CN108830316A - The end-to-end fault diagnosis of wind electric converter based on convolutional neural networks - Google Patents
The end-to-end fault diagnosis of wind electric converter based on convolutional neural networks Download PDFInfo
- Publication number
- CN108830316A CN108830316A CN201810580359.9A CN201810580359A CN108830316A CN 108830316 A CN108830316 A CN 108830316A CN 201810580359 A CN201810580359 A CN 201810580359A CN 108830316 A CN108830316 A CN 108830316A
- Authority
- CN
- China
- Prior art keywords
- convolutional neural
- neural networks
- fault diagnosis
- electric converter
- wind electric
- 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.)
- Pending
Links
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
-
- 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 belongs to wind electric converter fault diagnosis fields.To simplify deep learning in the process of fault diagnosis, improve the efficiency of diagnosis process, introduce convolutional Neural neural network model, it is directly driven with wind electric converter operation data, fault signature is extracted and expressed and is completed by convolutional Neural neural network, end-to-end fault diagnosis is realized, artificial extraction fault signature process is avoided, optimizes convolutional Neural pessimistic concurrency control.Method based on convolutional Neural neural network achieves better classification results in wind electric converter fault diagnosis.
Description
Background technique
The invention belongs to wind electric converter intelligent diagnostics fields, are related to a kind of wind electric converter intelligent diagnosing method.
Summary of the invention
Wind turbines currently on the market are mainly by squirrel-cage, permanent magnet direct-driven and double-fed type composition.Wherein, in addition to mouse cage
Formula, permanent magnet direct-driven and double-fed wind generator unit must all guarantee the electric energy generated in the side such as frequency and amplitude using current transformer
Face meets the requirement of power grid networking.Wind turbines face hot humid, greasy dirt all in than relatively rugged environment under normal conditions
The threat of dust, current transformer also needs to bear high current in this case, the switch state of high-low pressure and high frequency, therefore easily
It breaks down.According to statistics, in wind power system, the frequency that current transformer failure occurs is only second to pitch-controlled system failure.
There are many intelligent method for the current transformer failure being currently directed to, for example are based on support vector machine method, based on expert
The method etc. of system, since these conventional methods need to combine some feature extracting methods, such as wavelet analysis, algorithmic procedure is uncomfortable
For inline diagnosis, it is therefore desirable to the method that can be realized end-to-end fault diagnosis.
This patent is based on this, main using the end-to-end on-line fault diagnosis of the wind electric converter of convolutional neural networks
Technical point is：
(1) it fully considers the coupled relation certainly existed between current transformer three-phase current signal, designs convolutional neural networks.
It is moved using multiple local filters along signal and convolution operation is carried out to signal, just obtained after the completion of operation defeated
Enter signal characteristic figure of the signal under different convolution kernels.
Convolution operation is substantially that operation is filtered to signal, does so the spy that can extract signal under different frequency bands
Sign enriches the expression of signal, provides condition for subsequent Accurate classification.Also, convolution operation realizes part connection and weight
It is shared.Subsequent in each convolutional layer is a down-sampled layer.The effect of down-sampled layer has two o'clock, first is that dropping to characteristic pattern
Dimension, second is that keeping the scale invariability of feature to a certain extent.Down-sampled layer is operated by pondization and carries out dimensionality reduction to characteristic pattern,
Partial deformation and the displacement of characteristic pattern can also be offset while dimensionality reduction.
Three-phase current fault-signal is arranged in 3 × 3000 2D signal, is identified using the advantage of convolutional neural networks
Relating Characteristic between fault-signal.In order to intuitively compare the difference between unlike signal, by the good failure of component arrangement
Signal amplitude Linear Mapping is indicated in the form of grayscale image to 0 to 255 section, because the height of 3 × 3000 images is too small, will be schemed
Image width height is rearranged to 30 × 300.
The fault-signal grayscale image of different faults type has differences, therefore can use the outstanding figure of convolutional neural networks
As recognition capability diagnoses wind electric converter failure.
(2) wind electric converter operation history data fault simulation data, direct training convolutional neural networks, optimization mind are based on
Optimized network parameter.
Propagated forward
Assuming that the training set of convolutional neural networks is (X, Y), wherein X is input, and Y is target output.After training set determines,
The training step of convolutional neural networks is as follows：
1) random initializtion
Before starting training, need to initialize the parameter of convolutional neural networks.The weight of convolution kernel and biasing, tail
The parameter of the full articulamentum in portion can all select at the time of initialization with it is some it is different close to 0 small random number, parameter is not
It can be normally carried out with guarantee training, and small random number can then guarantee that network will not enter saturation shape because weight is excessive
State, so as to cause failure to train.
2) reality output is calculated
Small lot training set is constructed from training sample, and training set is inputted into convolutional neural networks by input layer, is passed through
The successively formula calculating as described above of convolutional neural networks, until reaching the last layer.It is compared after obtaining result with label, really
Determine error.Quantify training result and the direct error size of label usually using entropy function is intersected at present.
Backpropagation
After the completion of forward-propagating, since network initial value is set at random, so obtained result is with desired output
There is a certain error, and the training of convolutional neural networks seeks to the value that network parameter is updated by these errors, makes convolution
Neural network can be fitted the distribution of sample data well.Convolutional neural networks generally use BP algorithm and carry out undated parameter, volume
There are convolutional layer, the special constructions such as down-sampling layer for product neural network, therefore cannot directly use BP algorithm completely, need to convolution
Layer and down-sampling layer carry out specially treated.Each layer of back-propagation algorithm will be specifically introduced below.
1) the reversed derivation of full articulamentum
Firstly, derivative of the calculating target function L about the last layer logits value z+1 (j), use intersect entropy function as
Loss function, as shown in formula (4.6)：
Objective function is determined to the inverse of weight and biasing, as shown in following formula (2) by chain type Rule for derivation
Then calculating target function is about using ReLU as the hiding layer unit of activation primitive and logits valueDerivative,
As shown in following formula (3)
It obtainsAfterwards, can similarly be acquired by formula (2) objective function aboutWithDerivative.
2) the reversed derivation of pond layer
Example is turned to maximum pond.Only have maximum value and next layer of neuron in the region of pond to deposit in maximum pond layer
It is connecting, therefore in error back propagation, it is only necessary to which, to maximum neuron derivation is worth, other neuron derivatives are 0.Specifically
Shown in calculation formula following (4)
3) the reversed derivation of convolutional layer
Convolutional layer backpropagation is similar with full articulamentum, first asks objective function about the derivative of logits value in convolutional layer,
Due to equally using ReLU as activation primitive,
The derivative calculations of convolutional layer input value are
Objective function is about the derivative of convolution kernel
Start with from each layer of convolutional neural networks, main thought be by increase network sparsity and randomness come
It avoids convolutional neural networks from over-fitting occur, enhances its generalization ability, to avoid the problem that over-fitting.
Detailed description of the invention
Fig. 1 is the grayscale image of wind electric converter fault-signal.
Fig. 2 is the end-to-end diagnostic flow chart of wind electric converter failure based on convolutional neural networks.
Specific embodiment
(1) data are collected, including the fault sample data in operational process and the fault simulation data based on model；
(2) according to the coupled characteristic between the fault mode of current transformer and three-phase circuit, convolutional neural networks are determined
Structural parameters, including the neural network number of plies and network node quantity；
(3) the initial learning parameter that setting convolutional neural networks learn；
(4) it is based on data-driven, according to learning parameter, training network；
(5) different learning parameters are adjusted, optimizes and determines network structure.
Claims (3)
1. a kind of end-to-end method for diagnosing faults of wind electric converter based on convolutional neural networks, it is characterised in that：
(1) corresponding convolutional Neural is designed with the associate feature between operation data based on current transformer three-phase circuit coupled characteristic
Network model；
(2) sample data based on history run and analogue simulation data, to the structure and parameter of convolutional neural networks model into
Row optimum choice.
2. it is moved using multiple local filters along signal and convolution operation is carried out to signal according to the data of wind electric converter,
Signal characteristic figure of the input signal under different convolution kernels has just been obtained after the completion of operation, it thus can be by convolutional neural networks
It is used for fault diagnosis in the outstanding processing capacity of image, to obtain outstanding fault identification and diagnosis effect.
3. operation data can directly be connected with convolutional neural networks model by this patent, fault diagnosis flow scheme is simplified, is convenient for
The realization of on-line fault diagnosis.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810580359.9A CN108830316A (en) | 2018-06-05 | 2018-06-05 | The end-to-end fault diagnosis of wind electric converter based on convolutional neural networks |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810580359.9A CN108830316A (en) | 2018-06-05 | 2018-06-05 | The end-to-end fault diagnosis of wind electric converter based on convolutional neural networks |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108830316A true CN108830316A (en) | 2018-11-16 |
Family
ID=64144576
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810580359.9A Pending CN108830316A (en) | 2018-06-05 | 2018-06-05 | The end-to-end fault diagnosis of wind electric converter based on convolutional neural networks |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108830316A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111709190A (en) * | 2020-06-24 | 2020-09-25 | 国电联合动力技术有限公司 | Wind turbine generator operation data image identification method and device |
| CN112202196A (en) * | 2020-08-18 | 2021-01-08 | 广西大学 | Quantum deep reinforcement learning control method of doubly-fed wind generator |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104410098A (en) * | 2015-01-07 | 2015-03-11 | 上海电机学院 | Doubly-fed asynchronous generator set low voltage ride through control system and control method thereof |
| CN104502758A (en) * | 2014-12-17 | 2015-04-08 | 西北工业大学 | Fault diagnosis method for aeronautical static inverter |
| CN105975931A (en) * | 2016-05-04 | 2016-09-28 | 浙江大学 | Convolutional neural network face recognition method based on multi-scale pooling |
| CN106249144A (en) * | 2016-08-16 | 2016-12-21 | 株洲中车时代电气股份有限公司 | Double-fed wind power generator interturn short-circuit failure diagnosing method and fault monitoring method |
| US20170122291A1 (en) * | 2015-10-29 | 2017-05-04 | General Electric Company | System and method for categorizing trip faults of a wind turbine power converter |
-
2018
- 2018-06-05 CN CN201810580359.9A patent/CN108830316A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104502758A (en) * | 2014-12-17 | 2015-04-08 | 西北工业大学 | Fault diagnosis method for aeronautical static inverter |
| CN104410098A (en) * | 2015-01-07 | 2015-03-11 | 上海电机学院 | Doubly-fed asynchronous generator set low voltage ride through control system and control method thereof |
| US20170122291A1 (en) * | 2015-10-29 | 2017-05-04 | General Electric Company | System and method for categorizing trip faults of a wind turbine power converter |
| CN105975931A (en) * | 2016-05-04 | 2016-09-28 | 浙江大学 | Convolutional neural network face recognition method based on multi-scale pooling |
| CN106249144A (en) * | 2016-08-16 | 2016-12-21 | 株洲中车时代电气股份有限公司 | Double-fed wind power generator interturn short-circuit failure diagnosing method and fault monitoring method |
Non-Patent Citations (1)
| Title |
|---|
| 张亮: "基于小波神经网络的风电变流器故障诊断研究", 《中国优秀硕士学位论文全文数据库 工程科技Ⅱ辑》 * |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111709190A (en) * | 2020-06-24 | 2020-09-25 | 国电联合动力技术有限公司 | Wind turbine generator operation data image identification method and device |
| CN112202196A (en) * | 2020-08-18 | 2021-01-08 | 广西大学 | Quantum deep reinforcement learning control method of doubly-fed wind generator |
| CN112202196B (en) * | 2020-08-18 | 2022-04-29 | 广西大学 | Quantum deep reinforcement learning control method of doubly-fed wind generator |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Nayak et al. | ECNet: An evolutionary convolutional network for automated glaucoma detection using fundus images | |
| CN108009594B (en) | A kind of image-recognizing method based on change grouping convolution | |
| CN110929765B (en) | Batch-imaging-based convolution self-coding fault monitoring method | |
| WO2020092143A1 (en) | Self-attentive attributed network embedding | |
| CN108875906B (en) | A kind of multiple dimensioned convolutional neural networks learning method gradually to add up | |
| CN107704958A (en) | A kind of thermal power plant's generated energy Forecasting Methodology of multivariable modeling | |
| CN111861906A (en) | Pavement crack image virtual augmentation model establishment and image virtual augmentation method | |
| CN108830316A (en) | The end-to-end fault diagnosis of wind electric converter based on convolutional neural networks | |
| He et al. | Alzheimer's disease diagnosis model based on three-dimensional full convolutional DenseNet | |
| CN111861886A (en) | Image super-resolution reconstruction method based on multi-scale feedback network | |
| CN110490248A (en) | A kind of converters method for diagnosing faults, terminal device and storage medium | |
| CN117313251B (en) | Train transmission device global fault diagnosis method based on non-hysteresis progressive learning | |
| CN117292330B (en) | Intelligent monitoring system suitable for time sequence data operation and maintenance | |
| CN115860113B (en) | Training method and related device for self-countermeasure neural network model | |
| CN111340133A (en) | Image classification processing method based on deep convolutional neural network | |
| Deshpande et al. | Detection of Plant Leaf Disease by Generative Adversarial and Deep Convolutional Neural Network | |
| CN113361494B (en) | Self-service method and self-service system based on face recognition | |
| CN115546862A (en) | Expression recognition method and system based on cross-scale local difference depth subspace characteristics | |
| CN115578325A (en) | Image anomaly detection method based on channel attention registration network | |
| CN112308208B (en) | Transformer fault diagnosis method based on deep learning model | |
| CN113269702A (en) | Low-exposure vein image enhancement method based on cross-scale feature fusion | |
| CN115166415A (en) | Power distribution network fault diagnosis method and system of self-adaptive graph convolution neural network | |
| CN114118149A (en) | Induction motor fault diagnosis system based on finite element simulation and symmetric feature migration | |
| Wang et al. | Reliable joint segmentation of retinal edema lesions in oct images | |
| Ying et al. | Efficient multi-objective evolutionary neural architecture search for U-Nets with diamond atrous convolution and Transformer for medical image segmentation |
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 | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20181116 |
|
| WD01 | Invention patent application deemed withdrawn after publication |