CN112801151A - Wind power equipment fault detection method based on improved BSMOTE-Sequence algorithm - Google Patents

Abstract

The invention discloses a wind power equipment fault detection method, which is a fan data set sampling strategy model designed by improving BorderlinesMOTE based on an improved BSMOTE-Sequence fan fault sampling strategy. When selecting which sample points to synthesize, BorderlineSMOTE selects K ' few-class neighbor samples through a KNN algorithm in the first step, and randomly selects K ' samples from the K ' samples in the second step. According to the method, the second step of randomly selecting the samples is improved into the step of selecting the samples according to the time sequence characteristics, the mode of generating new samples comprehensively considers the space distance and the time sequence rule, the generation of noise points can be effectively reduced, the problem of unbalance of a fan data set is solved, and the time sequence rule of the data set is not damaged. In addition, the invention also combines the Tomek Links technology, can effectively filter noise samples and overlapping samples among classes in a data set, thereby improving the efficiency and accuracy of subsequent classifier training and avoiding model overfitting.

Description

Wind power equipment fault detection method based on improved BSMOTE-Sequence algorithm

Technical Field

The invention relates to the field of wind power equipment fault data mining and detection, in particular to a wind power equipment fault detection method based on an improved BSMOTE-Sequence algorithm.

Background

As a driven fluid machine, the wind driven generator has a complex internal structure and an abnormally severe working environment, is usually installed in complex tuyere areas such as wildlands, beaches and the like, is in an extremely cold and hot environment and an extremely temperature difference environment throughout the year, and is often impacted by strong gust wind, so that the reliability and the service life of components are weakened to a certain extent, the data processing and the detection of faults of the wind turbine generator are enhanced, and the wind driven generator has great significance for improving the reliability of power generation of the wind turbine generator and reducing the operation and maintenance cost.

Along with the continuous improvement of the reliability of the fan equipment, the probability of equipment failure is gradually reduced, the equipment normally operates most of the time, so that a large proportion of fan operation data sets are sample data of normal operation, the proportion of the sample data of failure is very small, and the data sets are unbalanced. The unbalanced data set can have negative influence on the fault detection effect of the fan component, and the classification result is usually biased to a large sample, so that how to process the fan data set before establishing a classification model is particularly important to eliminate the negative influence caused by the unbalance.

When an existing algorithm for processing an unbalanced data set considers which samples are synthesized, a core idea is to select a few classes of samples at random or select a few classes of samples based on clustering and Euclidean distance. The two selection ideas cannot be completely applied to the fan data with the time sequence characteristics, on one hand, the time sequence characteristics of the fan data set can be damaged, and information of the fan which is converted from a normal state to a fault state is lost; on the other hand, the model overfitting is easily caused by not filtering the overlapped samples among the classes and the noise samples, so that the classification precision is reduced.

Disclosure of Invention

The invention provides a wind power equipment fault detection method which is based on an improved BSMOTE-Sequence fan fault sampling strategy and mainly aims to solve the problem that fan data sets with time Sequence characteristics are unbalanced in category and reduce negative effects on fan part fault detection effects.

The sampling strategy is a new fan data set sampling strategy model designed by improving BorderlineSMOTE, and particularly, when BorderlineSMOTE selects which sample points to synthesize, K ' few-class neighbor samples are selected through a KNN algorithm in the first step, and K ' samples are randomly selected from the K ' samples in the second step. According to the method, the second step of randomly selecting the samples is improved into the step of selecting the samples according to the time sequence characteristics, the mode of generating new samples comprehensively considers the space distance and the time sequence rule, the generation of noise points can be effectively reduced, the problem of unbalance of a fan data set is solved, and the time sequence rule of the data set is not damaged. In addition, the invention also combines the Tomek Links technology, can effectively filter noise samples and overlapping samples among classes in a data set, thereby improving the efficiency and accuracy of subsequent classifier training and avoiding model overfitting.

The construction of the fan fault sampling strategy model comprises the following steps:

(1) for each sample P in the minority class sample set P⁽ⁱ⁾According to the formula

Calculate it to all other samples d^(j)The nearest Top K samples { d } are selected⁽¹⁾,d⁽²⁾,...,d^(K)}; where M represents the number of features of the sample,

represents p⁽ⁱ⁾The m-th feature of the sample,

denotes d^(j)The m-th feature of the sample, dist (p)⁽ⁱ⁾,d^(j)) Represents a sample p⁽ⁱ⁾And d^(j)The euclidean distance between.

(2) Let { d⁽¹⁾,d⁽²⁾,...,d^(K)In is Np⁽ⁱ⁾A plurality of majority samples according to the formula

And

judgment of p⁽ⁱ⁾Of type Cp⁽ⁱ⁾；Cp⁽ⁱ⁾When 0, 1 and-1 are taken, p is represented respectively⁽ⁱ⁾The samples are security class samples, boundary class samples, potential noise class samples.

(3) For each Cp ⁽ⁱ⁾1 of p⁽ⁱ⁾Samples according to the formula

Calculate it to all other minority class samples p^(j)The Euclidean distance of (c), select the K' samples of Top with the closest spatial distance { p⁽¹⁾,p⁽²⁾,...,p^(K'⁾}。

(4) From { p⁽¹⁾,p⁽²⁾,...,p^(K'⁾In the formula, t is Tp⁽ⁱ⁾-Tp^(j)Selecting the samples { p ] of TopK' with the smallest time span⁽¹⁾,p⁽²⁾,...,p^(K″)}; wherein Tp⁽ⁱ⁾Represents p⁽ⁱ⁾Temporal characteristics of the sample, Tp^(j)Represents p^(j)Temporal characteristics of the sample, t denotes p⁽ⁱ⁾And p^(j)The time span in between.

(5) According to the formula p^(i，j)＝p⁽ⁱ⁾+α*(p⁽ⁱ⁾-p^(j)) At p of⁽ⁱ⁾Sample and { p⁽¹⁾,p⁽²⁾,...,p^(K″)Synthesize new sample points between samples

Wherein p is^(i，j)Represents p⁽ⁱ⁾Sample and p^(j)New samples of sample synthesis, alpha represents a random number between (0, 1).

(6) And (4) adding all the new sample points synthesized in the step (5) into the minority sample set P.

(7) According to the formula

For each sample P in the minority class sample set P⁽ⁱ⁾Calculating it to each of the majority class samples n^(j)Distance (p) of⁽ⁱ⁾,n^(j))。

(8) If there is no any other sample point d^(k)So that the formula Distance (p)⁽ⁱ⁾,d^(k))＜Distance(p⁽ⁱ⁾,n^(j)) Or Distance (n)^(j),d^(k))＜Distance(p⁽ⁱ⁾,n^(j)) If true, then it is called (p)⁽ⁱ⁾,n^(j)) Is a Tomeklinks pair, and most of the class sample points are deleted from each Tomeklinks pair.

(9) According to the formula

And

and calculating the sample class proportion around each sample, finding out the noise class sample and deleting the noise class sample.

Drawings

FIGS. 1-1 and 1-2 are general flow charts of the present invention;

FIGS. 2-1, 2-2, 2-3, 2-4, and 2-5 are BSMOTE-Sequence sampling effect diagrams.

Detailed Description

A specific example is given below, and the technical solution and the achieved effect of the present invention can be better understood by combining the example.

The data sets used in this experiment are all from historical operating data of a certain wind power plant in Yunnan, as shown in Table 1. The training set had 691160 records with a positive to negative sample ratio of 3: 7. The test set had 24116 records with a positive to negative sample ratio of 4: 6. The training set and the testing set have 18 characteristics, namely time, outlet pressure of a lubricating oil filter screen of the gearbox, oil temperature of the gearbox, temperature of a cabin cabinet, winding voltage of U1 items, winding voltage of U2 items, winding voltage of U3 items, winding current of U1 items, winding current of U2 items, winding current of U3 items, cooling temperature of the generator, temperature of a slip ring of the generator, rotating speed of an impeller, wind speed 1, wind speed 2, wind direction 1 and wind direction 2.

TABLE 1 data set

In order to verify the effectiveness of the BSMOTE-Sequence algorithm in processing the unbalanced fan data set on the gearbox fault detection task, SVM, CNN and LSTM are respectively used as fault detection algorithms, and comparison experiments are carried out with unprocessed Borderline SMOTE, ADASYN, TomekLinks, SMOTE + ENN and SMOTE + Tomek Links algorithms.

Table 2 confusion matrices corresponding to the models of the present invention, where TP represents the number of samples predicted to be positive and truly positive, FP represents the number of samples predicted to be positive and truly negative, FN represents the number of samples predicted to be negative and truly positive, and TN represents the number of samples predicted to be negative and truly negative.

TABLE 2 confusion matrix

There are many evaluation criteria for classifier performance, such as Accuracy, Precision, Recall, F1-Score, ROC, AUC, G-mean, etc., however, if the single index of Accuracy, Precision, Recall, etc. is not suitable for unbalanced data set, because it can not effectively reflect the classification performance to minority classes, so this experiment uses F1-Score, AUC, G-mean as evaluation index, these indexes can comprehensively measure the classification performance.

The F1-Score is an evaluation index commonly used for measuring the two-classification model, and as the evaluation index is a harmonic average of the accuracy and the recall ratio, the accuracy and the recall ratio of the model can be comprehensively measured at the same time, and the evaluation index is very effective for a data set with unbalanced classes. In the classification task of unbalanced data sets, the F-Score is larger only if the precision rate and the recall rate are both larger, otherwise the value is closer to the smaller one, as shown in equation (1). Where Precision is the Precision rate, which represents the ratio of samples for which the true result is positive among the samples predicted to be positive, as shown in equation (2). Recall is the Recall ratio, representing the ratio of samples predicted to be positive among samples for which the true outcome is positive, as shown in equation (3).

AUC is the area under ROC (receiver operating characteristic curve), and its value is not greater than 1. The ROC curve is plotted with the false positive case rate FPR on the horizontal axis and the true case rate TPR on the vertical axis, where FPR is the ratio of samples predicted to be positive among the samples with true negative results, as shown in equation (4). The TPR is the ratio of samples predicted to be positive among samples whose true result is positive, as shown in equation (5). The AUC area can not be influenced by the category distribution, and the data set with unbalanced category distribution can be effectively evaluated and compared.

The G-mean index measures how many positive and negative examples are successfully predicted, respectively, and the value is not greater than 1, as shown in equation (6), where specificity is specificity, as shown in equation (7).

The BSMOTE-Sequence algorithm is compared with the original data direct classification, BorderlineSMOTE, ADASYN, Tomek Links, SMOTE + ENN, SMOTE + Tomek Links and other algorithms, and an SVM, CNN and LSTM are adopted as a classifier. The sampling rate N in the borderlinessmott algorithm is equal to the imbalance rate IR 0.5 and rounded up, and the nearest neighbor threshold K is equal to N3. The synthesis coefficient beta in the ADASYNN algorithm is equal to 1, a strategy for removing overlapped samples among classes in the Tomek Links algorithm is set to remove only most samples, and a strategy for rejecting ENN samples in the SMOTE + ENN algorithm is set to reject most samples if more than half of K adjacent points of the most samples do not belong to the most samples. In order to eliminate the influence of random factors, each algorithm takes the results of 20 experiments, and calculates the average values of F1-Score, AUC and G-mean, wherein the larger the average value is, the better the classification effect is, and the experimental results are shown in Table 3.

TABLE 3 comparison of F1-Score, AUC, G-mean values for different sampling algorithms

From the experimental results in the table, the following conclusions can be analyzed: compared with other algorithms, the BSMOTE-Sequence algorithm has certain improvement on F1-Score, AUC and G-mean. Secondly, only using a Tomek Links undersampling technology to process a fan data set, important characteristic information can be lost, the effect of a classification model is poor, and therefore F1-Score is the lowest; although its AUC was not the lowest, F1-Score was more sensitive than AUC for unbalanced datasets. When processing an unbalanced data set, not all sampling algorithms are suitable, and sometimes the sampling algorithms may be suitable for the opposite, so that the classification performance of the model can be improved only by specifically analyzing the characteristics of the data set and selecting the suitable sampling algorithms. The classification effect of the SVM is not as good as that of the CNN and the LSTM, because the SVM is a traditional machine learning algorithm, more feature engineering and model parameter optimization are needed to improve the classification effect. The classification effect of the LSTM is better than that of the CNN, because the LSTM can capture important features from time sequence data and perform associated modeling, the extraction of the time sequence features is more sufficient, and the importance of the time sequence rule is further considered when the time sequence data set is subjected to unbalanced processing.

Aiming at the problem of unbalanced fan data sets, the invention provides a BSMOTE-Sequence sampling algorithm by adopting the idea of reconstructing the data sets in the field of data mining and combining the time Sequence characteristics of the data sets. Its advantages are as follows: based on the characteristics of the data set, the time sequence rule of the data of the fan is fully utilized to generate a reasonable new sample, and the classification performance of the unbalanced data set is improved to a certain extent. Filtering the inter-class overlapping samples and the noise samples to avoid the negative influence of wrong classification; the algorithm adopts over-sampling and under-sampling technologies, and the problem caused by the independent use of a certain sampling method is avoided. And fourthly, systematically verifying the effectiveness of the BSMOTE-Sequence algorithm by using three classifiers of SVM, CNN and LSTM as a gearbox fault detection model. From the experimental results of the embodiment, for the unbalanced data set of the fan, the BSMOTE-Sequence algorithm provided by the invention is generally superior to other data sampling methods.

Claims (1)

1. The wind power equipment fault detection method based on the improved BSMOTE-Sequence algorithm comprises a fan fault sampling strategy model and is characterized in that the fan fault sampling strategy model is constructed by the following steps:

(1) for each sample P in the minority class sample set P⁽ⁱ⁾According to the formula

Calculate it to all other samples d^(j)The nearest Top K samples { d } are selected⁽¹⁾,d⁽²⁾,...,d^(K)}; where M represents the number of features of the sample,

represents p⁽ⁱ⁾The m-th feature of the sample,

denotes d^(j)The m-th feature of the sample, dist (p)⁽ⁱ⁾,d^(j)) Represents a sample p⁽ⁱ⁾And d^(j)The euclidean distance between;

(2) let { d⁽¹⁾,d⁽²⁾,...,d^(K)In is Np⁽ⁱ⁾A plurality of majority samples according to the formula

And

judgment of p⁽ⁱ⁾Of type Cp⁽ⁱ⁾；Cp⁽ⁱ⁾When 0, 1 and-1 are taken, p is represented respectively⁽ⁱ⁾The samples are safety class samples, boundary class samples and potential noise class samples;

(3) for each Cp⁽ⁱ⁾1 of p⁽ⁱ⁾Samples according to the formula

Calculate it to all other minority class samples p^(j)The Euclidean distance of (c), select the K' samples of Top with the closest spatial distance { p⁽¹⁾,p⁽²⁾,...,p^(K')}；

(4) From { p⁽¹⁾,p⁽²⁾,...,p^(K')In the formula, t is Tp⁽ⁱ⁾-Tp^(j)Selecting the samples { p ] of TopK' with the smallest time span⁽¹⁾,p⁽²⁾,...,p^(K”)}; wherein Tp⁽ⁱ⁾Represents p⁽ⁱ⁾Temporal characteristics of the sample, Tp^(j)Represents p^(j)Temporal characteristics of the sample, t denotes p⁽ⁱ⁾And p^(j)The time span in between;

(5) according to the formula p^(i，j)＝p⁽ⁱ⁾+α*(p⁽ⁱ⁾-p^(j)) At p of⁽ⁱ⁾Sample and { p⁽¹⁾,p⁽²⁾,...,p^(K”)Between samplesSynthesis of a New sample Point { p^(i,1),p^(i,2),...,p^(i,K”)}; wherein p is^(i，j)Represents p⁽ⁱ⁾Sample and p^(j)New samples of sample synthesis, alpha represents a random number between (0, 1);

(6) adding all the new sample points synthesized in the step (5) into a minority sample set P;

(7) according to the formula

For each sample P in the minority class sample set P⁽ⁱ⁾Calculating it to each of the majority class samples n^(j)Distance (p) of⁽ⁱ⁾,n^(j))；

(8) If there is no any other sample point d^(k)So that the formula Distance (p)⁽ⁱ⁾,d^(k))＜Distance(p⁽ⁱ⁾,n^(j)) Or Distance (n)^(j),d^(k))＜Distance(p⁽ⁱ⁾,n^(j)) If true, then it is called (p)⁽ⁱ⁾,n^(j)) Is a Tomeklins pair, and most sample points of the same type are deleted from each Tomeklins pair;

(9) according to the formula

And

and calculating the sample class proportion around each sample, finding out the noise class sample and deleting the noise class sample.

Images