CN114817856B - Beam-pumping unit fault diagnosis method based on structural information retention domain adaptation network - Google Patents

Abstract

The invention relates to the technical field of petroleum industry, in particular to a pumping unit fault diagnosis method based on a structural information retention domain adaptation network, which comprises the steps of collecting pumping unit indicator diagram data to obtain a source domain data set and a target domain data set, extracting source domain features and target domain features by using a feature generator, performing supervised training by using source domain features and source domain sample labels in the source domain to update the feature generator and a classifier, clustering the target domain features by using an unsupervised clustering algorithm to obtain target domain pseudo-labels, inputting the source domain features and the source domain labels and the target domain features and the target domain pseudo-labels into a local maximum mean difference measurement formula, and performing distribution difference calculation to optimize a feature generator. The structural information holding domain adaptation network provided by the invention can be used for fault diagnosis of the oil pumping unit, can reduce the excessive requirement of fault diagnosis model training on data annotation, and can improve the generalization performance of the fault diagnosis model in different application scenes.

Description

A method for fault diagnosis of oil pumping units based on structural information preserving domain adaptive network

Technical Field

The invention relates to the technical field of petroleum industry, and in particular to a method for diagnosing oil pump faults based on a structural information preserving domain adaptive network.

Background technique

Energy issues are major issues related to social stability and the healthy and sustainable development of the national economy, and oil occupies a very important position in the energy structure. According to different oil production methods, oil production can be divided into two methods: mechanical oil production and self-flowing oil production. The most commonly used method in my country is mechanical oil production. The oil pumping unit works thousands of meters underground all year round. There are problems such as harsh environment, complex working conditions, wear, corrosion, mechanical fatigue and functional failure, which bring hidden dangers and uncertainties to the oil pumping unit.

At present, the fault diagnosis of pumping units based on artificial intelligence algorithms has achieved good performance in simulation experiments. Machine learning algorithms (such as artificial neural networks (ANN), support vector machines (SVM), etc.) have difficulties in processing dimensional and dynamic monitoring data. Deep learning algorithms can solve more complex problems and have been widely used in fault diagnosis. Deep learning algorithms such as deep belief networks (DBN), sparse autoencoders (SAE) and convolutional neural networks (CNN) have outstanding performance in fault diagnosis.

However, most current research remains at the experimental simulation stage and has certain limitations. Most studies do not take into account the problems of insufficient model generalization ability and decreased recognition accuracy caused by different model application scenarios.

Summary of the invention

The purpose of the present invention is to provide a method for pumping unit fault diagnosis based on a structural information preserving domain adaptive network, aiming to solve the problem that most current studies do not take into account different model application scenarios, resulting in insufficient model generalization ability and reduced recognition accuracy.

To achieve the above object, the present invention provides a method for fault diagnosis of an oil pumping unit based on a structural information preserving domain adaptive network, comprising collecting the oil pumping unit indicator diagram data to obtain a source domain data set and a target domain data set;

Use feature generator to extract source domain features and target domain features;

In the source domain, supervised training is performed using source domain features and source domain sample labels to update the feature generator and classifier;

Use an unsupervised clustering algorithm to cluster the target domain features to obtain the target domain pseudo-label;

The source domain features and source domain labels, as well as the target domain features and target domain pseudo labels are input into the local maximum mean difference to calculate the distribution difference and optimize the feature generator.

The dynamometer diagram refers to a closed curve diagram consisting of discrete points of displacement and load.

Among them, the feature generator is a feature extraction model composed of a convolutional neural network.

The specific steps of obtaining the target domain pseudo label include:

Randomly initialize the membership matrix according to the data set;

Calculate the optimal membership matrix and cluster centers;

The pseudo labels of the target domain are obtained using the membership matrix and cluster centers.

The specific steps of calculating the optimal membership matrix and cluster centers include:

Update cluster centers using membership matrix;

Use the cluster center to calculate the Euclidean distance from the sample to the cluster center;

Update the membership matrix using the number of clusters, cluster center distance and fuzzy weight index;

Determine whether the iteration stop condition is met based on whether the maximum number of iterations and the error are reached;

Get the algorithm's best membership matrix and cluster center;

The specific steps of obtaining the pseudo label of the target domain are as follows:

Obtain sample cluster labels based on the membership matrix;

For each sample under each cluster label, use the classifier to obtain the classification label;

Count the number of each classification label;

Select the label with the largest number of classification labels as the label of the cluster category and update the cluster center;

Update the membership matrix using the new cluster centers;

Obtain the target domain pseudo-label according to the new membership matrix.

The present invention discloses a method for diagnosing oil pump faults based on a structural information preserving domain adaptive network, which can be used for fault diagnosis of oil pumps, can reduce the excessive requirements for data labeling in fault diagnosis model training, and can improve the generalization performance of the fault diagnosis model in different application scenarios.

BRIEF DESCRIPTION OF THE DRAWINGS

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

FIG1 is a flow chart of a method for diagnosing oil pumping unit faults based on a structural information preserving domain adaptive network according to the present invention.

FIG2 is a flowchart of the specific steps of acquiring pseudo labels in the target domain.

Figure 3 is the confusion matrix of the test set results for task D1→D2.

Figure 4 is the loss curve of SIP-DAN during the training process of task D1→D2.

Figure 5 is the target domain test accuracy curve during the training process of various methods.

Figure 6 is a t-SNE graph of the test set results for task D1→D2.

Detailed ways

Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and should not be construed as limiting the present invention.

The specific embodiments of the present invention are:

Please refer to the present invention, which provides a method for diagnosing oil pump faults based on a structure information preserving domain adaptation network (Structure Information Preserving Domain Adaptation Network, SIP-DAN), comprising:

S101 collects the pumping unit indicator diagram data to obtain a source domain data set and a target domain data set;

The pumping unit dynamometer data are collected and divided into source domain data set and target domain data set.

In order to avoid the influence of different sizes between data on the calculation, the original data of the dynamometer diagram are normalized. Each dynamometer diagram consists of a set of discrete points {( _xi , _yi )} of displacement x and load y.

The displacement normalization formula is as follows:

Where: M is the number of displacement samples; N is the number of displacement points in a displacement sample;

_xik is the kth sample value of the ith displacement sample before normalization;

is the kth sample value of the i-th displacement sample after normalization;

_xi,min = min{ _xik |1≤k≤N};

_xi,max = max{ _xik |1≤k≤N}.

The load normalization formula is as follows:

Where: M is the number of load samples; N is the number of load points in a load sample;

y _ik is the kth sample value of the ith load sample before normalization;

is the kth sample value of the i-th load sample after normalization;

_yi,min = min{ _yik |1≤k≤N};

_yi,max = max{ _yik |1≤k≤N}.

Use the normalized load and displacement to draw a dynamometer diagram with displacement as the horizontal axis and load as the vertical axis, and save it as a picture sample.

S102 uses a feature generator to extract source domain features F(x ^s ) and target domain features F(x ^t );

The feature generator is a convolutional neural network.

S103 uses source domain features and source domain sample labels to perform supervised training in the source domain to update the feature generator and the classifier;

The optimization goal of the fault classifier is to minimize the classification error under the supervision of the source domain data label. In this way, recognizable fault features can be extracted. Use F(·) to represent the output of the feature extractor and C(·) to represent the output of the classifier. The classification loss uses the cross entropy loss function:

Among them, y ^s represents the one-hot form of the source domain sample label; is the output of the classifier in the source domain,

S104 uses an unsupervised clustering algorithm to cluster the target domain features to obtain a target domain pseudo label;

Here, the fuzzy C-means (FCM) clustering algorithm is used to cluster the target domain features F(x ^t ) to obtain cluster labels, and then a classifier is used to assist in the alignment of cluster labels. The steps are:

S401 randomly initializes the membership matrix U=[u _ij ] _N×M .

Where N is the number of samples, M is the number of clusters, and _uij is the membership of the i-th (i=1,2,…,N) sample to the j-th (j=1,2,…,M) class, satisfying the following relationship:

S402 uses the membership matrix to calculate the best membership matrix and cluster center

(1) Calculate cluster centers:

Among them, _cj represents the jth cluster center, N represents the number of samples, _fi represents the ith sample, and m is the fuzzy weight index.

(2) Calculate the Euclidean distance from the i-th sample _fi to the j-th cluster center _cj :

(3) Update the membership matrix u _ij :

Among them, M is the number of clusters, d _ij is the distance from the ith sample to the jth cluster center, d _ik is the distance from the ith sample to the kth cluster center, and m is the fuzzy weight index.

(4) Determine whether any of the following iteration stop conditions are met

Condition 1: Determine whether the maximum number of iterations max_iter is reached

Condition 2: Judgment error

in, is the membership matrix of the current iteration,/> is the membership matrix of the previous iteration, and ε is the error threshold.

If one of condition 1 and condition 2 is met, the iteration can be stopped.

(5) If the iteration stop condition of (4) is not met, repeat (2) to (4). If the iteration stop condition of (4) is met, stop the iteration.

The algorithm obtains the optimal membership matrix U and cluster center C.

S403 uses the membership matrix and cluster centers to obtain the pseudo labels of the target domain

(1) According to the membership matrix U, the clustering category of each sample is obtained.

(2) Input the samples of the i-th (i=1, 2, …, k) cluster category into the classifier C to obtain the classification label of the classifier.

(3) Count the number of each category in the classification label.

(4) Update the cluster center of the current cluster category to: C _{new i} = C (max (counter)). Repeat (1) to (3) so that each cluster center is updated to obtain a new cluster center C _new , and use the updated cluster center C _new to update the membership matrix U according to the formula (3) in S402 to obtain a new membership matrix U _new .

Get pseudo labels for the target domain S105 inputs the source domain features and source domain labels, as well as the target domain features and target domain pseudo labels into the local maximum mean discrepancy (LMMD) to calculate the distribution difference and optimize the feature generator.

Obtain target domain pseudo labels Then, the source domain features F( ^xs ) and labels ^ys , as well as the target domain features F( ^xt ) and pseudo labels/> Enter LMMD to calculate the distribution difference.

The basic idea of LMMD is to divide the two domains into multiple subdomains by category labels, and then perform fine-grained domain adaptation on the relevant subdomains. Assume that the weight of each sample belonging to each class is w ^c . The definition of LMMD is as follows:

Where H is the reproducing kernel Hilbert space (RKHS) of a given feature kernel k; φ(·) represents the nonlinear mapping from the original feature space to RKHS; C is the number of fault categories; is the weight of the i-th sample in the source domain belonging to subclass c;/> is the weight of the jth sample in the target domain belonging to subclass c.

Since φ(·) cannot be calculated directly, the formula is expanded as follows:

The method to calculate the weight w ^c is as follows:

Where y _ic is the probability that the ith sample belongs to the cth class, and n is the number of samples. For source domain samples, the one-hot form of y ^s is used directly; for the target domain, the sample pseudo label y ^t obtained in step (5) is used.

In this embodiment:

(1) Dataset

Three data sets (named D1, D2, and D3) were formed according to different data sources. Each data set includes normal working conditions and 6 types of faults, numbered as Normal, Fault#1, Fault#2, Fault#3, Fault#4, Fault#5, and Fault#6. The specific situation of the data set is shown in Table 1. In general, most pumping units are in a normal working state, which makes it difficult to collect fault samples. In the analyzed data set, there are many normal samples, but there are relatively few samples of some fault types (such as gas lock). In order to reduce the impact of data imbalance, the number of samples for each working condition is limited to 500.

Table 1 Dataset samples

The typical dynamometer diagrams of each data set are shown in Table 2. Due to the inconsistent acquisition conditions of the original data of different data sets, the same working condition has different shapes in different data sets. The conditional factors may be different machine models, sensor models, oil well geology, etc., which will cause distribution differences between data sets, so that the classification ability of the model trained on one data set is reduced when it is applied to another data set for testing, and it does not have good generalization performance.

Table 2 Typical dynamometer diagram shapes for each data set

(2) Experimental setup

Six domain adaptation tasks from source domain to target domain are set on three datasets D1, D2 and D3 (D1→D2, D1→D3, D2→D1, D2→D3, D3→D1, D3→D2), and the details of the data samples are shown in Table 3. For example, D1→D2 means: D1 is a labeled source domain, and D2 is an unlabeled target domain.

Table 3 Domain adaptation task settings

In order to verify the effectiveness of the SIP-DAN method, several supervised learning algorithms and domain adaptation methods are compared with the proposed method. These methods include:

SVM: Support vector machine, as a classic method of machine learning, is widely used in various fields.

ResNet18: A typical deep residual convolutional neural network, used as a benchmark for evaluating the feature extraction capabilities of domain adaptation tasks. Because ResNet18 is a supervised learning network, only labeled source domain data is used for training, and unlabeled target domain data is not used.

DAN: A deep domain adaptation method based on distance measurement, using maximum mean discrepancy (MMD) as the distribution difference metric, is a global domain adaptation method.

D-CORAL: A deep domain adaptation method based on distance measurement, using CORAL as the distribution difference metric. It linearly transforms the source and target domains to align their second-order statistics, which is a global domain adaptation method.

For SVM, the image is read into the dimension of 100*100, and the dimension is directly flattened to 10000 as the input of SVM. The SVM under the sklearn library is used for experiments, and the parameter settings are: C=1.0, γ=auto, Kenel=RBF.

For SIP-DAN, the balance parameter is set dynamically, and the calculation formula is: λ=2/(1+exp(-10*ep/epoch))-1 (where ep is the current iteration step and epoch is the total number of iterations).

The above several deep model methods all use ResNet18 as the basic feature extraction network, and the general experimental parameters are set as follows:

Table 4 Experimental parameter settings

(3) Fault diagnosis performance comparison and analysis experiment

SIP-DAN and other methods are implemented on 6 domain adaptation tasks, and Table 5 summarizes the average diagnostic accuracy of all compared methods.

First, the comparison between the traditional machine learning method SVM and the deep method ResNet18 shows that the accuracy of cross-domain fault diagnosis of SVM is very low because the traditional method does not have a special feature extraction method for the dynamometer diagram. The deep method ResNet18 has a much higher accuracy than SVM after end-to-end adaptive feature extraction and training, which shows the superiority of the deep method.

DAN and D-CORAL are both representative methods of global domain adaptation. The average accuracy of the six tasks is 86.96% and 88.33% respectively, which is 1.57% and 3.17% higher than that of ResNet18 without domain adaptation. It can be found that global domain adaptation has limited improvement on the accuracy of the model. This is because the global domain adaptation method ignores the importance of category information in the domain adaptation process, resulting in no obvious class discriminant features being extracted.

SIP-DAN is a fine-grained domain adaptation proposed by the present invention. It takes into account the category information in the domain adaptation process and also considers the preservation of the target domain feature information, thus achieving a better accuracy. Among the 6 tasks, the best accuracy was achieved on 4 tasks, and the average accuracy was improved by 10.59% compared with ResNet18, which played a better effect of source domain knowledge transfer.

Table 5 Results of various methods

(4) Confusion Matrix Comparison

Taking the task D1→D2 as an example, the confusion matrix result of the target domain test set is shown in Figure 3. From the results of ResNet18, the distribution difference of the D1→D2 task is mainly reflected in the categories Normal and Falut#2. ResNet18 misidentifies 55% of Normal samples as Fault#1 and 13% of Normal samples as Fault#6; ResNet18 misidentifies 32% of Fault#2 samples as Fault#1 and 9% of Fault#2 samples as Fault#3. The misidentification situation was not improved after DAN domain adaptation. The recognition of Normal was improved after D-CORAL domain adaptation, but the recognition effect of Fault#2 was worse. As shown in (d), the proposed method SIP-DAN greatly improves the model fault diagnosis capability.

(5) Algorithm convergence analysis experiment

Taking task D1→D2 as an example, the change of loss value with the number of training iterations during SIP-DAN training is shown in Figure 4. It can be found that the classification loss converges very quickly at the beginning of training. At the beginning, the domain adaptation balance parameter λ is small, and the domain adaptation loss accounts for a small proportion of the total loss value, so the total loss is also small. As the training progresses, λ gradually increases, and the proportion of LMMD domain adaptation loss in the total loss increases. The total loss curve shows an increasing trend in the early stage of training. However, with the domain adaptation training of the network, the distribution difference (LMMD) between the source domain and the target domain becomes smaller and smaller, so the total loss curve shows a decreasing trend in the middle of training, and finally converges in the middle and late stages of training.

Taking task D1→D2 as an example, the changes of the test set accuracy with the number of training steps during the training process of the four methods are shown in Figure 5. It can be seen that DAN converges the slowest and has the lowest final accuracy. The convergence speeds of the other three methods are not much different, and the final accuracy is about 80%. The accuracy of SIP-DAN proposed in the present invention is the best, and the training process is stable without large fluctuations.

(6) Feature visualization experiment

In order to observe the feature distribution after domain adaptation, t-SNE is used to reduce the dimension of the last feature layer of the deep network to 2 dimensions and then visualized. The t-SNE dimension reduction visualization of the task D1→D2 test set is shown in Figure 6. It can be seen that there is an obvious offset in the feature distribution of ResNet18. After global domain adaptation of DAN and D-CORAL, the feature distribution is not completely aligned, but there is a certain effect. SIP-DAN can achieve better performance by aligning the distribution of subclasses. The alignment of the feature distribution of the 7 categories is significantly higher than that of the other three methods.

What is disclosed above is only a preferred embodiment of the present invention, and it certainly cannot be used to limit the scope of rights of the present invention. Ordinary technicians in this field can understand that all or part of the processes of the above embodiment and equivalent changes made according to the claims of the present invention still fall within the scope of the invention.

Claims (4)

1. A pumping unit fault diagnosis method based on a structural information retention domain adaptation network is characterized in that,

Collecting oil pumping unit indicator diagram data to obtain a source domain data set and a target domain data set;

Extracting source domain features and target domain features using a feature generator;

performing supervised training in a source domain by using source domain features and source domain sample labels to update a feature generator and a classifier;

clustering the target domain features by using an unsupervised clustering algorithm to obtain target domain pseudo tags;

Inputting the source domain features and the source domain labels and the target domain features and the target domain pseudo labels into a local maximum mean difference measurement formula to perform distribution difference calculation so as to optimize a feature generator;

the specific steps for obtaining the target domain pseudo tag comprise:

randomly initializing a membership matrix according to the data set;

Calculating an optimal membership matrix and a clustering center;

obtaining a pseudo tag of the target domain by using the membership matrix and the clustering center;

the specific steps of obtaining the pseudo tag of the target domain are as follows:

obtaining a sample clustering label according to the membership matrix;

Aiming at the samples under each cluster label, a classifier is utilized to obtain a classification label;

Counting the number of each classification label;

selecting the label with the largest number of classified labels as the label of the clustering class, and updating the clustering center;

updating the membership matrix by using a new cluster center;

And obtaining the target domain pseudo tag according to the new membership matrix.

2. The method for diagnosing the failure of the pumping unit based on the structural information retention domain adaptive network according to claim 1, wherein,

The indicator diagram refers to a closed graph formed by displacement and load discrete points.

3. The method for diagnosing the failure of the pumping unit based on the structural information retention domain adaptive network according to claim 1, wherein,

The feature generator is a feature extraction model formed by a convolutional neural network.

4. A pumping unit fault diagnosis method based on a structural information holding domain adaptive network as defined in claim 1, wherein,

The specific steps of calculating the optimal membership matrix and the clustering center comprise:

updating the clustering center by using the membership matrix;

Calculating the Euclidean distance from the sample to the clustering center by using the clustering center;

updating the membership matrix by using the cluster number, the cluster center distance and the fuzzy weight index;

Judging whether an iteration stop condition is met according to whether the maximum iteration times and the errors are reached;

obtaining the optimal membership matrix and clustering center of the algorithm.