^{*}

This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution license (http://creativecommons.org/licenses/by/3.0/).

This study presents a fault detection of roller bearings through signal processing and optimization techniques. After the occurrence of scratch-type defects on the inner race of bearings, variations of kurtosis values are investigated in terms of two different data processing techniques: minimum entropy deconvolution (MED), and the Teager-Kaiser Energy Operator (TKEO). MED and the TKEO are employed to qualitatively enhance the discrimination of defect-induced repeating peaks on bearing vibration data with measurement noise. Given the perspective of the execution sequence of MED and the TKEO, the study found that the kurtosis sensitivity towards a defect on bearings could be highly improved. Also, the vibration signal from both healthy and damaged bearings is decomposed into multiple intrinsic mode functions (IMFs), through empirical mode decomposition (EMD). The weight vectors of IMFs become design variables for a genetic algorithm (GA). The weights of each IMF can be optimized through the genetic algorithm, to enhance the sensitivity of kurtosis on damaged bearing signals. Experimental results show that the EMD-GA approach successfully improved the resolution of detectability between a roller bearing with defect, and an intact system.

As modern industries inevitably utilize a wide range of rotating machinery, the imperative of securing its safety during the service life has also escalated significantly. In particular, the maintenance and repair costs due to degradation of a system have increased, along with the functional complexity of the mechanical systems. In fact, sudden failure or structural defects may lead to catastrophic accidents. Thus, the potential threat to the economy and human losses has increased. In the case of rotating machinery, real-time vibration analysis and feature extraction techniques are required to identify the incipient defects, caused by fatigue, contamination, overload or poor maintenance. In this context, vibration-based preemptive condition monitoring schemes for rotating machinery have been actively studied since the 1970s [

One of the most widely used vibration-based techniques is to compare the trend of shifted harmonics or kurtosis, which evaluate the frequency and time-series measurements, to determine the presence of defects [

Among many signal processing techniques, deconvolution-type methods have attracted extensive attention in recent years. The minimum entropy deconvolution (MED) technique was first proposed by Wiggins, for extracting reflective information from seismic recordings [

Moreover, one of the modulation techniques, also known for Teager-Kaiser Energy Operator (TKEO) has been applied for monitoring abrupt change of energy in neurological signals [

Huang [

In this study, we investigate two different signal processing methods of vibration signals of bearing, with attempts to identify the characteristics of defects: MED [

In the 1980s, Wiggins first introduced the concept of Minimum Entropy Deconvolution (MED) that maximizes spike-like characteristics from noise-corrupted data points, or equivalently minimizes the entropy of signal components, using a linear operator [_{i}_{i}_{i}_{i}_{i}_{1}_{N}

Notably, the coefficient of the filter _{i}

To achieve this, the varimax norm needs to be differentiated with respect to filter coefficients, or

As depicted above, the MED algorithm minimizes the entropy of an arbitrary input signal through the deconvolution process, so that the kurtosis of the signal can be maximized.

The TKEO algorithm belongs to the category of nonlinear high-pass filters, which reduce the variation of low frequency background signals, while boosting transient components of a signal in the high frequency region. Eventually, transients and background signals can be easily separated through the TKEO. The first-order discrete time model of the TKEO is expressed as

Given an oscillatory data set, EMD iteratively separates high frequency components from the original data, by enforcing the condition of IMF. Specifically, the core of EMD is the sifting process, which first defines envelopes of upper and lower extremes, and iteratively subtracts the mean of both envelopes from the previous data set or residual, until it satisfies the necessary condition of IMF. The first condition of IMF begins with the fact that any oscillatory signal, regardless of stationarity or non-stationarity, can be decomposed into separate constituent functions that have a number of extrema (both maxima and minima) and zero-crossings that must be equal, or differ at most by one. The second condition is that the mean value of the envelope defined by the local maxima and the local minima should be zero at an arbitrary point on the IMF.

The process of EMD [_{1}_{1}_{1}

If _{1}_{1}

If the sifting process is successfully completed after the k-_{1k}

Or equivalently:

Again, as _{1}_{1} from the original _{1}_{2}

Eventually, the decomposition process ends when _{n} becomes a monotonic function (contains no more than two extrema), so that no more IMF can be deduced.

Thus, we found _{n}. Here, _{n} represents the mean trend of _{1}, c_{2}_{n}

This section describes an experimental setup and test procedure for a bearing monitoring system. A NJ202ECP roller bearing from SKF was used for the test, as shown in

Again, we measured the vibration signals from accelerometers mounted on the top of the bearing housing. Signals from the bearing with and without defect are compared, for verifying the fault detection methods. ^{14} data points (50 kHz sampling frequency), and we applied algorithms of MED and the TKEO on the segment. In total, five cases are considered; Case I for the original signal, Case II for only MED, and Case III for only the TKEO. We also connected two algorithms in series for Case IV, and switched the sequence for Case V (see

Having finished MED/TKEO-based signal processing, the results of Cases II–V are shown in

It is apparent that applying MED and the TKEO will significantly increase the kurtosis for the segments from damaged bearings, compared to healthy ones. Interestingly, the MED-TKEO provides a better separation capability than the TKEO-MED process. However, it needs to be also noted that the MED and TKEO processes are not very effective for D2 and D3, as shown in

In the previous section, we compared and discussed the effectiveness of MED and TKEO for diagnosing scratch-type defects on inner-race bearings. Here, we introduce a bearing diagnosis algorithm that combines empirical mode decomposition (EMD) with a genetic algorithm (GA), to maximize the damage-sensitive feature or kurtosis level of given bearing signal. In this study, we used EMD-GA to improve the detectability of bearing faults.

First, we calculated the IMFs of eight segments of signal,

This study presents a comparative study on the condition monitoring of roller bearings through signal processing and optimization techniques. Although it is widely known that the kurtosis values of a bearing with a defect are higher than those of signals from healthy bearings, in many cases the difference is not very obvious. This study suggests and compares two different signal processing techniques (MED and the TKEO), and their combinations, to enhance the resolution of kurtosis, for differentiating the condition of roller-bearing in terms of kurtosis. Experimental results indicate that combining MED and the TKEO successfully improves the resolution of kurtosis for scratch-type defects on the surface of the inner-race of bearings. Also, each segment of the bearing vibration signals can be decomposed into a linear combination of IMFs, using EMD. We employed a GA to optimize the weights of IMFs, to reconstruct the modified segment, providing improved kurtosis sensitivity toward bearing signals with defects on the inner-race. The laboratory study found that the EMD-GA effectively increased the kurtosis sensitivity up to six times on a damaged bearing, while only 40% growth was witnessed on the segment from a healthy bearing. Apparently, the performance of kurtosis-based fault detection for roller bearings can be significantly enhanced by proper selection of data processing and optimization techniques.

This work was supported by the Basic Science Research Program, through the National Research Foundation of Korea (NRF), funded by the Ministry of Education, Science and Technology (2012R1A1A2003787).

The authors declare no conflict of interest.

A systematic block diagram to illustrate the process of minimum entropy deconvolution (MED).

Acceleration measurements of a roller bearing with (

(

(

Vibration data points of a roller-bearing with defect (D1).

Vibration data points of a roller-bearing in healthy condition (H1).

Vibration data points after applying only MED signal processing on a damaged bearing (Case II).

Vibration data points after applying only TKEO signal processing on a damaged bearing (Case III).

Vibration data points after applying MED first, and then TKEO signal processing on a damaged bearing (Case IV).

Vibration data points after applying the TKEO first, and then MED signal processing on a damaged bearing (Case V).

Comparison of kurtosis values after applying different signal processing cases (I–V) on two healthy (H1, H2), and four damaged (D1–D4) bearing data sets.

A schematic block diagram for the process of sensitivity enhancement of a bearing defect through EMD-GA.

Comparison of the convergence rates of EMD-GA on a single healthy (H1–H4), and four damaged bearing cases (D1–D4).

Overall comparison of kurtosis values before and after applying EMD-GA on the bearing signals.

Five different Cases (I–V) of signal processing before calculating kurtosis.

Case | Signal Processing |
---|---|

I | Bearing signal → Kurtosis |

II | Bearing signal → MED → Kurtosis |

III | Bearing signal → TKEO → Kurtosis |

IV | Bearing signal → MED → TKEO → Kurtosis |

V | Bearing signal → TKEO → MED → Kurtosis |