A semi-supervised learning anomaly detection model for fracturing construction risk early warning method
Through the semi-supervised learning anomaly detection model, the feature extraction and classification is used by DBN and SVVD networks, and superspheres are constructed for risk detection, which solves the problem that risk judgment in fracturing construction depends on high-skilled experts and sample imbalance, achieving more Accurate risk warning and diagnostic capabilities.
Patent Information
- Application Number
- CN202011525451.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2020-12-22
- Publication Date
- 2024-12-27
- Estimated Expiration
- 2040-12-22
AI Technical Summary
In the existing technology, the degree of informatization and intelligence during fracturing construction is low, which leads to the on-site experts requiring high skills and experience to make risk judgments. Moreover, the performance of the machine learning model declines due to sample imbalance, so it is impossible to effectively detect construction risks and cannot be timely. Make an early warning.
Using semi-supervised learning anomaly detection model, feature extraction and SVVD network are used to perform abnormal classification through DBN network, combined with preprocessing and model fine-tuning, a supersphere with the smallest volume is constructed for detection in the feature space, and through the deep neural network model Real-time warning and diagnosis, updating models to improve accuracy.
It improves the real-time early warning capability of fracturing construction risks, reduces the skills and experience requirements for on-site experts, improves the performance and diagnostic capabilities of machine learning models, and can more accurately identify construction abnormal risks.
Smart Images
Figure CN114662811B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the application field of artificial intelligence technology in oil and gas fault diagnosis technology, and in particular to a semi-supervised learning anomaly detection model fracturing construction risk early warning method. Background Art
[0002] With the rapid development of my country's social economy, the demand for energy is getting higher and higher. However, due to the limited oil and gas resources in my country, the overload exploitation of crude oil in recent years has caused many oil fields to enter the late stage of exploitation. The difficulty of exploitation has continued to increase. Most oil fields use multiple fracturing methods on existing oil wells to increase oil production.
[0003] The existing fracturing construction process is complex, and various risks such as sand plugging often occur during the fracturing construction process. However, due to the low level of informatization and intelligence in the construction process, the main risks discovered and handled during the fracturing construction process are:
[0004] 1) On-site experts monitor various construction parameters in real time and rely on their own knowledge and experience to make risk judgments, which places very high demands on the skills and experience of on-site fracturing experts.
[0005] 2) Fracturing construction risks are discovered through machine learning. However, in production practice, few historical risk data samples are discovered and retained, and the number of samples used to train the machine learning algorithm model is unbalanced, making the structure of the machine learning model relatively simple and the model performance seriously degraded. As a result, this model cannot effectively discover construction risks at the construction site during on-site fracturing construction, and cannot promptly warn of abnormal risks in fracturing construction.
[0006] Therefore, how to provide an anomaly detection model that can timely warn of abnormal risks in fracturing construction has become an urgent problem to be solved. Summary of the invention
[0007] The present invention provides a fracturing construction risk warning method based on a semi-supervised learning anomaly detection model, which is used to solve the problem in the prior art that the machine learning model is simple and cannot effectively discover the construction risks at the construction site and provide timely warnings. It is also used to solve the problem in the prior art that the skills and experience of on-site fracturing experts are required to be high when providing anomaly risk warnings for fracturing construction.
[0008] In order to achieve the above-mentioned purpose, the technical solution of the present invention provides a semi-supervised learning anomaly detection model fracturing construction risk warning method, including: preprocessing the fracturing construction data. A deep neural network model is designed based on the DBN network and the SVVD network, including the DBN network extracting the features of the front network part of the pressure anomaly detection model, and the SVVD network performing anomaly classification. Specifically, a layer-by-layer unsupervised method is used to pre-train the DBN network in combination with the pre-processed fracturing construction data to obtain an RBM, and the SVVD network receives the feature vector output by the RBM as its input feature vector. The SVVD classifier is supervisedly trained and the DBN network is fine-tuned using the labeled fracturing construction data. After the training is completed, the accuracy of the algorithm model is tested using a test sample set, including constructing a hypersphere with the smallest volume in the feature space of the SVVD. The calculation formulas for the center a and radius R of the hypersphere are respectively:
[0009]
[0010] Among them, x v ∈SV,K(x i ,x j ) is the kernel function, which is equivalent to the inner product of the sample in the feature space, that is, K(x i ,x j )=Φ(x i ),Φ(x j ). The test samples are trained based on the deep neural network model, and the accuracy of the algorithm model is tested, including calculating the distance from the test sample to the center of the hypersphere, and judging whether the test sample is a normal sample based on the result. The deep neural network model calls the inference interface to warn and diagnose the real-time fracturing construction data collected on site. If the warning or diagnosis is wrong, the error is corrected through the operation interface. The deep neural network model updates the model according to the received error correction information.
[0011] As a preferred embodiment of the above technical solution, preferably, before collecting and preprocessing the fracturing construction data, it includes: collecting samples of historical fracturing construction data segments and marking the types of fracturing anomalies therein.
[0012] As a preferred embodiment of the above technical solution, preferably, the preprocessing includes: preprocessing the collected on-site construction data, preprocessing the test sample set, including interpolating the missing values of the data and then normalizing them.
[0013] As a preferred embodiment of the above technical solution, preferably, constructing a minimum volume hypersphere in the feature space of SVVD includes: mapping the fracturing operation data from the original space to the feature space through a nonlinear transformation function Φ:x→F. Constructing a minimum volume hypersphere must satisfy:
[0014]
[0015]
[0016] Wherein, R is the radius of the hypersphere, a is the center of the hypersphere, ξ is the relaxation factor, and C is a penalty parameter that weighs the volume of the hypersphere and the misclassification rate. The dual problem is obtained according to the Lagrangian algorithm, and after solving the dual problem, the Lagrangian coefficients corresponding to all samples are obtained, thereby obtaining the calculation formula for the center a and radius R of the hypersphere.
[0017] As a preferred embodiment of the above technical solution, preferably, the dual problem is:
[0018]
[0019]
[0020] Among them, α i is a sample X i The corresponding Lagrange coefficient, α in all training samples i Satisfying 0<α i The samples with <C are support vectors, and the sample set of training samples belonging to the support vectors is SV, then the calculation formula for the center and radius of the hypersphere is obtained.
[0021] As a preferred embodiment of the above technical solution, preferably, the test sample x is calculated test？ Its distance to the center of the hypersphere:
[0022]
[0023] If d≤R, the test sample is on or inside the hypersphere and is a normal sample; otherwise, it is an abnormal sample.
[0024] As a preferred embodiment of the above technical solution, preferably, a small number of negative samples are added to the positive training set to prevent overfitting. Suppose the labels of the positive samples and negative samples in the training set are y i = +1 and y j =-1, then the dual problem is,
[0025]
[0026]
[0027] 0≤α i ≤C1,y i =+1
[0028] 0≤α i ≤C2,y i =-1
[0029] Therefore, the calculation formulas for the center a and radius R are:
[0030]
[0031]
[0032] Test sample x test‘ Its distance to the center of the hypersphere:
[0033]
[0034] If d≤R, it means that the test sample is on or inside the hypersphere and is a normal sample; otherwise, it is an abnormal sample.
[0035] As a preferred embodiment of the above technical solution, after the fracturing construction data collected in real time on site is input into the deep neural network model, the deep neural network model calls the reasoning interface of the algorithm model to perform fracturing construction risk warning and visualize the results, and notifies the fracturing construction experts to confirm and correct the fault.
[0036] As a preferred embodiment of the above technical solution, preferably, the fracturing construction data corresponding to the diagnosis errors of the deep neural network model are classified and collected, and when the amount of data reaches a set threshold, the deep neural network model is trained and updated for the second time.
[0037] The technical solution of the present invention provides a method for early warning of fracturing construction risks using a semi-supervised learning anomaly detection model, including: using DBN (deep belief network) to extract features of the front network part of the pressure anomaly detection model, followed by an SVVD anomaly detection model for anomaly classification. Since DBN is a semi-supervised network, normal samples can be used to train the DBN network model, and then a limited number of abnormal samples can be used to train the entire model, ultimately achieving a practical purpose.
[0038] The advantage of the present invention is that DBN is used as a semi-supervised learning network for feature extraction. When training the DBN network, only a large number of normal samples existing in fracturing construction need to be used for training, thereby ensuring the performance accuracy of the DBN network, and then the output of the DBN network is used as the input of the SVVD network. During training, the abnormal samples of fracturing construction are subjected to SVVD network training and fine-tuning of the entire DBN network, thereby ensuring the performance of the entire abnormal risk warning model for fracturing construction and maximizing the diagnostic capability of the system. BRIEF DESCRIPTION OF THE DRAWINGS
[0039] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, a brief introduction is given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention, and a person of ordinary skill in the art can obtain other drawings based on these drawings without paying any creative work.
[0040] Figure 1 The process of the fracturing construction risk early warning system based on the anomaly detection model of the present invention is as follows: Figure 1 .
[0041] Figure 2 The process of the fracturing construction risk early warning system based on the anomaly detection model of semi-supervised learning of the present invention is as follows: Figure 2 .
[0042] Figure 3 The DBN network structure diagram is composed of stacked RBMs;
[0043] Figure 4 The data flow diagram of SVVD as a classifier of DBN;
[0044] Figure 5 The present invention is a flowchart of the application of the fracturing construction risk early warning system based on the anomaly detection model of semi-supervised learning in a specific scenario. DETAILED DESCRIPTION
[0045] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0046] Figure 1 A schematic diagram of a flow chart provided for an embodiment of the present invention, such as Figure 1 As shown, this embodiment provides a method for early warning of fracturing construction risks based on anomaly detection model using semi-supervised learning, which mainly includes:
[0047] Step 101: Collect historical fracturing construction data segment samples.
[0048] The historical fracturing construction data segment includes data such as oil pressure, casing pressure, displacement, sand ratio, etc., and the samples in the time period are labeled with the type of fracturing anomaly. Since the historical fault types of fracturing construction of a well or the oil wells in the entire block are very few, it is not enough to collect enough fault samples and fault types, so it is necessary to collect the fracturing construction data of the entire oil field and even other oil fields, and manually classify and label the faults that occur.
[0049] Step 102: pre-process the fracturing construction data.
[0050] The missing values of fracturing construction data such as oil pressure, casing pressure, displacement, sand ratio, etc. are filled and outliers are removed. Finally, the data of each latitude are normalized to facilitate rapid convergence and performance improvement of subsequent model training.
[0051] Step 103: Design a deep neural network model based on DBN and SVVD, and use the preprocessed fracturing construction data samples for training, tuning, and deployment.
[0052] Step 104, calling the reasoning interface based on the DBN and SVVD neural network models, diagnosing the fracturing construction data collected in real time on site, automatically issuing a fracturing construction risk warning, and confirming and correcting the fracturing construction experts.
[0053] Step 105: Return the corrected case data, perform correction training on the model, and repeat steps 103 to 105.
[0054] Specifically, the fracturing construction data corresponding to the system diagnosis errors that the fracturing construction experts believe in step 104 are classified and collected. When the number reaches a set threshold, the system starts a training program to automatically train and update the deep neural network model.
[0055] The technical solution of this application is now described in detail in conjunction with specific implementation methods. Figure 2 As shown:
[0056] Step 201: Collect historical fracturing construction data.
[0057] By collecting fracturing construction data for the entire oil field block and even the entire oil field, the collected data is classified and sorted by risk type after experts manually analyze the data for each time period. The same risk type is classified into the same folder, and different risk types are placed in different folders. At the same time, it is necessary to select normal fracturing construction data that is about 100 times the risk type and put it in the same folder.
[0058] Step 202: pre-process the fracturing construction data.
[0059] Not only the fracturing construction data suitable for model training is preprocessed, but also the data collected on site is preprocessed.
[0060] Preprocessing includes interpolation of missing values in the data, and finally, after normalization, the data is input into the semi-supervised learning anomaly detection neural network algorithm model established in the next stage for training or inference prediction.
[0061] Step 203: Construct a semi-supervised learning anomaly detection neural network algorithm model. The front stage of this algorithm model uses a DBN network for feature extraction, and the back stage uses an SVDD network for classification.
[0062] Step 204: Use the test sample set to test the accuracy of the algorithm model.
[0063] For steps 203 and 204, in detail,
[0064] For DBN network: Based on the characteristics of business data, the training sample set and test sample are divided into 8:2 ratio. During the training process, the training sample set is divided into training sample and verification sample in a ratio of 6:4. During the training, the DBN network is pre-trained first, and the parameters are learned by layer-by-layer unsupervised method. Figure 3 As shown in the figure, the DBN network is composed of l layers of RBM networks stacked together, the first l-1 layers are hidden layers, and the last output layer is the visualization layer. First, the data vector x and the first hidden layer h 1 As an RBM, train the parameters of this RBM (connecting x and h 1 The weights of x and h 1 The bias of each node, etc.), and then fix the parameters of this RBM and set h 1 As a visible vector, h 2 Treat it as a hidden vector, train the second RBM, get the parameters of the second RBM, fix these parameters, and train h 2 and h 3 The RBM constructed is analogous to that, and unsupervised pre-training of the entire DBN network is used to complete the pre-training of the entire DBN network.
[0065] Afterwards Figure 4 As shown in the figure, the SVDD network receives the output feature vector of the RBM as its input feature vector, and uses labeled data (normal data, classified abnormal data) to supervisely train the SVVD classifier and fine-tune the DBN network: the post-stage anomaly classification uses the SVVD network, and the features extracted by the DBN network are input into the SVVD network for abnormal risk warning of fracturing construction.
[0066] Specifically, there is a set of positive training data x∈R n×d, where n is the number of samples and d is the feature dimension. First, the data is mapped from the original space to the feature space through the nonlinear transformation function Φ:x→F, and then a hypersphere with the smallest volume is found in the feature space. In order to construct such a minimum volume hypersphere, SVDD solves the following optimization problem:
[0067]
[0068]
[0069] In the above formula, R is the radius of the hypersphere, a is the center of the hypersphere, ξ is the relaxation factor, and C is a penalty parameter that weighs the volume of the hypersphere and the misclassification rate. Combined with the Lagrange multiplier method, the dual problem of the original problem is:
[0070]
[0071]
[0072] In the above formula, α i is a sample X i The corresponding Lagrangian coefficients. After solving the dual problem, the Lagrangian coefficients corresponding to all samples can be obtained.
[0073] In all training samples, the Lagrange coefficient satisfies 0<α i The samples with <C are called support vectors. Assuming that the sample set belonging to the support vector in the training data set is SV, the calculation formulas for the center and radius of the hypersphere are:
[0074]
[0075]
[0076] In the above formula, x v ∈SV,K(x i ,x j ) is the kernel function, which is equivalent to the inner product of samples in the feature space, that is, K(x i ,x j )=Φ(x i ),Φ(x j ).
[0077] For the test sample x test？ , and its distance to the center of the hypersphere is:
[0078]
[0079] If d≤R, it means that the test sample is on or inside the hypersphere and is a normal sample; otherwise, it is an abnormal sample.
[0080] In fact, a small number of negative samples can be added to the positive training set to prevent overfitting. Assume that the labels of the positive and negative samples in the training set are y i = +1 and y j =-1, then the dual problem of the original optimization problem becomes:
[0081]
[0082]
[0083] 0≤α i ≤C1,y i =+1
[0084] 0≤α i ≤C2,y i =-1
[0085] At this time, the calculation formula for the center and radius of the hypersphere is:
[0086]
[0087]
[0088] Then, the test sample x test？ The distance to the center of the hypersphere is
[0089]
[0090] When d≤R, it means that the test sample is on or inside the hypersphere and is a normal sample; otherwise, it is an abnormal sample.
[0091] If the test sample set is the test sample x test？ If the ratio of normal samples to abnormal samples is greater than the threshold, the accuracy of the current abnormality detection algorithm model is qualified, and step 205 is executed; otherwise, step 203 is returned.
[0092] Step 205: Call the inference interface of the anomaly detection algorithm model based on semi-supervised learning to diagnose the risk type and probability corresponding to the fracturing operation at that moment.
[0093] The fracturing construction expert confirms and corrects the diagnosis result. Specifically, if the system diagnosis is wrong, the fracturing construction expert selects the error correction button on the interface and selects the risk type corresponding to that moment.
[0094] Step 206: The anomaly detection algorithm model collects feedback correction information and counts it. When the configured threshold is reached, the model training is restarted to replace the current online algorithm model.
[0095] Statistical analysis is performed on the accuracy of diagnosis and each risk type is counted. When the number of diagnostic errors reaches the configured threshold, the background restarts the model training. After reaching the set training threshold, the new anomaly detection model replaces the existing online deployed model.
[0096] This application is now further described in conjunction with specific implementation scenarios. Figure 5 As shown:
[0097] Step 301: Collect historical fracturing construction data.
[0098] Step 302: Create a sample set of fracturing construction data.
[0099] Step 303: Design a semi-supervised anomaly detection model.
[0100] The specific design process is as shown in steps 203 and 204, which will not be described in detail here.
[0101] Step 304: Train a semi-supervised anomaly detection model.
[0102] Specifically, after the DBN network is trained using normal samples, the entire model is then trained using a limited number of abnormal samples, so that the DBN network learns the abnormal data and trains the abnormal classification ability of the SVVD network.
[0103] Step 305: deploy the semi-supervised anomaly detection model into the current system.
[0104] Step 306: Collect current fracturing construction data at the construction site.
[0105] Step 307: The semi-supervised anomaly detection model analyzes the current fracturing construction data and issues warning information.
[0106] Specifically, if the distance between the current fracturing operation data and the center of the hypersphere is less than or equal to R, it is normal data, otherwise it is abnormal data. If it is abnormal data, the SVVD network classifies the data as abnormal and then executes step 308.
[0107] Step 308: The expert determines whether the warning information is correct. If it is correct, execute step 309; otherwise, execute step 310.
[0108] Step 309: Enter the fracturing construction management workflow.
[0109] Step 310: Collect error warning information.
[0110] Mark, collect and classify erroneous warning information.
[0111] Step 311, determine whether the error warning information reaches a preset threshold, if so, execute step 312, otherwise end.
[0112] Step 312: Retrain the semi-supervised anomaly detection model.
[0113] Step 313, determine whether the upper online accuracy is reached, if so, execute step 314, otherwise return to step 312 to continue training.
[0114] Step 314: Execute step 305 after updating the current online model.
[0115] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A semi-supervised learning anomaly detection model fracturing construction risk early warning method, characterized in that: include: Collect samples of historical fracturing construction data segments and mark the types of fracturing anomalies therein; Preprocessing of fracturing construction data; Designing a deep neural network model based on a DBN network and an SVVD network, including: after using normal samples to extract features from the front network part of the DBN network pressure anomaly detection model, the SVVD network performs anomaly classification, specifically, using a layer-by-layer unsupervised method to pre-train the DBN network in combination with the normal samples in the pre-processed fracturing construction data to obtain an RBM, and after the SVVD network receives the feature vector output by the RBM as its input feature vector, the abnormal samples in the pre-processed fracturing construction data are input into the current deep neural network model for training; Then, the SVVD classifier and the DBN network are supervisedly trained using the labeled fracturing construction data. After the training is completed, the accuracy of the algorithm model is tested using the test sample set. A minimum hypersphere is constructed in the feature space of the SVVD: Training the test sample based on the deep neural network model and testing the accuracy of the algorithm model, including calculating the distance from the test sample to the center of the hypersphere, and judging whether the test sample is a normal sample according to the result; The deep neural network model calls the reasoning interface to warn and diagnose the fracturing construction data collected in real time on site, and if the warning or diagnosis is wrong, the error is corrected through the operation interface; wherein, after the fracturing construction data collected in real time on site is input into the deep neural network model, the deep neural network model calls the reasoning interface of the algorithm model to warn of the risk of fracturing construction and visualizes the results, notifying the fracturing construction experts to confirm and correct the fault; The deep neural network model updates the model according to the received error correction information, including classifying and collecting the fracturing construction data corresponding to the diagnosis errors of the deep neural network model. When the amount of data reaches a set threshold, the deep neural network model is trained and updated for the second time.
2. The method for early warning of fracturing construction risk using anomaly detection model based on semi-supervised learning according to claim 1, characterized in that: The preprocessing includes: preprocessing the collected on-site construction data, preprocessing the test sample set, including performing interpolation processing on missing values of the data and then performing normalization processing.