, with a time step of 8, corresponding to a 1-meter well interval. Each BiLSTM model in the high, mid, and low-frequency BiLSTM sequence processing module has three layers, with 8, 16, and 32 neurons per layer, respectively, extracting formation logging data features from both forward and backward directions.

[0029] When processing well logging sequence data, LSTM can only process data sequentially, thus only acquiring information above the corresponding sampling point depth, while information below that depth cannot be effectively captured. However, the well logging response at a certain depth sampling point is simultaneously influenced by the previous and subsequent depth sampling points. That is, if a certain depth sampling point is an oil-bearing reservoir, its previous and subsequent depth sampling points are also very likely to be oil-bearing reservoirs, and there is a bidirectional dependency between the well logging interpretation results and the upper and lower formations. Since LSTM cannot encode information from back to front, this unidirectional encoding mechanism limits its ability to process bidirectional dependencies in well logging data. Therefore, reversing the order of well logging data is particularly important to compensate for the limitations of LSTM in processing bidirectional dependencies, so that the data features captured at each depth sampling point simultaneously contain information about the upper and lower formations. Bi-directional LSTM (BiLSTM) combines a forward LSTM and a backward LSTM, forming a bidirectional LSTM based on LSTM. The BiLSTM structure is as follows. Figure 1As shown, a BiLSTM unit consists of two LSTM units: one is responsible for forward computation of the input data at the current time step, and the other is responsible for backward computation of the input data at the current time step. The vectors obtained from the two computations are concatenated to update the model weight matrix, generating a new vector containing forward and backward direction information of the input data as the final output of the BiLSTM.

[0030] 3.2: Constructing the FCNN feature integration and processing module: The forward and backward information extracted from three BiLSTM models are concatenated. After independently modeling high-frequency, mid-frequency, and low-frequency data, the results from these three models are further fused and used as input to the FCNN feature integration processing module. The FCNN feature integration processing module consists of two FCNN layers, with 128 and 64 neurons in each layer, respectively. The FCNN layer extracts and integrates feature information, and learns the complex nonlinear relationships of the data through the nonlinear mapping of the sigmoid activation function, mapping the FCNN output value between 0 and 1 to output the probability of identifying oil-bearing reservoirs.

[0031] A fully connected neural network (FCNN) consists of an input layer, one or more hidden layers, and an output layer. Each layer is fully connected to all neurons in the previous layer. The input layer receives an n-dimensional vector as input data. In the hidden layers, m neurons linearly combine the input data using an m×n weight matrix W and a bias b. Then, an activation function performs a non-linear transformation on the data, mapping the originally inseparable non-linear data in the low-dimensional linear space to a high-dimensional space, making it linearly separable. In the high-dimensional space, the complex non-linear characteristics of the data can be captured more effectively. This non-linear mapping significantly enhances its ability to represent complex data in the real world, thus enabling more accurate handling of various non-linear problems. The output layer maps the output of the hidden layers to a k-dimensional vector.

[0032] The structure of the BiLSTM-FCNN oil layer dessert recognition model is as follows: Figure 2 As shown, firstly, the high-frequency, mid-frequency, and low-frequency components of the well logging sequence after wavelet processing are standardized and preprocessed, and then converted into numerical vectors so that they can be input into the deep learning algorithm for processing. Secondly, the BiLSTM sequence processing module is used to fully mine the stratigraphic feature information contained in the well logging data. Figure 2 In this context, the input at time t is the original formation resistivity RT, which incorporates high-frequency, mid-frequency, and low-frequency components. t Output y t To determine the probability that this depth contains an oil-bearing reservoir, a forward LSTM is used to analyze the data along the forward direction of the logging sequence, using the latent variable h from the previous time step. t-1 Unit State Ct-1 The feature information of variables such as spontaneous potential and undisturbed formation resistivity above the extraction depth sampling point is obtained from the well logging data input at the current moment. Then, based on the latent variable h at that moment... t Unit state C t The input data at the next time step is used to determine the layer to which the sampling point belongs. The backward LSTM runs in the opposite direction of the logging sequence, capturing data features below the depth sampling point. Then, the vectors calculated by the forward LSTM and backward LSTM are concatenated to obtain the final output vector, and this process is repeated. Finally, the feature information is integrated by the FCNN feature integration processing module, and after nonlinear mapping, the neural network model can effectively learn the complex nonlinear relationships in the logging data, achieving deep characterization and accurate identification of oil sweet spots.

[0033] Due to the complexity of geological conditions, well logging curves show little variation at some depths, leading to similar values ​​in logging data from adjacent layers and making it difficult to extract layer features. Through this bidirectional analysis, BiLSTM-FCNN can take into account information from deeper formations, extracting more complete information from continuous well logging sequences and more fully integrating the influence of upper and lower formations on oil-bearing reservoir identification results, thereby better identifying oil-bearing reservoirs under complex geological conditions.

[0034] 3.3: Model Training: By combining the sub-components from well logging wavelet decomposition with the BiLSTM-FCNN oil sweet spot recognition model, multi-dimensional feature fusion is performed to generate the final oil sweet spot recognition result. The model is trained using the Adam optimizer with an initial learning rate of 0.05, and the network is trained using the training set until the loss function value is minimized.

[0035] Step 4: Identification of oil-based desserts.

[0036] The method described in this application was applied to 11 other wells in the Lower Wuerhe Formation reservoir of the Ma2 well area. Through comparative analysis of FCNN, RNN, and LSTM neural network models, the oil layer sweet spot identification method based on logging wavelet analysis and BiLSTM-FCNN was found to be more accurate, improving the accuracy by 1.10%–2.82%. Compared to other methods, it significantly reduced false identification and missed identification, especially demonstrating higher accuracy in layer boundary identification. Furthermore, the model exhibits strong generalization ability, accurately identifying oil layer sweet spots in other wells within the Ma2 study area, achieving an overall accuracy of 82%, which is of great significance for the sequential information mining of continuous logging data.

[0037] This application, through validation on data from other test wells in multiple study areas, demonstrates that the method has strong generalization ability under different geological conditions. Compared with other methods, it significantly reduces the false identification, missed identification, and under-identification of oil reservoir sweet spots, and is more sensitive in identifying the boundary between oil-bearing reservoirs and non-reservoir areas. Compared with traditional prediction models, the model in this application can effectively improve prediction accuracy and reduce errors caused by factors such as manual operation.

[0038] This application solves the problem of separating noise from effective features in raw data signals, the nonlinear relationship between reservoir heterogeneity and logging response characteristics, and improves the model's generalization ability under complex geological conditions.

[0039] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application 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 or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A method for identifying oil-based desserts that integrates wavelet time-frequency analysis and deep learning, characterized in that: Includes the following steps: Step 1: Collect a series of physical parameters continuously measured and recorded along the wellbore by different logging instruments during the logging process in the oil and gas field as logging data, form a logging curve sequence that varies with depth, and associate the measurement data of logging curves that vary with depth with the reservoir or non-reservoir results obtained after oil testing and production at the corresponding layers to form a structured dataset, and standardize the logging data. Step 2: Plot the cross plots of different logging curves, and use the cross plots to perform sensitivity analysis on the logging curves. Normalize the sensitive logging curves, and perform wavelet transforms on the normalized sensitive curves with different wavelet functions at the same decomposition scale and with the same wavelet function at different decomposition scales to determine the optimal wavelet function and the optimal decomposition scale. Step 3: Construct a BiLSTM-FCNN oil layer dessert recognition model; Step 4: Train the BiLSTM-FCNN oil layer sweet spot recognition model, and process the sensitive logging curve data using the determined optimal wavelet function and optimal decomposition scale, then input the processed data into the trained BiLSTM-FCNN oil layer sweet spot recognition model to output the oil layer sweet spot recognition result.

2. The method for identifying oil-based desserts by integrating wavelet time-frequency analysis and deep learning according to claim 1, characterized in that: The logging curves in step 1 include one or more of the following: spontaneous potential (SP) curve, undisturbed formation resistivity (RT) curve, intrusive zone resistivity (RI) curve, flushed zone resistivity (RXO) curve, sonic transit time (AC) curve, natural gamma ray (GR) curve, lithological density (DEN) curve, and neutron porosity (CNL) curve.

3. The method for identifying oil-based desserts by integrating wavelet time-frequency analysis and deep learning according to claim 1, characterized in that: In step 2, the normalized sensitive logging curves are subjected to wavelet transforms with different wavelet functions at the same decomposition scale and wavelet transforms with the same wavelet function at different decomposition scales. The results are compared and analyzed with the oil testing and production results to determine the wavelet function and decomposition scale that are most effective in identifying sweet spots in oil reservoirs.

4. The method for identifying oil-based desserts by integrating wavelet time-frequency analysis and deep learning according to claim 3, characterized in that: The sensitive logging curve is decomposed into high-frequency, mid-frequency and low-frequency components using the optimal wavelet function and decomposition scale. Frequency domain features are extracted and then input into the BiLSTM-FCNN oil layer sweet spot recognition model to extract time domain features.

5. The method for identifying oil-layer desserts by integrating wavelet time-frequency analysis and deep learning according to claim 4, characterized in that: The construction steps of the BiLSTM-FCNN oil layer dessert recognition model are as follows: 3.1: Constructing the BiLSTM sequence processing module: The high-frequency, mid-frequency, and low-frequency logging sequences processed by wavelet were input into three BiLSTM models respectively. 3.2: Constructing the FCNN feature integration and processing module: The forward and backward information extracted from the three BiLSTM models are concatenated. After independently modeling the high-frequency, mid-frequency, and low-frequency components, the results of the three are further fused and used as input to the FCNN feature integration processing module.

6. The method for identifying oil-based desserts by integrating wavelet time-frequency analysis and deep learning according to claim 5, characterized in that: The BiLSTM model is used to perform bidirectional processing on the logging sequences of high-frequency, mid-frequency, and low-frequency components, and the features of the logging data are extracted from both the front and back directions. The forward and backward outputs of the BiLSTM model are concatenated into a vector, which is then input into the FCNN feature integration module for further processing.

7. The method for identifying oil-based desserts by integrating wavelet time-frequency analysis and deep learning according to claim 5, characterized in that: The FCNN feature integration processing module integrates and nonlinearly maps the feature information extracted by the BiLSTM sequence processing module, mapping the FCNN output value between 0 and 1 to output the identification probability of oil-bearing reservoirs.

8. The method for identifying oil-based desserts by integrating wavelet time-frequency analysis and deep learning according to claim 5, characterized in that: The Adam optimizer was used to train the BiLSTM-FCNN oil layer dessert recognition model.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1-8.

10. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method according to any one of claims 1-8.

Images