Printer Friendly

An efficient approach for the diagnosis of faults in turbo pump of liquid rocket engine by employing FFT and time-domain features.

1. Introduction

Turbo pump is one of the major components in the propulsion system of space rocket engines. Turbo pump converts the low pressure propellant to high pressure propellant and supplies to the combustion chamber where it gets ignited. In this work, the vibration data-sets of Vikas engine turbo pump used in Indian Launch vehicles like PSLV, GSLV is taken for study. Turbo pump consists of three centrifugal pumps driven by a common turbine. The rotating assembly of Turbo Pump consists of impellers, bearings, shafts and seals. Turbo pump rotates at a speed of 10,000 RPM. The rotating assembly which rotates at its natural frequency is called critical speed. The natural frequency and the rotational speed are expressed in hertz and RPM, respectively. Kamijo (2013) has dealt with the turbo pump elaborately. In industrial pumps, the rotational speed is well below the critical speed. But in launch vehicle turbo pumps, the rotational speed is closer to the critical speed. Running the turbo pump closer to the critical speed is very dangerous when the vibration level of any one of the rotational components exceeds the normal value. The factors like rotor unbalance, shaft misalignment, wear and rub create more vibration.

Lurui et al. (2008) proposed transform processing techniques in which the vibration data in time domain is transformed to the angular domain in which the value of vibration at equally spaced angle of the rotating machine is obtained. It is then transformed to the order domain where the fundamental frequency and its multiples of frequency are related to the rotor blade weakness.

Lei et al. (2011) proposed the incremental clustering algorithm for extracting the optimised points from the data-set and also use the one-class support vector machine (SVM). Abnormality detection techniques were proposed by Hu et al. (2012) to find the variations in abnormal vibration from the normal vibration. The method distinguishes the sensor faults from turbo pump faults. Hong and Li (2013) proposed the protruding frequency method, in which the signal is divided into bands. In each band, the high protruding frequency is extracted as the feature, and root mean square (RMS) value for these protruding frequencies are used for fault detection.

Zhong et al. (2013) employed Lifting wavelet method for signal decomposition and least mean square error adaptive algorithm to reduce the error in classification of faults. Lei et al. (2013) applied the adaptive Gaussian threshold method for online monitoring of turbo pump faults and validated the method by post analysis using one-class SVM. Tharakan and Unnikrishnan (2015) compared the vibration level of turbo pump and thrust chamber in the diagnosis process.

The RMS value at characteristics frequency of the vibration is used for the health monitoring of the machine. It also uses other parameters for health monitoring like temperature, pressure and rotational speed. Lo, Wu and Whitehead (1993) used the amplitude of high peaks at the fundamental and its harmonics frequency in the power spectral densities to detect the abnormality in the turbo pump vibration using Neural network classifier.

In order to avoid the damage to the flight, the Turbo pump needs to be checked and qualified. Turbo pump is qualified for flight engine by dry running the Turbo pump to the 1/3 of the functional rotating speed and the vibration data acquired during its speed decay are used to check the health of the Turbo pump.

The turbo pump vibration data are corrupted by noise during the rising speed of the pump because VFP sensor is assembled close to the gas inlet source and so vibration data acquired during the rising speed are not used for the detection. Thus, vibration during the falling speed of the pump is acquired after closing the gas inlet to make it free of noise. The proposed method provides the steps to extract the vibration from the falling speed of turbo pump using the first natural frequency of the vibration. It detects the abnormality in the turbo pump operation with the simple time domain features like RMS, Crest Factor, Skew Factor and Kurtosis extracted in the noise-free decay region of vibration. The proposed method provides an accuracy of 100% for the fault detection of turbo pump used in liquid rocket engine using simple robust technique. The objective of the proposed work is to provide good accuracy for abnormality detection in vibration data with time-domain features. The motivation of the proposed work is to exploit the benefits of signal processing technique such as Fast Fourier Transform (FFT) and time domain features for vibration data analysis and to provide a robust technique than current state-of-the-art approaches.

The rest of this paper is organised as follows. Section 2 deals with the methodology involved in the fault detection process. The synchronisation process, feature extraction techniques and classification process are briefly summarised in this section. In Section 3, the experimental results are briefly discussed. Final conclusions are presented in Section 4.

2. Methodology

The block diagram of the proposed system is shown in Figure 1.

In the method proposed, the detection process is divided into synchronisation of the two vibration data [vibration in fuel turbo pump (VFP), vibration in oxidizer turbo pump (VOP)] with speed data, extraction of features between the decay speeds of2700 and 3000 RPM and classification of the data into either Normal data or Abnormal data. The block diagram of the proposed system is shown in Figure 1 explains the process involved in each module. Synchronisation of the vibration and speed signal is the main process in the synchronisation module. In feature extraction module, the features like RMS value, Kurtosis Factor, Crest Factor and Skew factor are extracted from the vibration data acquired between the speeds of 3000 and 2700 RPM. In classification module, the extracted features are used as the input to the SVM classifier for classification.

2.1. Synchronisation module

The steps involved in the synchronisation module are shown in the flow chart in Figure 2.

2.1.1. Measurement of vibration and speed signals

Two vibration sensors one for each pump (fuel and oxidiser) are assembled. The vibration on fuel pump is named as VFP and vibration on oxidiser pump is named as VOP. The vibration data are measured in the OROS analyser, and it is expressed in unit of G (acceleration due to gravity). The heavier object would require more force to generate the same acceleration as a light object. The speed of the turbo pump is measured by the magnetic sensor (non-contact type) and expressed in revolution per minute (RPM).

2.1.2. Splitting of vibration signals

The root mean square value for two vibration data is calculated. Identification of VFP (fuel pump) and VOP (Oxidiser pump) is done using the RMS value of the vibration level. VFP measurement is very closer to the gas inlet source of the turbine and hence the VFP vibration level is higher than VOP vibration level during startup of the pump.

2.1.3. Synchronisation of speed and vibration signals

The speed and vibration signals are acquired simultaneously, but through different acquisition system. Hence, both speed and vibration signals need to be synchronised for analysis. Theoretically, to synchronise speed and vibration signals with respect to time axis, the time instance at which the computed speed calculated from the vibration signal must be same as the time instant for the actual speed in speed signal. Hence, correction needs to be applied to the speed signal if the time instances are not same. If the actual speed time leads the computed speed, the correction time (difference) is reduced in the actual speed signal time domain values and if actual speed time lags the computed speed, the correction time (difference) is added in the actual speed signal time domain values. So, that the speed signal is synchronised with the vibration signal.

The synchronisation involves computation of the speed from the vibration signal by finding the first natural frequency (1N) of the vibration signal. The First natural frequency represents the speed in RPS (revolution per second). The time frame for first natural frequency calculation is selected from the starting point of decay speed of the turbo pump. Since the VFP sensor is assembled close to the gas inlet source most noise gets mixed with the vibration during the pump start up, and so it cannot be used for the 1N calculation. In order to extract the vibration from the falling speed of turbo pump, the time frame after the maximum peak of speed signal is needed in the vibration signal. The steps involved in detecting the time frame are

Step 1: Find the maximum vibration and its corresponding time is noted as the max time.

Step 2: Divide the vibration signals obtained after max time into blocks of 128 samples, and root mean square (RMS) is calculated at each block.

Step 3: Check condition that any root mean square (RMS) value calculated is less than 0.5 times the maximum vibration value. This step is to differentiate the region obtained is either noisy or decay region.

Step 4: After checking the condition, the first 2048 samples in the decay region is considered as are the time frame for 1 N calculation.

Step 5: Compute FFT to the vibration extracted at maximum time frame.

Step 6: The frequency within 20-80 Hz for which the amplitude is maximum is the first natural frequency.

Step 7: Calculate the speed at the maximum time frame from vibration using this first natural frequency

by

Speed = 60 X 1 N frequency (1)

Step 9: Compare the actual speed after the maximum speed (in the decay speed region) with the computed speed.

Step 10: Find the time at which the actual speed matches with the calculated speed and it is considered as the actual time.

Step 11: The difference between the actual time and max time is the time interval needed to be synchronised with the vibration signal. This difference of time interval is added to the vibration time domain for synchronisation.

The corrected vibration signal and the speed signal are synchronised, and this synchronisation is validated by computing the calculated speed from vibration and matching with the actual speed. This validation is performed by dividing the vibrations after max time frame into blocks of 2048 samples. Speed is calculated at each block by computation of first natural frequency at each block. This calculated speed at each block is compared with the actual speed.

2.2. Feature extraction module

In feature extraction module, the time features like root mean square (RMS), Kurtosis factor, Skew factor and Crest factor are extracted in each block of synchronised vibration signal which is of 128 samples. These features were extracted between the 3000 and 2700 RPM during speed decay period and the values are given to the classifier for detection. The flow chart for this module is given in Figure 3.

This feature extraction module involves two steps. In the first step, the vibration signal is divided into bocks of 256 samples for 0.1 s since the sampling frequency of the vibration signal is 2560 samples per second. The next step is finding the time domain features like RMS value, Crest factor, Kurtosis factor and Skew factor for every block of samples for every 0.1 s. Statistical characteristics of vibration signals will change in the form of energy or waveform. These changes can be described with time-domain features such as Mean, Stand Deviation, root mean square (RMS), Kurtosis Factor (KF), Crest Factor (CF) and Skew Factor (SF). Time-domain features have the merit of low cost of computation, and they are widely used for monitoring the turbo pump of the rocket engine in practical engineering. Blade damage, rub-impact and unbalance faults change the level and the waveform of vibration. The change in vibration level can be described with RMS value. RMS increases with the presence of increasing vibration level. RMS means root-mean-square of the signal, also known as the quadratic mean. As of Wikipedia definition, it is a statistical measure of the varying quantity magnitude. This measure gives additional information about the signal, namely the power of the signal.

The waveform of vibration signals can be described with statistical features such as Kurtosis Factor and Crest Factor (Mitchell et al. 2000). Kurtosis measures the degree of peak value of a signal distribution compared to a normal distribution. It describes the impulsive shape of the time signal. It is defined as the fourth static moment of the distribution of data (Carlo 1997).

As of Wikipedia definition, Skew Factor is a measure of the asymmetry of the probability distribution of a real valued random variable about its mean.

2.3.1. Calculation of decay time and rise time

The rise time is calculated as the time taken for the turbo pump to reach its maximum speed (3400 RPM) from zero. During ambient, speed measurement shows minor fluctuations. In order to find the time at this highest slope of speed signal, the difference between the adjacent speed points is calculated, and the first difference which has more than 5 RPM speed increase is the start time. The time at 3400 RPM is the max time. The rise time is calculated using the Equation (2):

Rise time = max time - start time (2)

The decay time is calculated as the time taken for the turbo pump from 3100 RPM to the speed at which the turbo pump stops rotating. When the pump comes to rest (i.e. at ambient), the speed measurement fluctuates with low amplitudes, and end time is calculated using the difference between the adjacent points and the time at very first difference which is less than 5 RPM is the end time. The decay time is calculated using the Equation (3)

Decay time = end time - max time (3)

2.4. Classification module

A support vector machine (SVM) is a binary classifier which determines the decision boundary which separates the features of Normal and Abnormal data in the multi-dimension space. The separating decision boundary is known as the hyperplane and this is obtained by the subset of features, and this is known as the support vector (Omar, Ngadi, and Jebur 2013). Linear discriminant functions can be used to discriminate between patterns belonging to two or more classes. There are two classes, one is normal and the other class is the faulty or abnormal. Patterns corresponding to these two classes are considered to be linearly separable, and there will be many lines separating the patterns belonging to class + 1 (normal) from those of class-1 (abnormal). The separating lines are expressed using the functional form shown in Equation (4)

f (x) = [w.sub.1][x.sub.1] + [w.sub.1][x.sub.2] + b (4)

In a d-dimensional space, the decision boundary is a hyperplane and can be represented by Equation (5)

f (x) = [w.sup.T]x + b (5)

where w and x are d-dimension weight and feature vectors, respectively, and b is the bias. The dimension for the feature used in the proposed work is 10 since 4 feature vectors for 2 vibrations are considered along with the rise and fall time. If we can find a weight vector w and a scalar b so that

[w.sup.T]x + b > 0 (6)

for all patterns x belonging to one class (say class + 1) and

[w.sup.T]x + b < 0 (7)

for all the patterns belonging to the other class (that is class 'X'). Equations (6) and (7) describe the separation of patterns into two classes.

The purpose of support vector classification is to devise a computationally efficient way of determining good separating hyper planes in a high-dimensional feature space. The SVM also uses high-dimensional feature space for large attributes formed by the non-linear mapping. A non-linear SVM is employed to find the best hyper plane from d-dimensional spaces that divide these data into a high-dimensional feature space where they may become linearly separable. Let S = {(x(k), y(k)), k = 1, 2,... n} be a linearly separable set of training data, where x(k) is the input data of the feature space, and y(k) = {-1, +1} is the class label. In the training phase, the training sets of feature vector are given as input to train the SVM classifier and the support vector, and the maximal margin hyperplane is identified for the training set. Three data-sets are used as a training set, two normal and other one being abnormal data (Omar, Ngadi, and Jebur 2013).

3. Results and discussion

The data-set for the proposed work is obtained from the Indian Space Research Organisation (ISRO), Mahendragiri. The data are real data. Rocket engine Turbo pump after assembly undergoes cold run test by supplying Nitrogen gas to drive the turbine. Turbo pump after passing this Cold run test will be delivered for flight. Cold run test is dry running of Turbo pump to the limited speed (i.e. up to 3400 RPM) without any pumping medium and to ensure the freeness of rotation without any friction. Vibrations are also acquired during the test.

The pump is run up to 3400 RPM considering the limitation in dry running of seals and bearings. The turbine drive pressure is withdrawn at 3400 RPM. At the time of withdrawal, some mechanical disturbances will occur and hence the vibration data near to 3400 RPM are not considered for analysis. Hence, the data from 3000 RPM to a range of10% (300 RPM) 2700 RPM are selected. The range is not allowed to exceed beyond 10% because mechanical vibrations tend to cease with respect to time.

The data-set contains 35 data, of which 27 data are normal data and 8 data are abnormal data. Five normal data and three abnormal data are used for training the classifier. The remaining 27 data containing 22 normal data and 5 abnormal data are used for the testing the classifier. Every data contain two vibration signals, (VFP, VOP) and speed signal.

3.1. Experimental results

The visualisation of the splitted raw vibration signals VOP, VFP and the speed signals is shown in Figure 4(a). The time frame for first natural frequency is calculated and shown in Figure 4(b). The time correction using the first natural frequency calculation is shown in Figure 5(a). The synchronised vibration data VOP, VFP, and speed data are shown in Figure 5(b).

The validation of the synchronisation is shown in Figure 6. The computed speed can be calculated only for first 20 instances. Due to the fluctuation in speed during ambient condition, computed speed cannot be calculated.

The start time in the rise time calculation and end time in the decay time calculation are based on the difference in adjacent speed values.

The detected max time, start time and end time are shown in Figure 7(a) and (b).

The RMS value for VOP and VFP turbo pumps for normal data is below 0.1 and 0.3 grms value, respectively. The RMS value for each block of data containing 256 samples extracted between the 3000 and 2700 RPM for VOP data for normal and abnormal data is shown in Figure 8(a) and (b), respectively.

The spec line in the graph shows the grms value limit for the normal data and this can be used for the diagnosis of the normal and the abnormal data. The Kurtosis value, RMS value, Crest Factor, Skew factor were extracted and given to the SVM classifier as input.

Kurtosis value, Crest value and Skew value extracted between 3000 and 2700 RPM for normal and abnormal data are shown in Figure 9(a)-(c), respectively.

In testing phase, 27 data are tested using the proposed method, in which no misclassification is found for 22 normal data and 5 abnormal data. The kernel function used in the SVM classifier is linear. Thus SVM classifier diagnoses the fault with 100% accuracy. True positives (TP) refer to the normal data-sets that were correctly labelled by the classifier. True negatives (TN) are the abnormal data-sets that were correctly labelled by the classifier. False positives (FP) are the normal data-sets that were incorrectly labelled as positive. False negatives (FN) are the abnormal data-sets that were mislabelled as negative.

The performance evaluation for the proposed work using SVM classifier is tabulated in Table 1.

Accuracy = [TP + TP/TP + TN + FP + FN] x 100% (8)

The accuracy of the proposed method is calculated using the Equation (8). Thus the proposed method yields the accuracy of 100% for the real-time data using simple robust technique and it is the advantage of the method. The optimisation is done for high accuracy, and this is considered as the score value and no disadvantage is found with this method. In the proposed method, the vibration data are divided into blocks of 128 samples. RMS value for each block of data is extracted and value is checked with the standard-specified value for more accuracy. Other features like kurtosis, skew factor and crest factor are also computed along with the RMS value in order to extract the peak information of the vibration signal in each block. From the results obtained it is found that the normal data have the highest rate of change in the values of kurtosis, skew factor and crest factor than those of abnormal data.

On the other hand, RMS values of protruding component of some frequency bands were extracted for fault detection by Hong and Li (2013).The method employs contribution coefficient and contribution limit to find protruding frequency components and setting the value to contribution coefficient plays a major role in reducing the sensitivity and increasing the complexity. Moreover, the method employed by Zhong et al. (2013) requires selection of proper values for the parameters in the algorithm and setting the threshold value of fault discriminate decision.

But in the proposed method simple time domain-features are extracted and classified into normal and abnormal data using SVM classifier with an accuracy of 100%.

4. Conclusion

In the proposed work, the fault detection method for the turbo pump was implemented using FFT and time domain features. The vibration of the turbo pump and speed data are obtained and synchronised using the first natural frequency. The vibration extracted between the speed of 3000 and 2700 RPM are analysed by the extraction of statistical time-domain features like RMS value, Kurtosis value, Crest Factor and Skew factor. The turbo pump historical data-sets obtained from the ISRO are used to train and test the SVM. The proposed method is robust enough to classify 22 normal and 5 abnormal data without any misclassification with the accuracy of 100%. This method prevents the damages to the turbo pump or to the whole rocket engine due to faults in turbo pump. The proposed work diagnoses the abnormalities of the turbo pump in the ground test itself.

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes on contributors

N. Aiswarya received her bachelor degree in Electronics and Communication Engineering from Anna University, Chennai in 2014. She completed her master degree in Applied Electronics in Anna University Chennai in 2016. Her main area of interest is signal processing.

S. Suja Priyadharsini received her Bachelor Degree in 2001 from Manonmaiam Sundaranar University, Tirunelveli, Tamilnadu. She completed her ME in Applied Electronics in Anna University, Chennai, Tamilnadu in 2006. She was conferred her PhD in Information and communication Engineering by Anna University, Chennai, Tamilnadu in 2015. She has been working as an assistant professor in Anna University Regional Campus, Tirunelveli, Tamilnadu since 2008. Her area of interests includes signal processing and soft computing.

K. S. Moni was awarded BE in Mechanical Engineering in 2005 and ME in Computer Aided Design in 2012. At present, he is a scientist/engineer in ISRO Propulsion Complex, Indian Space Research Organisation, Mahendragiri, India. His area of interest is vibration analysis and turbo machinery.

ORCID

S. Suja Priyadharsini [iD] http://orcid.org/0000-0002-3926-5263

References

Carlo, L. T. D. 1997. "On the Meaning and Use of Kurtosis." Psychological Methods 2: 292-307.

Hong, T., and H. Li. 2013. "Turbo Pump Fault Detection Algorithm Based on Protruding Frequency Components RMS and SVM." In 2013 IEEE International Conference on Mechatronics and Automation, Takamatsu, 1311-1316. doi: 10.1109/ICMA.2013.6618103.

Hu, L., N. Hu, B. Fan, and F. Gu. 2012. "Application of Novelty Detection Methods to Health Monitoring and Typical Fault Diagnosis of a Turbo Pump". In 25th International Congress on Condition Monitoring and Diagnostic Engineering, Journal of Physics, 12128-12138. doi: 10.1088/1742-6596/364/1/012128.

Kamijo, K. 2013. Research and Development of Rocket Turbo Pump. Sendai: Tohoku University Press.

Lei, H., H. Niaoqing, Q. Guojun, and G. Fengshou. 2011. "Turbopump Condition Monitoring Using Incremental Clustering and One-class Support Vector Machine." Chinese Journal of Mechanical Engineering 24: 474-489. ISSN-1000-9345.

Lei, H., H. Niaoqing, Z. Xinpeng, G. Fengshou, and G. Ming. 2013. "Novelty Detection methods for Online Health Monitoring and post data Analysis of Turbo Pumps." Journal of Mechanical Science and Technology 27: 1933-1942. doi: 10.1007/s12206-013-0508-x.

Lo, C.F., K. Wu, and B. A. Whitehead.1993. "Anomaly Detection of Turbo Pump Vibration in Space Shuttle Main Engine-using Statistics and Neural Networks." In 29th Joint Propulsion Conference and Exhibit, Monterey, CA, 179-182.

Lurui, X., H. Niaoqing, Q. Guojun, and G. Ming. 2008. "Typical Fault Diagnosis Method for High-speed Turbo Pump of Liquid Rocket Engine." Second International Symposium on Systems and Control in Aerospace and Astronautics: 1-5. doi: 10.1109/ISSCAA.2008.4776202.

Mitchell, L., M. Katherine, C. Robert, B. Carl, and M. Kenneth. 2000. "Review of Vibration Analysis Methods for Gearbox Diagnostics and Prognostics". In Proceedings of the 54th Meeting of the Society for Machinery Failure Prevention Technology, Virginia Beach, VA, 623-634.

Omar, S., A. Ngadi, and H. H. Jebur. 2013. "Machine Learning Techniques for Anomaly Detection: An Overview." International Journal of Computer Applications 79: 33-41.

"Root Mean Square." Accessed October 28, 2016. https://en.wikipedia.org/wiki/Root_mean_square

"Skew Factor." Accessed October 29, 2016. https://en.wikipedia.org/wiki/Skewness

Tharakan, T. J., and V. Unnikrishnan. 2015. Health Monitoring of Liquid Rocket Engine, 320-324. Thiruvananthapuram: Aeronautical Society of India-health Monitoring and Fault Detection in Aerospace System.

Zhong, F., H. Li, Q. Wu, and T. Hong. 2013. "A Fault Detection Algorithm for Turbo Pump Based on Lifting Wavelet and LMS." In IEEE International Conference on Mechatronics and Automation, Takamatsu, 348-353. doi: 10.1109/ICMA.2013.6617943.

N. Aiswarya (a), S. Suja Priyadharsini (a) [iD] and K. S. Moni (b)

(a) Department of Electronics and Communication Engineering, Anna University Regional Campus-Tirunelveli, Tirunelveli, India; (b) ISRO Propulsion complex, indian space research organisation, Mahendragiri, india

CONTACT s. suja Priyadharsini [??] sujapriya_moni@yahoo.co.in

KEYWORDS

time domain feature; root mean square; Kurtosis factor; crest factor; skew factor; support vector machine

ARTICLE HISTORY

Received 2 August 2016

Accepted 16 November 2016

https://doi.org/10.1080/14484846.2016.1264285
Table 1. Performance evaluation of SVM classifier for the
classification of vibration data.

Parameter name       Value

Total negative (N)    5
Total positive (P)   22
true positive (TP)   22
true negative (TN)    5
False positive (FP)   0
False negative (FN)   0
COPYRIGHT 2018 The Institution of Engineers, Australia
No portion of this article can be reproduced without the express written permission from the copyright holder.
Copyright 2018 Gale, Cengage Learning. All rights reserved.

Article Details
Printer friendly Cite/link Email Feedback
Author:Aiswarya, N.; Priyadharsini, S. Suja; Moni, K.S.
Publication:Australian Journal of Mechanical Engineering
Geographic Code:9INDI
Date:Oct 1, 2018
Words:4508
Previous Article:A simplification of railway vehicle lateral vibration model based on LQG control strategy.
Next Article:Experimental study using different heat transfer fluid of a packed bed thermal energy storage system during charging process.

Terms of use | Privacy policy | Copyright © 2020 Farlex, Inc. | Feedback | For webmasters