Intelligent fault diagnosis algorithm for fiber optic current transformer

Abstract

Affected by environmental factors, the performance of fiber optic current transformer (FOCT) will deteriorate over a long period of time. Intelligent fault diagnosis algorithm of Long-Short Term Memory (LSTM) combing with Support Vector Machine (SVM) is an effective way to deal with FOCT failures. According to the characteristics of LSTM, a signal prediction model in FOCT based on LSTM is proposed by analyzing the historical data. The residual signal can be obtained by the prediction signal and the observed signal. Set the residual threshold to determine whether the FOCT has fault. With the residual signal characteristics, a fault diagnosis model based on SVM is established. By analyzing the residual signal and extracting features, the diagnostic network can realize the pattern recognition and system fault diagnosis. Experiments demonstrate that the drift deviation fault, the ratio deviation fault and the fixed deviation fault can be diagnosed with an accuracy of 94.5%.

Keywords

Fiber optic current transformer fault feature fault diagnosis long-short term memory

1. Introduction

Fiber optic current transformer (FOCT) is a current sensor based on Faraday magneto-optic effect in power industry, and provides current information for relay protection, energy metering and control devices and so on. FOCT is widely used in intelligent, ultra-high voltage, and DC power grids, which has the advantages of large dynamic range, good frequency response characteristics and excellent insulation performance [1–3]. However, the working environment of the FOCT is complex, which affect the stability of the device. FOCT in actual work are prone to failure. To ensure the safe and stable operation of device, it is urgent to conduct research on online monitoring and fault diagnosis method of FOCT.

With the development of big data science and computer technology, fault diagnosis technology based on data method has been greatly developed [4]. Long-Short Term Memory (LSTM) is a time Recurrent Neural Network (RNN). It mainly solves the problem of gradient explosion and gradient disappearance in processing of nonlinear time by RNN. It is widely used to process and predict time series. Literature [5] proposed a deep learning fault diagnosis method based on LSTM to diagnose the chiller sensor deviation fault. Literature [6] proposed a power transformer operation state prediction method based on LSTM, which can predict the future transformer state. Various studies have shown that LSTM can be effectively applied to signal prediction and fault diagnosis.

The output signal is analyzed and the characteristic of fault signal is studied. The historical data is used to build a signal prediction model based on LSTM. The model can predict the output signal by a period of historical signal. When FOCT causes malfunction, output signal contains the fault characteristics. The residual signal determines if the device is faulty, which is obtained by the predicted signal and the observed signal. A fault diagnosis model based on Support Vector Machine (SVM) is established. The model can be used for online monitoring and fault diagnosis. Through simulation analysis, the model can accurately distinguish different faults, including fixed deviation, drift deviation and ratio deviation.

2. Fault signal characteristics of FOCT

Comparing with electromagnetic current transformer, FOCT has many advantages, which can meet the needs of smart grid and UHV. However, as an optical interference device, FOCT is susceptible to interference from temperature and electromagnetic fields. The factors that affect its stability are as follows:

The light source is susceptible to temperature during actual operation. And it will age when it is running for a long time. The deterioration of the light source affects the stability of FOCT.

The working environment of the FOCT is relatively harsh, with electromagnetic interference and a wide range of temperature changes. These factors can affect their performance and result in inaccurate measurements on the device.

The components of the signal processing unit will deteriorate under the influence of temperature, electronic circuit insulation aging, electromagnetic interference and other factors. This leads to inaccurate measurements.

The optical fiber in the optical path system is affected by temperature, and its own parameters will change, causing drift deviation of the device.

When the system fails, the output signal will change significantly [9,10]. Different faults contain different characteristics in both time domain and frequency domain. Online monitoring and fault diagnosis of the FOCT can be realized by analyzing the characteristics of the residual signal.

In practice, FOCT output signal is obtained by scaling the primary side current value. When the device fails, the fault signal is superimposed to the signal under normal conditions. The output signal expression of the current in the fault state is given by $\begin{eqnarray}\displaystyle y_{t}=kA\sin ({\omega}t+{\varphi})+e_{t}+f_{t} & & \displaystyle\end{eqnarray}$ (1) where y_t is output signal, A is current amplitude, ω and φ are frequency and phase, e_t and f_t are noise signal and fault signal respectively.

When the device has a fixed deviation fault, output signal has a constant fixed deviation relative to the normal signal. The fault signal is obtained by $\begin{eqnarray}\displaystyle f_{t}=b & & \displaystyle\end{eqnarray}$ (2) where b is the coefficient of the fixed deviation.

When the device has a ratio deviation fault, its ratio factor may change. The frequency and phase of the fault signal and the normal signal are consistent, and the amplitude changes proportionally. The fault signal is obtained by $\begin{eqnarray}\displaystyle f_{t}={\mu}\sin ({\omega}t+{\varphi}) & & \displaystyle\end{eqnarray}$ (3) where μ is the coefficient of the ratio deviation.

The materials of FOCT are related to changes in external temperature. As time goes by, some devices will age. The measured deviation of the signal changes with time, and the device generates a drift deviation fault. The fault signal is obtained by $\begin{eqnarray}\displaystyle f_{t}=c^{\ast }(t-t_{s}) & & \displaystyle\end{eqnarray}$ (4) where c is the coefficient of the drift deviation, and t_s is the moment when the fault occurs.

The output signal characteristics can reflect the fault status of device. Online monitoring and fault diagnosis can be performed by analyzing the characteristics of the fault signal.

3. Fault diagnosis model of FOCT

The data-driven method is to analyze the historical signal for condition monitoring and fault diagnosis. A prediction model based on LSTM is built by historical signal. When the prediction signal and the observed signal have a large deviation, the device is faulty. By analyzing the residual signal characteristics, online monitoring and fault diagnosis can be realized.

3.1. Long-short term memory

LSTM adds a cell state to the hidden layer of RNN, which is suitable for predicting long-term sequences and has strong memory. Each nerve unit of the LSTM contains three gates, including input gate, output gate, and forgetting gate. The input gate selectively records information to the current unit state, the forgetting gate determines whether to discard the information, and the output gate outputs the processed information to other units [5]. The LSTM transition equations are calculated by $\begin{eqnarray}\displaystyle \text{} & \text{} & \displaystyle f_{t}={\sigma}(W_{f}\cdot [h_{t-1},x_{t}]+b_{f})\end{eqnarray}$ (5) $\begin{eqnarray}\displaystyle \text{} & \text{} & \displaystyle i_{t}={\sigma}(W_{i}\cdot [h_{t-1},x_{t}]+b_{i})\end{eqnarray}$ (6) $\begin{eqnarray}\displaystyle \text{} & \text{} & \displaystyle c_{t}=f_{t}c_{t-1}+i_{t}(\tanh (W_{c}[h_{t-1},x_{t}]+b_{c}))\end{eqnarray}$ (7) $\begin{eqnarray}\displaystyle \text{} & \text{} & \displaystyle o_{t}={\sigma}(W_{o}[h_{t-1},x_{t}]+b_{o})\end{eqnarray}$ (8) $\begin{eqnarray}\displaystyle \text{} & \text{} & \displaystyle h_{t}=o_{t}\tanh (c_{t})\end{eqnarray}$ (9) where x_t, c_t, h_t are the input unit, cell state, and output unit of the neural unit at time t respectively, W_f, W_i, W_o, W_C is the weight matrices, b_f, b_i, b_o, b_C are the bias vectors.

3.2. Fault diagnosis model based on LSTM-SVM

The prediction model can predict the output signal at the next moment. When the predicted signal and the observed signal cause a large deviation, it is determined that the device is faulty. A fault diagnosis model based on SVM is used to distinguish fault diagnosis models. LSTM and SVM are cascaded to build fault diagnosis model [10]. The structure of the model is shown in Fig. 1.

Fig. 1.

The structure of FOCT prediction model based on LSTM.

The historical data is preprocessed at the input layer to construct a data set that meets the model requirements. To construct the prediction model accurately, the history data is normalized to a time series with zero mean and unit variance [6], which is calculated by $\begin{eqnarray}\displaystyle x(t_{i})^{\ast }=\frac{x(t_{i})-mean(x(t))}{std(x(t))}. & & \displaystyle\end{eqnarray}$ (10)

The prediction model can predict the value of the current time by historical data. Split the time series according to the step value $\begin{eqnarray}\displaystyle \text{} & \text{} & \displaystyle x_{train}=[x_{1},x_{2},\ldots ,x_{n-L}]=\left[\begin{array}{@{}cccc@{}}x(t_{1})^{\ast } & x(t_{2})^{\ast } & \ldots & x(t_{n-L})^{\ast }\\ x(t_{2})^{\ast } & x(t_{2})^{\ast } & \ldots & x(t_{n-L+1})^{\ast }\\ \ldots & \ldots & \ldots & \ldots \\ x(t_{L})^{\ast } & x(t_{L+1})^{\ast } & \ldots & x(t_{n-1})^{\ast }\end{array}\right]\end{eqnarray}$ (11) $\begin{eqnarray}\displaystyle \text{} & \text{} & \displaystyle y_{train}=[x(t_{2})^{\ast },x(t_{3})^{\ast },\ldots ,x(t_{n})^{\ast }].\end{eqnarray}$ (12)

The prediction model based on LSTM is obtained by training the segmented data. According to the historical signal $\{y_{real}(t_{i})\},i=m-1,m-2,\ldots ,m-L+1$ , the signal y_pred(t_m) predicted by the model at the current time can be obtained.

The residual signal can be obtained by calculating the observed signal and the predicted signal, which is given by $\begin{eqnarray}\displaystyle error(t_{i})=y_{pred}(t_{i})-y_{obs}(t_{i}). & & \displaystyle\end{eqnarray}$ (13)

The residual signal contains the fault information of FOCT. According to the feature of residual signal, the machine learning classification algorithm can be used to realize the fault diagnosis of the FOCT [11]. SVM is a machine learning method based on statistical learning theory. Its goal is using the structural risk minimization principle to construct an optimal decision function, which can solve the two classification problem.

Based on the time domain characteristics of the residual signal, the eigenvector is constructed by parameter mean value, variance, root mean square error and peak value. Fault diagnosis model based on SVM is constructed by the constructed eigenvector. The model can determine whether the device is faulty. The FOCT fault is indicated when there is a large deviation in the residual signal. According to the characteristics of residual signal, the fault diagnosis model can determine the type of fault.

4. Experiment

The historical output signal is acquired with a signal amplitude of 100 A and a sampling rate of 4000 Hz. The training data set is constructed by different time steps, which can be used to construct the prediction model. In order to better quantify the accuracy of the prediction model, the root mean square error (RMSE) and the mean absolute error (MAE) are introduced to measure the accuracy of the model prediction [12]. REMS and MAPE are calculated by $\begin{eqnarray}\displaystyle \text{} & \text{} & \displaystyle \mathit{RMSE}=\sqrt{\frac{1}{N}\mathop{\sum }_{i=1}^{n}(y_{pred}(t_{i})-y_{real}(t_{i}))^{2}}\end{eqnarray}$ (14) $\begin{eqnarray}\displaystyle \text{} & \text{} & \displaystyle \mathit{MAE}=\frac{1}{N}\mathop{\sum }_{i=1}^{N}|y_{pred}(t_{i})-y_{real}(t_{i})|.\end{eqnarray}$ (15)

The accuracy of model prediction is related to the historical data step size. When the step size is an integer multiple of the current signal period, the accuracy of the model prediction is significantly higher. And as the step size increases, the accuracy begins to decrease. Therefore, set the time step value to 20 to establish the signal prediction model. The test data set is used to judge the accuracy of the model prediction signal. Figure 2(a) shows the model prediction signal value and the observed value of the device, and Fig. 2(b) shows the residual curve. It can be seen that the prediction model can predict the output signal well.

Fig. 2.

Model prediction signal and observed signal of the device.

When the device is faulty, the fault signal is superimposed on the normal output signal. The residual signal also changes significantly. Simulation analysis is performed by superimposing different types of fault data (Eq. (3)–(5)) on the normal output signal of the FOCT.

Figure 3 shows the predicted value and the observed value when the device has a variable ratio deviation fault. It can be seen from the figure that the signal amplitude becomes larger, and the residual signal has a large range of oscillation.

Figure 4 shows the residual signal when the device has a fixed deviation fault. It can be seen from the figure that the fault signal has a constant deviation from the normal signal, and the residual signal suddenly increases and stabilizes at a constant value.

Figure 5 shows the residual signal when the device has a drift deviation fault. In the initial phase of the fault, the residual signal does not exceed the threshold. As time passes, the residual signal gradually increases beyond the threshold, exhibiting significant fault characteristics.

According to the simulation experiment, the residual threshold is set to 1 A. When the residual is less than 1 A, it is determined that the transformer is in a normal working state. When the residual exceeds the threshold, the residual signal of 1 second is acquired. And the fault diagnosis model is used to determine the type of fault.

Fig. 3.

The residual signal when the ratio deviation fault occurs.

Fig. 4.

The residual signal when the ratio deviation fault occurs.

Fig. 5.

The residual signal when the ratio deviation fault occurs.

The simulation of different faults is done by adjusting the severity of the fault. The eigenvector of the signal residual is constructed according to its time domain characteristics, including the mean, variance, root mean square, maximum and minimum value. A fault diagnosis model based on SVM is established by training the eigenvector, and the test signal is used to verify the accuracy of the model. When the system is running, there may be a momentary deviation between the predicted signal and the real signal, but the transformer does not malfunction.

The output of the fault classification model includes four categories, including no fault, drift deviation fault, ratio deviation fault and fixed deviation fault. The simulation results demonstrate that the model can classify faults with an accuracy of 94.5%.

5. Conclusion

Combining with the advantages that LSTM can deal with time correlation problem well, a fault diagnosis technology based on LSTM-SVM is proposed. By analyzing the typical fault model of FOCT, the characteristics of each fault signal are determined. The prediction model of the output signal based on LSTM is established by using the historical signal. The model can be used to predict the output signal by historical signal. And the residual signal between prediction signal and the observed signal can be obtained. The characteristics of residual signal is analyzed to determine if the device is faulty. When the residual signal exceeds the threshold, it is judged that the device is faulty. The FOCT diagnostic model is constructed by SVM, and the fault type of the FOCT is judged according to the time domain characteristics of the residual signal. The simulation results demonstrate that the method based on LSTM-SVM can accurately classify the three faults, and verify that the proposed method is effective and feasible. It provides an effective method for online monitoring and fault diagnosis of FOCT.

Footnotes

Acknowledgements

The project was supported by National Natural Science Foundation of China [51477028], Primary Research & Development Plan of Jiangsu Province [BE2018384].

References

Wang

, Optimization analysis and experiment study on measurement performance of all-fiber optical current transformer, Journal of Electrical Engineering & Technology33(17) (2018), 4146–4153.

Wang

Wei

Zhu

Liu

and Zhengqi

, Real-time modeling and online filtering of the stochastic error in a fiber optic current transducer, Measurement Science and Technology27(10) (2016), 105103, 1–6.

Wang

Wei

and Liu

, Statistical modeling of fiber optic current transducer, Measurement92(10) (2016), 288–294.

Teng

Zhao

and Ma

, Data driven fault diagnosis for segment-powered linear induction motor, Electric Machines and Control23(1) (2019), 1–8.

Yin

Zheng

and Liu

, Application of improved LSTM method in sensor fault detection of the chiller, Transactions of China Electrotechnical Society34(11) (2019), 2324–2332.

Dai

Song

Sheng

Jiang

, Prediction method for power transformer running state based on LSTM network, High Voltage Engineering44(4) (2018), 1099–1106.

Zhang

Cheng

, Research on online monitoring technology for active electronic current transformer, High Voltage Engineering42(1) (2016), 208–213.

Jiang

Wang

, Application of artificial neural network to fault diagnosis for optical voltage transformer, Proceedings of the CSU-EPSA30(6) (2018), 134–139.

Wang

Zhang

Liu

and Yang

, Research and application of electronic transformer fault diagnosis, Power System Protection and Control43(18) (2015), 74–79.

10.

Shang

Meng

Yang

and Cai

, LSTM-BP neural network based fault diagnosis for on-board equipment of Chinese train control system, Journal of Beijing Jiaotong University43(1) (2019), 54–62.

11.

Jegadeeshwaran

and Sugumaran

, Fault diagnosis of automobile hydraulic brake system using statistical features and support vector machines, Mechanical Systems and Signal Processing2015:52–53, 436--446.

12.

Wang

Xie

Wang

, Prediction of interaction between grid and wind farms based on PCA-LSTM model, Proceedings of the CSEE39(14) (2019), 4070–4081.