阅读说明：本技术 一种风力发电机组叶片状态监测方法及其系统 (Method and system for monitoring state of blade of wind generating set ) 是由 张云峰 赵爽 李民 黄久平 王德玉 孙亮 王秋强 步兵 刘钊辰 安阳 戚志强 于 2021-07-09 设计创作，主要内容包括：本本发明提供的一种风力发电机组叶片状态监测方法及其系统,能够判断叶片异常状态,其首先通过变分自编码判断叶片状态是否异常,再通过K-Means聚类算法对叶片异常数据进行分类,进而识别叶片的故障类型,该方法可对风力发电机组叶片实施状态监测与故障诊断,可以监测叶片在运行过程中出现的故障,从而降低由于突发事故产生的不必要损失及停机维护检修带来的发电量损失,减少叶片维修维护成本。(According to the method and the system for monitoring the blade state of the wind generating set, provided by the invention, the abnormal state of the blade can be judged, whether the blade state is abnormal or not is judged through variation self-coding, abnormal data of the blade is classified through a K-Means clustering algorithm, and then the fault type of the blade is identified.)

1. A method for monitoring the state of a blade of a wind generating set is suitable for a horizontal-axis wind generating set comprising a master control system and a variable pitch system,

step S1: the method comprises the following steps that through a plurality of sensor units arranged on blades of a wind generating set, the sensor units at least collect blade load signals and blade vibration signals in real time;

step S2: performing data normalization processing on the real-time blade load and vibration signals acquired in the step S1;

step S3: distinguishing and extracting abnormal data in the blade state characteristic vector of the load signal and the vibration signal after the data normalization processing in the step S2 by an unsupervised learning method in deep learning, and outputting a blade state signal after distinguishing and judging;

step S4: if abnormal data exists in the blade state feature vector, fault data classification is further carried out on the abnormal data in the blade state feature vector extracted in the step S3 by adopting a clustering method to obtain a specific fault category;

step S5: the blade status signal and the failure category output according to the processing results of step S3 and step S4.

2. The method for monitoring the condition of the blade of the wind generating set according to claim 1, wherein each sensor unit further comprises a load sensor and a vibration sensor, the load sensor is installed at the root position of the blade, the vibration sensor is installed at the middle position of the blade, and the vibration sensor is installed at a position which is spaced from the root position of the blade by about the entire length 1/3 of the blade.

3. A wind park blade condition according to claim 1The monitoring method is characterized in that, in step S2, the data normalization processing is performed on the blade load signal and the blade vibration signal acquired in real time, that is, the acquired data of the blade load signal and the blade vibration signal is transformed, so that the normalized data conforms to the standard normal distribution, that is, the mean value is 0, the standard deviation is 1, and the transformation function is

Where μ is the mean of all sample data and σ is the standard deviation of all sample data.

4. The method for monitoring the blade state of the wind generating set according to claim 3, wherein in the step S3, the method for distinguishing and extracting the blade state feature vector through the unsupervised learning method in the deep learning adopts one of a self-encoder or a variational self-encoder in the unsupervised learning method to distinguish abnormal data in the blade state feature vector, and the abnormal data is output after being distinguished and judged.

5. The method for monitoring the blade state of the wind generating set according to claim 1 or 4, wherein the blade state signal is divided into a blade normal signal and a blade abnormal signal.

6. The method for monitoring the blade state of the wind generating set according to claim 4, wherein in the step S3, the method of the variational self-encoder in the unsupervised learning method is adopted to distinguish abnormal data in the blade state feature vector, further comprising:

step S301: the variational self-encoder model adopts a convolutional neural network method to construct an encoder, and adopts a LeakyRelu activation function to construct an activation function;

step S302: the parameter reconstruction adopts Gaussian distribution processing;

step S303: the variational self-encoder model adopts a convolutional neural network method to construct a decoder, and the activation function adopts Relu and Sigmoid activation functions;

step S304: the loss function adopts KL divergence;

step S305: the optimizer employs the Adam algorithm.

7. The method for monitoring the blade condition of the wind generating set according to claim 6, wherein in the step S301, the encoder of the variational self-encoder model utilizes p_θ(z|x⁽ⁱ⁾) A construction is made wherein z ═ μ + σ ∈ epsilon, epsilon to N (0, i), μ is the mean value of all sample data, and σ is the standard deviation of all sample data.

8. The method for monitoring the blade condition of the wind generating set according to claim 7, wherein in the step S304, the encoding end of the variational self-encoder utilizes q_φ(z|x⁽ⁱ⁾) Posterior probability p of de-approximating real decoding end_θ(z|x⁽ⁱ⁾) To measure the similarity of two distributions, KL divergence calculation is adopted

Further obtain

Since the KL divergence is not negative, when the two distributions are in agreement, the KL divergence is 0, so log p_θ(x⁽ⁱ⁾)≥Wherein the content of the first and second substances,referred to as the lower bound of the variation of the log-likelihood function.

Direct optimizationlog p_θ(x⁽ⁱ⁾) Is infeasible, so that the lower bound of variation of the log-likelihood function is optimizedAccordingly, the optimized log-likelihood function is converted into an optimization

Calculation using the Monte Carlo methodExpectation that then

9. The method for monitoring the blade state of the wind generating set according to claim 8, wherein in step S4, the method for classifying the fault data by the fault data classification fault classification is performed on the abnormal data existing in the blade state feature vector by using a clustering method, and the clustering method is a K-Means clustering method, and the distance calculation is performed by using cosine similarity.

10. The method for monitoring the condition of the blade of the wind generating set according to claim 8, wherein the fault data categories are further subdivided into at least four fault data categories of blade icing, blade cracking, blade unbalance and blade breakage

11. A wind generating set blade condition monitoring system applying the wind generating set blade condition monitoring method of any one of claims 1 to 10, comprising a sensor unit, wherein the sensor unit further comprises a plurality of load sensors and vibration sensors which are arranged on the wind generating set blade and used for acquiring blade real-time load signals and blade vibration signals; the wind generating set blade state monitoring system is characterized in that the wind generating set blade state monitoring system further comprises a data processing unit which is electrically connected with the sensing unit and can perform data transmission and communication, and the data processing unit can perform data normalization processing on blade real-time load and vibration signals acquired by the load sensor and the vibration sensor, and comprises:

the blade state feature extraction unit is electrically connected with the data processing unit and can perform data transmission and communication, and is used for extracting blade state feature vectors of the load signals and the vibration signals processed by the data processing unit, distinguishing and extracting abnormal data in the blade state feature vectors according to the blade state feature vectors by an unsupervised learning method in deep learning, and outputting blade state signals after distinguishing and judging the abnormal data;

the blade fault classification unit is used for acquiring the blade state feature vector output by the blade state feature extraction unit and classifying fault data in abnormal data in the blade state feature vector by using a clustering method;

and the blade state output unit is used for outputting the blade state signals and further giving out specific fault categories according to the fault data category classification results.

12. The system for monitoring the blade state of the wind generating set according to claim 11, wherein the distinguishing and extracting of the abnormal data in the blade state feature vector by the unsupervised learning method in the deep learning is to distinguish the abnormal data in the blade state feature vector by using one of a self-encoder or a variational self-encoder in the unsupervised learning method, and output the blade state signal after distinguishing and judging the abnormal data.

13. The system of claim 12, wherein the blade status signal comprises a blade normal signal and a blade abnormal signal.

14. The system according to claim 12, wherein the blade fault classification unit classifies fault data types of abnormal data in the blade state feature vector by a K-Means clustering method to obtain a specific fault type.

15. A wind park according to any of claims 11-14, wherein said fault categories are further subdivided into at least four fault categories indicative of blade icing, blade cracking, blade imbalance, blade breakage.

Technical Field

The invention relates to the field of wind power generation equipment, in particular to a method and a system for monitoring the state of a blade of a wind generating set.

Background

The blade of the wind generating set is one of key components in the whole wind generating set, and due to the use environment, the self-operation characteristics and the like, the blade is easily influenced by various adverse factors such as various weather or people in the working process, the problems of cracks, abrasion, ice coating, imbalance and the like can not be avoided, if an effective means is not available, the abnormal state of the blade is detected in time, and the maintenance and the overhaul can cause various damages to the blade in time, the safe operation of the whole wind generating set is seriously threatened, so that the blade implementation state monitoring and fault diagnosis research of the wind generating set are of great significance. The state monitoring of the wind generating set blade can avoid the blade from possibly breaking down in the operation process, reduce unnecessary loss caused by accidents and power generation loss caused by shutdown maintenance, and reduce the maintenance cost of the wind generating set blade. The wind generating set blade on-line monitoring and fault diagnosis method can effectively prevent the generation of sudden and serious random events and provide theoretical support for the shutdown maintenance of the wind generating set blade.

In recent years, with the maturity and wide application of artificial intelligence technology, the application of artificial intelligence to the fault monitoring and diagnosis of the wind generating set blade is a hot spot of the current structural fault monitoring and diagnosis. At present, the identification of structural damage modes by artificial intelligence technology comprises: the mature technologies such as deep learning, machine learning, wavelet analysis and expert system are widely applied to diagnosis of structural damage conditions, but most of them do not achieve ideal effects. How to develop a set of effective method, device and system by using artificial intelligence technology becomes very slow.

Disclosure of Invention

The invention aims to provide a method and a system for monitoring the state of a blade of a wind generating set, aiming at the defects in the prior art.

In order to achieve the above object, the present invention provides a method for monitoring the state of a blade of a wind turbine generator system, which is suitable for a horizontal-axis wind turbine generator system including a main control system and a pitch system,

step S1: the method comprises the following steps that through a plurality of sensor units arranged on blades of a wind generating set, the sensor units at least collect blade load signals and blade vibration signals in real time;

step S2: performing data normalization processing on the real-time blade load and vibration signals acquired in the step S1;

step S3: distinguishing and judging the load signal and the vibration signal after the data normalization processing in the step S2 by an unsupervised learning method in deep learning, extracting abnormal data in the blade state characteristic vector, and outputting a blade state signal after distinguishing and judging;

step S4: if abnormal data exist in the blade state feature vector, further performing fault data classification on the abnormal data of the blade state feature vector extracted in the step S3 by adopting a clustering method;

step S5: the blade state signal or/and the fault category output according to the processing result of step S3 or step S4.

Further, each sensor unit further comprises a load sensor and a vibration sensor, the load sensor is installed at the root position of the blade, the vibration sensor is installed at the middle position of the blade, and the installation position of the vibration sensor is away from the root of the blade by the whole length 1/3 of the blade.

Further, in step S2, the data normalization processing is performed on the blade load signal and the blade vibration signal collected in real time, that is, the collected blade load signal and the collected blade vibration signal are processedThe data is transformed so that the normalized data conforms to the standard normal distribution, namely the mean value is 0, the standard deviation is 1, and the transformation function is

Where μ is the mean of all sample data and σ is the standard deviation of all sample data.

Further, in step S3, the distinguishing and extracting the blade state feature vector by the unsupervised learning method in the deep learning adopts one of a self-encoder or a variational self-encoder in the unsupervised learning method to distinguish normal data or abnormal data in the blade state feature vector, and outputs the blade state signal after distinguishing and judging the normal data or the abnormal data.

Further, the blade state signal is divided into a blade normal signal and a blade abnormal signal.

Further, in step S3, the method of using a variational self-encoder in an unsupervised learning method to distinguish abnormal data in the blade state feature vector further includes:

step S301: the variational self-encoder model adopts a convolutional neural network method to construct an encoder, and adopts a LeakyRelu activation function to construct an activation function;

step S302: the parameter reconstruction adopts Gaussian distribution processing;

step S303: the variational self-encoder model adopts a convolutional neural network method to construct a decoder, and the activation function adopts Relu and Sigmoid activation functions;

step S304: the loss function adopts KL divergence;

step S305: the optimizer employs the Adam algorithm.

Further, in step S301, the encoder of the variational self-encoder model utilizes p_θ(z|x⁽ⁱ⁾) A construction is made wherein z ═ μ + σ ∈ epsilon, epsilon to N (0, i), μ is the mean value of all sample data, and σ is the standard deviation of all sample data.

Further, in the step S304, the changeCoding end utilization q of a partitional encoder_φ(z|x⁽ⁱ⁾) Posterior probability p of de-approximating real decoding end_θ(z|x⁽ⁱ⁾) To measure the similarity of two distributions, KL divergence calculation is adopted

Further obtain

Since the KL divergence is non-negative, when the two distributions are in agreement, the KL divergence is 0, soWherein the content of the first and second substances,referred to as the lower bound of the variation of the log-likelihood function.

Direct optimization of log p_θ(x⁽ⁱ⁾) Is infeasible, so that the lower bound of variation of the log-likelihood function is optimizedAccordingly, the optimized log-likelihood function is converted into an optimization

Calculation using the Monte Carlo methodExpectation that then

Further, in the step S4, the method of clustering is used to classify the fault data types of the abnormal data in the blade state feature vector, and the method of clustering is a K-Means clustering method, and the cosine similarity is used to calculate the distance.

Further, the fault categories are at least further subdivided into four fault categories of blade icing, blade crack, blade unbalance and blade fracture

The invention also provides a wind generating set blade state monitoring system, which comprises a sensor unit, a control unit and a control unit, wherein the sensor unit further comprises a plurality of load sensors and vibration sensors which are arranged on blades of the wind generating set and used for acquiring real-time load signals and vibration signals of the blades; the data processing unit is electrically connected with the sensing unit and can perform data transmission and communication, and the data processing unit can perform data normalization processing on blade real-time load and vibration signals acquired by the load sensor and the vibration sensor; the blade state feature extraction unit is electrically connected with the data processing unit and can perform data transmission and communication, and is used for extracting blade state feature vectors of the load signals and the vibration signals processed by the data processing unit, distinguishing and extracting abnormal data in the blade state feature vectors according to the blade state feature vectors by an unsupervised learning method in deep learning, and outputting blade state signals after distinguishing and judging the abnormal data; the blade fault classification unit is used for acquiring the blade state feature vector output by the blade state feature extraction unit and classifying fault data in abnormal data in the blade state feature vector by using a clustering method; and the blade state output unit is used for outputting the blade state signals and further giving out specific fault categories according to the fault data category classification results.

Further, the distinguishing and extracting abnormal data in the leaf state feature vector by the unsupervised learning method in the deep learning is to specifically distinguish normal data or abnormal data in the leaf state feature vector by adopting one of a self-encoder or a variational self-encoder in the unsupervised learning method, and output the leaf state signal after distinguishing and judging the abnormal data.

Further, the blade state signal is composed of a blade normal signal and a blade abnormal signal.

Further, the blade fault classification unit classifies fault data types of abnormal data in the blade state feature vectors through a K-Means clustering method.

Further, the fault categories are at least further subdivided into four fault categories which are characterized by blade icing, blade cracking, blade unbalance and blade breakage.

In summary, the method and the system for monitoring the blade state of the wind generating set provided by the invention can judge the abnormal state of the blade, firstly judge whether the blade state is abnormal through variational self-coding, and then classify the abnormal data of the blade through a K-Means clustering algorithm to further identify the specific blade fault.

Drawings

FIG. 1 is a schematic diagram of a method for monitoring a state of a blade of a wind turbine generator system according to the present invention.

FIG. 2 is a schematic structural diagram of a variational self-encoder in the wind turbine generator system blade state monitoring method of the present invention.

Detailed Description

In order to explain technical contents, structural features, and objects and effects of the present invention in detail, the following embodiments are described in detail with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not limit the invention. It will be apparent to one skilled in the art that the present invention may be practiced without some of these specific implementation details. The following description of the embodiments is merely intended to provide a better understanding of the present invention by illustrating examples of the present invention.

Referring to fig. 1 and fig. 2, a method for monitoring the state of a blade of a wind turbine generator system according to the present invention is applicable to a horizontal-axis wind turbine generator system including a main control system and a pitch system, and includes,

step S1: the method comprises the following steps that through a plurality of sensor units arranged on blades of a wind generating set, the sensor units at least collect blade load signals and blade vibration signals in real time;

step S2: performing data normalization processing on the real-time blade load and vibration signals acquired in the step S1;

step S3: distinguishing and extracting abnormal data in the blade state characteristic vector of the load signal and the vibration signal after the data normalization processing in the step S2 by an unsupervised learning method in deep learning, and outputting a blade state signal after distinguishing and judging the abnormal data;

step S4: if abnormal data exists in the blade state feature vector, fault data classification is further carried out on the abnormal data of the blade state feature vector extracted in the step S3 by adopting a clustering method to carry out fault data classification on the abnormal data in the blade state feature vector so as to obtain a specific fault class;

step S5: the blade state signal or/and the fault category output according to the processing result of step S3 or step S4.

Each sensor unit further comprises a load sensor and a vibration sensor, the load sensor is arranged at the root position of the blade, the vibration sensor is arranged at the middle position of the blade, and the distance between the installation position of the vibration sensor and the root of the blade is 1/3 of the whole length of the blade.

In step S2, the data normalization processing is performed on the blade load signal and the blade vibration signal collected in real time, that is, the collected data of the blade load signal and the blade vibration signal is transformed, so that the normalized data conforms to a standard normal distribution, that is, the mean value is 0, the standard deviation is 1, and the transformation function is

Where μ is the mean of all sample data and σ is the standard deviation of all sample data.

In the step S3, the leaf state feature vector is distinguished and extracted by an unsupervised learning method in the deep learning, and one of a self-encoder and a variational self-encoder in the unsupervised learning method is used to distinguish the leaf state feature vector or abnormal data, and the leaf state signal is output after the abnormal data is distinguished and judged.

The blade state signals are divided into two signals, namely blade normal signals and blade abnormal signals.

In step S3, the method of using a diversity self-encoder in an unsupervised learning method to distinguish the blade state feature vector or the abnormal data further includes:

step S301: the variational self-encoder model adopts a convolutional neural network method to construct an encoder, and adopts a LeakyRelu activation function to construct an activation function;

step S302: the parameter reconstruction adopts Gaussian distribution processing;

step S303: the variational self-encoder model adopts a convolutional neural network method to construct a decoder, and the activation function adopts Relu and Sigmoid activation functions;

step S304: the loss function adopts KL divergence;

step S305: the optimizer employs the Adam algorithm.

In step S301, the encoder of the variational self-encoder model uses p_θ(z|x⁽ⁱ⁾) A construction is made wherein z ═ μ + σ ∈ epsilon, epsilon to N (0, i), μ is the mean value of all sample data, and σ is the standard deviation of all sample data.

The method for monitoring the blade condition of the wind generating set according to claim 7, wherein in the step S304, the encoding end of the variational self-encoder utilizes q_φ(z|x⁽ⁱ⁾) Posterior probability p of de-approximating real decoding end_θ(z|x⁽ⁱ⁾) To measure the similarity of two distributions, KL divergence calculation is adopted

Further obtain

Since the KL divergence is non-negative, when the two distributions are in agreement, the KL divergence is 0, soWherein the content of the first and second substances,referred to as the lower bound of the variation of the log-likelihood function.

Direct optimization of log p_θ(x⁽ⁱ⁾) Is infeasible, so that the lower bound of variation of the log-likelihood function is optimizedAccordingly, the optimized log-likelihood function is converted into an optimization

Calculation using the Monte Carlo methodExpectation that then

In the step S4, the method of clustering is used to classify the fault data types of abnormal data in the blade state feature vector, the clustering method is a K-Means clustering method, and distance calculation is performed by using cosine similarity.

The fault categories are at least further subdivided into four fault categories of blade icing, blade crack, blade unbalance and blade fracture

The invention also provides a wind generating set blade state monitoring system, which comprises a sensor unit, a control unit and a control unit, wherein the sensor unit further comprises a plurality of load sensors and vibration sensors which are arranged on blades of the wind generating set and used for acquiring real-time load signals and vibration signals of the blades; the data processing unit is electrically connected with the sensing unit and can perform data transmission and communication, and the data processing unit can perform data normalization processing on blade real-time load and vibration signals acquired by the load sensor and the vibration sensor; the blade state feature extraction unit is electrically connected with the data processing unit and can perform data transmission and communication, and is used for extracting blade state feature vectors of the load signals and the vibration signals processed by the data processing unit, distinguishing and extracting abnormal data in the blade state feature vectors according to the blade state feature vectors by an unsupervised learning method in deep learning, and outputting blade state signals after distinguishing and judging the abnormal data; the blade fault classification unit is used for acquiring the blade state feature vector output by the blade state feature extraction unit and classifying fault data in abnormal data in the blade state feature vector by using a clustering method; and the blade state output unit is used for outputting the blade state signals and further giving out specific fault categories according to the fault data category classification results.

The method for distinguishing and extracting the normal data or the abnormal data in the leaf state characteristic vector through the unsupervised learning method in the deep learning specifically comprises the steps of distinguishing the normal data or the abnormal data in the leaf state characteristic vector by adopting one of a self-encoder or a variational self-encoder in the unsupervised learning method, and outputting leaf state signals after distinguishing and judging the normal data or the abnormal data.

The blade state signal consists of a blade normal signal and a blade abnormal signal.

And the blade fault classification unit classifies the fault data of abnormal data in the blade state feature vector by a K-Means clustering method.

The fault categories are at least further subdivided into four fault data categories of blade icing, blade crack, blade imbalance and blade fracture.

In summary, the method and the system for monitoring the blade state of the wind generating set provided by the invention can judge the abnormal state of the blade, firstly judge whether the blade state is abnormal through variational self-coding, and then classify the abnormal data of the blade through a K-Means clustering algorithm to further identify the specific blade fault.

It is to be understood that the invention is not limited to the specific arrangements and instrumentality described above and shown in the drawings. A detailed description of known methods is omitted herein for the sake of brevity. In the above embodiments, several specific steps are described and shown as examples. However, the method processes of the present invention are not limited to the specific steps described and illustrated, and those skilled in the art can make various changes, modifications and additions or change the order between the steps after comprehending the spirit of the present invention.

It should also be noted that the exemplary embodiments mentioned in this patent describe some methods or systems based on a series of steps or devices. However, the present invention is not limited to the order of the above-described steps, that is, the steps may be performed in the order mentioned in the embodiments, may be performed in an order different from the order in the embodiments, or may be performed simultaneously.

As described above, only the preferred embodiments of the present invention are shown, and it is clearly understood by those skilled in the art that, for convenience and brevity of description, the specific working processes of the system, the module and the unit described above may refer to the corresponding processes in the foregoing method embodiments, and are not described herein again. It should be understood that the scope of the present invention is not limited thereto, and any equivalent modifications or substitutions that can be easily made by those skilled in the art within the technical scope of the present invention are included in the scope of the present invention.