Learning deep representation of imbalanced SCADA data for fault论文分析

提出的故障对象

风机叶片结冰问题

相关变量

Considering that blades icing accretion can lead to overloading and power generation loss,twenty two continuous variables related to fault information areselected from SCADA. They can be grouped into three categories,i.e., (1) wind parameters: such as wind speed and wind direction.They are measured directly through anemometer and wind vane;(2) energy-related parameters: these parameters are closelyrelated with power output of wind turbines, e.g., rotating speed of generator, pitch angle, pitch speed, drive current of pitch motor and so on; (3) temperature parameters: they reflect working condition and working status of wind turbines to a certain degree. They include environmental temperature, nacelle temperature,temperature of pitch motor, temperature of battery cabinet, and so on. Fig. 7(a–c) shows curves of some critical parameters, i.e.,active power and wind speed. The detailed information is listed in Table 2. From this table, it can be seen that these parameters contain rich information of wind turbine operation condition and external environment. The collected SCADA data covers multiple operation stages of wind turbines, such as the stage of tracking wind energy, the stage of constant rotating speed, and the stage of constant power. Note that we discard the shutdown stage where the wind speed is bigger than cut-out speed. Due to wind turbines working well at the most of time, the phenomenon of imbalance is extremely severe, which can be seen from Table 3 that lists the proportions of normal data and abnormal data with regard to each wind turbine. The amount of normal data is much more than ten times that of the abnormal data.

本文提出的方法

DNNs，The loss function of the proposed DNNs is selected as the triplet loss

模型细节

The feature extraction network is
built up based on the principle that is stated in Section 3.2, and
is programed by deep learning toolbox named Tensorflow. This
network is optimized by the back propagation algorithm (BP)
[32] and is trained on wind turbine 1 dataset.

创新点

The bypass structure（It can take local features and global features of SCADA data into consideration simultaneously） is introduced into it for considering local features and global features of original SCADA data. Basis of well-designed DNNs and tri-plet loss. It can preserve within-class data information and between-classes data information at the same time by learning deep representation in hypersphere, unlike traditional DNN based fault diagnosis methods directly exploiting class label information to build up discriminative models.

文献【13】故障特征

风速和输出功率间的关系，本文也是关注风速和输出功率间的关系来对风机进行故障诊断，将实际功率输出与模型期望输出之间的预测误差作为异常指标。

文献【13】模型构造

1.Being same as the ANNs structure in [13], the ANNs used in here is constructed with two hidden layers. The number of hidden neurons is 9. The input features are also wind speed ws, wind direction wd, and environmental temperature tout. The last two features make some contribution to the prediction of power curve, which has been validated by the [13]. The output is active power Pa of wind turbine. Then, the NBM-ANNs model is trained on the data of wind turbine 1 by BP algorithm [32]. The prediction error between the model’s output(expectation) and the real measurement can be used as an indicator for anomaly

本文的优点

Unlike traditional fault diagnosis methods that directly use data
label information or drop the fault data information, the proposed
method tries to learn a deep feature representation of each data
sample in a hypersphere.