CN115163424A

CN115163424A - Wind turbine generator gearbox oil temperature fault detection method and system based on neural network

Info

Publication number: CN115163424A
Application number: CN202210758500.6A
Authority: CN
Inventors: 安鸯; 路安晨; 钟晓刚; 钱峰
Original assignee: China National Software & Service Co ltd
Current assignee: China National Software & Service Co ltd
Priority date: 2022-06-29
Filing date: 2022-06-29
Publication date: 2022-10-11

Abstract

The invention relates to a wind turbine generator gearbox oil temperature fault detection method and system based on a neural network. The method comprises the following steps: reading the running data of the wind turbine generator in a certain period as an initial data set; screening data under 5 categories of gearbox oil temperature based on the initial data set to serve as a data set to be analyzed; determining the working state of the gearbox of the current fan by manually confirming the oil temperature-power scatter diagram of the gearbox in the data set to be analyzed so as to determine the label of the training data set; resampling the classes with less samples to balance the number of samples of each class; constructing a neural network model based on a deep residual error network, and training the neural network model by using a data set to be analyzed; and marking the fan data samples to be classified according to the trained depth residual error network model. The invention realizes the detection of the oil temperature fault of the fan gear box from the aspects of data driving and mode identification, and can indirectly improve the power generation performance of the wind turbine generator.

Description

Wind turbine generator gearbox oil temperature fault detection method and system based on neural network

Technical Field

The invention relates to the field of wind generating set gear box fault identification and detection, in particular to a wind generating set gear box oil temperature fault detection method and system based on a neural network, which can detect the wind generating set gear box oil temperature fault based on table data visualization and a computer vision model.

Background

In the current social development process, compared with traditional fossil energy, clean, low-pollution and sustainable regeneration wind power generation gradually leaves open a new way in the energy field. Because the environment of a wind power generation field is severe, a wind generating set needs to operate under various wind conditions for a long time, and a series of problems are gradually exposed in the current wind power industry, particularly the problem that the operating performance of the wind generating set is reduced due to errors of a control system of the in-service wind generating set. How to improve the running performance of the wind turbine generator under the influence of different wind conditions and the error of a wind turbine control system of the in-service wind turbine generator is one of the main problems faced by each wind power plant.

Wind power generation units are usually deployed in areas with poor natural conditions, and therefore faults possibly caused by long-time operation of the units need to be detected and diagnosed. Through statistics, the most easily-failed large component of the wind generating set is the gear box, so that the fault detection research on the gear box has great significance for ensuring the stable and safe operation of the wind generating set. The working condition of the gear box of the wind turbine generator is complex during operation, and the fault detection of the gear box is generally completed by additionally arranging an additional camera to shoot the working state of the gear box. From wind field operation and maintenance experience, the gearbox oil temperature abnormity usually indicates that the gearbox has a fault, but the monitoring of the gearbox oil temperature is difficult to realize through a camera, and the analysis and modeling can be completed only by depending on a temperature sensor and relevant data collected by an SCADA system built in a wind turbine generator.

However, the operation of wind power plants depends on the wind energy in the natural environment, and the operation of the plants is generally unstable. This results in the data collected by the SCADA system being also discrete and irregular as a whole, which makes it difficult and complicated to model the gearbox oil temperature versus average power. Aiming at the problem, the invention adopts a computer vision method to perform visual analysis on the gearbox oil temperature data so as to realize the abnormal detection of the gearbox oil temperature based on ResNet-18.

Disclosure of Invention

The invention provides a wind turbine generator gearbox oil temperature fault detection method and system based on a neural network, and aims to solve the problem that fault categories cannot be accurately distinguished due to the fact that modeling is difficult in gearbox oil temperature-average power fault detection in the prior art.

The wind turbine generator gearbox oil temperature fault detection method is used for carrying out fault identification based on a wind turbine generator gearbox oil temperature-average power scatter diagram. The method is realized based on a Supervisory Control And Data Acquisition (SCADA) system arranged in the unit. Firstly, selecting characteristic variables related to the oil temperature of the gearbox from an SCADA data set operated by a unit, taking the average power as an x axis, wherein the upper limit and the lower limit are respectively 100 and 3000 (kilowatt), taking the oil temperature of the gearbox as a y axis, and the upper limit and the lower limit are respectively 0 and 70 (centigrade), and drawing a scatter diagram of the oil temperature and the average power of the gearbox. Then, a domain expert analyzes a scatter diagram of the wind turbine generator, and the state of the gearbox is classified into 5 types according to the distribution of the scatter diagram. To this end, the production of the initial data set is done by means of a SCADA system only. Since the failure data is typically low in scale, the initial data set is resampled to balance the number of samples under each category. Then, based on the already-made data set, training is performed using a deep residual network. And finally, marking the test data through the trained model, so that the oil temperature fault of the gearbox is identified in a data-driven mode, and the purposes of improving the power generation performance of the wind turbine generator and improving the operation and maintenance efficiency of a wind field are achieved.

The technical scheme adopted by the invention for solving the technical problem is as follows:

a wind turbine generator gearbox oil temperature fault detection method based on a neural network comprises the following steps:

selecting two variables of the oil temperature and the average power of the gearbox from the original SCADA data set;

drawing a scatter diagram of the oil temperature of the gearbox and the average power by taking the average power as an x axis and the oil temperature of the gearbox as a y axis;

analyzing the scatter diagram, and dividing the scatter diagram into a plurality of categories;

constructing a neural network model based on a deep residual error network, and training the neural network model based on the deep residual error network by taking scatter diagrams of different categories as training data;

inputting the sample to be detected into a trained neural network model based on a deep residual error network to obtain a classification result of the sample to be detected, and obtaining a detection result of the oil temperature fault of the gearbox of the wind turbine generator system.

Further, taking the average power as an x axis, setting the upper limit and the lower limit as 100 and 3000 (kilowatt) respectively, taking the oil temperature of the gearbox as a y axis, and setting the upper limit and the lower limit as 0 and 70 (centigrade) respectively, drawing scatter diagrams of the oil temperature and the average power of the gearbox, and taking each scatter diagram as a sample.

Further, a professional analyzes the scatter diagram, and the original data are divided into 5 categories, wherein the category is a category 1 when the temperature is normal, the category 2 when the temperature is in a rapid change trend, the category 3 when the temperature is increased and overrun, the category 4 when the power is increased and the temperature is decreased, and the category 5 when the temperature is obviously limited up and down.

Further, since most samples belong to class 1, there is a large difference between the number of samples under each class, and in order to avoid the influence of class imbalance, the data set of the class with a small number is resampled, so that the number of samples under each class is approximate.

Further, a certain number of pictures are randomly extracted from each category to serve as a verification set, and the other pictures are used as training sets, namely the training sets or the test sets, and normalization processing is carried out to prevent the training stages from not converging. For example, a training data set is constructed by taking each scatter diagram as a sample, and 30% of samples are randomly extracted from the training data set and divided into verification data sets.

Further, the building of the neural network model based on the depth residual error network is to build a computer vision model. On the basis of a traditional convolution neural network, the model constructs identity mapping by adding a residual error structure, improves the phenomenon of gradient explosion or gradient disappearance in the training process of the neural network, and greatly improves the depth of the neural network. Wherein, the residual structure of the network can be expressed as:

y _l ＝h(x _l )+F(x _l ，W _l )

x _l+1 ＝f(y _l )

wherein x is _l And y _l Respectively representing the input and output of the ith residual unit, h (x) _l ) Denotes identity mapping, F denotes residual function, W _l Weight matrix representing the 1 st residual building block, f represents the activation function, x _l+1 Representing the output of the ith residual unit after the activation function.

Further, in order to enhance the robustness of the model, the method performs data enhancement processing such as overturning, gaussian noise adding, covering and the like on the data set.

Furthermore, the neural network model based on the deep residual error network mainly adopts a ResNet-18 network as a backbone model; the main structure of the ResNet-18 network comprises 1 convolutional layer, 1 maximal pooling layer, 8 residual structure modules, and finally output average pooling layers and full-link layers, wherein each residual structure comprises 2 convolutional layers of 3x 3; wherein, the 18 layers refer to 18 layers with weights, including convolution layers and full connection layers, and not including pooling layers, batch normalization layers and the like.

And further, after the model is built, the training data set is sent to the model for training. In the training stage, cross entropy is selected as a loss function, reLU is selected as an activation function, an SGD (Stochastic Gradient Descender) algorithm containing Momentum is selected as an optimizer, a cosine annealing algorithm is selected as an adjusting algorithm of learning rate, and an early-stopping strategy is selected to avoid an overfitting phenomenon. After each round of training is finished, inputting the verification set into the model which finishes the round of training, evaluating and recording the capability of the model, and obtaining the model which finishes the training by comparing the recorded optimal (minimum) loss function value.

Further, the samples to be tested are sent into the model which is verified, the classification result of each test sample is obtained through reasoning of the model, and the detection result of the oil temperature fault of the gearbox is obtained.

A wind turbine generator system gearbox oil temperature fault detection system based on a neural network comprises:

the data selection module is used for selecting two variables of the oil temperature and the average power of the gearbox of the wind turbine generator from the SCADA data set;

the scatter diagram drawing module is used for drawing a scatter diagram of the gearbox oil temperature and the average power by taking the average power as an x axis and the gearbox oil temperature as a y axis;

the scatter diagram classification module is used for analyzing the scatter diagram and classifying the scatter diagram into a plurality of categories;

the model training module is used for constructing a neural network model based on a deep residual error network and training the neural network model based on the deep residual error network by taking scatter diagrams of different categories as training data;

and the fault detection module is used for inputting the sample to be detected into the trained neural network model based on the deep residual error network to obtain the classification result of the sample to be detected, namely the oil temperature fault detection result of the gearbox.

The beneficial effects of the invention are as follows:

(1) The method is based on data driving, firstly, characteristic variables (power and gearbox oil temperature) related to the gearbox oil temperature are extracted from normal operation data of the wind turbine generator, then, a data set is manufactured by drawing a gearbox oil temperature-power scatter diagram, and further, labels are generated through manual marking and are divided into 5 categories. And inputting the data into a deep residual error network for training, and finally realizing the detection of the abnormal oil temperature of the gearbox through the trained deep residual error network. The fault detection of the gearbox of the wind power generation system is achieved from the data driving angle, the intelligent operation and maintenance application of the wind power plant data is completed, and the intelligent transformation in the wind power field is promoted.

(2) The method is based on SCADA data of wind turbine operation, is easy to be applied practically and has no special requirements on the wind turbine, the required characteristic column is acquired in real time by the existing sensor at present, additional equipment such as a camera, an acquisition card and the like is not needed, the operation and maintenance cost of a wind field is not increased, and the method has strong universality, expandability and mobility and has high theoretical research value and practical application value on the research of digitized and intelligent operation and maintenance technologies in the wind power field based on data analysis and machine learning algorithm.

(3) Aiming at the problem of operation and maintenance ubiquitous in the field of wind power generation at present, original table type time sequence data are converted into image data, neighborhood characteristic information in the image is fully utilized, identification work of abnormal oil temperature of a gearbox of the wind generation set is creatively realized through an algorithm model based on computer vision, and the method has positive promoting effect on the field of efficiency optimization and intelligent diagnosis of the wind generation set based on artificial intelligence technology.

(4) The fan gearbox oil temperature identification method based on the deep neural network model, namely the artificial intelligent algorithm, can effectively solve the problem of identification errors caused by insufficient experience when faults are identified through traditional threshold division, makes up for the defects in the traditional wind power operation and maintenance process, and realizes value mining and analysis of fan operation data.

(5) Because the traditional wind power fault monitoring system needs to operate data in real time based on a fan, and a wind power plant operates under the wind condition conditions with high uncertainty and high randomness constantly, the real-time operation data is influenced by the difference of operating environments and can influence the recognition of the oil temperature of a gear box. The method innovatively carries out data mining and pattern recognition on long-term data of the operation of the wind turbine, verifies the effectiveness and accuracy of the algorithm, and provides a valuable reference case for the research direction and verification result of intelligent operation and maintenance and performance improvement of the wind turbine in the field of wind power.

(6) Due to the flexibility of algorithm design and the self universality and expandability of the deep residual error network, the method can try to migrate to other technical problems to be solved of the intelligent operation and maintenance system of the wind power plant, and has strong expansibility and mobility.

Drawings

FIG. 1 is a sample of category 1, showing that the temperature is normal.

Fig. 2 is a sample of category 2 showing a trend of rapid temperature change.

FIG. 3 shows a category 3 example, in which the temperature tends to rise or overrun.

Fig. 4 shows a category 4 example, showing a power increase and a temperature decrease.

FIG. 5 is a sample of category 5 showing significant upper and lower temperature limits.

Fig. 6 shows a depth residual network structure.

Fig. 7 shows a residual structure block.

Fig. 8 is a graph of the change in accuracy of the model during the training phase, with the abscissa representing the number of iterations and the ordinate representing the accuracy.

Fig. 9 is a graph of the variation of the loss function value of the model during the training phase, with the abscissa representing the number of iterations and the ordinate representing the loss function value.

Detailed Description

In order to make the aforementioned objects, features and advantages of the present invention more comprehensible, the present invention is described in further detail with reference to the following specific embodiments and the accompanying drawings.

1. Data set:

the SCADA data of 31 wind turbines from 1/2020 to 5/2020 and 31/10/are acquired at an interval of 10 minutes, and each wind turbine has 21600 pieces of data.

And selecting data of two dimensions of gear oil temperature and average power. The data of each unit are respectively sampled and drawn into a scatter diagram by adopting a random sampling method which is uniformly distributed on (0.5, 1), taking the average power as an x axis, the upper limit and the lower limit of the average power are respectively 100 and 3000 kilowatts, taking the oil temperature of a gearbox as a y axis, and the upper limit and the lower limit of the gearbox oil temperature are respectively 0 and 70 ℃. The scatter diagrams of the 31 wind turbines are manually analyzed, and the states of the oil temperature of the gearbox are classified into 5 types according to the distribution of the scatter diagrams, as shown in table 1 and fig. 1 to 5.

TABLE 1 different states of the oil temperature of the gearbox

Category numbering	Class name
		1	Normal temperature
2	The temperature shows the rapid change trend
		3	The temperature tends to rise and exceed the limit
4	Power up and temperature down
		5	The temperature has obvious upper and lower limits

Through professional manual analysis, the oil temperature state of a gearbox belongs to class 1 in 31 wind turbines, and the classes with normal temperature include No. 1, 3, 4, 5, 6, 8, 10, 11, 13, 14, 15, 16, 18, 19, 20, 21, 23, 24, 25, 27, 28, 29 and 30 wind turbines; the system belongs to a class 2-a number 2 unit with a large temperature fluctuation range; the unit belongs to the category 3, namely a No. 7 unit with rising and overrun trends in temperature; the device belongs to 4 types, namely a No. 12,22,26 and 31 unit with power rising and temperature reduction; belongs to category 5-the unit with No. 9,17 has obvious upper and lower limits of temperature. Since most of the oil temperature states of the unit belong to the 1 category, the unit arrays of the 2, 3, 4 and 5 categories are repeatedly sampled to avoid the problem of unbalanced data categories. And (3) randomly sampling to generate 5400 scatter diagrams of 1 class, 4400 scatter diagrams of 2, 3, 4 and 5 classes as data sets, randomly extracting 400 scatter diagrams from each class as verification sets, and using the rest scatter diagrams as training sets.

2. Constructing a model:

the traditional convolutional neural network can improve the performance of the network by a method of deepening the network layer number; however, too deep network may hinder the transfer of gradient during training, resulting in the phenomenon of gradient explosion or gradient disappearance. In order to solve the problem, resNet provides a residual error structure, and the information loss in the network is reduced by constructing identity mapping, so that the gradient transmission is improved, the phenomena of gradient disappearance and gradient explosion are avoided, the network degradation is prevented, and the network effect is improved.

Fig. 6 shows a depth residual network structure with different numbers of layers. Considering the problems that the data mode is simple, the training time is too long due to an excessively complex model and the like, the lightweight ResNet-18 with 18 layers is adopted for classifying the scatter diagram of the oil temperature of the gearbox in the embodiment. The residual structure used by the model of the present invention is shown in fig. 7, where 64-d represents a convolutional layer with dimension 64,3 × 3,64 represents input dimension 64, convolutional kernel 3 × 3, relu represents an activation function, and [ ] represents residual connection.

ResNet-18 contains 1 convolutional layer, 1 max pooling layer, 8 residual structures, and the final output average pooling layer and the full-link layer. Wherein each residual error structure comprises 2 convolution layers of 3x3, and the size of the output characteristic diagram is reduced by half and the number of channels is doubled after each 2 residual error layers are passed.

3. Training of the model:

the size of the input scatter diagram is 128x128, and the input scatter diagram is read in a gray scale diagram format. After the convolution layer of the model, batch Normalization is used, a loss function is a cross entropy function, an activation function uses ReLU, an SGD optimizer containing Momentum is used for optimizing model parameters, momentum alpha is 0.9, an initial learning rate value is 0.1, a cosine annealing algorithm is set as a learning rate planning algorithm and an early-stop training strategy is adopted, and the Batch size of each training is 32. When the training is 1500 times, the training is finished.

According to fig. 8 and fig. 9, after 200 times of training, the recognition accuracy of the model on the training set rapidly increases, and the loss rapidly decreases, and after 800 times of training, the recognition accuracy approaches convergence, but the loss still fluctuates slightly, so the training is finished when the training is finished up to 1500 times.

4. Experiments and conclusions:

after the model training is completed, the verification is performed on the verification set, and the obtained confusion matrix of the number of samples of the predicted value and the true value is shown in table 2.

TABLE 2 confusion matrix of model identification results on validation set

From table 2, the accuracy of the model on the validation set was 98.56%. According to the confusion matrix, the misjudgment of the model is mainly concentrated between 2 types, large temperature fluctuation range and 4 types, temperature rising power reduction, and 5 types, obvious upper and lower limits of temperature are misjudged as 1 type, normal temperature. From the distribution of the scatter diagram, the scatter diagrams of class 2 and class 4 are relatively similar, while class 1 includes scatter diagrams of various distributions, and among them, there are scatter diagrams similar to class 5, which may be a main cause of misjudgment.

Another embodiment of the present invention provides a wind turbine generator gearbox oil temperature fault detection system based on a neural network, which includes:

The specific implementation process of each module is referred to the description of the method of the invention.

Another embodiment of the invention provides an electronic device (computer, server, smartphone, etc.) comprising a memory storing a computer program configured to be executed by a processor, and a processor, the computer program comprising instructions for performing the steps of the method of the invention.

Another embodiment of the invention provides a computer readable storage medium (e.g., ROM/RAM, magnetic disk, optical disk) storing a computer program which, when executed by a computer, performs the steps of the method of the invention.

The particular embodiments of the present invention disclosed above are illustrative only and not intended to be limiting as to the scope of the invention which is to be given the full breadth of the claims appended and any and all modifications and variations which may be apparent to those skilled in the art may be resorted to without departing from the spirit and scope of the invention. The invention should not be limited to the disclosure of the embodiments in the present specification, but the scope of the invention is defined by the appended claims.

Claims

1. A wind turbine generator gearbox oil temperature fault detection method based on a neural network is characterized by comprising the following steps:

selecting two variables of the oil temperature and the average power of the gearbox of the wind turbine generator from the SCADA data set;

drawing a scatter diagram of the oil temperature and the average power of the gearbox by taking the average power as an x axis and the oil temperature of the gearbox as a y axis;

analyzing the scatter diagram, and classifying the scatter diagram into a plurality of categories;

inputting the sample to be detected into a trained neural network model based on a deep residual error network to obtain a classification result of the sample to be detected, namely obtaining a detection result of the oil temperature fault of the gearbox.

2. The method of claim 1, wherein the average power is taken as an x-axis and the gearbox oil temperature is taken as a y-axis, wherein the x-axis has upper and lower limits of 100 kilowatts and 3000 kilowatts, respectively, and the y-axis has upper and lower limits of 0 degrees and 70 degrees, respectively.

3. The method of claim 1, wherein a certain number of pictures are randomly drawn in the scatter plot of each category as a validation set, and the others as training sets.

4. The method of claim 1, wherein the scatter plot is analyzed into a number of categories, including 5 categories, each of which is: the normal temperature is in category 1, the trend of rapid temperature change is in category 2, the trend of temperature rise and overrun is in category 3, the trend of power rise and temperature drop is in category 4, and the obvious upper and lower limits of temperature are in category 5.

5. The method of claim 1, wherein the data set for the fewer classes is resampled such that the number of samples under each class is similar.

6. The method of claim 1, wherein the deep residual network based neural network model primarily employs a ResNet-18 network as a backbone model.

7. The method of claim 1, wherein the training process of the neural network model based on the deep residual network comprises:

selecting cross entropy as a loss function, selecting ReLU as an activation function, selecting a random gradient descent algorithm containing momentum as an optimizer, selecting a cosine annealing algorithm as a learning rate regulation algorithm, and selecting and using an early-stopping strategy to avoid an overfitting phenomenon;

and after each round of training is finished, inputting the verification set into the model which is finished with training, evaluating and recording the capability of the model, and comparing the recorded optimal loss function value to obtain the model which is finished with training.

8. The utility model provides a wind turbine generator system gear box oil temperature fault detection system based on neural network which characterized in that includes:

9. An electronic apparatus, comprising a memory and a processor, the memory storing a computer program configured to be executed by the processor, the computer program comprising instructions for performing the method of any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that it stores a computer program which, when executed by a computer, implements the method of any one of claims 1 to 7.