CN108510001B - Wind driven generator blade defect classification method and classification system thereof - Google Patents
Wind driven generator blade defect classification method and classification system thereof Download PDFInfo
- Publication number
- CN108510001B CN108510001B CN201810296437.2A CN201810296437A CN108510001B CN 108510001 B CN108510001 B CN 108510001B CN 201810296437 A CN201810296437 A CN 201810296437A CN 108510001 B CN108510001 B CN 108510001B
- Authority
- CN
- China
- Prior art keywords
- defect
- blade
- resnet
- sample
- classification
- 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.)
- Active
Links
- 230000007547 defect Effects 0.000 title claims abstract description 206
- 238000000034 method Methods 0.000 title claims abstract description 51
- 238000005070 sampling Methods 0.000 claims abstract description 48
- 238000003066 decision tree Methods 0.000 claims abstract description 28
- 238000013145 classification model Methods 0.000 claims abstract description 14
- 238000010276 construction Methods 0.000 claims abstract description 9
- 238000000605 extraction Methods 0.000 claims description 33
- 238000012549 training Methods 0.000 claims description 20
- 230000005284 excitation Effects 0.000 claims description 9
- 230000004048 modification Effects 0.000 claims description 6
- 238000012986 modification Methods 0.000 claims description 6
- 230000008569 process Effects 0.000 claims description 6
- 239000000284 extract Substances 0.000 claims description 4
- 238000010248 power generation Methods 0.000 abstract description 3
- 230000009286 beneficial effect Effects 0.000 abstract description 2
- 238000011161 development Methods 0.000 abstract description 2
- 230000006870 function Effects 0.000 description 15
- 238000004364 calculation method Methods 0.000 description 7
- 238000012360 testing method Methods 0.000 description 7
- 238000001514 detection method Methods 0.000 description 6
- 239000004576 sand Substances 0.000 description 5
- 238000010586 diagram Methods 0.000 description 4
- 229910044991 metal oxide Inorganic materials 0.000 description 3
- 150000004706 metal oxides Chemical class 0.000 description 3
- 230000000903 blocking effect Effects 0.000 description 2
- 238000005336 cracking Methods 0.000 description 2
- 239000002184 metal Substances 0.000 description 2
- 150000003839 salts Chemical class 0.000 description 2
- 238000006467 substitution reaction Methods 0.000 description 2
- 230000004075 alteration Effects 0.000 description 1
- 150000001875 compounds Chemical class 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 230000003247 decreasing effect Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 239000013307 optical fiber Substances 0.000 description 1
- 230000000750 progressive effect Effects 0.000 description 1
- 238000012795 verification 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
- G06F18/21—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
- G06F18/2148—Generating training patterns; Bootstrap methods, e.g. bagging or boosting characterised by the process organisation or structure, e.g. boosting cascade
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
- Investigating Materials By The Use Of Optical Means Adapted For Particular Applications (AREA)
Abstract
The invention discloses a method and a system for classifying defects of blades of a wind driven generator, wherein the classification method comprises the following steps: extracting defect characteristics of the blade by utilizing the ResNet trained on the wind turbine blade image sample data set; obtaining blade defect type information with high frequency degree by using the defect characteristics of the blade extracted by ResNet; using the extracted leaf defect type information with high frequency degree for the construction of a decision tree, repeating iteration until convergence to obtain a defect classification model, and classifying the defect characteristics by using a high-frequency sampling-based Catboost method; the method can accurately detect and classify the defects of the blades, and is beneficial to the development of wind power generation.
Description
Technical Field
The invention relates to the technical field of computer vision, in particular to a method and a system for classifying defects of a wind driven generator blade.
Background
The blade of the wind driven generator is a key component for capturing wind energy, and during the working process, due to the factors of bad environment, complex working conditions and the like, the surface of the blade may generate various common defects such as spots (spots), scratches (scratches), sand holes (Blister), cracks (Crack), barking (Peeling), Cracking (Cracking), skins (Skinning), Icing (Icing), oil stains (greasydidrt) and the like, and the defects may seriously affect the efficiency and safety of wind power generation. As the quality guarantee period of the wind driven generator is gradually increased in recent years, the blade defects of the wind driven generator are frequent. The defect detection of the blade depends on manual completion for a long time, and due to the limitations of detection precision, speed, safety and the like, the detection has the problems of low efficiency, unreliable result, high cost and the like, so that the detection means is in urgent need of improvement. Under the background, a defect detection technology based on computer vision is particularly important, and especially in recent years, extraction of defect features based on a deep convolutional neural network is proved to be effective, however, the extracted defect features still have large dimensionality and redundancy, and the performance of the current classifier on the data is poor, so that classification of defects cannot be performed quickly and effectively, and further manual identification of defect categories is still needed.
Disclosure of Invention
The invention provides a method for classifying defects of blades of a wind driven generator, which aims to solve the problem that the defects of the blades of the wind driven generator are difficult to classify quickly and effectively in the prior art.
A method for classifying defects of a wind driven generator blade comprises the following steps:
s101, extracting defect characteristics of the blade by utilizing a ResNet (residual error network) trained on a sample data set of the blade image of the wind driven generator;
s102, obtaining blade defect type information with high frequency degree by using the defect characteristics of the blade extracted by ResNet;
s103, using the extracted blade defect type information with high frequency degree for the construction of a decision tree, repeating iteration until convergence to obtain a defect classification model, and classifying the defect characteristics by using a high-frequency sampling-based Catboost method.
A wind turbine blade defect classification system, comprising: the blade defect extraction module is used for extracting the defect characteristics of the blade by utilizing ResNet (residual error network) trained on the wind driven generator blade image sample data set; the blade defect type information acquisition module acquires blade defect type information with high frequency by using the defect characteristics of the blade extracted by ResNet; the model generation module is used for constructing the decision tree by using the extracted leaf defect category information with high frequency degree, and repeating iteration until convergence to obtain a defect classification model; and the classification module is used for classifying the defect characteristics of the defect by using the defect classification model and based on a high-frequency sampling Catboost method.
Compared with the prior art, the invention has the following advantages:
the invention discloses a method and a system for classifying defects of blades of a wind driven generator, wherein the method comprises the following steps: extracting defect characteristics of the blade by utilizing the ResNet trained on the wind turbine blade image sample data set; obtaining blade defect type information with high frequency degree by using the defect characteristics of the blade extracted by ResNet; using the extracted leaf defect type information with high frequency degree for the construction of a decision tree, repeating iteration until convergence to obtain a defect classification model, and classifying the defect characteristics of the leaf defect type information by using a high-frequency sampling-based Catboost method; the method can accurately detect and classify the defects of the blades, and is beneficial to the development of wind power generation.
Drawings
The above features and technical advantages of the present invention will become more apparent and readily appreciated from the following description of the embodiments thereof taken in conjunction with the accompanying drawings.
FIG. 1 is a schematic flow chart illustrating a method for classifying defects of a wind turbine blade according to an embodiment of the present invention;
FIG. 2 is a flowchart of an algorithm framework of a training and testing phase of a method for classifying defects of a wind turbine blade according to an embodiment of the present invention;
FIG. 3 is a graph of loss and error values trends during training and testing phases using ResNet;
FIG. 4 is a schematic view of a blade defect detection method for a wind turbine according to an embodiment of the present invention;
FIG. 5 is a frequency degree value calculated by the high-frequency sampling Catboost-based wind turbine blade defect classification method in the classification of defects such as spots, scratches, sand holes, cracks, barks, cracks, skins, icing and oil stains;
FIG. 6 is a configuration diagram of a wind turbine blade defect classification system according to an embodiment of the present invention;
FIG. 7 is a block diagram of a blade defect extraction module according to an embodiment of the present invention;
FIG. 8 is a block diagram of a blade defect type information acquisition module according to an embodiment of the present invention;
fig. 9 is a configuration diagram of a model generation module according to an embodiment of the present invention.
Detailed Description
Embodiments of a method for classifying a defect of a wind turbine blade and a system for classifying the same according to the present invention will be described with reference to the accompanying drawings. Those of ordinary skill in the art will recognize that the described embodiments can be modified in various different ways, or combinations thereof, without departing from the spirit and scope of the present invention. Accordingly, the drawings and description are illustrative in nature and not intended to limit the scope of the claims. Furthermore, in the present description, the drawings are not to scale and like reference numerals refer to like parts.
As shown in FIG. 1, the method for classifying the defects of the wind turbine blade comprises a training stage and a testing stage. Wherein, in the training phase, firstly, a pre-trained ResNet (residual error network) model on a large data set is used to extract the leaf defect characteristics of a training image. Then, the extracted feature data is used as input of a high-frequency sampling Catboost (a classifier for classifying the extracted features) algorithm, and the defects with high frequency are extracted by using the defect frequency degree in each sampling process to serve as data of the round of decision tree model training, so that the decision tree at this time can be more concentrated on training the data with the high defects, the interference of redundant defects is reduced, the recognition accuracy and the calculation resource efficiency are improved, and the optimal defect classification high-frequency sampling Catboost model is obtained. In the testing stage, the optimized ResNet model is used for extracting various defect characteristics of the tested blade image, and the Catboost model obtained in the training stage is used for classifying the defects, so that more accurate defect classification is obtained.
Fig. 2 is a schematic flow chart of a method for classifying a blade defect of a wind turbine according to an embodiment of the present invention, as shown in fig. 2, including the following steps:
S101: the method comprises the steps of extracting defect characteristics of blades by utilizing ResNet (residual error network) trained on a blade image sample data set of the wind driven generator, wherein blade images are obtained by shooting the blades of the wind driven generator by using an unmanned aerial vehicle, and the blade image sample data set of the wind driven generator is obtained by segmenting and annotating the blade images.
S102: and obtaining the blade defect type information with high frequency by using the defect characteristics of the blade extracted by ResNet. For example, if the frequency of spots, sand holes, and cracks among the defect features is high, the blade defect category information with high frequency obtained by using the defect features of the blade extracted by ResNet is the three defect categories of spots, sand holes, and cracks.
S103: and using the extracted blade defect type information with high frequency degree for the construction of a decision tree, repeating iteration until convergence to obtain a defect classification model, and classifying the defect characteristics of the blade defect type information by using a high-frequency sampling-based Catboost method.
In step S101, extracting the defect feature of the blade by using the ResNet trained on the wind turbine blade image sample data set specifically includes: modifying an input layer of ResNet, and adjusting the size of an input image to a preset value; replacing and improving the last layer of Softmax function of ResNet by a MultiTaskLoss loss function; training a ResNet model by utilizing the wind driven generator blade image sample data set so as to obtain a feature extraction model, and effectively extracting the defect features of the blades through the feature extraction model.
Further, replacing the modified ResNet last layer Softmax function with the MultiTaskLoss loss function includes: evaluating the performance of the model by using a MultiTaskLoss loss function, wherein the expression of the MultiTaskLoss loss function is as follows:
wherein, L ({ p)i},{ti}) is the size of the multiple target loss;
i is the sequential index value of the target in the small batch of samples,
piis the predicted probability that sample i belongs to the correct classification sample; p is a radical ofi *Is a reference standard, which is a positive sample if it is 1, and a negative sample if it is 0;
ti4 parameters representing the prediction bounding boxTransforming the vector of coordinates;
ti *is a reference standard box associated with the positive sample;
λ is a balance parameter which can be given by a regularization coefficient NclsAnd NregAdding corresponding weights;
Lclsis the log classification error between the two classes;
Lregis a regression loss and can only be activated by positive samples.
In an alternative embodiment, in step 102, the specific method for obtaining the blade defect category information with higher frequency by using the defect features extracted by ResNet is as follows: taking a blade image needing to extract features as an input of ResNet, calculating by six layers of convolution layers (Res-conv1, Res-conv2, Res-conv3, Res-conv4, Res-conv5 and Res-conv6) and one layer of fully-connected layer (Res-FC), and outputting excitation of each layer as a feature representation of the blade image; the low-level output of ResNet has more general characteristics, the high-level output thereof has more targeted characteristics, and Res-FC layer characteristic excitation is extracted to obtain blade defect type information with high frequency.
In an alternative embodiment, the classification of the defect features by using the high-frequency sampling-based Catboost method comprises the following steps:
(1) calculating the defect frequency degree of each column of features of the target based on the distinguishing degree of the defects to the target;
the defect frequency f (k) is determined by the following formula (1.2):
in the formula (1.2), the compound,
the intra-class variance of all samples representing defect k,
(2) Calculating the probability of unequal probability not putting back the sample based on the defect frequency degree, namely the probability of the defect being extracted;
probability of defect sampling PiDetermined by the following formula (1.9):
in equation (1.9), N is the total number of defects in the sample space;
(3) initializing sampling parameters: setting the sampling rate according to the tree to be P1In a layer sampling ratio of P2；
(4) Sample extraction for generating a decision tree: generating a random integer i between 1 and N (total number of defects), generating a random integer i between 1 and N
The random number z of (a); if P isi *If z is larger than z, selecting the ith defect, putting the ith defect into k, and otherwise, repeating the step;
where k is the defect sample space into which a decision tree is generated, Pi *Adjusted sampling probability for removing the defect (which needs to be selected) in the sample space; adjusted defect sampling probability P i *Determined by the following formula (1.10):
repeating the extraction until the set extraction sample ratio P1；
(5) Utilizing the method of step (4) with k as sample space at each layer of the generated decision tree, and extracting the proportion P2The sample of (4) is used for constructing the decision tree of the layer;
(6) and (3) repeating the sampling processes in the steps (4) and (5) when each decision tree is constructed until the overall objective function is converged by the algorithm, and finally obtaining an optimal defect classification high-frequency sampling Catboost model by controlling and controlling parameters such as Bayes bagging control intensity, post-loss iteration coefficients and the like. And classifying the defect characteristics of the optical fiber by using a high-frequency sampling-based Catboost method.
Further, the air conditioner is characterized in that,
determined by the following formula (1.3)
In formula (1.3), m is the number of possible classes;
in the formula (1.4), the metal salt,
an intra-class variance in the k-dimension for class i;
in the formula (1.5), the metal oxide,
represents the average value at defect k belonging to class i,
Siin order for the samples to belong to the i category,
nithe number of samples in the i-class,
x is the sample data and is the number of the sample,
x(k)is the value of the kth dimension of the sample data.
Further, the air conditioner is provided with a fan,
determined by the following formula (1.7):
in the formula (1.8), μ(k)Is the average of the kth dimension of all samples, n is the total number of all class samples, S is all sample sets;
fig. 3 is a graph of loss values and accuracy trends during the training and testing phase using ResNet. As can be seen from fig. 3, the training loss value and the test error value are greatly reduced when the training period exceeds 80 times. In the embodiment, pseudo codes of algorithms in training and testing stages of the wind turbine blade defect classification method based on the high-frequency sampling Catboost are shown as an algorithm 1:
Referring to fig. 4 and 5, in fig. 4, when the blade is detected, the defects on the blade 100 are correctly classified into oil stains according to the defect classification model. FIG. 5 is a frequency degree value calculated by the method for classifying the defects of the wind driven generator blade based on the high-frequency sampling Catboost in the embodiment of the invention when the defects such as spots, scratches, sand holes, cracks, barks, cracks, skins, icing and oil stains are classified.
The embodiment is a specific implementation of the wind turbine blade defect classification method based on the high-frequency sampling Catboost in the previous embodiment. In the implementation process, four key parameters are debugged (the prior blocking coefficient beta, the number gamma of post-damage iterations, the learning rate eta necessary for tree classification and the number k of trees are respectively), wherein the smaller the value of beta is, the slower the blocking is before the random arrangement of the defect data, and the value is set to be 1, namely, the value of fold _ probability _ block _ size (int) ═ 1; the value of γ represents the number of iterations after the loss function is minimized, and is set here to 18, i.e., d _ wait (int) ═ 18; decreasing the value of η increases the value of k, where η is set to 0.08; and then, CV10 verification is set in training to evaluate the quality of the algorithm parameters, and the optimal parameters can be obtained in convergence. And finally, the correct classification of the defects of the wind driven generator blade and the correct calculation of the frequent degree value are realized.
A wind driven generator blade defect classification system comprises a blade defect extraction module 12, a blade defect category information acquisition module 14, a model generation module 16 and a classification module 18.
The blade defect extraction module 12 extracts the defect features of the blade by using the ResNet (residual error network) trained on the wind turbine blade image sample data set. In particular, a blade image is obtained by shooting a wind turbine blade by using an unmanned aerial vehicle, and the sample data set of the wind turbine blade image is obtained by segmenting and annotating the blade image.
The blade defect type information obtaining module 14 obtains the blade defect type information with high frequency by using the defect features of the blade extracted by ResNet.
The model generation module 16 uses the extracted leaf defect category information with high frequency degree for the construction of the decision tree, and repeats iteration until convergence to obtain a defect classification model. The classification module 18 classifies the defect features based on the high-frequency sampling Catboost method by using the defect classification model.
In an optional embodiment, the blade defect extraction module 12 further includes a modification unit 122, a replacement unit 124, and a training unit 126, wherein the modification unit 122 modifies the input layer of ResNet to adjust the size of the input image to a preset value; the replacing unit 124 replaces the modified ResNet last layer Softmax function with a MultiTaskLoss loss function; and the training unit 126 trains a ResNet model by using the wind turbine blade image sample data set, so as to obtain a feature extraction model, and the defect features of the blades can be effectively extracted through the feature extraction model.
Further, replacing the modified ResNet last-layer Softmax function with a MultiTaskLoss loss function includes: evaluating the performance of the model by using a MultiTaskLoss loss function, wherein the expression of the MultiTaskLoss loss function is as follows:
wherein, L ({ p)i},{ti}) is the size of the multiple target loss;
i is the sequential index value of the target in the small batch of samples,
piis the predicted probability that sample i belongs to the correct classification sample; p is a radical ofi *Is a reference standard, which is a positive sample if it is 1, and a negative sample if it is 0;
tia vector of 4 parameterized coordinates representing a predicted bounding box;
ti *is a reference standard box associated with the positive sample;
λ is a balance parameter which can be given by a regularization coefficient NclsAnd NregAdding corresponding weights;
Lclsis the log classification error between the two classes;
Lregis a regression loss and can only be activated by positive samples.
In an optional embodiment, the blade defect type information obtaining module 14 further includes a convolution unit 142 and a selection unit 144, when a blade image needing feature extraction is used as an input ResNet, the convolution unit 142 performs calculation through six convolution layers (Res-conv1, Res-conv2, Res-conv3, Res-conv4, Res-conv5, Res-conv6) and a full connection layer (Res-FC), and an excitation output of each layer is a feature representation of the blade image. Since the lower layer output of ResNet is more general and the higher layer output is more targeted, the selection unit 144 extracts Res-FC layer feature excitation to obtain blade defect class information with high frequency.
In an optional embodiment, the model generation module 16 further includes a defect frequency calculation unit 162, a defect extraction probability calculation unit 164, an initialization unit 166, a sampling unit 168, and a decision tree construction unit 169. The defect frequency degree calculation unit 162 calculates the defect frequency degree of each column of features thereof based on the degree of discrimination of the defect with respect to the target.
The defect frequency f (k) is determined by the following formula (1.2):
in the formula (1.2), the metal oxide,
the intra-class variance of all samples representing defect k,
The defect extraction probability calculation unit 164 calculates the probability of unequal probability not putting back the sample, i.e., the size of the probability of the defect being extracted, based on the frequency of the defect;
probability of defect sampling PiDetermined by the following formula (1.9):
in equation 1.9), N is the total number of defects in the sample space.
The initialization unit 166 initializes sampling parameters: setting the sampling rate according to the tree to be P1In a layer sampling ratio of P2。
The sampling unit 168 extracts samples for generating a decision tree: generating a random integer i between 1 and N (total number of defects), generating a random integer i between 1 and N
The random number z of (a); if P isi *If z is larger than z, selecting the ith defect, putting the ith defect into k, and otherwise, repeating the step;
Wherein k is a decision tree generatedDrawn defect sample space, Pi *Adjusted sampling probability for removing the defect (which needs to be selected) in the sample space; adjusted defect sampling probability Pi *Determined by the following formula (1.10):
repeating the extraction until the set extraction sample ratio P1。
The sampling unit 168 samples repeatedly at each layer of the generated decision tree with k as the sample space and the extraction ratio is P2Is used for the construction of the decision tree of the layer.
And (3) when each decision tree is constructed, the decision tree construction unit 169 repeats the sampling process of the steps (4) and (5) until the overall objective function is converged by the algorithm, and finally, the optimal defect classification high-frequency sampling Catboost model is obtained by controlling parameters such as Bayesian bagging control strength, post-damage iteration coefficient and the like.
The defect classification model obtained above is used by the classification module 18 to classify the defect features based on the high-frequency sampling Catboost method.
Further, the air conditioner is provided with a fan,
determined by the following formula (1.3)
In formula (1.3), m is the number of possible classes;
in the formula (1.4), the metal salt,
an intra-class variance in the k-dimension for class i;
in the formula (1.5), the metal oxide,
represents the average value at defect k belonging to class i,
Siin order for the samples to belong to the i category,
niThe number of samples in the i-class,
x is the sample data and is the number of the sample,
x(k)is the value of the kth dimension of the sample data.
Further, the air conditioner is provided with a fan,
determined by the following formula (1.7):
in the formula (1.8), μ(k)Is the average of all samples in the kth dimension, n is the total number of all class samples, and S is all sample sets.
The embodiments in the present specification are described in a progressive manner, each embodiment focuses on differences from other embodiments, and the same and similar parts among the embodiments are referred to each other. For the system embodiment, since it is substantially similar to the method embodiment, the description is simple, and reference may be made to the partial description of the method embodiment for relevant points.
While preferred embodiments of the present invention have been described, additional variations and modifications in those embodiments may occur to those skilled in the art once they learn of the basic inventive concepts. Therefore, it is intended that the appended claims be interpreted as including preferred embodiments and all such alterations and modifications as fall within the scope of the embodiments of the invention.
The above description is only for the specific embodiments of the present invention, but the scope of the present invention is not limited thereto, and any person skilled in the art can easily conceive of the changes or substitutions within the technical scope of the present invention, and all the changes or substitutions should be covered within the scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.
Claims (7)
1. A method for classifying defects of a wind driven generator blade is characterized by comprising the following steps:
s101, extracting defect characteristics of the blade by using ResNet trained on an image sample data set of the wind driven generator blade, wherein the defect characteristics comprise:
modifying an input layer of ResNet, and adjusting the size of an input image to a preset value;
replacing and improving the last layer of Softmax function of ResNet by a MultiTaskLoss loss function;
training a ResNet model by utilizing the wind driven generator blade image sample data set so as to obtain a feature extraction model, and effectively extracting the defect features of the blades through the feature extraction model;
s102, obtaining blade defect type information with high frequency degree by using the defect characteristics of the blade extracted by ResNet;
s103, using the extracted leaf defect type information with high frequency degree for the construction of a decision tree, repeating iteration until convergence to obtain a defect classification model, classifying the defect characteristics by using a high-frequency sampling-based Catboost method,
the expression of the MultiTaskLoss loss function used for evaluating the performance of the model is as follows:
wherein, L ({ p)i},{ti}) is multi-purposeMarking the size of the loss;
i is the sequential index value of the target in the small batch of samples,
piIs the predicted probability that sample i belongs to the correct classification sample;
pi *is a reference standard, which is a positive sample if it is 1, and a negative sample if it is 0;
tia vector of 4 parameterized coordinates representing a predicted bounding box,
ti *is a reference standard box associated with the positive sample;
λ is a balance parameter given by a regularization factor NclsAnd NregAdding corresponding weights;
Lclsis the log classification error between the two classes;
Lregis a regression loss, and can only be activated by positive samples,
in step 102, a specific method for obtaining blade defect category information with higher frequency by using the defect features extracted by ResNet is as follows:
taking a blade image needing to be subjected to feature extraction as input of ResNet, calculating through six convolutional layers and one full-connection layer, and outputting excitation of each layer as feature representation of the blade image;
and extracting Res-FC layer characteristic excitation to obtain blade defect class information with high frequency.
2. The method for classifying the blade defects of the wind driven generator according to claim 1, wherein the extracted blade defect type information with high frequency is used for constructing a decision tree, iteration is repeated until convergence is achieved, so that a defect classification model is obtained, and the classification of the defect features by using a high-frequency sampling-based Catboost method comprises the following steps:
(1) Calculating the defect frequency degree of each column of characteristics of the target based on the distinguishing degree of the defects to the target;
the defect frequency f (k) is determined by the following formula:
wherein,
the intra-class variance of all samples representing defect k;
(2) calculating the probability of unequal probability not putting back the sample based on the defect frequency degree, namely the probability of the defect being extracted;
probability of defect sampling PiIs determined by the following formula:
wherein N is the total number of defects in the sample space;
(3) initializing sampling parameters: setting the sampling rate according to the tree to be P1In a layer sampling ratio of P2；
(4) Sample extraction for generating a decision tree: generating a random integer i between 1 and N, generating a random integer 1 to N
The random number z of (a); if it is not
Selecting the ith defect, putting the ith defect into k, and otherwise, repeating the step;
wherein k is a defect sample space into which a decision tree is extracted,
to remove the defect in the sample space, the adjusted sampling probability; after adjustmentDefect sampling probability of
Is determined by the following formula:
repeating the extraction until the set extraction sample ratio P1；
(5) Utilizing the method in step (4) with k as sample space at each layer of the generated decision tree, and extracting the proportion P 2The sample of (2) is used for constructing the layer of decision tree;
(6) and (3) repeating the sampling processes in the steps (4) and (5) when each decision tree is constructed until the overall objective function is converged by the algorithm, finally obtaining an optimal defect classification high-frequency sampling Catboost model by controlling parameters such as Bayesian bagging control strength, post-loss iteration coefficients and the like, and classifying the defect characteristics of the optimal defect classification high-frequency sampling Catboost model by using a high-frequency sampling-based Catboost method.
3. The method of classifying a wind turbine blade defect according to claim 2,
is determined by
Wherein m is the number of possible categories;
Siin order for the samples to belong to the i category,
nithe number of samples in the i-class,
x is the sample data and is the number of the sample,
x(k)is the value of the kth dimension of the sample data.
4. The method of classifying a wind turbine blade defect according to claim 2,
is determined by the following formula:
wherein, mu(k)Is the average of the kth dimension of all samples,
n is the total number of samples of all categories,
s is all sample sets.
5. A wind turbine blade defect classification system, comprising:
the blade defect extraction module extracts the defect characteristics of the blade by utilizing the ResNet trained on the wind driven generator blade image sample dataset, and comprises the following steps:
Modifying an input layer of ResNet, and adjusting the size of an input image to a preset value;
replacing and improving the last layer of Softmax function of the ResNet by using a MultiTaskLoss loss function;
training a ResNet model by utilizing the wind driven generator blade image sample data set so as to obtain a feature extraction model, and effectively extracting the defect features of the blades through the feature extraction model;
the blade defect type information acquisition module is used for acquiring blade defect type information with high frequency degree by using the defect characteristics of the blade extracted by ResNet;
the model generation module is used for constructing the decision tree by using the extracted leaf defect category information with high frequency degree, and repeating iteration until convergence to obtain a defect classification model;
a classification module, which classifies the defect characteristics of the defect by using the defect classification model and based on a high-frequency sampling Catboost method,
the expression of the MultiTaskLoss loss function used for evaluating the performance of the model is as follows:
wherein, L ({ p)i},{ti}) is the size of the multiple target loss;
i is the sequential index value of the target in the small batch of samples,
piis the predicted probability that sample i belongs to the correct classification sample;
pi *is a reference standard, which is a positive sample if it is 1, and a negative sample if it is 0;
tiA vector of 4 parameterized coordinates representing a predicted bounding box,
ti *is a reference standard box associated with the positive sample;
λ is a balance parameter which can be given by a regularization coefficient NclsAnd NregAdding corresponding weights;
Lclsis the log classification error between the two classes;
Lregis a regression loss, and can only be activated by positive samples,
in step 102, a specific method for obtaining blade defect category information with higher frequency by using the defect features extracted by ResNet is as follows:
taking a blade image needing to be subjected to feature extraction as input of ResNet, calculating through six convolutional layers and one full-connection layer, and outputting excitation of each layer as feature representation of the blade image;
and extracting Res-FC layer characteristic excitation to obtain blade defect class information with high frequency.
6. The wind turbine blade defect classification system of claim 5, wherein the blade defect extraction module further comprises a modification unit, a replacement unit, and a training unit,
the modification unit modifies an input layer of ResNet and adjusts the size of an input image to a preset value; the replacing unit replaces the last layer of Softmax function of the improved ResNet by a MultiTaskLoss loss function; and the training unit trains a ResNet model by utilizing the wind driven generator blade image sample data set so as to obtain a feature extraction model, and the defect features of the blades can be effectively extracted through the feature extraction model.
7. The wind turbine blade defect classification system according to claim 5, wherein the model generation module further comprises a convolution unit and a selection unit, when the blade image needing to be subjected to feature extraction is input into ResNet, the convolution unit is calculated through six layers of convolution layers and one layer of full-connection layer, and the selection unit is used for extracting full-connection layer feature excitation to obtain blade defect classification information with high frequency.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810296437.2A CN108510001B (en) | 2018-04-04 | 2018-04-04 | Wind driven generator blade defect classification method and classification system thereof |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810296437.2A CN108510001B (en) | 2018-04-04 | 2018-04-04 | Wind driven generator blade defect classification method and classification system thereof |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108510001A CN108510001A (en) | 2018-09-07 |
| CN108510001B true CN108510001B (en) | 2022-06-10 |
Family
ID=63380369
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810296437.2A Active CN108510001B (en) | 2018-04-04 | 2018-04-04 | Wind driven generator blade defect classification method and classification system thereof |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108510001B (en) |
Families Citing this family (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109492345B (en) * | 2019-01-10 | 2020-11-10 | 西安交通大学 | SENET-based turbine blade high cycle fatigue life prediction method |
| CN110111328A (en) * | 2019-05-16 | 2019-08-09 | 上海中认尚科新能源技术有限公司 | A kind of blade crack of wind driven generator detection method based on convolutional neural networks |
| CN110863935B (en) * | 2019-11-19 | 2020-09-22 | 上海海事大学 | Method for identifying attached matters of blades of ocean current machine based on VGG16-SegUnet and dropout |
| CN111784632A (en) * | 2020-05-25 | 2020-10-16 | 内蒙古工业大学 | Wind driven generator blade surface defect detection system and method based on deep learning |
| CN112633368B (en) * | 2020-12-21 | 2022-07-01 | 四川大学 | Flat vibration motor defect detection system and method based on improved multi-granularity cascade forest |
| CN113918677A (en) * | 2021-10-18 | 2022-01-11 | 智联(无锡)信息技术有限公司 | Data processing method and device based on knowledge graph automation link layering and computer readable medium |
Family Cites Families (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US9453500B2 (en) * | 2013-03-15 | 2016-09-27 | Digital Wind Systems, Inc. | Method and apparatus for remote feature measurement in distorted images |
| CN106127780B (en) * | 2016-06-28 | 2019-01-18 | 华南理工大学 | A kind of curved surface defect automatic testing method and its device |
| CN107144569A (en) * | 2017-04-27 | 2017-09-08 | 西安交通大学 | The fan blade surface defect diagnostic method split based on selective search |
| CN107392241B (en) * | 2017-07-17 | 2020-12-25 | 北京邮电大学 | Image target classification method based on weighted column sampling XGboost |
-
2018
- 2018-04-04 CN CN201810296437.2A patent/CN108510001B/en active Active
Also Published As
| Publication number | Publication date |
|---|---|
| CN108510001A (en) | 2018-09-07 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108510001B (en) | Wind driven generator blade defect classification method and classification system thereof | |
| KR102263397B1 (en) | Method for acquiring sample images for inspecting label among auto-labeled images to be used for learning of neural network and sample image acquiring device using the same | |
| CN110197205B (en) | Image identification method of multi-feature-source residual error network | |
| CN111563893B (en) | Grading ring defect detection method, device, medium and equipment based on aerial image | |
| EP3690741B1 (en) | Method for automatically evaluating labeling reliability of training images for use in deep learning network to analyze images, and reliability-evaluating device using the same | |
| CN111696075A (en) | Intelligent fan blade defect detection method based on double-spectrum image | |
| CN112800682B (en) | Feedback optimization fan blade fault monitoring method | |
| CN106201897B (en) | Software defect based on principal component distribution function predicts unbalanced data processing method | |
| CN107830996B (en) | Fault diagnosis method for aircraft control surface system | |
| CN109919252A (en) | The method for generating classifier using a small number of mark images | |
| CN108717149A (en) | Diagnosis Method of Transformer Faults based on M-RVM fusion dynamic weightings AdaBoost | |
| CN114049305B (en) | Distribution line pin defect detection method based on improved ALI and fast-RCNN | |
| CN113255661B (en) | Bird species image identification method related to bird-involved fault of power transmission line | |
| CN117746077B (en) | Chip defect detection method, device, equipment and storage medium | |
| CN113239865B (en) | Deep learning-based lane line detection method | |
| CN111343147A (en) | Network attack detection device and method based on deep learning | |
| CN115861190A (en) | Comparison learning-based unsupervised defect detection method for photovoltaic module | |
| CN115812210A (en) | Method and apparatus for enhancing performance of machine learning classification tasks | |
| CN112149825A (en) | Neural network model training method and device, electronic equipment and storage medium | |
| CN117496223A (en) | Light insulator defect detection method and device based on deep learning | |
| CN117274212A (en) | Bridge underwater structure crack detection method | |
| CN114511521A (en) | Tire defect detection method based on multiple representations and multiple sub-field self-adaption | |
| CN107480715A (en) | The method for building up and system of the transmission device failure predication model of hydroforming equipment | |
| CN109543571B (en) | Intelligent identification and retrieval method for special-shaped processing characteristics of complex products | |
| CN114821174B (en) | Content perception-based transmission line aerial image data cleaning method |
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 |