EMD and ANN based intelligent fault diagnosis model for transmission line

Abstract

In the presented work, an intelligent model for fault classification of a transmission line is proposed. Ten different types of faults (LAG, LBG, LCG, LABG, LBCG, LCAG, LAB, LBC, LCA and LABC) have been considered along with one healthy condition on a simulated transmission line system. Post fault current signatures have been used for feature extraction for further study. Empirical Mode Decomposition (EMD) method is used to decompose post fault current signals into Intrinsic Mode Functions (IMFs). These IMFs are used as input variables to an artificial neural network (ANN) based intelligent fault classification model. Relief Attribute Evaluator with Ranker search method is used to select the most relevant input variables for fault classification of a three-phase transmission line. Proposed approach is able to select most relevant input variables and gives better result than other combinations. Ours is a first attempt at using EMD for feature selection in fault classification of transmission lines.

Keywords

Empirical mode decomposition artificial neural network transmission line fault diagnosis feature selection

1 Introduction

Power quality disturbances pose a major challenge for industries that use electrical equipments. Power quality disturbances lead to interruption of supply and damage to the equipments. Unsymmetrical faults or short circuit faults such as three phase faults, line to line faults and line to ground faults are the main causes of power quality disturbances and can make the system unstable. A fault classifier should be fast and reliable for maintaining continuous power flow in power transmission systems. Proper algorithm selection for analysis of the faulty condition is an equally important aspect. Furthermore, an appropriate algorithm is a prerequisite for extracting the most significant feature from the system under study. The features so extracted could be used for detecting faults using a classifier.

Many researchers have proposed techniques for accurate fault classification. The techniques available in literature include: classical approaches [1] or impedance matching based methods [2], fuzzy logic based approach [3], and artificial intelligence based approaches [4]. A combination of EMD and support vector machines (SVM) for fault classification in power systems is introduced in [5]; classification of faults using fuzzy rule based approach is presented in [6 , 9] describe ability of neural networks for storing Q values. However, both neural networks and reinforcement learning are slow learning approaches and their combination can make the overall system sluggish. Heuristic fuzzy logic approach [10] increases the computational burden due to redundancy and complexity. Thus, a transparent and simple fuzzy logic based approach for accurate fault classification is the need of the hour.

In this work, feature extraction is done using EMD and ANN has been used for fault classification. EMD (like Fourier transform and wavelet decomposition) is used to break down the signal. Unlike, Fourier transform and wavelet decomposition, EMD approach makes no a priori assumptions about the composition of a signal. IMF’s are successively traced out between maxima and minima using a spline interpolation [11]. Then the best feature is selected using Relief Attribute Evaluator with Ranker search method of the WEKA software. Selected feature is then fed to the ANN classifier. For classification of faults, a multi layer perceptron (MLP) network is used. MLP is a nonlinear learning and modeling tool which can be used for the accurate classification of faults. Main focus is on classification of single line to ground faults, double line to ground faults, and three phase faults. Accurate classification of faults among ten fault types (LAG, LBG, LCG, LABG, LBCG, LCAG, LAB, LBC, LCA and LABC) has been achieved. In the presented approach, post fault currents and voltages have been predicted. Thereafter, the acquired signal is decomposed into IMF’s by EMD. Then IMF’s are partitioned using fuzzy logic and best features are selected and learned by ANN’s. Proposed scheme is evaluated on a 500KV, 50 Hz transmission line of 300 km length connected between two sources using MATLAB. Fault data is generated using SimPowerSystems toolbox in the SIMULINKenvironment.

2 Model description

Current waveforms corresponding to one healthy condition and 10 fault conditions have been simulated on a 500KV, 50 Hz transmission line of length 300 km connected between two sources (Fig. 1). Parameters of the proposed system are listed inTable 1.

3 Methodology

3.1 Empirical Mode Decomposition (EMD)

In literature, wavelet transforms (WT) and empirical mode decomposition (EMD) methods have been employed for processing of voltage and current signature signals. We have selected EMD technique for this purpose. Although WT has excellent properties in time and frequency spaces; it has some shortcomings, e.g., WT breaks the signal into frequency bands and the user has to set decomposition levels in every trial. WT has strong dependence on wavelet basis functions and fails to split high frequency band when the fault has some modulation. EMD is advantageous as it can process characteristic time scales imbedded in the signal. EMD breaks down the signal to a set of IMFs, without needing basis functions. The signal itself provides number of IMF’s and scaling factors. EMD has wide applicability as it can handle non linear and non stationary signals as well. We have used EMD technique for shredding a nonlinear signal y (t) to IMFs. Each IMF has to satisfy two conditions:

Zero crossing counts and number of extremum must either be equal or differ by at most one.

The mean value of maximum (minimum) envelopes has to be zero at any given point.

First, an upper and lower envelope is constructed on the signal. Then an averaging is performed on these envelopes for calculating the difference signal by subtracting average from the original signal. The resulting signal is inspected to see whether it satisfies conditions 1) and 2), as specified above. In case the difference signal fails to satisfy any of the two conditions, it is rechristened as the original signal and envelope and difference signal calculation are repeated. This procedure is continued till the resulting signal satisfies conditions that make it an IMF. We give a step by step procedure for calculating IMF’s from a given signal:

Take the input signal y (t).

Calculate the extreme values, i.e., maxima and minima for the signal y (t).

Use cubic spline interpolation for connecting these maximum and minimum points.

Estimate an upper envelope e_m (t) and lower envelope e_t (t).

Generate mean value of the minima-maxima envelopes: a (t) = [e_m (t) + e_t (t)]/2.

Calculate the difference signal by subtracting the mean signal from the original signal: H₁ (t) = y (t) - a (t).

Test whether H₁ (t) satisfies conditions 1) and 2). If yes, then H₁ (t) is the 1st IMF.

If 7) is not true, H₁ (t) is not an IMF and H₁ (t) rechristened as the original signal, and steps 1 to 6 are repeated.

Repeating this procedure k-times; H₁ (k) become an IMF orH₁ (k) = H_1(k-1) - a (t).

Smallest temporal scale is defined as y (t) : Ω₁ (t) = H_1k (t), where Ω₁ (t) is the 1st IMF of the original signal.

Residue is calculated as ψ₁ (t) = y (t) - Ω₁ (t). We consider ψ₁ (t) as the original signal and repeat the above procedure for calculating 2nd IMF.

Repeat the entire process n-times to get n IMFs from the signal y (t). We terminate the process when ψ₁ (t) turns out to be a monotonic function from which no more IMFs can be obtained.

The EMD process decomposes y (t) to: $y (t) = \sum_{l = 1}^{L} Ω_{l} (t) + ψ_{L} (t)$ (1) where L = number of IMFs; Ω_l (t) = l^th IMF and ψ_L (t) =final residue. All like components or IMFs (1) should give a meaningful local frequency and unlike IMFs correspond to different frequencies. Equation (1) can be represented as: $y (t) \approx \sum_{l = 1}^{L} A_{l} (t) cos [φ_{l} (t)]$ (2)

The MATLAB code for EMD can be accessed at http://perso.ens-lyon.fr/patrick.flandrin/emd.html. We describe the EMD technique in Fig. 3.

We also give energy distributions of IMFs, from which we can alienate normal and unbalance fault conditions of a transmission line (Fig. 4(a-d)). We also calculate energy entropies using (3) and present them in Fig. 4(a-d). From Fig. 4, it is evident that entropy for unbalance fault differs from normal fault in terms of IMFs. Energy entropy is defined as: $E_{e} = - \sum_{n = 1}^{N} p_{n} log p_{n}$ (3) where, p_n = E_n/ - E is the percentage value of energy for the n^th IMF in the signal, $E = \sum_{n = 1}^{N} E_{n}$ is the energy of given continuous-time signal x (t). For discrete-time signals x (t) we have: $E_{s} = 〈 x (t), x (t) 〉 = \int_{- \infty}^{\infty} {| x (t) |}^{2} .$

From Fig. 4, we infer that energy distribution of IMFs changes with the type of fault. This shows that energy of IMFs can be used as an additional time-frequency feature for transmission line fault identification.

3.2 Most relevant input variable selection

To reduce computational complexity and for faster processing, we prune the IMF’s obtained from the EMD analysis to 4 most relevant ones [8]. Relief attribute evaluator with ranker search method has been used to identify most relevant input variables. Algorithm selects variables based on most appropriate features embedded in the data. This limits the number of input variables thereby reducing the computational burden.

Relief attribute evaluator calculates usefulness of an attribute by sampling it repeatedly. Value of an attribute belonging to a certain class is different from the one belonging to another class. With ranker search, we can assign the attributes according to their individual values. Relief attribute evaluator can tolerate noise and is unaffected by feature extraction. Application of this attribute evaluator gives us best four IMF’s (IMF3, IMF4, IMF6 and IMF8), out of the 12 available IMFs.

3.3 Artificial Neural Network (ANN) classifier

Neural network (NN) is approximation architecture with a number of neurons connected together via weights and biases. NN has the capability to correctly identify a non linear mapping between input and output and stores it in the form of weights and biases. The weights and biases of the network are adapted via a training algorithm.

Most relevant IMF’s (selected in the preceding section) are fed to the ANN, which is trained to identify fault corresponding to the input IMFs. We have used feed forward back propagation NN with gradient descent algorithm. The network analyses the inputs and attempts to classify a given fault: a trial and error procedure. Ideally, ANN would require large fault data for classifying faults. As we do not have access to such a large data set for real transmission lines, we have used simulation of faults on a transmission line model. NN learning is terminated when a specified value of mean square error (MSE) is achieved.

Use of back propagation algorithm in NN learning is a well established and successful method. NN architecture with back propagation is shown in Fig. 5. The weight update of NN using back propagation can be represented as: $W_{k + 1} \to W_{k} + α_{k} g_{k}$ (4) where, W_k =weight and bias vector of the NN at instant k, g_k =current gradient and α_k =learning rate.

As input samples are presented to the NN with associated targets (training phase), NN begins adjusting its weights and biases. Error between ANN output and actual fault value is used for tuning the free parameters of the network via back propagation of error signal. This process is repeated for all the samples. Eventually, the network is able to identify correct fault type from input data. For better understanding of ANN implementation in practical applications of engineering, reader may refer [8, 12–17 , 8, 12–17].

4 Proposed approach

Current signals I_a , I_b , I_c (for 3 phases) are processed using EMD to generate a set of IMFs. These IMFs are then pruned to a lower set of most relevant IMFs by the relief attribute evaluator. NN is trained on these pruned set of IMFs to identify faults. Next, NN is tested on a new set of data samples. Finally, we validate the ANN (obtained after training and testing) on a separate fault data set. This is explained in Fig. 6.

5 Results and discussion

For fault classification, a fault is created at 0.1 sec and cleared out at 0.25 second. Simulation is run for 10 second with a sampling frequency 12.5 KHz for each case. Simulation of the model generates 1, 25,000 data samples for each case. This dataset is decomposed into IMFs, which are used as input variables to the ANN. We use the first 1500 IMFs data samples (each case) for our study, i.e., a data set of order 16500×11. Out of this dataset, 70% samples (randomly chosen) are fed to the neural network during the training phase. Rest of the data samples are used for testing and validation.

Figures 7–9 give results obtained by using the proposed ANN classification approach. As evident from Fig. 7, our approach achieves a classification accuracy of 99.09%. Figure 8 gives histogram plots for the instances used for classification. Figure 9 gives MSE plot for the ANN corresponding to training, testing and validation phases. Finally, Table 2 gives matrix for 16,500 samples, indicating how each fault got classified by our ANN based approach.

6 Conclusions

This paper presented a combined EMD and ANN technique for fault classification for transmission lines. Relief attribute evaluator is used to extract relevant features enabling reduction of input data cardinality. The technique achieved an overall accuracy of 99.095%. Accuracy for the training and testing phases is 99.08 %. Selection of most relevant input parameters using relief attribute evaluator with ranker search method is a useful and promising fault diagnosis tool for transmission lines.

References

Aggarwal

R.K.

, Aslan

and Johns

A.T.

, New concept in fault location for overhead distribution systems using superimposed components, IET Gener Transm Distrib144(3) (1997), 309–316.

Jung

C.K.

, Lee

J.B.

, Wang

X.H.

and Song

Y.H.

, Wavelet based noise cancellation technique for fault location on underground power cables, Electric Power Systems Research77(10) (2007), 1349–1362.

Sadinezhad

and Joorabian

, A new adaptive hybrid neural network and fuzzy logic based fault classification approach for transmission lines protection, in Int Conf on Power and Energy Conference (2008), 895–900.

Dalstein

and Kulicke

, Neural network approach to fault classification for high speed protective relaying, in Power Delivery, IEEE Transactions on10(2) (1995), 1002–1011. doi: https://dx-doi-org.web.bisu.edu.cn/10.1016/j.asej.2015.08.005

Babu

N.R.

and Mohan

B.J.

, Fault classification in power systems using EMD and SVM, Ain Shams Engineering JournalISSN 2090–4479 (2015), doi: https://dx-doi-org.web.bisu.edu.cn/https://dx-doi-org.web.bisu.edu.cn/10.1016/j.asej.2015.08.005 (In Press).

Samantaray

S.R.

, A systematic fuzzy rule based approach for fault classification in transmission lines, Applied Soft Computing, ISSN 1568–494613(2) (2013), 928–938. doi: https://dx-doi-org.web.bisu.edu.cn/10.1016/j.asoc.2012.09.010

Lin

L.-J.

, Self-improvement Based On Reinforcement Learning, Planning and Teaching, in Machine Learning Proceedings, ISBN 9781558602007, edited by Lawrence Birnbaum

and Gregg Collins

, Morgan Kaufmann, San Francisco (CA) (1991), 323–327. doi: https://dx-doi-org.web.bisu.edu.cn/10.1016/B978-1-55860-200-7.50067-2

Malik

and Mishra

, Artificial neural network and empirical mode decomposition based imbalance fault diagnosis of wind turbine using TurbSim, FAST and Simulink IET Renewable Power Generation2016. DOI: 10.1049/iet-rpg.2015.0382 (In Press).

Rummery

and Niranjan

, On-line Q Learning using connectionnist systems, 1994. Accessed on 21 July 2016, https://dx-doi-org.web.bisu.edu.cn/10.1016/B978-1-55860-200-7.50067-2

10.

Youssef

O.A.S.

, Combined fuzzy-logic wavelet-based fault classification technique for power system relaying, in Power Delivery, IEEE Transactions on19(2) (2004), 582–589. doi: 10.1109/TPWRD.2004.826386

11.

Damerval

, Meignen

and Perrier

, A fast algorithm for bidimensional EMD, in Signal Processing Letters, IEEE, 12(10) (2005), 701–704doi: 10.1109/LSP.2005.855548

12.

Rahi

O.P.

, Yadav

A.K.

, Malik

, Azeem

and Kr

, Power System Voltage Stability Assessment through Artificial Neural Network, Procedia Engineering30 (2012), 53–60. doi: 10.1016/j.proeng.2012.01.833

13.

Malik

and Mishra

, Feature selection using RapidMiner and classification through probabilistic neural network for fault diagnostics of power transformer, in Proc IEEE India Conference (INDICON) (2014), pp. 1–6. doi: 10.1109/INDICON.2014.7030427

14.

Malik

and Mishra

, Application of LVQ network in fault diagnosis of wind turbine using Turbsim, fast and simulink, Michael Faraday IET International Summit83 (2015), 474–480. doi: 10.1049/cp.2015.1679

15.

Kumar

and Malik

, Generalized Regression Neural Network Based Wind Speed Prediction Model for Western Region of India, Procedia Computer Science93 (2016), 26–32. doi: 10.1016/j.procs.2016.07.177

16.

Kumar

, Sharma

and Malik

, Learning Vector Quantization Neural Network Based External Fault Diagnosis Model for Three Phase Induction Motor Using Current Signature Analysis, Procedia Computer Science93 (2016), 1010–1016. doi: 10.1016/j.procs.2016.07.304

17.

Malik

and Mishra

, Application of Probabilistic Neural Network in Fault Diagnosis of Wind Turbine Using FAST, TurbSim and Simulink, Procedia Computer Science58 (2015), 186–193. doi: 10.1016/j.procs.2015.08.052