Lightning protection system fault diagnosis method and wind power blade lightning protection system

Abstract

The invention belongs to the technical field of wind power blade lightning protection, and discloses a lightning protection system fault diagnosis method, which is characterized in that a data sound card collector is arranged at the root of a blade and is used for collecting sound signals in a blade cavity, converting the collected sound signals into digital signals, and sending the digital signals to a server; the fault monitoring, fault early warning and diagnosis module monitors the lightning protection system in real time by using the digital signal through a deep learning fault diagnosis algorithm; meanwhile, a current loop is arranged at the blade root and used for collecting lightning current, the state of a down lead of the lightning protection system in the service period of the blade is judged by assisting a fault monitoring, early warning and diagnosis module by comparing the lightning current with existing actual lightning current data, and finally the real-time state of the lightning protection system is responded to through an upper computer. The fault monitoring, early warning and diagnosis system can monitor the state of the lightning protection system in the service period in real time, can repair or replace the down lead fractured due to mechanical stress generated by deformation in time, improves the reliability of the lightning protection system, and reduces the risk of lightning protection failure.

Description

Technical Field

[0001] The invention relates to a lightning protection system fault diagnosis method and a wind turbine blade lightning protection system, belonging to the technical field of wind turbine blade lightning protection. Background Art

[0002] With the continuous development of the wind power industry, the blades of wind turbines are getting longer and longer, and the number of lightning damages to wind turbine blades and other components is increasing. Since wind turbine blades are expensive to manufacture and repair, once damaged by lightning, it will cause huge economic losses, so it is necessary to carry out reasonable lightning protection design for the blades.

[0003] In the prior art, the lightning protection down conductor is fixed in a straight line from the starting position of the blade web to the blade root flange, and the down conductor is in a taut state. During the operation of the wind turbine, the blades will vibrate, and the down conductor will deform with the vibration of the blades. If the wire is in a taut state, it will be torn off, and the copper nose connected to the flange will also be pulled off during the deformation process, which will cause the lightning protection system of the blade to fail, causing the wind turbine to be damaged by lightning.

[0004] Long-term operation of the wind turbine will cause the down conductor to break or the effective flow cross-section to decrease, resulting in lightning protection failure. Therefore, it is very necessary to invent a lightning protection system fault diagnosis method that can realize real-time monitoring of the operating status of the wind power lightning protection system during the service period of the blades, thereby reducing the risk of mechanical stress caused by deformation of the down conductor in the later stage, which may lead to breakage and cause lightning protection system failure. Summary of the invention

[0005] In view of the deficiencies of the above-mentioned prior art, the disclosed lightning protection system fault diagnosis method uses the fault detection, early warning and diagnosis module to monitor, warn and diagnose the working status of the wind turbine blade lightning protection system during its service by using the signals collected by the sound card collector and the current loop; it can realize real-time monitoring of the operating status of the wind turbine lightning protection system during the service of the blades, reduce the risk of failure of the lightning protection system due to mechanical stress caused by deformation of the down conductor in the later stage, resulting in fracture, thereby improving the safety performance of the wind turbine operation.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is:

[0007] Disclosed is a fault diagnosis method for a lightning protection system. A data sound card collector is installed at the root of a blade to collect sound signals in a blade cavity, and the collected sound signals are converted into digital signals. A fault monitoring, fault warning and diagnosis module uses the digital signals to monitor the lightning protection system in real time through a deep learning fault diagnosis algorithm. At the same time, a current loop is arranged at the root of the blade to collect lightning current. By comparing with the existing actual lightning current data, the auxiliary fault monitoring, warning and diagnosis module determines the state of the lightning protection system lead-down line during the service period of the blade, and finally responds to the real-time state of the lightning protection system on a host computer.

[0008] Furthermore, the data acquisition module is connected to the sound card collector and the current loop, and the data analysis module is connected to the data acquisition module for converting the sound signal into a digital signal.

[0009] Further, the data analysis module includes a data preprocessing algorithm, and the data preprocessing algorithm includes the following steps;

[0010] S1: Normalize the data and map it to the range of 0-1;

[0011] S2: Detect and remove abnormal mutation values ​​from the normalized data to reduce the computational complexity and the impact of abnormal mutation values ​​on fault warning and diagnosis errors;

[0012] S3: Overlap sampling is performed on the data processed in step S2 to enhance the data and highlight the data features.

[0013] Furthermore, the fault warning and diagnosis module is provided with a deep learning fault warning and diagnosis algorithm. The network model adopted by the deep learning fault warning and diagnosis algorithm is a multi-channel weighted fusion capsule network, and data feature extraction is performed through multi-channel weighted convolution and capsule routing.

[0014] Furthermore, the deep learning fault warning and diagnosis algorithm uses units with different convolution kernel sizes in each channel to form different receptive fields for feature extraction. The size of the receptive field changes with the change of the convolution kernel, and the two have the following relationship: Where n represents the number of convolutional pooling layers; rf n-1 Represents the receptive field size of the n-1th layer; rf n Represents the size of the receptive field of layer n; s n-1 represents the step length of n-1 layers; k n-1 Represents the size of n-1 convolution layers; rf last Represents the size of the receptive field band of the last layer, which is equal to the convolution kernel k of this layer last size.

[0015] Furthermore, after feature extraction through different convolution kernels, the output of each channel is: in is the output of the lth convolutional layer of channel j; The i-th feature input of the l-1th convolution layer of channel j, with a total of k feature inputs; is the convolution kernel of the l-1th layer of channel j.

[0016] Furthermore, after feature extraction based on the dynamic receptive field, the kurtosis value is used as the weight to measure the difference in importance. Assuming k jrepresents the kurtosis value of the vibration signal on channel j, m represents the number of channels, and the weight of the channel is: After obtaining the weights of each channel, they need to be weighted fused: Where X l is the output of the lth convolutional layer; is the feature output of the l-1th convolutional layer of channel j; is the bias of the l-1th layer of channel j; is the weight of the l-1th layer of channel j; f l-1 (x) is the activation function of the l-1th convolutional layer; m is the number of channels.

[0017] Furthermore, the capsule network includes a primary capsule layer and a digital capsule layer, the digital capsule layer includes a dynamic routing algorithm, and the idea of ​​the capsule network is divided into three stages. The first stage is completed in the primary capsule layer by weighted summation of neurons; its calculation formula is as follows: U j|i =W ij U i , where U i is the output of the i-th neuron in the previous convolutional network; W ij is the output of the neuron U i Multiply the weight matrix; U j|i is the prediction vector; the second stage is the prediction vector U obtained in the first stage j|i Multiply by c ij Perform weighted summation to obtain the output vector S j , the calculation formula is as follows: j =∑ i c ij *U j|i , where c ij is a coupling coefficient, which is determined by the dynamic routing algorithm; in the third stage, the intermediate vector S obtained in the second stage is j The output vector V of the routing capsule layer is obtained by performing nonlinear transformation through the nonlinear activation vector compression function Squashing. j , compress the modulus value of the input vector to between (0,1), and the calculation formula is as follows: in, Represents the total input vector S j The reduction factor of Represents the total input vector S j The unit vector of .

[0018] Another object of the present invention is to use the above-mentioned fault diagnosis method to detect the lightning protection system of wind turbine blades.

[0019] Compared with the prior art, the present invention has the following beneficial effects:

[0020] Compared with the existing technology, the present invention can realize real-time monitoring of the operating status of the wind power lightning protection system during the service period of the blades, reduce the risk of failure of the lightning protection system due to mechanical stress caused by deformation of the down conductor in the later stage, and improve the safety performance of wind turbine operation. The fault monitoring, early warning and diagnosis system can monitor the status of the lightning protection system in real time during its service, and can promptly repair or replace the down conductor that is broken due to mechanical stress caused by deformation, thereby improving the reliability of the lightning protection system and reducing the risk of lightning protection failure. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 Schematic diagram of a lightning protection system fault monitoring, early warning and diagnosis system for the lightning protection system fault diagnosis method of the present invention.

[0022] Figure 2 Schematic diagram of a multi-channel weighted fusion convolutional network for the lightning protection system fault diagnosis method of the present invention.

[0023] Figure 3 The present invention is a schematic diagram of the fault monitoring, early warning and diagnosis system flow of the lightning protection system fault diagnosis method.

[0024] Figure 4 The figure is a schematic diagram of the specific structure of the lightning protection system fault monitoring and diagnosis system of the lightning protection system fault diagnosis method of the present invention.

[0025] Figure 5 The present invention is a schematic diagram of the failure warning and diagnosis process in the lightning protection system fault monitoring and diagnosis system of the lightning protection system fault diagnosis method.

[0026] Among them, 1-blade, 2-down conductor, 3-current loop, 4-belly, 5-cavity, 6-sound card collector, 7-inner plate, 8-data analysis module, 9-fault monitoring, early warning and diagnosis module, 10-host computer, 11-reference data test platform. DETAILED DESCRIPTION

[0027] For the convenience of explaining and understanding the present invention, the following Figure 1-Figure 5 The embodiments of the present invention are described in detail.

[0028] The implementation scheme of the lightning protection system fault diagnosis method of the present invention includes: 1) a data acquisition module, a sound card collector and a current loop are installed at the root of the blade, the sound card collector is used to collect the sound signal in the blade cavity, and the current loop is used to collect the lightning current; 2) a data analysis module, which converts the collected sound signal into a digital signal; 3) a fault monitoring, early warning and diagnosis module, which includes a digital signal preprocessing algorithm and a deep learning fault diagnosis algorithm, and uses a reference database under the normal state of the lightning protection system to train a network model; 4) a host computer, which is used to respond to the real-time status of the lightning protection system.

[0029] It includes the following steps:

[0030] The first step: firstly arrange the sound card collector 6 in the inner cavity 5 of the blade 1 and arrange the current loop 3 in the down conductor 2 in the lightning protection system;

[0031] Step 2: Connect the sound card collector 6 and the current loop 3 to the monitoring controller, collect the sound signal and the current signal through the data acquisition module, and perform data preprocessing on them;

[0032] Step 3: The processed data is transmitted to the fault warning and diagnosis module 9, and the deep learning network model is trained multiple times using the data in the normal state to realize the fault warning and diagnosis of the wind turbine blade lightning protection system;

[0033] Step 4: The real-time monitored engine active mount operation status is reflected in the host computer 10. The specific implementation is described as follows.

[0034] Example 1

[0035] like Figure 1 As shown, the lightning protection system fault diagnosis method of this embodiment includes a data acquisition module, a data analysis module 8, a fault monitoring, early warning and diagnosis module 9 and a host computer 10. Specifically, a sound card collector 6 is installed at the root of the blade to collect the sound signal in the blade cavity, and the collected sound signal is converted into a digital signal. The fault monitoring, early warning and diagnosis module 9 uses the digital signal to monitor the lightning protection system in real time through a deep learning fault diagnosis algorithm; at the same time, a current loop 3 is arranged at the root of the blade to collect lightning current. By comparing with the lightning current data of the wind farm throughout the year, the auxiliary fault monitoring, early warning and diagnosis module 9 determines the state of the lightning protection system lead-down line 2 during the service period of the blade, and finally responds to the real-time state of the lightning protection system in the host computer 10. The host computer 10 is connected to the fault monitoring, early warning and diagnosis module 9. The host computer 10 is used to respond to the real-time state, and is also used to parse the real-time state information. When the failure type information is obtained, the information contained in the current failure type is displayed and the early warning response is issued; if the engine active suspension does not fail, only the current normal state information is displayed. The fault monitoring, early warning and diagnosis system process is as follows Figure 3 shown.

[0036] In this embodiment, the data acquisition module is connected to the sound card collector 6 and the current loop 3, and the data analysis module 8 is connected to the data acquisition module to convert the sound signal into a digital signal. The data analysis module 8 includes a data preprocessing algorithm, and the data preprocessing algorithm includes the following steps:

[0037] S1: Normalize the data and map it to the range of 0-1;

[0038] S2: Detect and remove abnormal mutation values ​​from the normalized data to reduce the computational complexity and the impact of abnormal mutation values ​​on fault warning and diagnosis errors;

[0039] S3: Overlap sampling and data enhancement are performed on the data processed in step S2 to highlight the data features. The network model used by the deep learning algorithm is a multi-channel weighted fusion capsule network. Data features are extracted through multi-channel weighted convolution and capsule routing. The network model structure is as follows: Figure 2 shown.

[0040] In this embodiment, the fault warning and diagnosis module 9 has a deep learning fault warning and diagnosis algorithm. The network model used by the deep learning fault warning and diagnosis algorithm is a multi-channel weighted fusion capsule network, and data feature extraction is performed through multi-channel weighted convolution and capsule routing. The deep learning fault warning and diagnosis algorithm uses units with different convolution kernel sizes in each channel to form different receptive fields for feature extraction. The size of the receptive field changes with the change of the convolution kernel, and the two have the following relationship:

[0041] Where n represents the number of convolutional pooling layers; rf n-1 Represents the receptive field size of the n-1th layer; rf n Represents the size of the receptive field of layer n; s n-1 represents the step length of n-1 layers; k n-1 Represents the size of n-1 convolution layers; rf last Represents the size of the receptive field band of the last layer, which is equal to the convolution kernel k of this layer last size.

[0042] After feature extraction through different convolution kernels, the output of each channel is:

[0043] ;in is the output of the lth convolutional layer of channel j; The i-th feature input of the l-1th convolution layer of channel j, with a total of k feature inputs; is the convolution kernel of the l-1th layer of channel j.

[0044] After feature extraction based on the dynamic receptive field, the kurtosis value is used as the weight to measure the difference in importance. Assuming k j represents the kurtosis value of the vibration signal on channel j, m represents the number of channels, and the weight of the channel is: After obtaining the weights of each channel, they need to be weighted fused: Where X l is the output of the lth convolutional layer; is the feature output of the l-1th convolutional layer of channel j; is the bias of the l-1th layer of channel j; is the weight of the l-1th layer of channel j; f l-1 (x) is the activation function of the l-1th convolutional layer; m is the number of channels.

[0045] The capsule network includes a primary capsule layer and a digital capsule layer. The digital capsule layer includes a dynamic routing algorithm. The idea of ​​the capsule network is divided into three stages. The first stage is completed in the primary capsule layer by weighted summation of neurons. The calculation formula is as follows: U j|i =W ij U i , where U i is the output of the i-th neuron in the previous convolutional network; W ij is the output of the neuron U i Multiply the weight matrix; U j|i is the prediction vector; the second stage is the prediction vector U obtained in the first stage j|i Multiply by c ij Perform weighted summation to obtain the output vector S j , the calculation formula is as follows: j =∑ i c ij *U j|i , where c ij is a coupling coefficient, which is determined by the dynamic routing algorithm; in the third stage, the intermediate vector S obtained in the second stage is j The output vector V of the routing capsule layer is obtained by performing nonlinear transformation through the nonlinear activation vector compression function Squashing. j , compress the modulus value of the input vector to between (0,1), and the calculation formula is as follows: in, Represents the total input vector S j The reduction factor of Represents the total input vector S j The unit vector of .

[0046] The core idea of ​​capsule network is dynamic routing algorithm. The core work of dynamic routing is to find the optimal coupling coefficient c ij , c ij The second operation stage of the capsule network plays a very critical role. It determines the connection between the current capsule layer and the previous capsule layer by adjusting the initial vector b. ij Softmax normalization is performed, and the initial vector is obtained by the prediction vector U j|i And the output vector V j Perform iterative updates, and the calculation formula is as follows: b ij =bij +V j *U j|i .

[0047] In each forward propagation, b ij is initialized to 0, and the coupling coefficient c is calculated ij The initialization value of V is then calculated by the forward propagation of the network. j Each forward propagation updates b ij The value of c is then updated ij The value of S is further updated j The value of V j Through this step-by-step iterative update, a set of optimal coupling coefficients is found.

[0048] The monitoring, early warning and diagnosis system in this embodiment also includes a reference data testing platform 11. The reference data testing platform 11 is used to establish a reference database, and transmit the established reference database to the monitoring, early warning and diagnosis module 9 to train the multi-channel weighted fusion capsule network multiple times and save the optimal training model parameters. When the signal during the service of the lightning protection system is collected and transmitted to the monitoring, early warning and diagnosis module, the trained model can be directly used for real-time status monitoring, early warning and diagnosis. The reference database can collect normal service data of each wind farm throughout the year.

[0049] To summarize, the original data is first processed into multiple samples of the same length, and the processed samples are divided into training set and test set in a ratio of 8:2. The training set is first sent to the network model for model training, and then the test set is used for verification. The model parameters with the best fault warning and diagnosis effects are determined, and the real-time data collected during the service of the lightning protection system is subsequently sent to the network model to determine the real-time status of the downconductor of the lightning protection system.

[0050] The wind turbine blade lightning protection system detected by the fault diagnosis method of this embodiment is as follows: Figure 4As shown, the blade 1 includes a down conductor 2, a current loop 3, a web 4, a cavity 5 and a sound card collector 6. The sound card collector 6 is attached to the inner plate 7 of the blade 1, and the down conductor 2 is attached to the surface of the web 4; the current loop 3 is fixed to the outside of the down conductor 2. During the operation of the blade 1, a large amount of sound is generated in the cavity 5 and collected by the sound card collector 6, and its natural frequency will shift as the effective flow area of ​​the down conductor 2 becomes smaller; at the same time, the current loop 3 is used to collect the lightning current in the down conductor 2, and the lightning current data of the wind field in four seasons are compared to assist in judgment. The sound card collector 6 and the current loop 3 transmit the real-time collected sound signals and currents to the data analysis module 8 for conversion, and obtain the unprocessed original sound signals and current original signals, and then the obtained original signals are normalized to eliminate abnormal values, etc., and the original data is enhanced by overlapping sampling.

[0051] The data analysis module 8 transmits the converted signal to the fault monitoring, early warning and diagnosis module 9, which transmits the enhanced data set to the trained multi-channel weighted fusion capsule network model for fault early warning and diagnosis. Finally, the real-time status of the lightning protection system fault diagnosis method is transmitted to the host computer 10 for display. When the host computer 10 obtains the failure information, it warns and diagnoses the current failure phenomenon.

[0052] Figure 5 This is a schematic diagram of the failure warning and diagnosis process in the lightning protection system fault diagnosis method. The failure forms include: lightning protection system failure evolution stage 1 (corresponding to threshold 1), lightning protection system failure evolution stage 2 (corresponding to threshold 2), and down conductor breakage.

[0053] The fault detection, early warning and diagnosis module 9 uses the signals collected by the sound card collector 6 and the current loop 3 to monitor, warn and diagnose the working state of the wind turbine blade lightning protection system during service. When the real-time signal of the lightning protection system during service is input into the network model, it will judge whether its working state is normal according to the characteristics of the input signal. If the output tensor is 0 or other constants, the model outputs the down conductor broken and fails, and synchronously displays its current state on the host computer 10; if the RMS value of the tensor matrix output by the network model is greater than the threshold 1 and less than the threshold 2, the model outputs the lightning protection system failure evolution stage 1 and issues an early warning prompt in time, and synchronously displays its current state on the host computer 10; if the RMS value of the tensor matrix output by the network model is greater than the threshold 2, the model outputs the lightning protection system failure evolution stage 2 and issues an early warning prompt in time, and synchronously displays its current state on the host computer 10. The above-mentioned failure warning situation can be carried out simultaneously with the fault diagnosis.

[0054] The present invention can monitor the operating status of the wind power lightning protection system in real time during the service period of the blades, reduce the risk of failure of the lightning protection system due to mechanical stress caused by deformation of the down conductor in the later stage, and improve the safety performance of the wind turbine operation. The fault monitoring, early warning and diagnosis system can monitor the status of the lightning protection system in real time during the service period, and can promptly repair or replace the down conductor that is broken due to mechanical stress caused by deformation, thereby improving the reliability of the lightning protection system and reducing the risk of lightning protection failure.

[0055] The above are only embodiments of the present invention. The invention is not limited to the field involved in this implementation case. The common knowledge such as the known specific structure and characteristics in the scheme is not described in detail here. It should be pointed out that for those skilled in the art, several deformations and improvements can be made without departing from the content of the present invention, which should also be regarded as the protection scope of the present invention, and these will not affect the effect of the implementation of the present invention and the practicality of the patent. The scope of protection required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to interpret the content of the claims.

Claims

1. A lightning protection system fault diagnosis method, characterized in that: A data sound card collector is installed at the root of the blade to collect sound signals in the blade cavity and convert the collected sound signals into digital signals. The fault monitoring, fault warning and diagnosis modules use digital signals to monitor the lightning protection system in real time through deep learning fault diagnosis algorithms. At the same time, a current loop is arranged at the root of the blade to collect lightning current. By comparing it with the existing actual lightning current data, the fault monitoring, warning and diagnosis modules are assisted to judge the status of the lightning protection system lead-down line during the service of the blade, and finally respond to the real-time status of the lightning protection system on the host computer.

2. The lightning protection system fault diagnosis method according to claim 1, characterized in that: The data acquisition module is connected to the sound card collector and the current loop, and the data analysis module is connected to the data acquisition module for converting the sound signal into a digital signal.

3. The lightning protection system fault diagnosis method according to claim 2, characterized in that: The data analysis module includes a data preprocessing algorithm, and the data preprocessing algorithm includes the following steps: S1: Normalize the data and map it to the range of 0-1; S2: Detect and remove abnormal mutation values ​​from the normalized data to reduce the computational complexity and the impact of abnormal mutation values ​​on fault warning and diagnosis errors; S3: Overlap sampling is performed on the data processed in step S2 to enhance the data and highlight the data features.

4. The lightning protection system fault diagnosis method according to claim 1, characterized in that: The fault warning and diagnosis module is equipped with a deep learning fault warning and diagnosis algorithm. The network model adopted by the deep learning fault warning and diagnosis algorithm is a multi-channel weighted fusion capsule network, and data feature extraction is performed through multi-channel weighted convolution and capsule routing.

5. The lightning protection system fault diagnosis method according to claim 4, characterized in that: The deep learning fault warning and diagnosis algorithm uses units with different convolution kernel sizes in each channel to form different receptive fields for feature extraction. The size of the receptive field changes with the change of the convolution kernel, and the two have the following relationship: Where n represents the number of convolutional pooling layers; rf n-1 Represents the receptive field size of the n-1th layer; rf n represents the size of the receptive field of layer n; s n-1 represents the step length of n-1 layers; k n-1 Represents the size of n-1 convolution layers; rf last Represents the size of the receptive field band of the last layer, which is equal to the convolution kernel k of this layer last size.

6. The lightning protection system fault diagnosis method according to claim 5, characterized in that: After feature extraction through different convolution kernels, the output of each channel is: in is the output of the lth convolutional layer of channel j; The i-th feature input of the l-1th convolution layer of channel j, with a total of k feature inputs; is the convolution kernel of the l-1th layer of channel j.

7. The lightning protection system fault diagnosis method according to claim 6, characterized in that: After feature extraction based on the dynamic receptive field, the kurtosis value is used as the weight to measure the difference in importance. Assuming k j represents the kurtosis value of the vibration signal on channel j, m represents the number of channels, and the weight of the channel is: After obtaining the weights of each channel, they need to be weighted fused: Where X l is the output of the lth convolutional layer; is the feature output of the l-1th convolutional layer of channel j; is the bias of the l-1th layer of channel j; is the weight of the l-1th layer of channel j; f l-1 (x) is the activation function of the l-1th convolutional layer; m is the number of channels.

8. The lightning protection system fault diagnosis method according to claim 4, characterized in that: The capsule network includes a primary capsule layer and a digital capsule layer. The digital capsule layer includes a dynamic routing algorithm. The idea of ​​the capsule network is divided into three stages. The first stage is completed in the primary capsule layer by weighted summation of neurons. The calculation formula is as follows: U j|i =W ij U i , where U i is the output of the i-th neuron in the previous convolutional network; W ij is the output of the neuron U i Multiply the weight matrix; U j|i is the prediction vector; the second stage is the prediction vector U obtained in the first stage j|i Multiply by c ij Perform weighted summation to obtain the output vector S j , the calculation formula is as follows: j =∑ i c ij *U j|i , where c ij is a coupling coefficient, which is determined by the dynamic routing algorithm; in the third stage, the intermediate vector S obtained in the second stage is j The output vector V of the routing capsule layer is obtained by performing nonlinear transformation through the nonlinear activation vector compression function Squashing. j , compress the modulus value of the input vector to between (0,1), and the calculation formula is as follows: in, Represents the total input vector S j The reduction factor of Represents the total input vector S j The unit vector of .

9. A wind turbine blade lightning protection system detected by the fault diagnosis method according to any one of claims 1 to 8.

Images