CN112836719A - Indicator diagram similarity detection method fusing two classifications and three groups - Google Patents
Indicator diagram similarity detection method fusing two classifications and three groups Download PDFInfo
- Publication number
- CN112836719A CN112836719A CN202011443534.3A CN202011443534A CN112836719A CN 112836719 A CN112836719 A CN 112836719A CN 202011443534 A CN202011443534 A CN 202011443534A CN 112836719 A CN112836719 A CN 112836719A
- Authority
- CN
- China
- Prior art keywords
- indicator diagram
- data
- sample
- similarity
- test
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000010586 diagram Methods 0.000 title claims abstract description 181
- 238000001514 detection method Methods 0.000 title claims abstract description 56
- 238000012549 training Methods 0.000 claims abstract description 57
- 238000000034 method Methods 0.000 claims abstract description 37
- 238000012795 verification Methods 0.000 claims abstract description 33
- 238000005086 pumping Methods 0.000 claims abstract description 30
- 238000004519 manufacturing process Methods 0.000 claims abstract description 10
- 238000012360 testing method Methods 0.000 claims description 80
- 230000002159 abnormal effect Effects 0.000 claims description 22
- 238000006073 displacement reaction Methods 0.000 claims description 21
- 238000001914 filtration Methods 0.000 claims description 21
- 238000005070 sampling Methods 0.000 claims description 13
- 230000006870 function Effects 0.000 claims description 11
- 230000008569 process Effects 0.000 claims description 10
- 238000013528 artificial neural network Methods 0.000 claims description 9
- 230000014759 maintenance of location Effects 0.000 claims description 7
- 238000013527 convolutional neural network Methods 0.000 claims description 6
- 230000007774 longterm Effects 0.000 claims description 6
- 238000007781 pre-processing Methods 0.000 claims description 6
- 230000008859 change Effects 0.000 claims description 5
- 238000011176 pooling Methods 0.000 claims description 4
- 238000012545 processing Methods 0.000 claims description 4
- 238000003860 storage Methods 0.000 claims description 4
- 238000004364 calculation method Methods 0.000 claims description 3
- 238000010606 normalization Methods 0.000 claims description 3
- 238000009877 rendering Methods 0.000 claims description 2
- 238000013135 deep learning Methods 0.000 abstract description 12
- 238000004458 analytical method Methods 0.000 abstract description 7
- 230000004927 fusion Effects 0.000 abstract description 7
- 239000000523 sample Substances 0.000 description 68
- 239000003921 oil Substances 0.000 description 7
- 238000005516 engineering process Methods 0.000 description 6
- 238000010200 validation analysis Methods 0.000 description 4
- 238000003745 diagnosis Methods 0.000 description 3
- 238000000605 extraction Methods 0.000 description 3
- 238000013145 classification model Methods 0.000 description 2
- 239000013068 control sample Substances 0.000 description 2
- 239000010779 crude oil Substances 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000013459 approach Methods 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 238000013096 assay test Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 239000004576 sand Substances 0.000 description 1
- 238000007619 statistical method Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Abstract
The invention discloses a method for detecting the similarity of indicator diagrams fusing two classifications and three groups, which comprises the steps of classifying indicator diagram sample sets according to the similarity of the diagrams, and establishing a training set and a verification set of a two classification and three group model; then, respectively inputting a deep learning network based on the second classification and the triple for training to obtain a second classification similarity detection model and a triple similarity detection model; and finally, performing fusion analysis on the detection results of the graph trend similarity of the real-time indicator diagram of the pumping well and the standard indicator diagram under the stable working condition according to the trained model, and alarming the fault indicator diagram and informing field personnel. The method can reduce the proportion of wrong similarity matching pairs of the model, improve the accuracy rate of similarity matching and enhance the working condition analysis and production guidance control capability of the oil pumping well.
Description
Technical Field
The invention relates to the field of fault detection of a rod-pumped well, in particular to a method for performing data fusion on the detection results of similarity of a two-classification indicator diagram and a three-component indicator diagram and outputting comprehensive diagnosis on whether the rod-pumped well has a fault according to two models.
Background
In the field of crude oil exploitation, a pumping unit indicator diagram records the change of load and displacement in a single stroke, and the method is the most direct and effective means for analyzing the working condition of a pumping unit well. In recent years, the method for detecting the abnormal working condition of the rod-pumped well based on the indicator diagram data is widely applied to oil extraction and geological engineering. However, due to the influence of conditions such as geological conditions, most pumping units are difficult to approach the indicator diagram under the standard working condition in the engineering principle under the influence of environmental factors. In consideration of engineering, currently, in the industry, an indicator diagram of a pumping unit in long-term steady operation is generally used as a standard indicator diagram of the equipment, most of similar indicator diagrams are analyzed and filtered by a graph similarity detection method, and the equipment is judged to have a fault only when the indicator diagram which is not similar to the indicator diagram under a steady working condition appears. Compared with common graph similarity methods such as an overlap area method, a statistical analysis method and the like, the deep learning similarity detection technology has strong generalization capability, has high similarity discrimination accuracy for indicator diagrams with local inconsistency and similar graph trends, can detect the changes of the shapes and the trends of the indicator diagrams, and has strong application value.
In the study of a deep learning-based graph similarity analysis method, most scholars at home and abroad improve the correctness of neural network diagnosis by optimizing neural network parameters, and a remarkable result is achieved in the aspect. Meanwhile, with the development of the neural network technology, a new neural network structure and a new model are also applied to the graph similarity detection technology, so that the speed and the accuracy of the similarity detection technology are improved. However, for the artificial intelligence graph similarity detection technology of a single network, the accuracy improvement degree brought by the optimization of the network parameter weight value and the network structure is still limited.
In order to solve the problem, multiple deep learning graph similarity technologies can be combined in a data fusion mode, the proportion of wrong similarity matching pairs is reduced, and the accuracy of a similarity matching model is improved, so that the detection precision of abnormal working conditions of the pumping well is effectively improved, the false alarm and the false alarm of the system are reduced, the reliability and the stability of the indicator diagram fault detection system of the pumping well are enhanced, and the guidance of scientific and stable production of the oil field is promoted.
Disclosure of Invention
Aiming at the defects, the invention provides a method for detecting the similarity of indicator diagrams fusing two classifications and three groups, which comprises the steps of constructing an indicator diagram similarity two classification sample set and a three group comparison sample set, inputting the sample sets into a deep learning network respectively for training to obtain a two classification and three group similarity detection model, and carrying out fusion analysis on the graph trend similarity detection results of the real-time indicator diagram of the pumping well and the standard indicator diagram under a stable working condition to achieve the aims of accurately detecting the faults of the pumping well and optimizing the working condition analysis and production guidance of the pumping well.
The invention discloses a method for detecting the similarity of indicator diagrams by fusing two classifications and three groups, which comprises the following steps:
(1) acquiring original indicator diagram two-dimensional data of a plurality of pumping unit devices within a duration T, and performing abnormal data filtering processing on the original indicator diagram two-dimensional data to obtain an original data sample set S:
wherein s isi(Xi,Fi) Representing the original data set of the indicator diagram sample, N in total, XiDisplacement data representing the ith indicator diagram sample, FiRepresenting load data of an ith indicator diagram sample, wherein M represents the number of sampling points of displacement x and load f in each original sample data set of the indicator diagram, and j represents a jth data point of the sample data;
(2) setting the indicator diagram coordinate system of each pumping unit equipment and using the indicator diagram coordinate systemMaximum load f of production parametermaxMaximum range of vertical axis, maximum stroke xmaxFor the maximum range of the horizontal axis, the coordinate normalization is carried out on the original two-dimensional data of the indicator diagram, the original two-dimensional data are mapped to the n multiplied by n grids and the lower rounding operation is carried out, and the pixel point data corresponding to the displacement and load data after the ith indicator diagram sample is gridded are obtained
(3) According to
From the first data point of the indicator diagram two-dimensional coordinate points
Starting to draw, connecting the data points with the next point, and repeating the steps until the data points are connected to the last point in pairs
Connecting the last point and the first point in the two-dimensional coordinate points of the indicator diagram to form an indicator diagram closed curve, and repeating the steps to obtain an indicator diagram sample set after the original data sample set S is processed
(4A) Based on the similarity of indicator diagram graphs, the indicator diagram sample set is used
Set of subsamples divided into k different classes
Equalizing the categories with less samples through a data enhancement method, and then adopting a set-out method to collect indicator diagram samples
Division into training sets
And verification set
Inputting ResNet50 deep learning network, introducing L2NORM layer, triple selection layer and triple loss layer after full connection layer, training to obtain similarity detection model M based on triplestriplet；
(4B) Determining a long-term steady working condition indicator diagram calibrated by a process expert in each subsample set as a standard indicator diagram, namely an anchor point set A ═ { a }p1,2,., k }, and respectively selecting similar and dissimilar samples for each anchor point to perform superposition drawing to generate a two-classification positive sample set
The label tau is 1; sum negative sample set
The label tau is 0, and a two-classification total sample set is formed
Then, a retention method is adopted to classify the two general sample sets
Division into training sets
And verification set
Inputting the convolutional neural network for training and verification, training the loss function by adopting SoftmaxWithLoss to obtain a two-classification similarity detection model Mbinary；
(5) Made by the technologyHome slave duration TtestDetermining the current standard indicator diagram of the pumping unit equipment in the indicator diagram under the stable working condition in the period, and taking the indicator diagram as the current anchor point a of the equipmenttest；
(6) Acquiring indicator diagram data of the oil pumping unit in real time, filtering abnormal data and preprocessing the abnormal data to generate an indicator diagram stest；
(7A) Sample pairs (a)test，stest) Inputting a trained triple similarity detection model MtripletTesting is carried out to obtain the characteristic vector v after the full connection layeraAnd vtestThe unit feature vector is obtained by L2NORM layer
And
calculating the inner product of the two to obtain cosine similarity cos thetavavtestFurther, the Euclidean distance xi is obtained, if xi exceeds the calibration threshold ThL2Outputs a dissimilar tag τtripletOtherwise, a similarity label τ is outputtriplet＝1；
(7B) A is totestAnd stestN x n image input training two-classification similarity detection model M drawn by superpositionbinaryTesting to obtain classification label taubinary；
(8) Fusing the results of the two-classification and three-tuple tests, if tautriplet＝τbinaryIf the real-time indicator diagram is 0, the real-time indicator diagram is not similar to the standard indicator diagram, namely, an abnormal fault occurs in the current working condition, an alarm needs to be given and field personnel needs to be informed to take measures; otherwise, the current working condition is considered to be stable or the test result is uncertain, and the alarm is not required at present.
Specifically, the processing of filtering the abnormal data of the two-dimensional data of the historical indicator diagram includes the following steps:
(1-1) filtering the original indicator diagram data with displacement or load of any dimension data being empty, i.e. filtering the original indicator diagram data
Or
Wherein XiDisplacement data representing the ith indicator diagram sample, FiLoad data representing an ith indicator diagram sample;
(1-2) filtering the original indicator diagram data of which the displacement is inconsistent with the sampling point of the load data due to the partial missing of data of any dimension of the displacement or the load, namely:
wherein M isxNumber of sampling points, M, representing displacement x in original sample data set of indicator diagramfThe sampling point number of the load f in the original sample data set of the indicator diagram is represented, and j represents the jth data point of the sample data;
(1-3) filtering the original indicator diagram data with the displacement or load of any one-dimensional data being all zero or approximate zero, namely:
Xi＝{xij|xij0, j 1,2,.., M }, or abs (max (x) -min (x))<ε，
Or Fi＝{fij|f ij0, j 1,2,.., M }, or abs (max (f) -min (f))<ε
Wherein M represents the number of sampling points of displacement x and load f in an original sample data set of each indicator diagram, abs () is an absolute value function, max () is a maximum function, min () is a minimum function, and epsilon represents a constant close to zero, such indicator diagrams are represented as a straight line or a line with small fluctuation, and are generally data sampled when a well is shut in;
(1-4) filtering the maximum stroke f of any dimension of displacement or load data exceeding the production parametermaxOr maximum load xmaxThe original data of the indicator diagram, namely max (X)>xmaxOr max (F)>fmax；
And (1-5) filtering the original data of the indicator diagram which is abnormal due to equipment failure in the two-dimensional data of the historical indicator diagram.
Specifically, the training to obtain the similarity detection model based on the triples includes the following steps:
(4A-1) training set of Picture Format
And verification set
Marking, converting into an out-of-order list, marking a picture path and a type label p on each row in the list, and converting the list into an LMDB format commonly used by a large neural network data set;
(4A-2) defining a convolutional neural network structure, firstly defining 2 input layers of a training set and a verification set, then defining 50 convolutional layers, 50 ReLU layers, 2 global pooling layers and 1 full-connection layer of ResNet50, and finally introducing an L2NORM layer, a triple selection layer and a triple loss layer.
(4A-3) defining solver, including basic learning rate r, learning strategy P, learning rate change degree gamma, threshold Th for entering next training process, maximum iteration time tmaxWeight retention number mu, penalty factor lambda and storage snapshot model mode tsThe like;
(4A-4) training set of LMDB Format
And verification set
Inputting ResNet50 network to obtain feature vector after full connection layer
Solving the Unit feature vector normalized by L2 through L2NORM layer
(4A-5) randomly selecting a category p from the k different categories of indicator diagram sample categories as an original indicator diagram category set at a triple selection layer, and randomly selecting the category p from the k different categories of indicator diagram sample categories
Taking each image as an original indicator diagram anchor point set:
to ApEach anchor point in
Randomly selecting samples s within a class pplAs a positive sample
Randomly selecting the division
Out-of-class arbitrary sample classes
Inner sqlAs a negative sample
Get the triad
Repeating the steps for all the categories to obtain the triple set
(4A-6) introducing a triplet loss layer after the triplet selection layer, the loss function being:
wherein alpha represents the triple loss margin value, when the value in + represents [ ] is greater than zero, the value is taken as loss, otherwise, the loss is taken as zero;
(4A-7) when the z-th triplet is lost
When the value is more than zero, calculating the parameter gradient according to a derivative gradient formula as follows:
carrying out back propagation by using the gradient to update network parameters, and obtaining a similarity detection model M based on the triples after completing iterative trainingtriplet。
Specifically, the step of respectively selecting similar and dissimilar samples from each anchor point to perform superposition drawing to generate a two-classification positive and negative sample set comprises the following steps:
(4B-1) for the anchor point a of the p-th class indicator diagrampTo, for
Class internal division anchor point apSampling the outer indicator diagram without putting back to form an indicator diagram positive sample pair (a)p，spl) Wherein:
splrepresenting the p-th set of subsamples
The first indicator diagram image in the inner,
contain l in totalpA placard image; all the sub-sample sets are not intersected with each other and are combined into a whole setEach subsample set corresponds to a similar characteristic category, indicator diagrams in the categories are similar to each other, and indicator diagrams in the categories are not similar to each other;
(4B-2) adding apAnd splRendering the same n x n image by superposition, wherein apUsing the anchor color ΘaDrawing, splUsing the test color thetasDrawing, repeating the steps (1) and (2) on the full k types of indicator diagrams to obtain an indicator diagram similarity positive sample set
(4B-3) for the anchor point a of the p-th type indicator diagrampAnd is to remove
Out-of-class arbitrary sample classes
The indicator diagram obtained by sampling without putting back forms an indicator diagram negative sample pair (a)l，sql) And superposing the images on the same n multiplied by n image, repeating the steps on the full k types of indicator diagrams to obtain an indicator diagram similarity negative sample set
Specifically, the triple similarity detection model MtripletThe procedure for performing the test was as follows:
(7A-1) slave duration T according to the Process experttestDetermining the current standard indicator diagram of the pumping unit equipment in the indicator diagram under the stable working condition in the period, and taking the indicator diagram as the current anchor point a of the equipmenttest；
(7A-2) acquiring indicator diagram data of the oil pumping unit in real time, filtering abnormal data and preprocessing the abnormal data to generate an indicator diagram stest；
(7A-3) sample pairs (a)test，stest) Inputting a trained triple similarity detection model MtripletTesting is carried out to obtain the characteristic vector v after the full connection layeraAnd vtestDisclosure of the inventionSolving unit feature vector by passing L2NORM layer
And
(7A-4) calculation
And
inner product of (1) to obtain cosine similarity cos thetavavtestAnd calculating the Euclidean distance xi:
(7A-5) comparing the Euclidean distance with a calibration threshold ThL2And outputting the triple similarity detection result tautriplet：
Advantageous effects
The invention adopts the technical scheme, discloses a method for detecting the similarity of indicator diagrams by fusing two classifications and three groups, and has the following beneficial effects: and respectively constructing a deep learning network framework based on the two-classification and the three-tuple, generating a training set and a verification set, inputting the training set and the verification set into corresponding networks, performing feature extraction and training on curve trend similarity to obtain a model, analyzing a detection result by using a data fusion mode, reducing the proportion of wrong similar matching pairs of the model, and accurately judging whether the pumping unit fails. Compared with the traditional detection method, the method improves the accuracy of the similarity detection model, has strong generalization performance, and has important value for controlling the safety and stable production of the pumping unit.
Drawings
FIG. 1 is a flow chart of a method for detecting similarity of indicator diagrams fusing two classifications and three groups according to the present invention;
FIG. 2(a) is a loss rate of a triple-based similarity diagnostic model on a training set;
FIG. 2(b) is a graph of the loss rate of the similarity diagnostic model on the validation set based on triplets;
FIG. 3 is an anchor point in 34 categories of a training set and a validation set of triples;
FIG. 4(a) is a graph of the loss rate of a similarity diagnostic model on a training set based on two classes;
FIG. 4(b) is a graph of the loss rate of the similarity diagnostic model on the validation set based on two classes;
FIG. 5(a) is an example of positive and negative samples in a binary training set;
FIG. 5(b) is an example of positive and negative samples in a binary authentication set;
FIG. 6 anchors against a test set;
FIG. 7(a) is an example of a test sample for a triplet of models in a control test set;
FIG. 7(b) is an example of a test sample for a two-class model in a control test set;
detailed description of the preferred embodiment
The following describes the implementation effect of the method in the fault diagnosis of the rod-pumped well by using a specific operation flow with reference to the accompanying drawings and specific examples. It is to be understood that these examples are for illustrative purposes only and are not intended to limit the scope of the present invention, which is to be given the full breadth of the appended claims and any and all equivalent modifications thereof which may occur to those skilled in the art upon reading the present specification.
And selecting the original indicator diagram data of eight pumping units of a certain crude oil extraction enterprise between 2019 and 2020 and 6 months to verify the feasibility of the indicator diagram similarity detection method combining the two classifications and the triple. The general system flow chart of the invention is shown in fig. 1, and the specific implementation steps are as follows:
(1) abnormal data filtering and preprocessing are carried out on the two-dimensional data of the historical indicator diagram, and 4691 groups of abnormal data are eliminated to obtain an indicator diagram sample set
For a total of 114269 indicator diagrams. With a first set of valid data S of the device 1 in a set S of raw data samples1(X1,F1) For example, the set of data has M — 200 samples:
(2) maximum load f of the apparatus 1max96.60kN, maximum stroke xmaxIn the problem of fine-grained classification such as similarity detection, in order to take account of image feature quantity and calculation quantity, n is generally set to 224, and is used as the size of an input layer of a deep learning neural network. Pixel data corresponding to displacement and load data after gridding
The following were used:
(3) adopting an OpenCV curve drawing method, connecting the two-dimensional coordinate points of the indicator diagram according to the end-to-end rule according to
Drawing is started from the first data point (0,123) in the two-dimensional coordinate points of the indicator diagram and is connected with the next pointAnd so on, connecting the two phases of the two-dimensional data point to the last point (0,121); and connecting the last point and the first point in the two-dimensional coordinate points of the indicator diagram to form an indicator diagram closed curve.
(4) Based on the similarity of indicator diagram graphs, indicator diagram sample sets can be obtained
The data is divided into 42 sub-sample sets of different categories, 34 of which are selected for generating a training set and a validation set, and the remaining 8 categories are used for generating a real-time data test set in a supplementary mode. The following is for the 34 selected categories
The description is as follows:
the categories with less sample number are equalized by adding a data enhancement method such as random disturbance, after data enhancement, the total sample number is 174398, and the sample number l under each categorypThe number of the cells is 3894 at minimum, 9228 at maximum, and the overall magnitude is relatively balanced. The sub-sample sets are mutually non-intersecting and are combined into a whole set, each sub-sample set corresponds to a similar characteristic category, indicator diagrams in the categories are similar to each other, and indicator diagrams in the categories are not similar to each other:
using a leaving method to carry out a set of indicator diagram samples
Is divided into
Randomly selecting a part of each category as a verification set
Verification set
The number of samples under each category is at least 114 and at most 349, and the overall magnitude is relatively balanced. Verification set
Total number of samples 6008, and the remaining 168390 as training set
(4A) Training set of three-tuple models
And verification set
Inputting a ResNet50 deep learning network, introducing an L2NORM layer, a triple selection layer and a triple loss layer after a full connection layer, and training to obtain a similarity detection model M based on triplestriplet：
Firstly, a training set of picture formats is combined
And verification set
Marking is carried out, the data are converted into an out-of-order list, each row in the list is marked with a picture path and a type label p, and then the list is converted into an LMDB format commonly used by a large data set of a neural network.
Secondly, a convolutional neural network structure is defined, which comprises 2 input layers, 50 convolutional layers, 50 ReLU layers, 2 global pooling layers, 1 full-link layer, L2NORM layer, triple selection layer and triple loss layer of the training set and the verification set.
Then, defining a solver, including a basic learning rate r, a learning strategy P, a learning rate change degree gamma, a threshold Th for entering the next training process, and a maximum iteration number tmaxWeight retention number mu, penalty factor lambda and storage snapshot model mode tsVerification set iterative test interval tAnd the parameters are equal, and the specific parameters are shown in table 3:
TABLE 1 ternary model solver parameters
| Parameter(s) | r | P | γ | Th | tmax | μ | λ | ts | t |
| Numerical value | 0.0001 | multistep | 0.1 | 150000 | 400000 | 0.9 | 0.0005 | 4000 | 4000 |
Training set to LMDB format
And verification set
Inputting a ResNet50 network, introducing an L2NORM layer after a full connection layer to obtain a unit feature vector obtained by L2 normalization, then performing online selection through a triplet selection layer, finally solving a triplet loss objective function, performing back propagation by using a gradient to update network parameters, and completing iterative training. Visualizing the relation result of the model loss rate along with the training iteration number, wherein the training set result is shown in fig. 2(a), the verification set result is shown in fig. 2(b), and the model has better accuracy and lower loss rate when the iteration number is 136000, so that the model is selected as a triple similarity detection model Mtriplet
(4B) Determining a long-term steady working condition indicator diagram calibrated by a process expert in each subsample set as a standard indicator diagram, namely an anchor point set A ═ { a }p1, 2.., k }, as shown in fig. 3. Respectively selecting similar and dissimilar samples from each anchor point to perform superposition drawing to generate two-classification positive and negative sample sets
(tag. tau. is 1) and
(tag. tau. is 0) together form a two-class total sample set
And classifying the two general sample sets by a retention method
Division into training sets
And verification set
Anchor point a with class 1 indicator diagram1For example, pair
Class internal division anchor point a1The outer indicator diagram is sampled without putting back, the first sampling forms the indicator diagram positive sample pair (a)1，s14) The positive sample of 224 × 224 obtained by the overlay drawing is shown in fig. 4 (a); selecting and removing
Out-of-class arbitrary sample classes
Here selecting
To pair
The inner indicator diagram is sampled without being put back, and the first sampling forms an indicator diagram negative sample pair (a)l，s69) The negative sample of 224 × 224 obtained by the overlay drawing is shown in fig. 4 (b). Wherein a is1Using the anchor color ΘaPlotted as (255,0,0), s14And s69All adopt test color thetasPlotted as (0,0,255). Repeating the steps on all 34 types of indicator diagrams to obtain a two-classification total sample set
Wherein the positive sample set
Label 1, total 58659 sheets; negative sample set
The label is 0, and the total number is 53683.
Training set of binary classification models
And verification set
Inputting ResNet50 deep learning network for training verification, training to obtain a two-classification similarity detection model M by adopting SoftmaxWithLoss as a loss functionbinary：
Firstly, a training set of picture formats is combined
And verification set
Marking is carried out, the data are converted into an out-of-order list, each row in the list is marked with a picture path and a label tau, and then the list is converted into an LMDB format commonly used by a large data set of a neural network.
Secondly, defining a ResNet50 deep learning network, including 2 input layers, 50 convolutional layers, 50 ReLU layers, 2 global pooling layers and 1 full-link layer of a training set and a verification set, and finally adopting a SoftmaxWithLoss function as a loss function.
Then, defining a solver, including a basic learning rate r, a learning strategy P, a learning rate change degree gamma, a threshold Th for entering the next training process, and a maximum iteration number tmaxWeight retention number mu, penalty factor lambda and storage snapshot model mode tsAnd the verification set iteration test interval t and other parameters, wherein the specific parameters are shown in table 2:
TABLE 2 solver parameters
| Parameter(s) | r | P | γ | Th | tmax | μ | λ | ts | t |
| Numerical value | 0.0001 | multistep | 0.1 | 50000 | 100000 | 0.9 | 0.0005 | 2000 | 4000 |
Training set to LMDB format
And verification set
Inputting ResNet50 network for training verification, and visualizing the relationship result of model loss rate with training iteration number, wherein the training set result is as shown in FIG. 5(a)The result of the verification set is shown in fig. 5(b), and the model has better accuracy and lower loss rate when the number of iterations is 96000, so the model is selected as the two-class similarity judgment model Mbinary。
(5) And (3) carrying out real-time fault detection on the indicator diagram of the oil pumping unit by adopting an indicator diagram similarity detection method fusing two classifications and three groups:
taking indicator original data of 7 months in 2020 as real-time working condition data, firstly determining indicator diagram of long-term stable working condition of pumping unit equipment as standard indicator diagram according to expert system knowledge, and calibrating anchor points a of each equipmenttestAnd according to the above-mentioned rule making abnormal data filtering and data preprocessing, drawing and generating real-time indicator diagram image stest. In order to verify the generalization ability of the model, 8 classes of indicator diagram images which are not used for training are introduced here to add test data, and the anchor point set of the test data is shown in fig. 6.
A control sample set based on the triplet and binary similarity detection models is then generated. With anchor points a of a certain equipmenttestAnd its real-time indicator diagram stestFor example, the two are combined into a sample pair (a)test，stest) As shown in fig. 7(a), a test sample of the triple model can be obtained; the two images are superposed to form the same 224 x 224 image, wherein atestUsing the anchor color ΘaPlotted as (255,0,0), stestUsing the test color thetasPlotted as (0,0,255), as shown in fig. 7(b), a test sample of the binary model can be obtained. The total number of the two test sample sets was 8833.
Then, a control sample set is used for testing, and the test sample pair set is input into the trained triple similarity detection model MtripletTesting is carried out to obtain the characteristic vector v after the full connection layeraAnd vtestThe unit feature vector is obtained by L2NORM layer
And
computing
And
inner product of (2) to obtain cosine similarity
And calculates the euclidean distance ξ. Calibration threshold ThL20.9, comparing Euclidean distance xi with calibration threshold ThL2If xi exceeds the calibration threshold ThL2Outputs a dissimilar tag τtripletOtherwise, a similarity label τ is outputtripletThe results of 8833 test specimens are shown in table 3:
TABLE 3 test results of the three tuple model
| Item | The positive sample is identified as a negative sample | Identifying negative samples as positive samples | Total error match logarithm | Rate of accuracy |
| Numerical value | 270 | 308 | 578 | 93.46％ |
Meanwhile, inputting the superposed and drawn test sample set into a trained two-classification similarity detection model MbinaryAnd (6) carrying out testing. If the test result τ isbinary1, the real-time indicator diagram is similar to the standard indicator diagram; if the test result τ isbinaryAnd 0, the real-time indicator diagram is not similar to the standard indicator diagram.
The results of 8833 test specimens are shown in table 4:
TABLE 4 test results of two classification models
| Item | The positive sample is identified as a negative sample | Identifying negative samples as positive samples | Total number of error classifications | Rate of accuracy |
| Numerical value | 87 | 149 | 263 | 97.33％ |
And finally, fusing the results of the two-classification and three-tuple tests if tautriplet＝τbinaryIf the current working condition is not stable or the test result is uncertain, the current working condition does not need to be carried outAnd (6) alarming. The test results are shown in table 5 after fusion analysis:
TABLE 5 fusion assay test results
The analysis shows that the indicator diagram similarity detection method combining the two classifications and the triple has a fault detection function, can combine two deep learning similarity detection methods to reduce the proportion of model mismatching pairs, improve the similarity matching accuracy rate, and effectively identify the abnormal conditions different from the long-term stable working condition of the pumping unit. Therefore, the fault of the oil pumping unit can be accurately detected through the method, the fault is prevented from influencing the yield and the benefit of the oil field, the safe and stable production is threatened, the generalization performance of the model can be effectively improved, and the development and deployment difficulty and the labor intensity of field personnel are reduced.
Based on the embodiments of the present invention, other embodiments obtained by a person of ordinary skill in the art without any creative effort belong to the protection scope of the present invention.
Claims (5)
1. A indicator diagram similarity detection method fusing two classifications and three groups is characterized by comprising the following steps:
(1) acquiring original indicator diagram two-dimensional data of a plurality of pumping unit devices within a duration T, and performing abnormal data filtering processing on the original indicator diagram two-dimensional data to obtain an original data sample set S:
S＝{si(Xi,Fi)|i＝1,2,...,N}，
wherein s isi(Xi,Fi) Representing the original data set of the indicator diagram sample, N in total, XiDisplacement data representing the ith indicator diagram sample, FiRepresenting the load data of the ith indicator diagram sample, and M representing the displacement x and the load of each indicator diagram original sample data setThe number of sampling points of the load f, j represents the jth data point of the sample data;
(2) setting the indicator diagram coordinate system of each pumping unit equipment and the maximum load f of the production parametersmaxMaximum range of vertical axis, maximum stroke xmaxFor the maximum range of the horizontal axis, the coordinate normalization is carried out on the original two-dimensional data of the indicator diagram, the original two-dimensional data are mapped to the n multiplied by n grids and the lower rounding operation is carried out, and the pixel point data corresponding to the displacement and load data after the ith indicator diagram sample is gridded are obtained
(3) According to
From the first data point of the indicator diagram two-dimensional coordinate points
Starting to draw, connecting the data points with the next point, and repeating the steps until the data points are connected to the last point in pairs
Connecting the last point and the first point in the two-dimensional coordinate points of the indicator diagram to form an indicator diagram closed curve, and repeating the steps to obtain an indicator diagram sample set after the original data sample set S is processed
(4A) Based on the similarity of indicator diagram graphs, the indicator diagram sample set is used
Set of subsamples divided into k different classes
Equalizing the categories with less samples through a data enhancement method, and then adopting a set-out method to collect indicator diagram samples
Division into training sets
And verification set
Inputting the three-element similarity detection model M into a convolutional neural network, introducing an L2NORM layer, a triple selection layer and a triple loss layer behind a full connection layer, and training to obtain a triple-based similarity detection model Mtriplet；
(4B) Determining a long-term steady working condition indicator diagram calibrated by a process expert in each subsample set as a standard indicator diagram, namely an anchor point set A ═ { a }p1,2,., k }, and respectively selecting similar and dissimilar samples for each anchor point to perform superposition drawing to generate a two-classification positive sample set
The label tau is 1; sum negative sample set
The label tau is 0, and a two-classification total sample set is formed
Then, a retention method is adopted to classify the two general sample sets
Division into training sets
And verification set
Inputting the convolutional neural network for training and verification, training the loss function by adopting SoftmaxWithLoss to obtain a two-classification similarity detection model Mbinary；
(5) From duration T by a process experttestDetermining the current standard indicator diagram of the pumping unit equipment in the indicator diagram under the stable working condition in the period, and taking the indicator diagram as the current anchor point a of the equipmenttest；
(6) Acquiring indicator diagram data of the oil pumping unit in real time, filtering abnormal data and preprocessing the abnormal data to generate an indicator diagram stest；
(7A) Sample pairs (a)test，stest) Inputting a trained triple similarity detection model MtripletTesting is carried out to obtain the characteristic vector v after the full connection layeraAnd vtestThe unit feature vector is obtained by L2NORM layer
And
calculating the inner product of the two to obtain cosine similarity cos thetavavtestFurther, the Euclidean distance xi is obtained, if xi exceeds the calibration threshold ThL2Outputs a dissimilar tag τtripletOtherwise, a similarity label τ is outputtriplet＝1；
(7B) A is totestAnd stestN x n image input training two-classification similarity detection model M drawn by superpositionbinaryTesting to obtain classification label taubinary；
(8) Fusing the results of the two-classification and three-tuple tests, if tautriplet＝τbinaryWhen the value is equal to 0, thenThe real-time indicator diagram is considered to be dissimilar to the standard indicator diagram, namely, when the current working condition has an abnormal fault, alarming is needed and field personnel are informed to take measures; otherwise, the current working condition is considered to be stable or the test result is uncertain, and the alarm is not required at present.
2. The indicator diagram similarity detection method fusing the two classifications and the three groupings according to claim 1, wherein the abnormal data filtering processing of the two-dimensional data of the historical indicator diagram comprises the following steps:
(1-1) filtering indicator diagram original data with displacement or load of any dimension data being empty;
(1-2) filtering the original indicator diagram data which are caused by inconsistent sampling points of displacement and load data due to partial missing of data of any dimension of the displacement or the load;
(1-3) filtering original indicator diagram data of which any one-dimensional data of displacement or load is all zero or approximately zero;
(1-4) filtering the maximum stroke f of any dimension of displacement or load data exceeding the production parametermaxOr maximum load xmaxThe original data of the indicator diagram;
and (1-5) filtering abnormal indicator diagram original data caused by the failure of production equipment in the historical indicator diagram two-dimensional data.
3. The indicator diagram similarity detection method fusing the two classifications and the three groupings according to claim 1, wherein the training to obtain the similarity detection model based on the three groupings comprises the following steps:
(4A-1) training set of Picture Format
And verification set
Marking, converting into an out-of-order list, marking a picture path and a type label p on each row in the list, and converting the list into an LMDB format commonly used by a neural network large-scale data set；
(4A-2) defining a convolutional neural network structure, firstly defining 2 input layers of a training set and a verification set, then defining 50 convolutional layers, 50 ReLU layers, 2 global pooling layers and 1 full-connection layer of ResNet50, and finally introducing an L2NORM layer, a triple selection layer and a triple loss layer.
(4A-3) defining solver, including basic learning rate r, learning strategy P, learning rate change degree gamma, threshold Th for entering next training process, maximum iteration time tmaxWeight retention number mu, penalty factor lambda and storage snapshot model mode tsThe like;
(4A-4) training set of LMDB Format
And verification set
Inputting ResNet50 network to obtain feature vector after full connection layer
Solving the Unit feature vector normalized by L2 through L2NORM layer
(4A-5) randomly selecting a category p from the k different categories of indicator diagram sample categories as an original indicator diagram category set at a triple selection layer, and randomly selecting the category p from the k different categories of indicator diagram sample categories
Taking each image as an original indicator diagram anchor point set:
to ApEach anchor point in
Randomly selecting samples s within a class pplAs a positive sample
Randomly selecting the division
Out-of-class arbitrary sample classes
Inner sqlAs a negative sample
Get the triad
Repeating the steps for all the categories to obtain the triple set
(4A-6) introducing a triplet loss layer after the triplet selection layer, the loss function being:
wherein alpha represents the triple loss margin value, when the value in + represents [ ] is greater than zero, the value is taken as loss, otherwise, the loss is taken as zero;
(4A-7) when the z-th triplet is lost
When the value is more than zero, calculating the parameter gradient according to a derivative gradient formula as follows:
carrying out back propagation by using the gradient to update network parameters, and obtaining a similarity detection model M based on the triples after completing iterative trainingtriplet。
4. The indicator diagram similarity detection method fusing the two classifications and the triple according to claim 1, wherein the step of respectively selecting similar and dissimilar samples for each anchor point to perform superposition drawing to generate a two-classification positive and negative sample set comprises the following steps:
(4B-1) for the anchor point a of the p-th class indicator diagrampTo, for
Class internal division anchor point apSampling the outer indicator diagram without putting back to form an indicator diagram positive sample pair (a)p，spl) Wherein:
splrepresenting the p-th set of subsamples
The first indicator diagram image in the inner,
contain l in totalpA placard image;
(4B-2) adding apAnd splRendering the same n x n image by superposition, wherein apUsing the anchor color ΘaDrawing, splUsing the test color thetasDrawing, for all kRepeating the steps (1) and (2) to obtain a positive sample set of the similarity of the indicator diagrams
(4B-3) for the anchor point a of the p-th type indicator diagrampAnd is to remove
Out-of-class arbitrary sample classes
The indicator diagram obtained by sampling without putting back forms an indicator diagram negative sample pair (a)l，sql) And superposing the images on the same n multiplied by n image, repeating the steps on the full k types of indicator diagrams to obtain an indicator diagram similarity negative sample set
5. The indicator diagram similarity detection method fusing the two classifications and the three groups according to claim 1, wherein the three group similarity detection model MtripletThe procedure for performing the test was as follows:
(7A-1) slave duration T according to the Process experttestDetermining the current standard indicator diagram of the pumping unit equipment in the indicator diagram under the stable working condition in the period, and taking the indicator diagram as the current anchor point a of the equipmenttest；
(7A-2) acquiring indicator diagram data of the oil pumping unit in real time, filtering abnormal data and preprocessing the abnormal data to generate an indicator diagram stest；
(7A-3) sample pairs (a)test，stest) Inputting a trained triple similarity detection model MtripletTesting is carried out to obtain the characteristic vector v after the full connection layeraAnd vtestThe unit feature vector is obtained by L2NORM layer
And
(7A-4) calculation
And
inner product of (2) to obtain cosine similarity
And calculating the Euclidean distance xi:
(7A-5) comparing the Euclidean distance with a calibration threshold ThL2And outputting the triple similarity detection result tautriplet：
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011443534.3A CN112836719B (en) | 2020-12-11 | 2020-12-11 | Indicator diagram similarity detection method integrating two classifications and triplets |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011443534.3A CN112836719B (en) | 2020-12-11 | 2020-12-11 | Indicator diagram similarity detection method integrating two classifications and triplets |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112836719A true CN112836719A (en) | 2021-05-25 |
| CN112836719B CN112836719B (en) | 2024-01-05 |
Family
ID=75923526
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011443534.3A Active CN112836719B (en) | 2020-12-11 | 2020-12-11 | Indicator diagram similarity detection method integrating two classifications and triplets |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112836719B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113537010A (en) * | 2021-07-05 | 2021-10-22 | 华北电力大学（保定） | Fifteen-phase asynchronous motor rolling bearing fault diagnosis method based on single-channel diagram data enhancement and migration training residual error network |
| CN113780403A (en) * | 2021-09-07 | 2021-12-10 | 中国石油化工股份有限公司 | Oil well indicator diagram fault diagnosis and interpretation method and device |
Citations (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106709511A (en) * | 2016-12-08 | 2017-05-24 | 华中师范大学 | Urban rail transit panoramic monitoring video fault detection method based on depth learning |
| US20170308770A1 (en) * | 2016-04-26 | 2017-10-26 | Xerox Corporation | End-to-end saliency mapping via probability distribution prediction |
| EP3336763A1 (en) * | 2016-12-14 | 2018-06-20 | Conti Temic microelectronic GmbH | Device for classifying data |
| CN108805151A (en) * | 2017-05-05 | 2018-11-13 | 杭州微禾迅科技有限公司 | A kind of image classification method based on depth similitude network |
| WO2019128367A1 (en) * | 2017-12-26 | 2019-07-04 | 广州广电运通金融电子股份有限公司 | Face verification method and apparatus based on triplet loss, and computer device and storage medium |
| CN110889398A (en) * | 2019-12-19 | 2020-03-17 | 南通大学 | Multi-modal image visibility detection method based on similarity network |
| CN111144548A (en) * | 2019-12-23 | 2020-05-12 | 北京寄云鼎城科技有限公司 | Method and device for identifying working condition of pumping well |
| CN111275113A (en) * | 2020-01-20 | 2020-06-12 | 西安理工大学 | Skew time series abnormity detection method based on cost sensitive hybrid network |
| CN111295669A (en) * | 2017-06-16 | 2020-06-16 | 马克波尔公司 | Image processing system |
| CN111553193A (en) * | 2020-04-01 | 2020-08-18 | 东南大学 | Visual SLAM closed-loop detection method based on lightweight deep neural network |
| CN111723675A (en) * | 2020-05-26 | 2020-09-29 | 河海大学 | Remote sensing image scene classification method based on multiple similarity measurement deep learning |
| CN111783590A (en) * | 2020-06-24 | 2020-10-16 | 西北工业大学 | Multi-class small target detection method based on metric learning |
-
2020
- 2020-12-11 CN CN202011443534.3A patent/CN112836719B/en active Active
Patent Citations (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20170308770A1 (en) * | 2016-04-26 | 2017-10-26 | Xerox Corporation | End-to-end saliency mapping via probability distribution prediction |
| CN106709511A (en) * | 2016-12-08 | 2017-05-24 | 华中师范大学 | Urban rail transit panoramic monitoring video fault detection method based on depth learning |
| EP3336763A1 (en) * | 2016-12-14 | 2018-06-20 | Conti Temic microelectronic GmbH | Device for classifying data |
| CN108805151A (en) * | 2017-05-05 | 2018-11-13 | 杭州微禾迅科技有限公司 | A kind of image classification method based on depth similitude network |
| CN111295669A (en) * | 2017-06-16 | 2020-06-16 | 马克波尔公司 | Image processing system |
| WO2019128367A1 (en) * | 2017-12-26 | 2019-07-04 | 广州广电运通金融电子股份有限公司 | Face verification method and apparatus based on triplet loss, and computer device and storage medium |
| CN110889398A (en) * | 2019-12-19 | 2020-03-17 | 南通大学 | Multi-modal image visibility detection method based on similarity network |
| CN111144548A (en) * | 2019-12-23 | 2020-05-12 | 北京寄云鼎城科技有限公司 | Method and device for identifying working condition of pumping well |
| CN111275113A (en) * | 2020-01-20 | 2020-06-12 | 西安理工大学 | Skew time series abnormity detection method based on cost sensitive hybrid network |
| CN111553193A (en) * | 2020-04-01 | 2020-08-18 | 东南大学 | Visual SLAM closed-loop detection method based on lightweight deep neural network |
| CN111723675A (en) * | 2020-05-26 | 2020-09-29 | 河海大学 | Remote sensing image scene classification method based on multiple similarity measurement deep learning |
| CN111783590A (en) * | 2020-06-24 | 2020-10-16 | 西北工业大学 | Multi-class small target detection method based on metric learning |
Non-Patent Citations (3)
| Title |
|---|
| YONGGUANG LI ET.AL: "Defects Detection of Catenary Suspension Device Based on Image Processing and CNN", 《2019 IEEE SYMPOSIUM SERIES ON COMPUTATIONAL INTELLIGENCE (SSCI)》 * |
| 何果财等: "基于图像三元组挖掘的无监督视觉表示学习", 《计算机学报》, vol. 41, no. 12 * |
| 唐之刚: "基于卷积神经网络的故障示功图诊断方法研究", 《中国优秀硕士学位论文全文数据库（电子期刊）》, no. 01 * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113537010A (en) * | 2021-07-05 | 2021-10-22 | 华北电力大学（保定） | Fifteen-phase asynchronous motor rolling bearing fault diagnosis method based on single-channel diagram data enhancement and migration training residual error network |
| CN113537010B (en) * | 2021-07-05 | 2023-09-01 | 华北电力大学（保定） | Fifteen-phase asynchronous motor rolling bearing fault diagnosis method based on single-channel diagram data enhancement and migration training residual error network |
| CN113780403A (en) * | 2021-09-07 | 2021-12-10 | 中国石油化工股份有限公司 | Oil well indicator diagram fault diagnosis and interpretation method and device |
| CN113780403B (en) * | 2021-09-07 | 2024-04-26 | 中国石油化工股份有限公司 | Fault diagnosis and interpretation method and device for oil well indicator diagram |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112836719B (en) | 2024-01-05 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112508105B (en) | Fault detection and retrieval method for oil extraction machine | |
| CN109858509A (en) | Based on multilayer stochastic neural net single classifier method for detecting abnormality | |
| CN111639678B (en) | INS/GPS integrated navigation fault detection and diagnosis method based on integrated neural network | |
| CN112766301B (en) | Oil extraction machine indicator diagram similarity judging method | |
| CN112560919A (en) | Man-machine asynchronous recognition method based on one-dimensional interpretable convolutional neural network | |
| CN112419299A (en) | Bolt loss detection method, device, equipment and storage medium | |
| CN112070134A (en) | Power equipment image classification method and device, power equipment and storage medium | |
| CN114509266B (en) | Bearing health monitoring method based on fault feature fusion | |
| CN112836719A (en) | Indicator diagram similarity detection method fusing two classifications and three groups | |
| CN109145956B (en) | Scoring method, scoring device, computer equipment and storage medium | |
| CN104794729A (en) | SAR image change detection method based on significance guidance | |
| CN111122811A (en) | Sewage treatment process fault monitoring method of OICA and RNN fusion model | |
| CN111124896B (en) | Metamorphic test system for primary and secondary peak ratio calculation algorithm | |
| CN115546223A (en) | Method and system for detecting loss of fastening bolt of equipment under train | |
| Chou et al. | SHM data anomaly classification using machine learning strategies: A comparative study | |
| CN107038143A (en) | Belt conveyer scale method for diagnosing faults based on improved multilayer artificial immune network model | |
| CN115221963B (en) | Data-driven nuclear-grade pipeline resonance fault detection method and system | |
| CN111415326A (en) | Method and system for detecting abnormal state of railway contact net bolt | |
| CN116338545A (en) | Method, system, equipment and medium for identifying metering error state of current transformer | |
| CN113033845B (en) | Construction method and device for power transmission resource co-construction and sharing | |
| Yang et al. | A semantic information decomposition network for accurate segmentation of texture defects | |
| CN112508946B (en) | Cable tunnel anomaly detection method based on antagonistic neural network | |
| Dunphy et al. | Autonomous crack detection approach for masonry structures using artificial intelligence | |
| CN114418929A (en) | Weld defect identification method based on consistency multi-scale metric learning | |
| CN114120043A (en) | Method for detecting abnormal pumping well based on production dynamic data and indicator diagram |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |