A data-driven fault diagnosis method for electric submersible screw pump production system

Abstract

The present invention discloses a fault diagnosis method for an electric submersible screw pump oil production system based on data drive, which belongs to the field of oilfield exploitation technology and includes the following steps: step 1, obtaining the active power signal of the electric submersible screw pump and performing data preprocessing; step 2, constructing a feature extractor based on the contrast learning idea and a one-dimensional convolutional neural network to extract the active power signal feature; step 3, constructing an electric submersible screw pump fault diagnosis model based on an extreme gradient boosting algorithm, and using the active power signal feature as the input of the electric submersible screw pump fault diagnosis model to obtain the diagnosis result of the electric submersible screw pump fault. The present invention is suitable for the exploitation of complex oil wells, can improve the fault diagnosis accuracy, and thus improve the oilfield exploitation efficiency and equipment service life.

Description

Technical Field

[0001] The invention belongs to the technical field of oilfield exploitation, and in particular relates to a fault diagnosis method for an electric submersible screw pump oil production system based on data drive. Background Art

[0002] The screw pump plays an irreplaceable role in the exploitation of sand-containing and heavy oil reservoirs with its strong adaptability to well fluids. The screw pump lifting technology is based on the principle of single-screw hydraulic machinery. Through the cycloid internal meshing helical gear, a continuous sealed cavity is formed to lift the crude oil to the ground. In the early days, the screw pump mainly worked in the form of ground drive. The drive motor was installed at the wellhead. The torque was transmitted through the bare rod to indirectly drive the downhole screw pump to rotate. The starting torque was large, and there was also eccentric wear of the rod and pipe in the inclined well. In order to eliminate the problems of eccentric wear of the rod and pipe in the production of ground-driven screw pumps, the electric submersible screw pump oil production system came into being. The electric submersible screw pump oil production system mainly includes the ground part and the downhole part. The ground part mainly includes transformers, frequency converters, and junction boxes. The downhole part mainly includes oil drain valves, check valves, cables, screw pumps, flexible shafts, motor protectors, and submersible motors. The technicians installed the drive motor underground, thereby changing the ground drive of the screw pump to the downhole drive. Under this driving mode, the screw pump is driven by the flexible shaft to achieve rotation. Therefore, there is no need to use a sucker rod, eliminating problems related to the sucker rod, such as eccentric wear of the rod and tube and breakage of the sucker rod.

[0003] At present, many oil fields have successfully applied the electric submersible screw pump oil production system. Although the electric submersible screw pump oil production system eliminates the problem of eccentric wear of the rod and pipe, other faults may occur during the operation, such as motor burnout, bearing damage, oil pipe leakage, oil pipe wax deposition, etc., which will reduce the operating efficiency and service life of the equipment; in addition, with the change of oil well conditions, the unreasonable production parameters of the electric submersible screw pump oil production system will also reduce the operating efficiency of the electric submersible screw pump oil production system. The engineering significance of the fault diagnosis of the electric submersible screw pump is to timely and accurately detect the potential faults of the electric submersible screw pump oil production system, reduce the downtime caused by the fault, and ensure the safety of equipment operation.

[0004] The existing method for fault diagnosis of electric submersible screw pump oil production system mainly relies on traditional manual experience and expert system, which has low diagnostic accuracy and low efficiency. Summary of the invention

[0005] In order to solve the above problems, the present invention proposes a data-driven fault diagnosis method for an electric submersible screw pump oil production system. For the active power signal of the electric submersible screw pump, a contrastive learning feature extraction method is used for feature extraction. The contrastive learning feature extraction method aims to minimize the distance between positive sample pairs and maximize the distance between negative sample pairs. Through contrastive learning of sample pairs, deep and high-dimensional features of small sample data are obtained. In addition, this method is completely data-driven, which reduces the subjectivity of feature extraction.

[0006] The technical solution of the present invention is as follows:

[0007] A data-driven fault diagnosis method for an electric submersible screw pump oil production system comprises the following steps:

[0008] Step 1, obtaining the active power signal of the electric submersible screw pump and performing data preprocessing;

[0009] Step 2: Based on the contrastive learning idea and one-dimensional convolutional neural network, a feature extractor is constructed to extract the active power signal features;

[0010] Step 3: construct an electric submersible screw pump fault diagnosis model based on the extreme gradient boosting algorithm, and use the active power signal characteristics as the input of the electric submersible screw pump fault diagnosis model to obtain the diagnosis result of the electric submersible screw pump fault.

[0011] Furthermore, the data preprocessing in step 1 includes missing value supplementation, outlier detection, and data noise reduction; the specific process is:

[0012] Step 1.1, obtain the current oilfield historical electric submersible screw pump well active power signal data from the electric submersible screw pump production system, and use the linear interpolation method to fill the missing values ​​in the active power signal data. The interpolation method is as follows:

[0013] (1);

[0014] in, is the missing point; , are two different data points; , is the horizontal coordinate of different data points, , is the vertical coordinate of different data points; For missing points The missing values ​​to be filled;

[0015] Step 1.2: Use box plot to detect outliers in active power signal data; for sequences with length A set of data is arranged in ascending order, and the position of the lower quartile, the median, the upper quartile, the upper edge and the lower edge are calculated respectively, and then a box plot is generated. Data points exceeding the upper or lower edge of the box of the box plot are judged as abnormal points, that is, abnormal values; the specific formula is as follows:

[0016] (2);

[0017] (3);

[0018] (4);

[0019] (5);

[0020] (6);

[0021] in, The lower quartile location; is the median location; Upper quartile location; is the upper edge; for the lower edge;

[0022] Step 1.3: Use wavelet decomposition technology to perform noise reduction on active power signal data.

[0023] Furthermore, the specific process of step 1.3 is as follows:

[0024] Step 1.3.1, perform continuous wavelet signal decomposition on the original active power signal data, and use a high-pass filter and a low-pass filter to decompose the signal into a high-frequency approximate part and a low-frequency approximate part each time; the decomposition process is a tree structure, and the specific decomposition process is: first, the original active power signal data is decomposed into a first high-frequency approximate part and a first low-frequency approximate part through a high-pass filter and a low-pass filter; then, the first high-frequency approximate part is further decomposed into a second high-frequency approximate part and a second low-frequency approximate part using a high-pass filter and a low-pass filter; at the same time, the first low-frequency approximate part is also further decomposed into a third high-frequency approximate part and a third low-frequency approximate part using a high-pass filter and a low-pass filter; and so on, until the last decomposition is completed; wherein, the original active power signal data is a one-dimensional signal; the low-pass filter allows low-frequency signals to pass through and blocks high-frequency signals; the high-pass filter is opposite to the low-pass filter, allowing high-frequency signals to pass through and blocking low-frequency signals;

[0025] For continuous wavelet signal decomposition, the wavelet decomposition coefficient calculation formula is as follows:

[0026] (7);

[0027] in, is the wavelet decomposition coefficient; is the scaling factor; is the translation factor; is the original active power signal data; is the wavelet function; For time;

[0028] Step 1.3.2, use the maximum and minimum thresholds to perform noise reduction on the wavelet decomposition coefficients; pre-set the maximum threshold and minimum threshold. If the current wavelet decomposition coefficient is greater than the maximum threshold, the current wavelet decomposition coefficient is determined to be a signal and retained; if the current wavelet decomposition coefficient is less than the minimum threshold, the current wavelet decomposition coefficient is determined to be noise and removed; the calculation formula of the maximum threshold or the minimum threshold is as follows:

[0029] (8);

[0030] in, is the maximum threshold or the minimum threshold; is the signal length; when hour, is the maximum threshold, when hour, is the minimum threshold;

[0031] Step 1.3.3, perform wavelet signal reconstruction. Wavelet signal reconstruction is the inverse process of wavelet signal decomposition, that is, using the denoised wavelet decomposition coefficients to restore the signal. The specific calculation of wavelet signal reconstruction is as follows:

[0032] (9);

[0033] in, is the restored noise reduction signal;

[0034] Step 1.4: Divide the preprocessed data set into training set and test set in proportion.

[0035] Furthermore, the specific process of step 2 is as follows:

[0036] Step 2.1, introduce the contrastive learning idea to build a contrastive learning framework; the contrastive learning idea is to learn by comparing the similarities of sample pairs, so that the feature extractor can ensure that the distance between samples of the same category in the feature space is minimized and the distance between samples of different categories is maximized;

[0037] Step 2.2: Under the contrastive learning framework, a one-dimensional convolutional neural network is used in the feature extractor to extract the active power signal features of the submersible screw pump, and automatically learn the effective features in the one-dimensional data; the feature extractor includes six convolutional pooling combinations connected in sequence, each combination includes a convolutional layer and a pooling layer.

[0038] Furthermore, the specific process of step 2.1 is: traverse the training set to form a positive sample set and negative sample set , each sample in the positive sample set is composed of two samples of the same category, and each sample in the negative sample set is composed of two samples of different categories;

[0039] set up Indicates the first The first sample in the sample pair is of the category ; The features extracted after the feature extractor are as follows:

[0040] (10);

[0041] in, for Features extracted after passing through the feature extractor; It is the feature extraction operation;

[0042] Calculate the distance between positive sample pairs, the formula is as follows:

[0043] (11);

[0044] in, is the first The distance between pairs of samples; is the first The first sample in a pair of samples; is the first The second sample in a pair of samples;

[0045] Calculate the distance between negative sample pairs, the formula is as follows:

[0046] (12);

[0047] in, is the negative sample set The distance between pairs of samples; is the negative sample set The first sample in a pair of samples; is the negative sample set The second sample in a pair of samples.

[0048] Furthermore, the specific process of step 2.2 is as follows:

[0049] Step 2.2.1, divide the active power signals of different fault samples into several positive and negative sample pairs; bring the positive and negative sample pairs into 6 convolution pooling combinations for convolution pooling operations, obtain a feature map after the convolution pooling operation, flatten the feature map and pass it to the fully connected layer for classification, and then pass it to the output layer;

[0050] Step 2.2.2: Input the positive and negative sample pairs into the feature extractor, and perform convolution layer and pooling layer calculations in sequence; the calculation formula of the convolution layer is:

[0051] (13);

[0052] in, For the The first convolutional layer The output of a neuron; is the activation function; , are the different neuron numbers of different convolutional layers; is the first convolutional layer in the previous The set of all features of a neuron; For the In the convolutional layer The output corresponding to each neuron; is the first convolutional layer in the previous The neuron is connected to the In the convolutional layer The weights of the neurons; For the In the convolutional layer The bias of each neuron;

[0053] There are two pooling methods for the pooling layer, namely average pooling and maximum pooling. For signal data containing noise, average pooling is used, and for signal data containing strong edge features, maximum pooling is used. The specific calculation formula is:

[0054] (14);

[0055] in, For the In the pooling layer The output of the pooling window; To find the maximum value; To find the average value; It is a variable index used to traverse the elements in the sliding window; is the size of the sliding window; For the In the pooling layer The pooling window starts at Output: is the pooling stride;

[0056] Step 2.2.3: Send the output of the feature extractor to the fully connected layer, and finally extract the active power signal feature. The calculation formula is:

[0057] (15);

[0058] in, is the first Outputs; is the input to the fully connected layer individual data; is transposed; is the first The neuron is connected to the In the fully connected layer The weights of the neurons; For the In the fully connected layer The bias of each neuron;

[0059] Step 2.2.4: Convert the sample distance in contrastive learning into a sample pair classification problem, and use cross entropy as the loss function for network training. The cross entropy loss function is as follows:

[0060] (16);

[0061] in, is the total number of samples in the training set; For the The loss of sample pairs; is the category index; is the total number of categories to which the sample belongs; For the The symbolic function of the class, when If the true category of a sample pair is the same as the model classification result, the value is 1, otherwise the value is 0; For the model The sample pairs are divided into The probability of the class.

[0062] Furthermore, in step 3, the calculation formula of the electric submersible screw pump fault diagnosis model constructed by using the extreme gradient boosting algorithm is as follows:

[0063] (17);

[0064] In the formula, is the final output result of the model, which is in the form of numerical codes of 0-6. Each numerical code corresponds to a working condition type. When the output numerical code is 0, 1, 2, 3, 4, 5, and 6, the corresponding diagnostic results of the electric submersible screw pump fault are normal, low working parameters, oil pipe leakage, oil pipe wax deposition, stator debonding, flexible shaft breakage, and stator swelling. The active power signal features are extracted in step 2 and used as the input of this model; is the number of decision trees in the model; No. The output of a decision tree.

[0065] Beneficial technical effects brought by the present invention: The method of the present invention is based on data-driven, and under the basic framework of comparative learning, a one-dimensional convolutional neural network feature extractor is trained by small samples to extract the invariant features of the active power signal of the electric submersible screw pump, and establish a fault diagnosis model for the electric submersible screw pump. The present invention is suitable for mining operations in complex oil wells, can improve the accuracy of fault diagnosis, and thus improve the mining efficiency of oil fields and the service life of equipment, and has broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 The present invention is a flow chart of a fault diagnosis method for an electric submersible screw pump oil production system based on data drive.

[0067] Figure 2 Schematic diagram of the box plot structure of the present invention.

[0068] Figure 3 Schematic diagram of the comparative learning framework of the present invention.

[0069] Figure 4 It is the network structure and feature extraction flow chart of the feature extractor under the comparative learning framework of the present invention.

[0070] Figure 5 This is the diagnostic confusion matrix diagram of the contrastive learning-XGB fault diagnosis model combination in the experiment of the present invention.

[0071] Figure 6 This is the diagnosis confusion matrix diagram of the three-layer wavelet packet-XGB fault diagnosis model combination in the experiment of the present invention.

[0072] Figure 7 This is a diagnostic confusion matrix diagram of the time domain statistics-DNN fault diagnosis model combination in the experiment of the present invention. DETAILED DESCRIPTION

[0073] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0074] like Figure 1 As shown in the figure, a fault diagnosis method for an electric submersible screw pump oil production system based on data drive is used to collect production parameters of electric submersible screw pump wells, and to mine and analyze the information contained therein through data features, and to realize data closed loop by combining knowledge application artificial intelligence, that is, to perform fault diagnosis on the electric submersible screw pump based on data drive, which specifically includes the following steps:

[0075] Step 1: Obtain the active power signal of the electric submersible screw pump and perform data preprocessing, including missing value supplementation, outlier detection, and data noise reduction; the specific process is as follows:

[0076] Step 1.1, obtain the current oilfield historical electric submersible screw pump well active power signal data from the electric submersible screw pump production system, and use the linear interpolation method to fill the missing values ​​in the active power signal data. The interpolation method is as follows:

[0077] (1);

[0078] in, is the missing point; , are two different data points; , is the horizontal coordinate of different data points, , is the vertical coordinate of different data points; For missing points The missing values ​​to be filled;

[0079] Step 1.2: Use box plot to detect abnormal values ​​of active power signal data. The box plot structure diagram is as follows: Figure 2 As shown. For a sequence length of For a set of data, arrange them in ascending order, and calculate the position of the lower quartile, the median, the upper quartile, the upper edge, and the lower edge respectively. By calculating these statistics, a box plot is generated. The box plot contains most of the normal data, and the data points that exceed the upper or lower edge of the box are judged as abnormal points, that is, abnormal values. The specific formula is as follows:

[0080] (2);

[0081] (3);

[0082] (4);

[0083] (5);

[0084] (6);

[0085] in, The lower quartile location; is the median location; Upper quartile location; is the upper edge; for the lower edge;

[0086] Step 1.3: Use wavelet decomposition technology to reduce the noise of active power signal data to improve the accuracy of the data. The specific process is as follows: Figure 3 As shown:

[0087] Step 1.3.1, perform continuous wavelet signal decomposition on the original active power signal data, and use a high-pass filter and a low-pass filter to decompose the signal into a high-frequency approximation part and a low-frequency approximation part each time; the decomposition process can be understood as a tree structure, and the specific decomposition process is: first, the original active power signal data is decomposed into a first high-frequency approximation part and a first low-frequency approximation part through a high-pass filter and a low-pass filter; then, the first high-frequency approximation part is further decomposed into a second high-frequency approximation part and a second low-frequency approximation part using a high-pass filter and a low-pass filter; at the same time, the first low-frequency approximation part is also further decomposed into a third high-frequency approximation part and a third low-frequency approximation part using a high-pass filter and a low-pass filter; and so on, until the last decomposition is completed. Among them, the original active power signal data is a one-dimensional signal; the low-pass filter allows low-frequency signals to pass through and blocks high-frequency signals; the high-pass filter is opposite to the low-pass filter, allowing high-frequency signals to pass through and blocking low-frequency signals.

[0088] For continuous wavelet signal decomposition, the wavelet decomposition coefficient calculation formula is as follows:

[0089] (7);

[0090] in, is the wavelet decomposition coefficient; is the scaling factor; is the translation factor; is the original active power signal data; is the wavelet function; For time;

[0091] Step 1.3.2, use the maximum and minimum thresholds to perform noise reduction on the wavelet decomposition coefficients; pre-set the maximum threshold and minimum threshold. If the current wavelet decomposition coefficient is greater than the maximum threshold, the current wavelet decomposition coefficient is determined to be a signal and retained; if the current wavelet decomposition coefficient is less than the minimum threshold, the current wavelet decomposition coefficient is determined to be noise and removed; noise reduction can retain the main components of the signal and remove noise. The calculation formula for the maximum threshold or minimum threshold is as follows:

[0092] (8);

[0093] in, is the maximum threshold or the minimum threshold; is the signal length; when hour, is the maximum threshold, when hour, is the minimum threshold;

[0094] Step 1.3.3, perform wavelet signal reconstruction. Wavelet signal reconstruction is the inverse process of wavelet signal decomposition, that is, using the denoised wavelet decomposition coefficients to restore the signal. The specific calculation of wavelet signal reconstruction is as follows:

[0095] (9);

[0096] in, is the restored noise reduction signal;

[0097] In order to obtain the highest signal-to-noise ratio, the symN wavelet system is used to reduce the noise of the active power signal data. In the embodiment of the present invention, the sym4 wavelet system is specifically used to reduce the noise of the active power signal data, where 4 is the order.

[0098] Step 1.4: Divide the preprocessed data set into a training set and a test set according to a certain ratio (usually in the range of 7:3 to 9:1). In the embodiment of the present invention, some examples of data set division are shown, as shown in Table 1:

[0099] Table 1 Examples of data set division

[0100] .

[0101] Step 2: Based on the contrastive learning idea and one-dimensional convolutional neural network, a feature extractor is constructed to accurately extract fault features. The specific process is as follows:

[0102] Step 2.1, the present invention introduces the contrastive learning idea to build a contrastive learning framework. The characteristic of the contrastive learning idea is to construct positive and negative sample pairs, learn by comparing the similarity of sample pairs, so that the feature extractor ensures that the distance between samples of the same category in the feature space is minimized and the distance between samples of different categories is maximized.

[0103] Assume that the training set contains samples, the categories to which the samples belong are By traversing the training set, two subtask sets can be formed, namely the positive sample set and negative sample set Each sample in the positive sample set is composed of two samples of the same category, and each sample in the negative sample set is composed of two samples of different categories. After passing through the feature extractor, the distance between samples of the same category in the feature space is minimized, and the distance between samples of different categories is maximized.

[0104] The schematic diagram of the contrast learning framework is as follows Figure 3 As shown, Figure 3 middle is the first Sample pairs, is the first Sample pairs, Indicates the first The first sample in the sample pair is of the category ; Indicates the first The second sample in the sample pair is of the ; Indicates the negative sample set The first sample in the sample pair is of the category ; Indicates the negative sample set The second sample in the sample pair is of the . The features extracted after the feature extractor are as follows:

[0105] (10);

[0106] in, for Features extracted after passing through the feature extractor; It is the feature extraction operation;

[0107] Figure 3 In , , can be calculated in the same way as formula (10). , , They are , , Features extracted after passing through the feature extractor;

[0108] Calculate the distance between positive sample pairs, the formula is as follows:

[0109] (11);

[0110] in, is the first The distance between pairs of samples; is the first The first sample in a pair of samples; is the first The second sample in a pair of samples;

[0111] In order to ensure the convergence of the algorithm, the distance between negative sample pairs is restricted and the distance between negative sample pairs is calculated. The formula is as follows:

[0112] (12);

[0113] in, is the negative sample set The distance between pairs of samples; is the negative sample set The first sample in a pair of samples; is the negative sample set The second sample in a pair of samples;

[0114] Step 2.2: Under the contrastive learning framework, a one-dimensional convolutional neural network (1D-CNN) is used in the feature extractor to extract the active power signal features of the submersible screw pump and automatically learn the effective features in the one-dimensional data. The network structure and feature extraction process of the feature extractor are shown in the figure. Figure 4 As shown in the figure, the feature extractor includes six sequentially connected convolutional pooling combinations, each of which contains a convolutional layer and a pooling layer. The specific process of feature extraction based on the feature extractor under the contrastive learning framework is as follows:

[0115] Step 2.2.1, divide the active power signals of different fault samples such as oil pipe leakage and low parameters into several positive and negative sample pairs. 720 positive and negative sample pairs, each sample pair contains 2 samples (i.e. corresponding to Figure 4720*2 in the convolution layer), and then bring it into the 6 convolution pooling combinations for convolution pooling operations. For example, the first convolution layer has 32 convolution kernels, each convolution kernel extracts different features, and the output feature map has 32 channels. The output feature map size is 64*16 (that is, corresponding to Figure 4 32@64*16 in the example), the first pooling layer uses a 2*2 window to downsample the input feature map. After the convolution pooling operation, a feature map of size 4*1 is obtained, which is flattened and passed to the fully connected layer for classification, and then to the output layer. When the true category of the sample pair is the same as the model classification result, the value is 1, otherwise the value is 0;

[0116] Step 2.2.2: Input the positive and negative sample pairs into the feature extractor, and perform convolution layer and pooling layer calculations in sequence; the calculation formula of the convolution layer is:

[0117] (13);

[0118] in, For the The first convolutional layer The output of a neuron; is the activation function; , are the different neuron numbers of different convolutional layers, is a neuron in the previous convolutional layer, For the A neuron in a convolutional layer; is the first convolutional layer in the previous The set of all features of a neuron; For the In the convolutional layer The output corresponding to each neuron; is the first convolutional layer in the previous The neuron is connected to the In the convolutional layer The weights of the neurons; For the In the convolutional layer The bias of each neuron;

[0119] There are two pooling methods for the pooling layer, namely average pooling and maximum pooling. For signal data containing noise, average pooling may be more effective because it smoothes out the noise. For signal data containing strong edge features, maximum pooling is usually more appropriate. The specific calculation formula is:

[0120] (14);

[0121] in, For the In the pooling layer The output of the pooling window; To find the maximum value; To find the average value; It is a variable index used to traverse the elements in the sliding window; is the size of the sliding window; For the In the pooling layer The pooling window starts at Output: is the pooling stride;

[0122] Step 2.2.3: Send the output of the feature extractor to the fully connected layer, and finally extract the active power signal feature. The calculation formula is:

[0123] (15);

[0124] in, is the first Outputs; is the input to the fully connected layer individual data; is transposed; is the first The neuron is connected to the In the fully connected layer The weights of the neurons; For the In the fully connected layer The bias of each neuron;

[0125] Step 2.2.4: Convert the sample distance in contrastive learning into a classification problem of sample pairs. After the model is forward propagated, use cross entropy as the loss function for network training. Optimize the performance of the feature extractor through feedback. The cross entropy loss function as follows:

[0126] (16);

[0127] in, is the total number of samples in the training set; For the The loss of sample pairs; is the category index; is the total number of categories to which the sample belongs; For the The symbolic function of the class, when If the true category of a sample pair is the same as the model classification result, the value is 1, otherwise the value is 0; For the model The sample pairs are divided into The probability of the class.

[0128] The pooling layer of the present invention selects the maximum pooling, and the detailed output information of each layer of the feature extractor and the fully connected layer is shown in Table 2:

[0129] Table 2 Output of each layer of feature extractor

[0130] .

[0131] The present invention Figure 4 In the embodiment of , only one fully connected layer is used. In practical applications, multiple fully connected layers can also be connected in series. As shown in Table 2, the sample feature dimension extracted based on the contrastive learning method is 640, where (None, 498, 32) represents (batch size, feature map length, number of channels).

[0132] Step 3: construct a high-precision fault diagnosis model for the submersible screw pump based on the extreme gradient boosting algorithm, and use the active power signal features extracted by the feature extractor as the input of the fault diagnosis model for the submersible screw pump to obtain the diagnosis result of the fault of the submersible screw pump;

[0133] The extreme gradient boosting algorithm (eXtreme Gradient Boosting, Xgboost) belongs to the distributed forward additive model. All weak learners are summed in each iteration. The model calculation formula is as follows:

[0134] (17);

[0135] In the formula, is the final output result of the model, which is in the form of numerical codes of 0-6. Each numerical code corresponds to a working condition type. For specific comparison, see Table 1 for working condition types and corresponding numerical codes. When the output numerical codes are 0, 1, 2, 3, 4, 5, and 6, the corresponding diagnostic results of the electric submersible screw pump fault are normal, low working parameters, tubing leakage, tubing wax deposition, stator debonding, flexible shaft breakage, and stator swelling. The active power signal features are extracted in step 2 and used as the input of this model; is the number of decision trees in the model; No. The output of a decision tree.

[0136] In order to prove the feasibility and superiority of the present invention, the method of the present invention is compared with the existing method, and specifically, three fault diagnosis model combinations are used for comparative experiments. The method of combining contrastive learning feature extraction and Xgboost algorithm of the present invention is referred to as contrastive learning-XGB fault diagnosis model combination, the method of combining wavelet packet energy feature extraction and Xgboost algorithm is referred to as 3-layer wavelet packet-XGB fault diagnosis model combination, and the method of combining signal time domain statistical feature extraction and deep learning neural network is referred to as time domain statistics-DNN fault diagnosis model combination.

[0137] Among them, wavelet packet energy feature extraction belongs to the existing technology, which decomposes the low-frequency part of the signal and further refines the high-frequency part; the three-layer wavelet packet uses three-layer wavelet packets to extract energy features. Time domain statistical feature extraction belongs to the existing technology, which is a method of extracting features using the waveform, amplitude, and frequency of the signal. Deep-Learning Neural Network (DNN) belongs to the existing technology and is mainly composed of input layer, hidden layer, and output layer. The first layer is the input layer, which is responsible for receiving the original data; the last layer is the output layer, which outputs the network operation results; the middle layers are all hidden layers, which are used to extract and transform the original data. In the DNN network structure, all layers of neurons are connected in a fully connected manner. The neurons in each layer are connected to all neurons in the next layer, and the data transmission method is consistent with the fully connected layer in the convolutional neural network.

[0138] Common evaluation indicators for classification models include accuracy, precision, recall, and F1 value. The experimental part of this invention selects these evaluation indicators for model evaluation. The calculation basis of these evaluation indicators is the confusion matrix. The types of confusion matrices of the three combination methods are as follows: Figure 5-Figure 7 As shown, the columns represent the prediction categories of the fault diagnosis model, and the elements on the diagonal are the number of samples whose predicted results are the same as the actual results. It can be seen that contrastive learning-XGB is the optimal fault diagnosis model combination.

[0139] The comparison of calculation indicators of the three fault diagnosis model combinations is shown in Table 3:

[0140] Table 3 Comparison of calculation indicators of three fault diagnosis model combinations

[0141] .

[0142] In Table 3, the numerical codes 0-6 represent the operating conditions of normal, low operating parameters, oil pipe leakage, oil pipe wax deposition, stator debonding, flexible shaft breakage, and stator swelling fault labels.

[0143] Among them, the accuracy is the percentage of samples correctly classified by the model to the total number of samples, and the calculation formula is as follows:

[0144] (18);

[0145] in, is the accuracy value; Indicates that the true label is predicted as positive class; Indicates that the actual label is predicted as a negative class; Indicates that the actual label is a negative class and is predicted as a positive class; Indicates that the actual label is positive and is predicted as negative;

[0146] Precision, also known as precision, refers to the probability that the true label of a sample predicted as a positive class is a positive class sample. The calculation formula is as follows:

[0147] (19);

[0148] in, is the precision value;

[0149] Recall rate refers to the probability that a sample that is actually a positive class is diagnosed as a positive class. The calculation formula is as follows:

[0150] (20);

[0151] in, is the recall value;

[0152] In practical applications, the accuracy and recall of the model influence and restrict each other. The F1 value is a weighted sum of the accuracy and recall of the model, and is also a comprehensive reflection of the model performance. The calculation formula is as follows:

[0153] (twenty one);

[0154] in, is the F1 value;

[0155] The evaluation index results are analyzed. The F1 value is an indicator that comprehensively considers the precision and recall rate. It has the greatest weight in judging the quality of the model, so it is given priority.

[0156] It can be seen from Table 3 that the comparative learning-XGB fault diagnosis model combination of the present invention has the best overall performance on the test set. The average recall rate of the comparative learning-XGB fault diagnosis model combination of the present invention is 91.2%, which is 6.1% higher than that of the three-layer wavelet packet-XGB fault diagnosis model combination and 7.2% higher than that of the time domain statistics-DNN fault diagnosis model combination; the average precision rate of the comparative learning-XGB fault diagnosis model combination of the present invention is 93.32%, which is 4.3% higher than that of the three-layer wavelet packet-XGB fault diagnosis model combination and 6.03% higher than that of the time domain statistics-DNN fault diagnosis model combination; the average F1 value of the comparative learning-XGB fault diagnosis model combination of the present invention is 91.8%, which is 5.28% higher than that of the three-layer wavelet packet-XGB fault diagnosis model combination and 6.8% higher than that of the time domain statistics-DNN fault diagnosis model combination.

[0157] Although the average diagnostic recall rate, precision rate and F1 value of the three-layer wavelet packet-XGB fault diagnosis model combination and the time domain statistics-DNN fault diagnosis model combination are all higher than 80%, their diagnostic effects on tubing leakage and tubing wax deposition in the test set are not ideal, with diagnostic recall rates of 79.41%, 79.41%, 55.17% and 72.41%, respectively. The diagnostic recall rate, precision rate and F1 value of the comparative learning-XGB fault diagnosis model combination of the present invention for these two types all reach above 88%.

[0158] In addition, for the test set data, the comparative learning-XGB fault diagnosis model combination of the present invention has a diagnosis F1 value of more than 90% for normal, low operating parameters, wax deposition in the oil pipe, and stator swelling; and a diagnosis F1 value of more than 80% for all working conditions.

[0159] In summary, the electric submersible screw pump fault diagnosis model established by using the active power signal features extracted by contrastive learning and the Xgboost algorithm not only performs well in overall performance, but also has the most stable and reliable effect when diagnosing different working conditions.

[0160] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A data-driven fault diagnosis method for an electric submersible screw pump oil production system, characterized in that: The steps include: Step 1, obtaining the active power signal of the electric submersible screw pump and performing data preprocessing; Step 2: Based on the contrastive learning idea and one-dimensional convolutional neural network, a feature extractor is constructed to extract the active power signal features; Step 3: construct an electric submersible screw pump fault diagnosis model based on the extreme gradient boosting algorithm, and use the active power signal characteristics as the input of the electric submersible screw pump fault diagnosis model to obtain the diagnosis result of the electric submersible screw pump fault.

2. The fault diagnosis method for electric submersible screw pump oil production system based on data drive according to claim 1 is characterized in that: The data preprocessing in step 1 includes missing value supplementation, outlier detection, and data noise reduction; the specific process is: Step 1.1, obtain the current oilfield historical electric submersible screw pump well active power signal data from the electric submersible screw pump production system, and use the linear interpolation method to fill the missing values ​​in the active power signal data. The interpolation method is as follows: (1)； in, is the missing point; , are two different data points; , is the horizontal coordinate of different data points, , is the vertical coordinate of different data points; For missing points The missing values ​​to be filled; Step 1.2: Use box plot to detect outliers in active power signal data; for sequences with length A set of data is arranged in ascending order, and the position of the lower quartile, the median, the upper quartile, the upper edge and the lower edge are calculated respectively, and then a box plot is generated. Data points exceeding the upper or lower edge of the box of the box plot are judged as abnormal points, that is, abnormal values; the specific formula is as follows: (2)； (3)； (4)； (5)； (6)； in, The lower quartile location; is the median location; Upper quartile location; is the upper edge; for the lower edge; Step 1.3: Use wavelet decomposition technology to perform noise reduction on active power signal data.

3. The fault diagnosis method for electric submersible screw pump oil production system based on data drive according to claim 2 is characterized in that: The specific process of step 1.3 is as follows: Step 1.3.1, perform continuous wavelet signal decomposition on the original active power signal data, and use a high-pass filter and a low-pass filter to decompose the signal into a high-frequency approximate part and a low-frequency approximate part each time; the decomposition process is a tree structure, and the specific decomposition process is: first, the original active power signal data is decomposed into a first high-frequency approximate part and a first low-frequency approximate part through a high-pass filter and a low-pass filter; then, the first high-frequency approximate part is further decomposed into a second high-frequency approximate part and a second low-frequency approximate part using a high-pass filter and a low-pass filter; at the same time, the first low-frequency approximate part is also further decomposed into a third high-frequency approximate part and a third low-frequency approximate part using a high-pass filter and a low-pass filter; and so on, until the last decomposition is completed; wherein, the original active power signal data is a one-dimensional signal; the low-pass filter allows low-frequency signals to pass through and blocks high-frequency signals; the high-pass filter is opposite to the low-pass filter, allowing high-frequency signals to pass through and blocking low-frequency signals; For continuous wavelet signal decomposition, the wavelet decomposition coefficient calculation formula is as follows: (7)； in, is the wavelet decomposition coefficient; is the scaling factor; is the translation factor; is the original active power signal data; is the wavelet function; For time; Step 1.3.2, use the maximum and minimum thresholds to perform noise reduction on the wavelet decomposition coefficients; pre-set the maximum threshold and minimum threshold. If the current wavelet decomposition coefficient is greater than the maximum threshold, the current wavelet decomposition coefficient is determined to be a signal and retained; if the current wavelet decomposition coefficient is less than the minimum threshold, the current wavelet decomposition coefficient is determined to be noise and removed; the calculation formula of the maximum threshold or the minimum threshold is as follows: (8)； in, is the maximum threshold or the minimum threshold; is the signal length; when hour, is the maximum threshold, when hour, is the minimum threshold; Step 1.3.3, perform wavelet signal reconstruction. Wavelet signal reconstruction is the inverse process of wavelet signal decomposition, that is, using the denoised wavelet decomposition coefficients to restore the signal. The specific calculation of wavelet signal reconstruction is as follows: (9)； in, is the restored noise reduction signal; Step 1.4: Divide the preprocessed data set into training set and test set in proportion.

4. The fault diagnosis method for electric submersible screw pump oil production system based on data drive according to claim 3 is characterized in that: The specific process of step 2 is: Step 2.1, introduce the contrastive learning idea to build a contrastive learning framework; the contrastive learning idea is to learn by comparing the similarities of sample pairs, so that the feature extractor can ensure that the distance between samples of the same category in the feature space is minimized and the distance between samples of different categories is maximized; Step 2.2: Under the contrastive learning framework, a one-dimensional convolutional neural network is used in the feature extractor to extract the active power signal features of the submersible screw pump, and automatically learn the effective features in the one-dimensional data; the feature extractor includes six convolutional pooling combinations connected in sequence, each combination includes a convolutional layer and a pooling layer.

5. The fault diagnosis method for electric submersible screw pump oil production system based on data drive according to claim 4 is characterized in that: The specific process of step 2.1 is: traverse the training set to form a positive sample set and negative sample set , each sample in the positive sample set is composed of two samples of the same category, and each sample in the negative sample set is composed of two samples of different categories; set up Indicates the first The first sample in the sample pair is of the category ; The features extracted after the feature extractor are as follows: (10)； in, for Features extracted after passing through the feature extractor; It is the feature extraction operation; Calculate the distance between positive sample pairs, the formula is as follows: (11)； in, is the first The distance between pairs of samples; is the first The first sample in a pair of samples; is the first The second sample in a pair of samples; Calculate the distance between negative sample pairs, the formula is as follows: (12)； in, is the negative sample set The distance between pairs of samples; is the negative sample set The first sample in a pair of samples; is the negative sample set The second sample in a pair of samples.

6. The fault diagnosis method for electric submersible screw pump oil production system based on data drive according to claim 5 is characterized in that: The specific process of step 2.2 is as follows: Step 2.2.1, divide the active power signals of different fault samples into several positive and negative sample pairs; bring the positive and negative sample pairs into 6 convolution pooling combinations for convolution pooling operations, obtain a feature map after the convolution pooling operation, flatten the feature map and pass it to the fully connected layer for classification, and then pass it to the output layer; Step 2.2.2: Input the positive and negative sample pairs into the feature extractor, and perform convolution layer and pooling layer calculations in sequence; the calculation formula of the convolution layer is: (13)； in, For the The first convolutional layer The output of a neuron; is the activation function; , are the different neuron numbers of different convolutional layers; is the first convolutional layer in the previous The set of all features of a neuron; For the In the convolutional layer The output corresponding to each neuron; is the first convolutional layer in the previous The neuron is connected to the In the convolutional layer The weights of the neurons; For the In the convolutional layer The bias of each neuron; There are two pooling methods for the pooling layer, namely average pooling and maximum pooling. For signal data containing noise, average pooling is used, and for signal data containing strong edge features, maximum pooling is used. The specific calculation formula is: (14)； in, For the In the pooling layer The output of the pooling window; To find the maximum value; To find the average value; It is a variable index used to traverse the elements in the sliding window; is the size of the sliding window; For the In the pooling layer The pooling window starts at Output: is the pooling stride; Step 2.2.3: Send the output of the feature extractor to the fully connected layer, and finally extract the active power signal feature. The calculation formula is: (15)； in, is the first Outputs; is the input to the fully connected layer individual data; is transposed; is the first The neuron is connected to the In the fully connected layer The weights of the neurons; For the In the fully connected layer The bias of each neuron; Step 2.2.4: Convert the sample distance in contrastive learning into a sample pair classification problem, and use cross entropy as the loss function for network training. The cross entropy loss function is as follows: (16)； in, is the total number of samples in the training set; For the The loss of sample pairs; is the category index; is the total number of categories to which the sample belongs; For the The symbolic function of the class, when If the true category of a sample pair is the same as the model classification result, the value is 1, otherwise the value is 0; For the model The sample pairs are divided into The probability of the class.

7. The data-driven fault diagnosis method for electric submersible screw pump oil production system according to claim 6 is characterized in that: In step 3, the calculation formula of the electric submersible screw pump fault diagnosis model constructed by the extreme gradient boosting algorithm is as follows: (17)； In the formula, is the final output result of the model, which is in the form of numerical codes of 0-6. Each numerical code corresponds to a working condition type. When the output numerical code is 0, 1, 2, 3, 4, 5, and 6, the corresponding diagnostic results of the electric submersible screw pump fault are normal, low working parameters, oil pipe leakage, oil pipe wax deposition, stator debonding, flexible shaft breakage, and stator swelling. The active power signal features are extracted in step 2 and used as the input of this model; is the number of decision trees in the model; No. The output of a decision tree.

Images