Dimension reduction and self-coding fan temperature abnormity early warning method and system and medium

Abstract

The invention relates to a dimensionality reduction and self-coding fan temperature anomaly early warning method and system and a medium, and provides the dimensionality reduction and self-coding fan temperature anomaly early warning method for solving the problems of high false alarm rate and missing report rate in the prior art, and the dimensionality reduction and self-coding fan temperature anomaly early warning method comprises the following steps: data acquisition and feature construction: obtaining a sample feature matrix; feature dimension reduction and manifold construction based on manifold learning: performing dimension reduction on feature vectors, and mapping high-dimensional features to a low-dimensional manifold space; building an anomaly detection model based on an importance weighted auto-encoder; and acquiring operation data of the wind turbine generator in real time, inputting the operation data into the trained model to calculate an abnormal score, and if the abnormal score exceeds a set threshold value, triggering early warning. Real-time and accurate criteria are provided for wind turbine generator temperature state early warning, probability distribution of data can be estimated more accurately, effective processing and modeling of complex wind turbine generator temperature data are achieved, the adaptability and accuracy of an early warning system are improved, and the false alarm rate and the missing report rate are reduced.

Description

Technical Field

[0001] The present invention relates to the field of wind turbine condition monitoring and fault diagnosis, and more particularly to a method, system, and medium for early warning of abnormal wind turbine temperature using dimensionality reduction and autoencoding. This method enables real-time temperature monitoring of key wind turbine components (such as the generator, gearbox, and pitch control system) and provides early warning of potential temperature anomalies, thereby improving wind turbine operational reliability and safety and reducing operation and maintenance costs. Background Art

[0002] Wind turbines often operate in harsh environmental conditions and are prone to various failures. Abnormal temperature is one of the key signs of wind turbine failure. Traditional temperature monitoring methods usually set fixed thresholds to generate alarms, but this method has the following problems: 1. Difficulty in setting thresholds: The temperature of wind turbines is affected by many factors (such as ambient temperature, wind speed, power, etc.), and fixed thresholds are difficult to adapt to various working conditions.

[0003] 2. High false alarm rate: Due to fluctuations in environmental factors, false alarms are easily triggered.

[0004] 3. Risk of missed reports: For some slowly developing temperature anomalies, fixed thresholds may not be able to detect them in time.

[0005] 4. Failure to consider data correlation: Traditional methods do not fully utilize the correlation between the temperature data of various components of wind turbines.

[0006] 5. Difficulty in processing high-dimensional data: Wind turbine SCADA systems generate large amounts of high-dimensional data. Traditional dimensionality reduction methods fail to combine knowledge from the wind power field to achieve processing that is more consistent with the mechanism background.

[0007] Therefore, a smarter and more accurate temperature anomaly warning method is needed that can adapt to different working conditions, reduce false alarm and missed alarm rates, and make full use of the correlation between data. Summary of the Invention

[0008] The technical solution of the present invention aims to solve the above problems and provides a fan temperature abnormality warning method based on dimensionality reduction and self-encoding, comprising the following steps: S1, Data Acquisition and Feature Construction: Data without downtime failures within a set time period is obtained from the wind turbine SCADA system as the initial data set. The initial data is cleaned, preprocessed, and feature processed. Each data with the same timestamp or processed features is combined to obtain a feature vector. The feature vectors of all time periods constitute the sample feature matrix. S2, feature dimensionality reduction and manifold construction based on manifold learning: Use the Dt-SNE (Density-t-distributed Stochastic Neighbor Embedding) algorithm, which is weighted by the relevance of wind turbine components, to reduce the dimensionality of the feature vectors obtained in S1, mapping the high-dimensional features to a low-dimensional manifold space to obtain a low-dimensional representation matrix; S3, Anomaly Detection Model Construction Based on Importance Weighted Autoencoder: Use the results of S2 to train an Importance Weighted Autoencoder (IWAE) model, map the data to multiple latent variables, reconstruct the input data from the latent variables, learn the probability distribution of the data by minimizing the IWAE loss function, and then train the model. Use the trained model to calculate the anomaly score and determine the anomaly score threshold based on the test data. S4 collects the operating data of the wind turbine in real time and inputs it into the trained model to calculate the anomaly score. If the anomaly score exceeds the set threshold, an early warning is triggered.

[0009] Furthermore, the step S1 specifically includes: S1.1. Collect temperature data of key components from the wind turbine SCADA system, including generator bearing temperature, gearbox oil temperature, pitch motor temperature, ambient temperature, wind speed, and power. Perform missing value and outlier processing on each type of data in turn. Align the pre-processed monitoring data based on the acquisition timestamp (by minute) to form a sample matrix. When obtaining the initial data set from the wind turbine SCADA system, select a unit of the same type that has not experienced a shutdown failure in the past three months, and select the data of its normal power generation status in its SCADA system as the initial data.

[0010] S1.2, add three features: the rate of change of each component temperature, the temperature difference between each component and the ambient temperature, and the sliding average of each temperature variable, and merge the sample matrix and the three features to obtain the sample feature matrix.

[0011] Furthermore, the missing values ​​are filled by linear interpolation; the outliers in the initial data set are detected and eliminated by boxplot method, that is, the monitoring data of each type of monitoring quantity are , make a judgment:

[0012] Q1 and Q3 are the lower quartile and upper quartile, respectively; for The elements in represent the original values ​​of the monitoring data. Represents the result of monitoring data after judgment and processing. Monitoring data , with P monitoring quantities, N is the length of the time series, p ranges from 1-P, n ranges from 1-N, and after processing, it forms ; Align the pre-processed monitoring data based on the acquisition timestamp (by minute) and then form a sample matrix:

[0013] in, , p takes the value of 1-P, representing the result of a certain monitoring quantity (such as active power) after preprocessing.

[0014] Furthermore, the step S2 includes: S2.1, each row of the sample feature matrix obtained in step S1 represents a sample point. A correlation matrix is ​​constructed based on the physical connections and thermal conductivity relationships between the components of the wind turbine. Each element in the correlation matrix represents the correlation coefficient between any two components. Component weights are calculated based on the correlation matrix to obtain the overall correlation between the component represented by any sample point and other components. S2.2, incorporate component weights into the conditional probability calculation to obtain the probability that any two sample points in the sample feature matrix are adjacent; S2.3, initialize a low-dimensional matrix, where each row in the low-dimensional matrix represents a sample point, and use t-distribution to calculate the similarity between any two sample points in the low-dimensional matrix to obtain the similarity probability; S2.4, symmetrize the probability of any two sample points being adjacent to each other in the sample feature matrix to obtain the joint probability, and optimize the low-dimensional matrix by minimizing the KL divergence of the probability distribution of the joint probability and the similarity probability; S2.5, use the gradient descent method to iteratively optimize the low-dimensional matrix, where each iteration updates the position of each sample point in the low-dimensional matrix according to the gradient of the KL divergence with respect to the similarity between any two sample points, until the iteration stops and the low-dimensional matrix is ​​obtained.

[0015] Furthermore, step S3 includes: S3.1 defines the structure of the importance-weighted autoencoder (IWAE), which maps a low-dimensional matrix to multiple latent variables, each of which follows a Gaussian distribution, and outputs the mean and logarithmic variance of the Gaussian distribution through the encoder; S3.2, reconstruct each latent variable into an output with the same dimension as the low-dimensional matrix, define the loss function between the low-dimensional matrix and the network parameters, and use the negative log-likelihood as the anomaly score, then train the model, calculate the anomaly score using the trained model, and determine the threshold of the anomaly score based on the test data.

[0016] Furthermore, the anomaly score expression is: Where: is the anomaly score; K is the number of latent variables mapped after dimensionality reduction, and k is the kth latent variable.

[0017] Furthermore, step S4 includes: performing the same preprocessing and feature engineering as the training data on the new real-time SCADA data to obtain .

[0018] The present invention also provides a dimension reduction and self-encoding fan temperature abnormality early warning system, comprising: The monitoring data processing module is used to obtain data without shutdown failures within a set time period from the wind turbine SCADA system as the initial data set, clean, preprocess and feature process the initial data to obtain feature vectors and construct a sample feature matrix; a manifold construction module connected to the monitoring data processing module, configured to reduce the dimensionality of the feature vectors obtained in the monitoring data processing module using a Dt-SNE algorithm weighted by the relevance of wind turbine components, mapping the high-dimensional features to a low-dimensional manifold space, and obtaining a low-dimensional representation matrix; an autoencoder building module, connected to the manifold building module, for training an importance-weighted autoencoder model using the reduced-dimensionality data under normal operating conditions, mapping the data to a plurality of latent variables, reconstructing the input data from the latent variables, learning the probability distribution of the data by minimizing the loss function of the importance-weighted autoencoder model, and then training the model, calculating an anomaly score using the trained model, and determining a threshold for the anomaly score; The real-time warning module is connected to the autoencoder building module and is used to collect the operating data of the wind turbine in real time, input it into the trained model to calculate the anomaly score, and trigger an early warning if the anomaly score exceeds a set threshold.

[0019] A wind turbine temperature anomaly warning system based on manifold learning and importance-weighted autoencoders includes a memory and one or more processors. The memory stores executable code. When the one or more processors execute the executable code, they are used to implement the dimension reduction and autoencoding wind turbine temperature anomaly warning method.

[0020] A computer-readable medium stores a program, which, when executed by a processor, implements the dimension reduction and self-encoding fan temperature abnormality early warning method.

[0021] Beneficial effects: The present application provides a dimensionality reduction and autoencoding wind turbine temperature anomaly warning method, system and medium. The warning method pre-processes the original data and then performs Dt-SNE dimensionality reduction on it to achieve low-dimensional manifold mapping of temperature-related high-dimensional features. An importance-weighted autoencoder model is constructed based on data under normal operating conditions and its anomaly score threshold is calculated to provide real-time and accurate judgment criteria for wind turbine temperature status warning; manifold learning is combined with importance-weighted autoencoder to reduce the high-dimensional, nonlinear wind turbine temperature data to a low-dimensional manifold space, revealing the inherent structure of the data. By introducing importance weighting, the probability distribution of the data can be more accurately estimated, thereby achieving effective processing and modeling of complex wind turbine temperature data; the probabilistic characteristics of the IWAE model are utilized to comprehensively consider the reconstruction error and the potential space probability density to calculate the anomaly score, thereby improving the adaptability and accuracy of the warning system and reducing the false alarm rate and missed alarm rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 This is a flow chart of Example 1; Figure 2 This is a visualization result diagram; Figure 3 This is a schematic diagram of abnormal fan temperature warning; Figure 4 This is a structural diagram of Example 2. DETAILED DESCRIPTION

[0023] The following is a clear and complete description of the technical solutions in the embodiments of the present invention. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0024] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. Those skilled in the art can also derive other drawings based on these drawings without inventive effort.

[0025] Example 1 A dimension reduction and self-encoding fan temperature abnormality warning method, such as Figure 1 As shown, the following steps are included: S1, Data Acquisition and Feature Construction: Obtain data from the wind turbine SCADA system that has no downtime failures within a set time period as the initial data set. Clean, preprocess, and feature-process the initial data to obtain feature vectors and construct a sample feature matrix. This includes: S1.1, collect temperature data of key components from the wind turbine SCADA system, including generator bearing temperature, gearbox oil temperature, pitch motor temperature, ambient temperature, wind speed, and power. Perform missing value processing and outlier processing on each type of data in turn, align the pre-processed monitoring data based on the acquisition timestamp (by minute) to form a sample matrix; when obtaining the initial data set from the wind turbine SCADA system, select a unit of the same type that has not experienced a shutdown failure in the past three months, and select its normal power generation status data in its SCADA system as the initial data. Missing values ​​are filled using linear interpolation; outliers are detected and eliminated using the box plot method, that is, for each type of monitoring data, , make a judgment:

[0026] Q1 and Q3 are the lower quartile and upper quartile, respectively; for The elements in represent the original values ​​of the monitoring data. Represents the result of monitoring data after judgment and processing. Monitoring data , with P monitoring quantities, N is the length of the time series, p ranges from 1-P, n ranges from 1-N, and after processing, it forms ; Align the pre-processed monitoring data based on the acquisition timestamp (by minute) and then form a sample matrix:

[0027] in, , p takes the value of 1-P, representing the result of a certain monitoring quantity (such as active power) after preprocessing.

[0028] S1.2, add three features: the rate of change of each component temperature, the temperature difference between each component and the ambient temperature, and the sliding average of each temperature variable, and merge the sample matrix and the three features to obtain the sample feature matrix. Add three features: Temperature change rate, calculate the change rate of each component temperature, that is, the first-order difference

[0029] Key temperature difference, calculate the temperature difference between each component and the ambient temperature,

[0030] Temperature sliding average, calculate the sliding average of each temperature variable (window size is )

[0031] Where, is the temperature change rate, is the data of the p-th monitoring quantity at the n-th moment, is the data of the p-th monitoring quantity at the n-1th moment, is the critical temperature difference, Represents the value of the ambient temperature monitoring quantity at the nth moment. is the sliding mean temperature, is the window size, The value is 1-n; The original sample matrix and the features obtained by feature engineering ( , , ) to obtain the final sample feature matrix .

[0032] S2, feature dimensionality reduction and manifold construction based on manifold learning: Use the Dt-SNE (Density-t-distributed Stochastic Neighbor Embedding) algorithm, which is weighted by the relevance of wind turbine components, to reduce the dimensionality of the feature vectors obtained in S1, mapping the high-dimensional features to a low-dimensional manifold space to obtain a low-dimensional representation matrix; this includes: S2.1, each row of the sample feature matrix obtained in step S1 represents a sample point. A correlation matrix is ​​constructed based on the physical connection and heat conduction relationship between the components of the wind turbine. Each element in the correlation matrix represents the correlation coefficient between any two components. The component weights are calculated based on the correlation matrix to obtain the overall correlation between the component represented by any sample point and other components. Specifically: The sample feature matrix obtained by step S1 is Each line Represents a sample point (all features at a time point), which is an m-dimensional vector located in a high-dimensional feature space.

[0033] According to the physical connection and heat conduction relationship of each component of the wind turbine, a correlation matrix is ​​constructed , by the element Composition, indicating parts and components The correlation coefficient between them (range 0-1, 1 means perfect correlation, 0 means no correlation, u and v are both 1-Q, Q represents the number of components). , where each monitoring quantity Corresponding to a specific component f, the weight represented by component f :

[0034] S2.2, integrate the component weights into the conditional probability calculation to obtain the probability that any two sample points in the sample feature matrix are adjacent; first integrate the component weights into the sample points and ( The matrix Row and The distance calculation formula of row is:

[0035] Then use the above weighted distance formula to perform conditional probability Calculation, conditional probability representation choose The probability of being its neighbor:

[0036] in Therefore The standard deviation of the Gaussian distribution centered at . Representatives As the center, is the Gaussian kernel function value with standard deviation, which is used to measure “proximity”.

[0037] S2.3, initialize a low-dimensional matrix, where each row in the low-dimensional matrix represents a sample point, and use t-distribution to calculate the similarity between any two sample points in the low-dimensional matrix to obtain the similarity probability; Randomly initialize low-dimensional matrices ,set up Each line Represents the coordinates of a sample point in the low-dimensional space, where is the dimension of the low-dimensional space (usually 2 or 3). In the low-dimensional space, the t distribution is used to calculate the and The similarity between :

[0038] in, yes and The square of the Euclidean distance between them. represents all different pairs of data points ( and ,in ), this sum acts as a normalization constant in the formula to ensure that the sum of similarities for all pairs of points in the low-dimensional space is stable.

[0039] Finally, the low-dimensional representation is optimized by minimizing the KL divergence (Kullback-Leibler divergence) of the probability distribution in the high-dimensional space and the low-dimensional space :

[0040] in, In high-dimensional space and The joint probability of The data points calculated previously and The similarity between them can be expressed by the conditional probability Symmetrization yields:

[0041] Where: The meaning is that in a high-dimensional space, given Select Point The conditional probability of . The meaning is the Kullback-Leibler divergence (KL divergence) between the probability distribution P in the high-dimensional space and the probability distribution Q in the low-dimensional space. The meaning of is the total number of data points.

[0042] S2.4, symmetrize the probability of any two sample points being adjacent to each other in the sample feature matrix to obtain the joint probability, and optimize the low-dimensional matrix by minimizing the KL divergence of the probability distribution of the joint probability and the similarity probability; S2.5, use the gradient descent method to iterate the optimized low-dimensional matrix, in which each step of the iteration updates the position of each sample point in the low-dimensional matrix according to the gradient of the KL divergence with respect to the similarity between any two sample points, until the iteration stops and the low-dimensional matrix is ​​obtained. Iterate, each step of the iteration will be based on the KL divergence The gradient of each point is updated Location:

[0043] Where, The meaning is the low-dimensional space point pair and The square of the Euclidean distance between them.

[0044] There are two conditions for stopping iterations: reaching the preset maximum number of iterations (recommended 1000 times); and the loss function (KL divergence) converging (the rate of change is less than 1%). If either of these conditions is met, the iteration is stopped. The formula for the rate of change of KL divergence is:

[0045] Where S represents the number of iterations. The meaning of is the value of KL divergence calculated at the Sth iteration. The meaning of is the value of KL divergence calculated at the S-1th iteration.

[0046] After Dt-SNE dimension reduction, Mapped to a low-dimensional space, a low-dimensional representation matrix is ​​obtained (two-dimensional).

[0047] S3, Anomaly Detection Model Construction Based on Importance Weighted Autoencoder: Use reduced-dimensional data from normal operating conditions to train an Importance Weighted Autoencoder (IWAE) model, map the data to multiple latent variables, and reconstruct the input data from the latent variables. The model is trained by minimizing the IWAE loss function to learn the probability distribution of the data. The trained model is used to calculate anomaly scores and determine the anomaly score threshold based on test data. This includes: S3.1 defines the structure of the Importance Weighted Autoencoder (IWAE), which maps a low-dimensional matrix to multiple latent variables, each of which follows a Gaussian distribution, and outputs the mean and log-variance of the Gaussian distribution through the encoder. S3.2, reconstruct each latent variable into an output with the same dimension as the low-dimensional matrix, define the loss function between the low-dimensional matrix and the network parameters, and use the negative log-likelihood as the anomaly score, then train the model, calculate the anomaly score using the trained model, and determine the threshold of the anomaly score based on the test data.

[0048] Sampling from the K Gaussian distributions of the output respectively, we get K latent variables:

[0049] in represents the diagonal covariance matrix, and the elements on the diagonal are The square of . is a latent variable, represents a normal distribution, represents the unit matrix, indicating that the covariance matrix has a specific structure.

[0050] Each latent variable Reconstruct them into outputs with the same dimension as the input y. Then start training and define the loss function as follows:

[0051] in, is the posterior distribution defined by the encoder. The meaning is to sample K latent variables from the distribution Take expectations. The meaning is that in the parameter Under the defined model, the data points Hedi latent variable samples The joint probability of . It means the loss function used to train the model in the importance-weighted autoencoder.

[0052] is a joint probability distribution, which can be decomposed into

[0053] is the prior distribution, and the standard normal distribution is usually chosen. Is the conditional probability distribution defined by the decoder, usually assumed to be a Gaussian distribution. In actual calculations, in order to avoid numerical underflow, the logarithmic form of the loss function is usually calculated:

[0054] According to the output of the decoder and the assumed distribution (Gaussian distribution) are used for calculations.

[0055]

[0056] in yes Dimensions, is the variance of the decoder output. The meaning is No. dimensional values ​​and the decoder is based on latent variables Output reconstructed data The squared difference between the values ​​of the d-th dimension of . Meaning that the encoder is based on the latent variable Generated reconstructed data No. Dimension values.

[0057] It can be calculated based on the probability density function of the standard normal distribution:

[0058] in is the dimension of the latent variable.

[0059] It can be calculated based on the mean and variance of the encoder output and the probability density function of the Gaussian distribution:

[0060] in, and are the mean and variance of the encoder output, respectively. elements; Use Adam, SGD and other optimization algorithms to train network parameters (decoder parameters) and (Encoder parameters). Usually use a smaller learning rate (such as 0.001 or 0.0001) and a larger batch size (such as 64 or 128) Input the test data y into the trained model and calculate the k reconstruction values ​​and the average error:

[0061] The meaning is test data The corresponding reconstructed value.

[0062] Calculate the negative log-likelihood of the data y as the anomaly score:

[0063] In actual calculation, use: Based on the anomaly score distribution of the test data, the 95% quantile is set as the threshold limit.

[0064] S4, real-time collection of wind turbine operation data, input into the trained model to calculate the anomaly score, if the anomaly score exceeds the set threshold, trigger an early warning. Mapping to the low-dimensional manifold space, we get . The mapped data Input to the IWAE model: The encoder calculates the mean of k groups and log variance .

[0065] from K latent variables are sampled in .

[0066] The decoder is based on each Reconstruct output.

[0067] Calculate the anomaly score:

[0068] if ,That is, the alarm is triggered when the anomaly score exceeds the threshold value calculated from the test data.

[0069] Example 2 A dimension reduction and self-encoding fan temperature abnormality warning system, such as Figure 4 Shown, including: The monitoring data processing module is used to obtain data without shutdown failures within a set time period from the wind turbine SCADA system as the initial data set, clean, preprocess and feature process the initial data to obtain feature vectors and construct a sample feature matrix; a manifold construction module connected to the monitoring data processing module, configured to reduce the dimensionality of the feature vectors obtained in the monitoring data processing module using a Dt-SNE algorithm weighted by the relevance of wind turbine components, mapping the high-dimensional features to a low-dimensional manifold space, and obtaining a low-dimensional representation matrix; an autoencoder building module, connected to the manifold building module, for training an importance-weighted autoencoder model using the reduced-dimensionality data under normal operating conditions, mapping the data to a plurality of latent variables, reconstructing the input data from the latent variables, learning the probability distribution of the data by minimizing the loss function of the importance-weighted autoencoder model, and then training the model, calculating an anomaly score using the trained model, and determining a threshold for the anomaly score; The real-time warning module is connected to the autoencoder building module and is used to collect the operating data of the wind turbine in real time, input it into the trained model to calculate the anomaly score, and trigger an early warning if the anomaly score exceeds a set threshold.

[0070] Example 3 A dimensionality reduction and self-encoding fan temperature anomaly warning system includes a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, they are used to implement the dimensionality reduction and self-encoding fan temperature anomaly warning method described in Example 1.

[0071] Example 4 A computer-readable medium stores a program thereon, which, when executed by a processor, implements the dimension reduction and self-encoding fan temperature abnormality warning method described in Example 1.

[0072] Application Examples A power generation group in my country has deployed 74 wind turbines with a total installed capacity of 301.2 MW, along with a supporting 220 kV substation. Full grid connection is planned for 2021. The complex and volatile marine environment can easily impact the safety of equipment and structures. Currently, manual methods are used to identify abnormal equipment temperatures. This process is impacted by factors such as accessories, ship schedules, manpower, number of turbines, and turbine type, resulting in a long detection cycle. Therefore, intelligent temperature anomaly diagnosis is necessary to prevent serious wind turbine accidents.

[0073] The wind turbine is equipped with monitoring related to the unit temperature, including but not limited to generator bearing temperature monitoring, gearbox oil temperature monitoring, pitch motor temperature monitoring, ambient temperature monitoring, wind speed monitoring, power monitoring, etc.

[0074] Based on the above-mentioned wind turbine temperature anomaly warning method and system, the monitoring data processing module function of module 10 is first used to obtain the key temperature parameters and operating parameters of the unit from September 2005 to December 2005, as follows: The ambient temperature is monitored by an air temperature probe outside the cabin with a measurement accuracy of 0.1°C; The generator bearing temperature is monitored by a pt100 resistive sensor with a measurement accuracy of 0.1°C; The gearbox oil temperature is monitored by a pt100 resistive sensor with a measurement accuracy of 0.1°C; The temperature of the pitch motor is monitored by a pt100 resistive sensor with a measurement accuracy of 0.1°C; Wind speed was monitored by an ultrasonic anemometer with a measurement accuracy of 0.1 m / s; The power is monitored by the power module with a measurement accuracy of 0.1Kw The above measurement data are combined based on the acquisition timestamp (accurate to the minute), the outliers are detected and eliminated using the box plot method, and judged according to the formula in Example 1 to obtain a processed sample matrix of length N .

[0075]

[0076] in, , p=1-6, representing ambient temperature, generator bearing temperature, gearbox oil temperature, pitch motor temperature, wind speed and power respectively. Add three features: Temperature change rate, calculate the change rate of generator bearing temperature, gearbox oil temperature, pitch motor temperature, and ambient temperature respectively:

[0077] Key temperature difference, calculate the temperature difference between each component temperature and the ambient temperature:

[0078] Temperature sliding average, calculate the sliding average of each temperature variable. The window size can be set to 10 based on experience. Then:

[0079] The final initial sample feature matrix as follows:

[0080] The preprocessed sample feature matrix Dimensionality reduction, including: Set the parameters of the Dt-SNE model: perplexity = 30, iterations = 1000, learning rate = 200.

[0081] Using the Dt-SNE model Map to two-dimensional space to obtain the reduced dimensionality data matrix :

[0082] Using the manifold dimension reduction data under normal working conditions to construct an importance-weighted autoencoder model: 1. Set up the encoder: input layer (2 nodes), two hidden layers (16 nodes and 8 nodes respectively), activation function is ReLU. The output layer is two parallel layers, outputting the mean and log variance (each has 4 nodes, corresponding to the dimensionality of the latent variable).

[0083] 2. Set up the decoder: input layer (4 nodes), two hidden layers (8 nodes and 16 nodes respectively), activation function is ReLU. Output layer (2 nodes), activation function is linear function.

[0084] 3. Set the sampling frequency to 5.

[0085] 4. Training parameters: Adam optimizer, learning rate 0.001; epoch = 100, batch_size = 64 5. After training is completed, calculate the training data The negative log-likelihood of is used as the anomaly score and the 95% quantile is set as the static threshold.

[0086] The new real-time SCADA data is subjected to the same preprocessing and feature engineering as the training data, and the obtained , part of which reads:

[0087] After all data is visualized, Figure 2 As shown, the horizontal axis of all curves is the data index, which represents the relative position of the data in the time dimension, and the vertical axis unit of the temperature-related curves is ℃. Figure 2 (1) is the ambient temperature curve; Figure 2 (2) is the temperature curve of the generator bearing; Figure 2 (3) in the figure is the gearbox oil temperature curve; Figure 2 (4) is the temperature curve of the pitch motor; Figure 2 (5) is the wind speed curve, and the vertical axis unit is m / s; Figure 2 (6) is the active power curve, and the vertical axis unit is KW; Figure 2 (7) in the figure is the generator bearing temperature_delta curve (i.e. the generator bearing temperature change rate curve); Figure 2 (8) is the gearbox oil temperature_delta curve (i.e., the gearbox oil temperature change rate curve); Figure 2 (9) is the temperature_delta curve of the pitch motor (i.e., the temperature change rate curve of the pitch motor); Figure 2 (10) is the ambient temperature_delta curve (i.e. the ambient temperature change rate curve); Figure 2 (11) is the generator bearing_diff temperature curve (i.e., the temperature difference curve of the generator bearing temperature); Figure 2 (12) in the figure is the gearbox oil temperature_diff curve (i.e., the temperature difference curve of the gearbox oil temperature); Figure 2 (13) in the figure is the temperature_diff curve of the pitch motor (i.e., the temperature difference curve of the pitch motor); Figure 2 (14) in the figure is the generator bearing temperature_avg curve (i.e., the sliding average curve of the generator bearing temperature); Figure 2 (15) in the figure is the gearbox oil temperature_avg curve (i.e., the sliding average curve of the gearbox oil temperature); Figure 2 (16) in it is the pitch motor temperature_avg curve (i.e. the sliding average curve of the pitch motor temperature); Figure 2 (17) in the figure is the ambient temperature_avg curve (i.e., the sliding average curve of the ambient temperature).

[0088] Use the trained Dt-SNE model to transform real-time data Mapping to the low-dimensional manifold space, we get . The mapped data Input to the IWAE model and calculate the anomaly score, if , that is, the anomaly score exceeds the threshold calculated from the test data and triggers an alarm. Figure 3 The example shown is calculated The red dotted line represents the anomaly score threshold given by the training model. It can be seen that the temperature of the entire unit has been deteriorating recently, especially around 200-250, where the anomaly score has obviously exceeded the threshold, which requires attention.

[0089] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be included therein. Any reference sign in a claim should not be construed as limiting the claim to which it relates.

Claims

1. A fan temperature abnormality warning method based on dimensionality reduction and self-encoding, characterized in that: The steps include: S1, obtain data without downtime failures within a set time period as the initial data set, clean, preprocess and feature process the initial data to obtain feature vectors, and construct a sample feature matrix based on the feature vectors; S2, using the Dt-SNE algorithm weighted by the correlation of wind turbine components to reduce the dimension of the feature vector obtained in S1, mapping the high-dimensional features to a low-dimensional manifold space to obtain a low-dimensional representation matrix; S3 uses the results of S2 to train an importance-weighted autoencoder model, maps the data to multiple latent variables, and reconstructs the input data from the latent variables. By minimizing the loss function of the importance-weighted autoencoder model, the probability distribution of the data is learned, and the model is trained. The trained model is used to calculate the anomaly score and determine the threshold of the anomaly score; S4 collects the operating data of the wind turbine in real time and inputs it into the trained model to calculate the anomaly score. If the anomaly score exceeds the set threshold, an early warning is triggered.

2. The fan temperature abnormality warning method based on dimensionality reduction and self-encoding according to claim 1 is characterized in that: The step S1 specifically includes: S1.

1. Collect temperature data of key components, including generator bearing temperature, gearbox oil temperature, pitch motor temperature, ambient temperature, wind speed, and power. Perform missing value and outlier processing on each type of data. Align the pre-processed monitoring data based on the acquisition timestamp to form a sample matrix. S1.2, add three features: the rate of change of each component temperature, the temperature difference between each component and the ambient temperature, and the sliding average of each temperature variable, and merge the sample matrix and the three features to obtain the sample feature matrix.

3. The fan temperature abnormality warning method based on dimensionality reduction and self-encoding according to claim 2 is characterized in that: In step S1, the linear interpolation method is used to fill the missing values ​​in the initial data set; the box plot method is used to detect and eliminate outliers in the initial data set.

4. The fan temperature abnormality warning method based on dimensionality reduction and self-encoding according to claim 2 is characterized in that: The step S2 comprises: S2.1, each row of the sample feature matrix obtained in step S1 represents a sample point. A correlation matrix is ​​constructed based on the physical connections and thermal conductivity relationships between the components of the wind turbine. Each element in the correlation matrix represents the correlation coefficient between any two components. Component weights are calculated based on the correlation matrix to obtain the overall correlation between the component represented by any sample point and other components. S2.2, incorporate component weights into the conditional probability calculation to obtain the probability that any two sample points in the sample feature matrix are adjacent; S2.3, initialize a low-dimensional matrix, where each row in the low-dimensional matrix represents a sample point, and use t-distribution to calculate the similarity between any two sample points in the low-dimensional matrix to obtain the similarity probability; S2.4, symmetrize the probability of any two sample points being adjacent to each other in the sample feature matrix to obtain the joint probability, and optimize the low-dimensional matrix by minimizing the KL divergence of the probability distribution of the joint probability and the similarity probability; S2.5, use the gradient descent method to iteratively optimize the low-dimensional matrix, where each iteration updates the position of each sample point in the low-dimensional matrix according to the gradient of the KL divergence with respect to the similarity between any two sample points, until the iteration stops and the low-dimensional matrix is ​​obtained.

5. The fan temperature abnormality warning method based on dimensionality reduction and self-encoding according to claim 1 is characterized in that: The step S3 comprises: S3.1 defines the structure of the importance-weighted autoencoder, which maps a low-dimensional matrix to multiple latent variables, each of which follows a Gaussian distribution, and outputs the mean and logarithmic variance of the Gaussian distribution through the encoder; S3.2, reconstruct each latent variable into an output with the same dimension as the low-dimensional matrix, define the loss function between the low-dimensional matrix and the network parameters, and use the negative log-likelihood as the anomaly score, then train the model, use the trained model to calculate the anomaly score, and determine the threshold of the anomaly score.

6. The fan temperature abnormality warning method based on dimensionality reduction and self-encoding according to claim 5 is characterized in that: The anomaly score expression is: Where: K is the number of potential variables obtained after dimensionality reduction, represents the conditional probability distribution defined by the decoder, represents the prior distribution, represents the posterior distribution defined by the encoder.

7. A dimension reduction and self-encoding fan temperature abnormality warning system, characterized in that: include: The monitoring data processing module is used to obtain data without downtime failures within a set time period as the initial data set, clean, preprocess and feature process the initial data to obtain feature vectors and construct a sample feature matrix; a manifold construction module connected to the monitoring data processing module, configured to reduce the dimensionality of the feature vectors obtained in the monitoring data processing module using a Dt-SNE algorithm weighted by the relevance of wind turbine components, mapping the high-dimensional features to a low-dimensional manifold space, and obtaining a low-dimensional representation matrix; an autoencoder building module, connected to the manifold building module, for training an importance-weighted autoencoder model using the reduced-dimensionality data under normal operating conditions, mapping the data to a plurality of latent variables, and reconstructing the input data from the latent variables, learning the probability distribution of the data by minimizing the loss function of the importance-weighted autoencoder model, and then training the model, calculating an anomaly score using the trained model, and determining a threshold for the anomaly score; The real-time warning module is connected to the autoencoder building module and is used to collect the operating data of the wind turbine in real time, input it into the trained model to calculate the anomaly score, and trigger an early warning if the anomaly score exceeds a set threshold.

8. A dimension reduction and self-encoding fan temperature abnormality warning system, characterized in that: It includes a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, it is used to implement the dimension reduction and self-encoding fan temperature abnormality warning method described in any one of claims 1 to 6.

9. A computer-readable medium having a program stored thereon, which, when executed by a processor, implements the dimension reduction and self-encoding fan temperature abnormality early warning method according to any one of claims 1 to 6.

Images