CN111814849B - A fault early warning method for key components of wind turbines based on DA-RNN - Google Patents

Abstract

The invention discloses a fault early warning method for key components of wind turbine generators based on a recurrent neural network DA-RNN with a dual attention mechanism. This method is based on data sets collected from the Supervisory Control and Data Acquisition (SCADA) system under normal operating conditions of wind turbines. It designs a preprocessing process and selects the DA‑RNN model for real-time estimation of variables. Through multi-threshold setting and discrimination criterion design, the judgment results are output. sequence, and give the final warning result based on the judgment result sequence. In the fault early warning method of the present invention, the preprocessing process is designed for different types of noise data, which provides a reliable data basis; the DA-RNN model comprehensively considers the influence of relevant variables and historical information, and assigns different weights to ensure the accuracy of variable estimation. Accuracy; multiple threshold settings and discrimination criterion design avoid a single 0-1 judgment, making the final early warning result more robust; ultimately realizing key component failure early warning, reducing unit downtime, saving operation and maintenance costs, and having strong Theoretical and practical.

Description

A fault early warning method for key components of wind turbines based on DA-RNN

Technical field

The invention relates to a DA-RNN-based fault early warning method for key components of wind turbines. Based on the wind turbine normal operating status data set, a standardized data preprocessing process is designed, and DA-RNN is selected as a variable estimation model to estimate target variables in real time. Based on the real-time Run the residuals and design a fault warning strategy based on multi-threshold settings combined with multi-discrimination criterion selection to provide fault warning for key components.

Background technique

With global pollution and the increasing scarcity of traditional fossil energy, the development of clean energy has attracted widespread attention. Wind energy has developed rapidly due to its clean and pollution-free advantages. The wind power industry has thus become one of the new renewable energy industries being vigorously developed at home and abroad. one. At present, my country's total installed capacity of wind turbines ranks among the top in the world. However, the rapid development of the wind power market in recent years has also led to insufficient preparation during the research and development period, and the operation and maintenance costs of wind turbines remain high.

The high failure rate of wind turbines is the main factor leading to high operation and maintenance costs. Wind turbines are complex systems composed of multiple components and subsystems. The units usually operate in remote areas such as suburban plains, mountainous areas, and coastal areas. The operating environment is harsh and changeable. The key is Component failure will cause the entire machine to be shut down for maintenance, resulting in a large amount of economic losses. Therefore, it is of great significance to realize the initial identification of abnormalities in key components, avoid early abnormalities from evolving into catastrophic failures, and achieve early warning of failure of key components to perform predictive maintenance, which is of great significance to reducing operation and maintenance costs and realizing intelligent operation and maintenance of wind farms. However, it is difficult for existing variable estimation models for fault early warning to comprehensively consider the impact of relevant variables and historical information, and it is difficult for existing simple early warning strategies to ensure the accuracy of early warning results. Therefore, selecting a more accurate variable estimation model and designing a robust early warning strategy are of great significance to achieve accurate fault early warning.

Contents of the invention

The purpose of the present invention is to provide a fault early warning method for key components of wind turbine generators through accurate estimation of target variables and designing a robust early warning strategy, and propose a fault early warning method for key components of wind turbine generators based on DA-RNN. This method selects the data set of the normal operating status of the wind turbine, first considers different noise data types to design the data preprocessing process, and then selects DA-RNN as the variable estimation model to comprehensively consider the impact of relevant variables and historical information to estimate the target variable in real time. The accuracy of the estimation is ensured, and a robust early warning strategy is designed, taking into account different early warning result requirements and different abnormal characteristics, ensuring the flexibility and accuracy of the early warning results. This method can be extended to key components with temperature measurement points in the wind turbine to achieve early warning of faults of key components. It has practical value and strong scalability.

The purpose of the present invention is achieved through the following technical solution: a DA-RNN-based wind turbine key component fault early warning method, which method includes the following steps:

1) Select the wind turbine for which fault warning is to be carried out, and obtain N pieces of operating data under normal operating conditions recorded in the SCADA system of the wind turbine. The key components of the wind turbine include gearboxes, generators, pitch systems, etc. Select the wind turbine to be used for fault warning in the SCADA system. The temperature variable measured at the temperature measuring point of the early warning component is used as the target variable y, and all variables related to the temperature of the component are used as related variables X to construct an initial training set.

2) Offline training phase, based on the initial training set Design the data preprocessing process. The preprocessing steps include the elimination and interpolation of isolated abnormal points, the elimination and interpolation of data during wind turbine shutdown maintenance based on operation and maintenance records, and the interpolation of missing values. The preprocessed training set [X _train , y _train ] as input to the variable estimation model for model training;

3) Select the recurrent neural network DA-RNN model based on the dual attention mechanism as the variable estimation model, select the sliding window length W, and the time within the sliding window is expressed as t _w , w=T-W+1, T-W+2, ...,T, where T is the current moment, and the input attention mechanism is designed to allocate relevant variables t _w at each moment To the target variable/> The weight of influence/> Refactor related variables/> As the encoder input, the encoder part is composed of multiple LSTM units. The input of each LSTM unit is the reconstructed relevant variable at a moment in the sliding window, and the output of the encoder part is the hidden vector h; then the temporal attention mechanism is designed to allocate the sliding window Hidden vectors at different historical moments within the target variable at the current moment/> The weight of influence/> The semantic vector c is obtained, and a linear regression model is used to integrate the semantic vector and the historical value of the target variable as the input of the decoder. The decoder part is composed of multiple LSTM units, and the output is the hidden state of the previous moment t _T-1 at the current moment, and then through linear The function obtains the estimated value of the target variable at the current moment/>

4) Subtract the model output at the corresponding moment from the actual value of the target variable in the training set, that is, the estimated value of the target variable, to obtain the estimated residual sequence of the training set, and obtain the mean μ _train and standard deviation σ _train of the residual sequence;

5) Multi-threshold setting is performed based on the training set residual sequence. The multi-threshold setting is the mean μ _train of the estimated residual sequence of the training set plus or minus k times the standard deviation σ _train , which are respectively used as the upper and lower limits of the residual sequence threshold, where the upper limit U _r ( k)=μ _train +kσ _train , lower limit L _r (k)=μ _train -kσ _train ; the higher the upper limit of the threshold, or the lower the lower limit, the fewer the number of data points exceeding this threshold in the online application phase, corresponding to the fault warning results The false positive rate will decrease and the false negative rate will increase;

6) In the online application phase, based on the real-time running data point d, based on the DA-RNN model trained in the offline phase, the estimated residual value r _d of the actual measured value of the data point d minus the model estimated value is obtained;

7) Select a value of k and determine the upper and lower limits of the threshold. If r _d exceeds the threshold corresponding to the current k value, that is, is greater than the upper threshold U _r (k) or less than the lower threshold L _r (k), calculate the continuous values before the data point d The number of data points that exceed the threshold count1(k), and the percentage value of the number of data points that exceed the threshold in the time range of the previous day of data point d to the total number of data points in one day count2(k)%, if r _d is in Between the upper threshold U _r (k) and the lower threshold L _r (k), two judgment results 0 are output;

8) Set the multi-discrimination criterion. The multi-discrimination criterion is the continuous over-limit discrimination criterion combined with the percentage over-limit discrimination criterion. Under the current k value, set the threshold parameter S(k) of the continuous over-limit discrimination criterion and the threshold parameter P of the percentage over-limit discrimination criterion. (k)%. When the condition count1(k)≥S(k) is met, the judgment result of the k-S(k) early warning strategy combination is recorded as 1. When the condition is not met, the judgment result is recorded as 0. When the condition count2(k)≥ When P(k), the judgment result of the k-P(k) early warning strategy combination is recorded as 1, and when the conditions are not met, the judgment result is recorded as 0;

9) For all other values of k, repeat steps 7) and 8) to obtain the discrimination result sequence. Based on the discrimination result sequence, determine whether to give a final alarm for the real-time data point d.

Further, in step 2), the preprocessing process of offline training includes the following steps:

a) Isolated abnormal points are usually recording errors caused by sensor abnormalities. They are judged through the operating mechanism. The judgment conditions are: for the temperature variable, the value is greater than 150 degrees or less than 0 degrees; for the wind speed variable, the value is greater than the cut-out wind speed of the unit or less than Cut into the wind speed; for power variables, the value is greater than the rated power of the unit or is a negative value; when the above conditions are met, the data is judged to be an isolated abnormal point and will be eliminated;

b) Combined with the operation and maintenance records, when the wind turbine unit is in operation and maintenance, the unit is in a shutdown state. The data recorded by the SCADA system is usually 0 value or the system default value. This value cannot represent the normal operating status of the unit, so shutdown maintenance and inspection are not required. Data during the period will be deleted;

c) In order to ensure the time continuity of the data, the deleted data and missing values in the SCADA system are interpolated. The interpolation method is average interpolation, that is, the average value of the data 1 hour before the variable interpolation position is used as the current time interpolation. value.

Further, in step 3), the model input is the relevant variables in the sliding window Where n is the number of relevant variables, the input to the attention mechanism part is the relevant variable The calculation process is as follows:

Among them v _en , W _en , U _en are the parameters that need to be learned. At time t _w , based on the influence weight Reconstruct the relevant variables to obtain the reconstructed relevant variables at time t _w , which is expressed as:

The reconstructed vector is used as the input of the encoder part. The encoder is composed of multiple LSTM units, and the output is the hidden state h and the memory unit s. At time t _w , the LSTM update consists of the forget gate f _t , the input gate i _t , and the output gate o _t is decided, and its update rules are as follows:

Among them, W _f , W _i , W _o , W _s , b _f , b _i , _bo , and b _s are the parameters that need to be learned, σ is the sigmoid function, and ⊙ represents the multiplication of corresponding elements;

The input of the temporal attention mechanism is the hidden state h, the decoder hidden state output h′ and the memory unit output s′, and the weight of the i-th hidden state is output. The calculation process is as follows:

Among them, v _de , W _{de and} U _de are the parameters that need to be learned. The hidden state is reconstructed based on the weight β to obtain the semantic vector, which is expressed as:

Then use a linear regression model to integrate the semantic vector and the historical value of the target variable to obtain the input of the decoder. Expressed as:

in For the parameters that need to be learned;

The decoder is also composed of LSTM units. The update rules of forget gate f _t ′, input gate i _t ′, and output gate o _t ′ are as follows:

Among them, W′ _f ,W′ _i ,W′ _o ,W′ _s ,b′ _f ,b′ _i ,b′ _o ,b′ _s are the parameters that need to be learned;

Finally, the estimated value of the T target variable at the current moment is obtained through a linear function. As follows:

Among them, W _T and b _T are the parameters that need to be learned.

Furthermore, in step 5), it can be assumed that the training set residual sequence obeys the normal distribution. Based on the characteristics of the normal distribution, the value of k in the multi-threshold setting is [1.5, 2, 2.5, 3].

Further, in the step 8), the multi-discrimination criterion corresponds to different abnormal characteristics, the continuous over-limit criterion corresponds to the abnormal characteristics of continuous outliers, the percentage over-limit criterion corresponds to the abnormal situation of large data fluctuations, and the continuous over-limit criterion corresponds to the abnormal situation of large data fluctuations. The judgment criterion threshold parameter _Sk and the percentage overrun judgment criterion threshold parameter P _k % are set to where C is a constant.

Further, in step 9), for the real-time data point d, each k value outputs two 0-1 judgment results, and the final output is a 0-1 sequence with a dimension of 2k. When the number of 1s in the sequence When it is greater than k, an alarm will be given for the real-time data point. When the number of 1s in the sequence is less than or equal to k, no alarm will be given.

Compared with the existing technology, the present invention has the following innovative advantages and significant effects:

1) Design a standardized data preprocessing process for different types of noise data existing in the data set during normal operation, ensuring that the constructed training set can accurately represent the normal operating status of the wind turbine and provide a data basis for accurate estimation of target variables;

2) Select DA-RNN as the variable estimation model, comprehensively consider the influence of relevant variables and historical information, and determine its different impact weights on target variable estimation through a double-layer attention mechanism, ensuring the reliability and accuracy of the model;

3) Design a robust fault early warning strategy. Multi-threshold settings can comprehensively consider the requirements of early warning results for different false alarm rates and false negative rates. Two discrimination criteria provide a more comprehensive early warning for different abnormal characteristics, based on the judgment result sequence. The final early warning results are given, and different combinations of early warning strategies are comprehensively considered to ensure the robustness and accuracy of the early warning results;

4) The present invention is a key component failure early warning method for temperature parameters. This process is applicable to all key components of wind turbines with corresponding temperature measurement points and is scalable.

Description of drawings

Figure 1 is a flow chart of the fault early warning method for key components of wind turbines according to the present invention;

Figure 2 is a data diagram before preprocessing of selected target variables in the embodiment of the present invention;

Figure 3 is a data diagram after preprocessing of selected target variables in the embodiment of the present invention;

Figure 4 is a schematic diagram of the encoder that adds the input attention mechanism to the DA-RNN model selected by the present invention;

Figure 5 is a schematic diagram of the decoder that adds a temporal attention mechanism to the DA-RNN model selected by the present invention;

Figure 6 is a diagram of the target variable estimation results of the variable estimation model according to the embodiment of the present invention;

Figure 7 is a schematic diagram of multiple threshold settings according to an embodiment of the present invention;

Figure 8 is a diagram of early warning results according to the embodiment of the present invention.

Detailed ways

In order to make the above objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

Many specific details are set forth in the following description to fully understand the present invention. However, the present invention can also be implemented in other ways different from those described here. Those skilled in the art can do so without departing from the connotation of the present invention. Similar generalizations are made, and therefore the present invention is not limited to the specific embodiments disclosed below.

As shown in Figure 1, this application proposes a DA-RNN-based fault early warning method for key components of wind turbines, including:

1) Select the wind turbine for which fault warning is to be carried out, and obtain N pieces of operating data under normal operating conditions recorded in the SCADA system of the wind turbine. The key components of the wind turbine include gearboxes, generators, pitch systems, etc. Select the wind turbine to be used for fault warning in the SCADA system. The temperature variable measured at the temperature measuring point of the early warning component is used as the target variable y, and all variables related to the temperature of the component are used as related variables X to construct an initial training set.

2) Offline training phase, based on the initial training set Design the data preprocessing process. The preprocessing steps include the elimination and interpolation of isolated abnormal points, the elimination and interpolation of data during wind turbine shutdown maintenance based on operation and maintenance records, and the interpolation of missing values. The preprocessed training set [X _train , y _train ] as input to the variable estimation model for model training;

3) Select the recurrent neural network DA-RNN model based on the dual attention mechanism as the variable estimation model, select the sliding window length W, and the time within the sliding window is expressed as t _w , w=T-W+1, T-W+2, ...,T, where T is the current moment, and the input attention mechanism is designed to allocate relevant variables t _w at each moment To the target variable/> The weight of influence/> Refactor related variables/> As the encoder input, the encoder part is composed of multiple LSTM units. The input of each LSTM unit is the reconstructed relevant variable at a moment in the sliding window, and the output of the encoder part is the hidden vector h; then the temporal attention mechanism is designed to allocate the sliding window Hidden vectors at different historical moments within the target variable at the current moment/> The weight of influence/> The semantic vector c is obtained, and a linear regression model is used to integrate the semantic vector and the historical value of the target variable as the input of the decoder. The decoder part is composed of multiple LSTM units, and the output is the hidden state of the previous moment t _T-1 at the current moment, and then through linear The function obtains the estimated value of the target variable at the current moment/>

4) Subtract the model output at the corresponding moment from the actual value of the target variable in the training set, that is, the estimated value of the target variable, to obtain the estimated residual sequence of the training set, and obtain the mean μ _train and standard deviation σ _train of the residual sequence;

5) Multi-threshold setting is performed based on the training set residual sequence. The multi-threshold setting is the mean μ _train of the estimated residual sequence of the training set plus or minus k times the standard deviation σ _train , which are respectively used as the upper and lower limits of the residual sequence threshold, where the upper limit U _r ( k)=μ _train +kσ _train , lower limit L _r (k)=μ _train -kσ _train ; the higher the upper limit of the threshold, or the lower the lower limit, the fewer the number of data points exceeding this threshold in the online application phase, corresponding to the fault warning results The false positive rate will decrease and the false negative rate will increase;

6) In the online application phase, based on the real-time running data point d, based on the DA-RNN model trained in the offline phase, the estimated residual value r _d of the actual measured value of the data point d minus the model estimated value is obtained;

7) Select a value of k and determine the upper and lower limits of the threshold. If r _d exceeds the threshold corresponding to the current k value, that is, is greater than the upper threshold U _r (k) or less than the lower threshold L _r (k), calculate the continuous values before the data point d The number of data points that exceed the threshold count1(k), and the percentage value of the number of data points that exceed the threshold in the time range of the previous day of data point d to the total number of data points in one day count2(k)%, if r _d is in Between the upper threshold U _r (k) and the lower threshold L _r (k), two judgment results 0 are output;

8) Set the multi-discrimination criterion. The multi-discrimination criterion is the continuous over-limit discrimination criterion combined with the percentage over-limit discrimination criterion. Under the current k value, set the threshold parameter S(k) of the continuous over-limit discrimination criterion and the threshold parameter P of the percentage over-limit discrimination criterion. (k)%. When the condition count1(k)≥S(k) is met, the judgment result of the k-S(k) early warning strategy combination is recorded as 1. When the condition is not met, the judgment result is recorded as 0. When the condition count2(k)≥ When P(k), the judgment result of the k-P(k) early warning strategy combination is recorded as 1, and when the conditions are not met, the judgment result is recorded as 0;

9) For all other values of k, repeat steps 7) and 8) to obtain the discrimination result sequence. Based on the discrimination result sequence, determine whether to give a final alarm for the real-time data point d.

Further, in step 3), part of the process of the encoder and input attention mechanism is shown in Figure 4, and the input is the relevant variable in the sliding window Where n is the number of relevant variables, the input to the attention mechanism part is the relevant variable The calculation process is as follows:

Among them v _en , W _en , U _en are the parameters that need to be learned. At time t _w , based on the influence weight Reconstruct the relevant variables to obtain the reconstructed relevant variables at time t _w , which is expressed as:

The reconstructed vector is used as the input of the encoder part. The encoder is composed of multiple LSTM units, and the output is the hidden state h and the memory unit s. At time t _w , the LSTM update consists of the forgetting gate f _t , the input gate i _t , and the output gate o _t is decided, and its update rules are as follows:

Among them, W _f , W _i , W _o , W _s , b _f , b _i , _bo , and b _s are the parameters that need to be learned, σ is the sigmoid function, and ⊙ represents the multiplication of corresponding elements;

Part of the process of the decoder and temporal attention mechanism is shown in Figure 5. The input of the temporal attention mechanism is the hidden state h, the decoder hidden state output h′ and the memory unit output s′, and the weight of the i-th hidden state is output. The calculation process is as follows:

Among them, v _de , W _{de and} U _de are the parameters that need to be learned. The hidden state is reconstructed based on the weight β to obtain the semantic vector, which is expressed as:

Then use a linear regression model to integrate the semantic vector and the historical value of the target variable to obtain the input of the decoder. Expressed as:

in For the parameters that need to be learned;

The decoder is also composed of LSTM units. The update rules of forget gate f _t ′, input gate i _t ′, and output gate o _t ′ are as follows:

Among them, W′ _f ,W′ _i ,W′ _o ,W′ _s ,b′ _f ,b′ _i ,b′ _o ,b′ _s are the parameters that need to be learned;

Finally, the estimated value of the T target variable at the current moment is obtained through a linear function. As follows:

Among them, W _T and b _T are the parameters that need to be learned.

An example of the present application is given below, and the specific steps for implementing this example are described in detail with reference to Table 1, Table 2, and Figures 2-8.

This embodiment provides a fault warning for a wind turbine in a wind farm that has suffered damage to the non-driving end bearing assembly of the generator. The unit suffered damage to the non-driving end bearing of the generator on April 17, 2017. The SCADA system of the wind turbine was selected from 2016 to The data collected in 2017 are used for fault warning. The data sampling interval of the SCADA system is 5 minutes. The data information lasts for 16 months. The time range is from 2016.01.01 00:00:00 to 2017.04.30 23:55:00. Power generation is selected. The temperature measured at the non-drive end bearing temperature measuring point of the machine is the target variable, and all other operating parameters of the generator and system parameters that affect the target variable value are used as relevant variables. The specific variables of the data set are shown in Table 1:

Table 1 Target variables and related variables of a wind turbine in a wind farm

The implementation data set of the wind turbine generator non-driving end bearing component failure early warning method in this embodiment is the 16-month operating data of the above-mentioned wind turbine generator. The implementation steps of the method are as follows:

1) Obtain the operating data set under normal operating conditions recorded in the SCADA system of the wind turbine. The data set includes the target variable generator non-drive end bearing temperature and all related variables. Select the data in normal operating conditions in the previous 12 months, that is, 2016.01 .01 00:00:00 to 2016.12.31 23:55:00 to construct the initial training set Construct the initial test set using data from the last 4 months, that is, 2017.01.01 00:00:00 to 2017.04.30 23:55:00

2) For the initial training set All variables in the data are preprocessed. The preprocessing steps include the elimination and interpolation of isolated abnormal points, the elimination and interpolation of data during fan shutdown and maintenance based on operation and maintenance records, and the interpolation of missing values. In this embodiment, cut in The wind speed is 2m/s, the rated wind speed is 14m/s, the cut-out wind speed is 25m/s, and the rated power is 1500kW. The test set/> As a real-time operating data set for online applications, the training set plus test set data before preprocessing of the target variable generator non-drive end bearing temperature is shown in Figure 2. Missing values and data during shutdown all appear as 0 values. Preprocessing The final training set and test set data are shown in Figure 3;

3) Select the DA-RNN model as the variable estimation model, select the sliding window length W = 5, and the time in the sliding window as t _w , w = T-4, T-3,..., T, and input the attention mechanism through design Assign relevant variables at each time t _w To the target variable/> The weight of influence/> Refactor related variables/> As the encoder input, the encoder part is composed of multiple LSTM units. The input of each LSTM unit is the reconstructed relevant variable at a moment in the sliding window, and the output of the encoder part is the hidden vector h; then the temporal attention mechanism is designed to allocate the sliding window Hidden vectors at different historical moments within the target variable at the current moment/> The weight of influence/> The semantic vector c is obtained, and a linear regression model is used to integrate the semantic vector and the historical value of the target variable as the input of the decoder. The decoder part is composed of multiple LSTM units, and the output is the hidden state of the previous moment t _T-1 at the current moment, and then through linear The function obtains the estimated value of the target variable at the current moment/> After training the model through the training set, input the test set to obtain the real-time estimated value of the test set. The variable estimation results are shown in Figure 6. The black circle is the actual operating value of the target variable corresponding to the fault confirmation time;

4) Subtract the model output at the corresponding moment from the actual value of the target variable in the training set, that is, the estimated value of the target variable, to obtain the estimated residual sequence of the training set, and obtain the mean μ _train and standard deviation σ _train of the residual sequence;

5) Multi-threshold setting is performed based on the training set residual sequence. The multi-threshold setting is the mean μ _train of the estimated residual sequence of the training set plus or minus k times the standard deviation σ _train , which are respectively used as the upper and lower limits of the residual sequence threshold, where the upper limit U _r ( k)=μ _train +kσ _train , the lower limit L _r (k)=μ _train -kσ _train , where the value of k is [1.5, 2, 2.5, 3]. In Figure 7, the black solid lines are the upper limits of multiple thresholds. The black dotted lines are the lower limits of multiple thresholds;

6) The test set is used as a data set in the online application phase. Based on the real-time running of data point d in the test set, through the DA-RNN model trained in the offline phase, the estimated residual value r of the actual measured value of data point d minus the model estimated value is obtained. _d ;

7) Select a value of k and determine the upper and lower limits of the threshold. If r _d exceeds the threshold corresponding to the current k value, that is, is greater than the upper threshold U _r (k) or less than the lower threshold L _r (k), calculate the continuous values before the data point d The number of data points that exceed the threshold count1(k), and the percentage value of the number of data points that exceed the threshold in the time range of the previous day of data point d to the total number of data points in one day count2(k)%, if r _d is in Between the upper threshold U _r (k) and the lower threshold L _r (k), two judgment results 0 are output;

8) Set the multi-discrimination criterion. The multi-discrimination criterion is the continuous over-limit discrimination criterion combined with the percentage over-limit discrimination criterion. Under the current k value, set the threshold parameter S(k) of the continuous over-limit discrimination criterion and the threshold parameter P of the percentage over-limit discrimination criterion. (k)%, in this embodiment the parameters are set to When the condition count1(k)≥S(k) is met, the judgment result of the kS(k) early warning strategy combination is recorded as 1. When the condition is not met, the judgment result is recorded as 0. When the condition count2(k)≥P(k) is met, the judgment result is recorded as 1. The judgment result of this kP(k) early warning strategy combination is recorded as 1, and when the conditions are not met, the judgment result is recorded as 0;

9) Repeat steps 7) and 8) for all other values of k. Each k value outputs two 0-1 judgment results, and finally outputs a 0-1 sequence for the data point d. In this embodiment, the sequence The dimension is 8. When the number of 1 in the sequence is greater than 4, an alarm is given for the data point d. When the number of 1 in the sequence is less than or equal to 4, no alarm is given. This operation is performed for all data points in the test set. Figure 8 shows the final warning result. The black asterisks are the data points that gave the alarm, and the black circles are the data points at the fault confirmation time. The earliest alarm time is 2017.03.23 02:00:00. The judgment results at this time point are as follows: Show:

Table 2 Judgment results of earliest alarm time

Parameter selection Percent over limit Continuously exceeding the limit k=1.5,S=P=60 0 1 k=2,S=P=45 1 1 k=2.5,S=P=36 1 1 k=3,S=P=30 0 1

The judgment result sequence is [0,1,1,1,1,1,0,1], and the number of 1's is 6, so an alarm is finally given at this moment, and an early warning is implemented 24 days before the fault occurs.

The invention's wind turbine key component fault early warning method mainly includes target variable and related variable selection, data preprocessing, variable estimation model training, real-time operation residual acquisition, robust early warning strategy design and other links. Figure 1 is the specific process of key component failure warning of the present invention. Figure 2 and Figure 3 are data diagrams before and after preprocessing of the selected target variables in the embodiment of the present invention. Figure 4 is the DA-RNN model selected by the present invention adding input attention. Schematic diagram of the encoder of the force mechanism. Figure 5 is a schematic diagram of the decoder of the DA-RNN model selected by the present invention adding the temporal attention mechanism. Figure 6 is a diagram of the target variable estimation results of the variable estimation model in the embodiment of the present invention. Figure 7 is a diagram of the target variable estimation results of the present invention. A schematic diagram of multi-threshold settings of the embodiment. Figure 8 is an early warning result diagram of the embodiment of the present invention. The result shows that the present invention can achieve accurate alarm before a fault occurs, and the result is effective and reliable.

The above embodiments are only examples of the present invention. Although the best examples and drawings of the present invention are disclosed for illustrative purposes, those skilled in the art can understand that: without departing from the spirit and scope of the present invention and the appended claims, , various substitutions, changes and modifications are possible. Therefore, the present invention should not be limited to the disclosure of the preferred embodiments and drawings.

Claims (6)

1. A fault early warning method for key components of a wind turbine generator based on DA-RNN is characterized by comprising the following steps:

1) Selecting a wind turbine generator to be subjected to fault early warning, acquiring N pieces of operation data recorded in a SCADA (supervisory control and data acquisition) system of the fan in a normal operation state, and selecting a temperature variable measured at a temperature measuring point of a component to be early warned in the SCADA system as a target variable y, wherein the temperature variable is related to the temperature of the componentWith variables as related variables X, constructing an initial training set

2) Offline training stage, based on initial training setDesigning a data preprocessing flow, wherein the preprocessing step comprises the steps of eliminating and interpolating isolated abnormal points, eliminating and interpolating data during fan shutdown maintenance based on operation and maintenance records, interpolating missing values, and preprocessing a training set [ X ] _train ,y _train ]Performing model training as input of a variable estimation model;

3) Selecting a circulating neural network DA-RNN model based on a double-attention mechanism as a variable estimation model, selecting a sliding window length W, and expressing the moment in the sliding window as t _w W=t-w+1, T-w+2, &..t, where T is the current time, each time T is assigned by designing an input attention mechanism _w Related variableFor the target variable->Influence weight of->Reconstruction related variable +.>As encoder input, the encoder part is a plurality of LSTM units, the input of each LSTM unit is a reconstruction related variable at one moment in the sliding window, and the encoder part outputs as a hidden vector h; then designing a time attention mechanism to distribute hidden vectors of different historical moments in the sliding window to target variables of the current moment +.>Influence weight of->Obtaining a semantic vector c, integrating the semantic vector and a target variable history value by using a linear regression model as input of a decoder, wherein the decoder is divided into a plurality of LSTM units and outputs the LSTM units as a time t before the current time _T-1 Obtaining the estimated value of the target variable at the current moment through a linear function>

4) Subtracting the model output at the corresponding moment from the actual value of the target variable in the training set, namely the estimated value of the target variable, obtaining an estimated residual sequence of the training set, and obtaining the average value mu of the residual sequence _train Standard deviation sigma _train ；

5) Multi-threshold setting is carried out based on the residual sequence of the training set, and the multi-threshold setting is that the average value mu of the estimated residual sequence of the training set is set _train Plus or minus k times standard deviation sigma _train Respectively used as the upper and lower limits of the residual sequence threshold, wherein the upper limit U _r (k)＝μ _train +kσ _train Lower limit L _r (k)＝μ _train -kσ _train ；

6) In the online application stage, based on real-time operation data point d and on DA-RNN model trained in the offline stage, an estimated residual value r of the actual measured value of the data point d minus the estimated value of the model is obtained _d ；

7) Selecting a value of k, determining upper and lower thresholds, if r _d Exceeding the corresponding threshold of the current k value, i.e. greater than the upper threshold limit U _r (k) Or less than the threshold lower limit L _r (k) Calculating the number of data points count1 (k) which continuously exceed the threshold value before the data point d, and calculating the percentage value count2 (k)% of the number of data points which exceed the threshold value in the time range of day before the data point d, if r _d At the upper threshold limit U _r (k) And a lower threshold limit L _r (k) Outputting two judgment results 0;

8) Setting multiple discriminants, wherein the multiple discriminants are continuous overrun discriminants combined with percentage overrun discriminants, setting a continuous overrun discriminant threshold parameter S (k) and a percentage overrun discriminant threshold parameter P (k)%, recording the judgment result of the k-S (k) early warning strategy combination as 1 when the condition count1 (k) is more than or equal to S (k) is met, recording 0 when the condition count2 (k) is not met, recording 1 when the condition count2 (k) is more than or equal to P (k) is met, and recording 0 when the condition is not met;

9) Repeating the step 7) and the step 8) for all other values of k to obtain a judging result sequence, and judging whether a final alarm is given for the real-time data point d based on the judging result sequence.

2. The method for early warning of faults of key components of a wind turbine generator system based on DA-RNN according to claim 1, wherein in the step 2), the pretreatment flow of offline training comprises the following steps:

a) The isolated abnormal point is usually a recording error caused by sensor abnormality, and is judged by an operation mechanism, and the judgment conditions are as follows: for temperature variables, the value is greater than 150 degrees or less than 0 degrees; for the wind speed variable, the numerical value is larger than the cut-out wind speed of the unit or smaller than the cut-in wind speed; for the power variable, the numerical value is larger than the rated power of the unit or is a negative value; when the conditions are met, the data are judged to be isolated abnormal points and are removed;

b) When the wind turbine generator is in an operation, maintenance and overhaul period, the wind turbine generator is in a shutdown state, and the data recorded by the SCADA system is usually 0 value or a system default value, wherein the value cannot represent the normal operation state of the wind turbine generator, so that the data in the shutdown, maintenance and overhaul period are removed;

c) In order to ensure the time continuity of the data, the missing values in the removed data and the SCADA system are interpolated, wherein the interpolation method is mean value interpolation, namely, the mean value of the data 1 hour before the variable interpolation position is taken as the interpolation value at the current moment.

3. The method for early warning of faults of key components of a wind turbine generator based on DA-RNN as claimed in claim 1, wherein in the step 3), the model input is related to a sliding windowVariable(s)Wherein n is the number of related variables, the input attention mechanism part inputs the related variables X, the hidden state output h and the memory unit output s of the encoder, and the influence weight of the kth related variable on the target variable is output->The calculation process is as follows:

wherein v is _en ,W _en ,U _en For the parameter to be learned, at time t _w Based on impact weightsReconstructing the related variable to obtain a time t _w Is expressed as:

the reconstructed vector is input as an encoder part, the encoder is composed of a plurality of LSTM units, and the reconstructed vector is output as a hidden state h and a memory unit s, at a time t _w The update of LSTM is performed by forget gate f _t Input gate i _t Output gate o _t The decision, its update rule is as follows:

wherein W is _f ,W _i ,W _o ,W _s ,b _f ,b _i ,b _o ,b _s For the parameter to be learned, σ is a sigmoid function, and the corresponding elements are multiplied by the ";

the input of the time attention mechanism is a hidden state h, a decoder hidden state output h 'and a memory unit output s', and the weight of the ith hidden state is outputThe calculation process is as follows:

wherein v is _de ,W _de ,U _de Reconstructing the hidden state based on the weight beta to obtain a semantic vector for the parameter to be learned, wherein the semantic vector is expressed as follows:

after which the wire is usedIntegrating semantic vectors and historical values of target variables by using a sexual regression model to obtain input of a decoderExpressed as:

wherein the method comprises the steps ofIs a parameter to be learned;

the decoder is also composed of LSTM units, forgetting the gate f _t ' input gate i _t ' output door o _t The update rule of' is as follows:

wherein W' _f ,W′ _i ,W′ _o ,W′ _s ,b′ _f ,b′ _i ,b′ _o ,b′ _s Is a parameter to be learned;

finally, obtaining the current time T through a linear functionEstimated value of target variableThe following is shown:

wherein W is _T ,b _T Is a parameter to be learned.

4. The method for early warning faults of key components of a wind turbine generator system based on DA-RNN according to claim 1, wherein in the step 5), the training set residual sequence is assumed to be subjected to normal distribution, and k in the multi-threshold setting is [1.5,2,2.5,3] based on normal distribution characteristics.

5. The method for early warning faults of key components of a wind turbine generator system based on DA-RNN as claimed in claim 1, wherein in the step 8), the multiple discriminant criteria correspond to different abnormal characteristics, the continuous overrun discriminant criteria correspond to abnormal characteristics with continuous abnormal values, the percentage overrun discriminant criteria correspond to abnormal conditions with large data fluctuation, and the continuous overrun discriminant criteria threshold parameter S _k Threshold parameter P of percentage overrun criterion _k % is set asWherein C is a constant.

6. The method for early warning faults of key components of a wind turbine generator system based on DA-RNN according to claim 1 is characterized in that in the step 9), for a real-time data point d, two 0-1 judgment results are output for each k value, a 0-1 sequence is finally output, the dimension is 2k dimensions, when the number of 1 in the sequence is greater than k, an alarm is given to the real-time data point, and when the number of 1 in the sequence is less than or equal to k, no alarm is given.