Classification of Misalignment and Unbalance Faults Based on Vibration analysis and KNN Classifier Ashkan Nejadpak, Student Member, IEEE, Cai Xia Yang*, Member, IEEE Mechanical Engineering Department, University of North Dakota Grand Forks, ND ashkan.nejadpak@und.edu; caixia.yang@engr.und.edu Abstract In this paper, we present a fault detection and classification technique for shaft misalignment and rotor imbalance which are common problems in rotating machinery based on vibration analysis and K-Nearest Neighbor (KNN) classifier. The vibration of the rotary machine was measured from both bearing housings along radial (vertical and horizontal) directions by four industrial accelerometers while machinery is in operating. Measured signals were acquired using NI PCI- 4498 and were further analyzed and processed using MATLAB program. Three operation conditions, i.e., health, unbalance and misalignment are simulated on a machine fault simulator and repeated six times for each situation with varied operation speed. Amplitude of acceleration at operating speed, and multi times of operating speed were selected as vibration signatures/features for distinguish different operation conditions. The training data set was generated based on sixteen cases, and the operation conditions were grouped into three categories. Five testing measured data sets were analyzed using proposed KNN classifier and over 95% accuracy rate was achieved which demonstrated the proposed technique is reliable. Index Terms Fault detection and classification, k-nearest neighbor rule, rotating machinery. I. INTRODUCTION Imbalance, misalignment of two mating shafts, wear or failure of some components induced cyclic excitation forces will cause vibration of the machine. They are the most common cases of machinery malfunction [1-3]. Since different mechanical faults generate different signatures in measured vibration data, and vibration retains a unique signature which shows how the system is responding to its operating conditions, vibration analysis is capable and be widely used for machine condition monitoring to detect fault, determine the faulty type, and severity of faulty systems. By comparing the signals of a machine running in normal and faulty conditions, the detection of faults is possible. Fault detection is the common practice of identifying machine faults before they become a problem. Effective fault detection techniques can help industries reduce maintenance expense, increase equipment reliability and uptime, and reduce the usage of expensive test equipment [4-6]. Condition based predictive maintenance (PdM) is proved reliable, cost-effective technique for monitoring and diagnosing machinery faults. Conducting controlled experiments help to gain an in-depth understanding of different vibration signatures and to recognize the signatures of various machine faults. Vibration monitoring techniques are classified as time-domain analysis, frequency-domain analysis and quefrency-domain analysis. There have been several attempts to use supervised or unsupervised machine learning approaches for fault diagnoses of bearing with statistical and frequency based features [6-12]. Principal Component Analysis (APC) [6-8] and KNN method [9-12] have been used to fault classification. However, PCA makes assumption of the linearity of the data set, while KNN method is a nonlinear classifier and makes no assumption of the linearity of the data set. Using a simulation example which was generated from a nonlinear process model, He and Wang [9,1] demonstrated that nonlinearity has a negative effect on PCA fault detection effectiveness, and KNN is capable for detecting all fault samples in their research. In this paper, KNN as a non-parametric supervised machine learning technique is performed for solving data mining problems for fault classification of misalignment and unbalance. Proposed KNN algorithm is based on Euclidean distance in same class, hence error due to improper selection of nearest neighbor K can be compensated. Experimental results show potential improvement in classification with modified KNN and optimized feature selection. Unbalance and misalignment of rotating shafts are common cause of vibration in a rotating machinery. In this work, experimental investigation is conducted on a machinery fault simulator. K nearest neighbor (KNN) technique is used for identifying unbalance and misalignment faults from normal operation condition. The training dataset for KNN is generated through controlled predetermined amount of unbalanced mass and misalignment at different levels, as well as ideal situation. Several key features are selected based on results of spectrum analysis of measured vibration signals from ideal, unbalance and misaligned conditions. To improve classification accuracy, features selected from training dataset were weighted according to their influence levels on distinguishing fault types. The classification results shows that using proposed weighted KNN approach, improved effective and accurate of detecting and classifying for imbalanced and misaligned operation condition at early stage are achieved. The remainder of this paper is organized as follows: After a brief discussion of the fundamentals of fault detection and knn
method in Section II, the experiment setup, experimental system faults and data collection process are described in section III. Frequency domain vibration analysis results based on measure from Individual faults are generated and tested, and the frequency analysis of the effects of are provided in Section IV. Finally, the paper concludes with a summary recapping the main advantages of the proposed method and future work in Section V. II. EFFECTS OF UNBALANCE AND MISALIGNMENT AND KNN METHOD The goal of this study was to find evidence regarding vibration patterns associated with specific misalignment and unbalance faults. The severity and type of each fault condition can be assessed based on the amplitudes of the corresponding peaks as well as their respective locations on the frequency spectrum. Additionally, certain types of faults can be determined based on the location where data was recorded. Because some faults display a higher level of severity if the accelerometer was placed on various locations of the equipment. To determine the vibration spectra and signatures that occur due to different levels of unbalance and parallel misalignment, multiple controlled experiments were conducted on a machinery fault simulator. In pattern recognition and classification, the knn method classifies a new unknown sample by examining its distances to the K nearest neighboring training samples in the feature space [9-12]. This method has been used in statistical estimation and pattern recognition since 197 s, this algorithm stores all available cases as training data set, then classifies new cases based on a similarity measure by a majority vote of its neighbors, with the case being assigned to the class most common amongst its k nearest neighbors. Different distance functions could be used for the similarity measure. In the paper, the Euclidean distance is used. III. EXPERIMENT SETUP In a condition monitoring system, acquire and process the data related to the condition of the machine is the most important stage. In this research, three operation conditions were studied in a series of experiments including normal condition, unbalanced load on one disk, and parallel misalignment of shaft with amount of 5 mils and 1 mils (1 mil is.1inches) were studied. The image of experimental setup is shown in Figure 1. This setup consists of a three phase, 4 poles induction motor, an AC drive for the motor, an amplifier, an analog to digital converter (ADC), four accelerometers located at two bearing housing along vertical (Z axis) and radial (Y axis) directions. The output data was analyzed using a NI-PXI 4498 module on the data acquisition board. Each of these experiments were conducted using the machinery fault simulator operating at 2Hz (12 RPM) and 25Hz (15 RPM). The vibration data acquisition is performed using a National Instrument PCI-4498. The data acquisition software is developed using Sound and Vibration suit. The data is further graphed using an FFT in MATLAB. In order to make sure that only steady-state of the signal was used, a Hanning window was selected to the acquired time-domain signal for fast Fourier Transform (FFT). Fig. 1. Experiment setup IV. FREQUENCY ANALYSIS It has been observed that the vibration amplitude under abnormal operating conditions are comparable to the normal operating condition. Additionally, faults caused by unbalance and misalignment can be detected and classified through the time and frequency domain analysis of the vibration signals. In order to characterize the effects of unbalanced load torque and misalignment on the frequency profile of the rotating machinery vibration, the following tests were performed when motor was operated at 12 RPM and 15 RPM: a. Health condition: Two disks on the shaft are perfectly balanced, shafts are accurately aligned. The response of vibration under health/normal operating condition is shown in Figure 2. For the figure, one can observe that the highest amplitude occurs at 1x RPM..5.45.4.35.3.25.2.15.1.5 Health Fig. 2. Vibration response under health operating condition
b. Indicator of unbalance operation condition Unbalance is the un-equal distribution of weight around the center of rotation. To study the unbalanced rotor condition, a screw with nut was mounted on different angles and locations of one rotor. This unbalanced mass generated unbalancing force induced an increased vibration, which shows as a large amplitude peak at 1x RPM. In the unbalanced condition, the disk is imposed to an increased force of: F U = m U e ω 2 (1) Where F U denotes the unbalanced force, m U denotes the unbalanced mass, e represents the distance from the unbalanced mass to the center of rotation, and ω is the running speed of the rotor. From Eq. (1), the amplitude of vibrations would increase with the operating speed of the motor, and the measured signal from the accelerometer and current transducer validate this statement..35.3.25.2.15.1.5 Unbalance Fig. 3. Vibration response under unbalance operating condition Unbalanced rotors will have a large response amplitude peak when operation frequency is the natural frequency of the system, this peak can be observed from frequency domain by transform vibration signal from time domain using Fast Fourier Transform (FFT). Comparing unbalance condition with health operating condition in spectrum displacement, one can observe that the amplitudes of the acceleration at the exact operation speed (1X), three times of operation speed (3X) and seven times of operation speed (7X) have influence on classification on machine operation conditions. However, the most dominant difference happened at 1X, therefore, the amplitude value at 1X would be the relevant indicator to distinguish unbalance condition from health condition, the fault happened in system will be detected, and be identified as unbalance fault. c. Indicator of misalignment operation condition Experiment on parallel misalignments at amount of 5 mil and 1 mil were conducted, vibration signal measured by four accelerometers were recorded and analysis. The amplitude vs frequency is shown in Figure 4..45.4.35.3.25.2.15.1.5 Misallignment Fig. 4. Vibration response under misalignment operating condition Comparing misalignment condition with health condition in spectrum displacement, one can observe that the amplitudes of the acceleration at multi-times of operating speed (6X, 7X and 8X) have influence on classification on machine operation conditions for misalignment. However, the most dominant difference happened at 6X. Therefore, the amplitude value at 6X would be the relevant indicator to distinguish misalignment condition from health condition. One limitation on misalignment detection is the impossibility to detect the severity of misalignment using vibration analysis, because the relation between the amount of misalignment and the level/amplitude of vibration has not been found. d. Generation of training dataset Features used for the classification or recognition task will play an important role in the choice of the classifier. For a specific set of features a classifier may perform better, but for another set, it might not provide the best solution. The training dataset was generated from all features selected from each experiment. The amplitudes of acceleration at 1x RPM, 2x RPM, until 6x RPM are evaluated and imported to the training dataset. K=3 is closed as the number of nearest neighbors, Euclidean distance of these three nearest neighbor to each sample from the training dataset were calculated. Given arbitrary testing sample, the proposed KNN algorithm provide fast and accurate classification based on majority vote rule. The classification results was shown in Table 1. V. DISCUSSION AND FUTURE WORK This paper presents a method for detecting and classifying unbalance and horizontal parallel misalignment faults for rotating machinery based on analysis of measured vibration signals using KNN technique from health condition. Operations under normal/health and fault conditions have been studied. It was shown that using the vibration profile, different level of
faults can be detected and classified. The result of this evaluation will be used to classify the signature of fault on the machine vibration. This can be used to acquire a complete machine vibration profile that may predict occurrence of damage in different parts of machine. This methodology can help to detect and identify the defective machineries at early stages. Although KNN is not a new technique, to the best of our knowledge, there has not been reported in literature on detection and classification of unbalance and misalignment faults in rotary machine. Detection and identification of combined fault situations will be our future research direction. ACKNOWLEDGMENT The authors would like to thank Tristan Plante and Santosh Paudyal for conducted part of experiment. The work reported in this paper was funded by ND EPSCoR New Faculty Start-up Award 437-2725-UND1985 and ND EPSCoR Advanced Undergraduate Research Award. REFERENCES [1] SpectraQuest, Inc., Applied vibration analysis training manual & laboratory exercises, Richmond: SpectraQuest, Inc. [2] S.S. Rao, Mechanical Vibrations, 6th Ed., Pearson Prentice Hall, 211 [3] Spectra Quest, InC., Machinery Fault Simulator Operation Manual. [4] Q.H. Alsafasfeh, Pattern Recognition for Fault Detection, Classification, and Localization in Electrical Power Systems, (21). Dissertations. Paper 494. [5] G. Verdier and A. Ferreira. (21).Adaptive Mahalanobis distance and k-nearest neighbor rule for fault detection in semiconductor manufacturing. IEEE Trans. Semiconductor Manufacturing. Available:doi:1.119/TSM.21.265531 [6] J. Starmer, StatQuest: Principle Component Analysis (PCA) clearly explained, 215. [7] E. L. Russell, L. H. Chiang and R. D. Braatz, Fault detection in industrial processes using canonical variate analysis and dynamic principal component analysis, Elsevier, 2. [8] T. Plante, L. Stanley, A. Nejadpak, and C. Yang, Vibration Based Fault Detection using Principal Component Analysis, Proceedings of IEEE AUTOTESTCON 216 [9] Q.P. He and J. Wang, Fault detection using the k-nearest neighbor rule for semiconductor manufacturing processes, IEEE Trans. Semiconductor Manufacturing, vol.2, no.4, pp.345-354, Nov. 27 [1] J.M. Johnson and A. Yadav, Fault detection and classification technique for HVDC transmission lines using KNN, International Conference on ICT for Sustainable Development ICT4SD 216 [11] Q. He and J. Wang, Fault Detection Using the k-nearest Neighbor Rule for Semiconductor Manufacturing Processes, IEEE Transactions on Semiconductor Manufacturing ( Volume: 2, Issue: 4, Nov. 27 ) Page(s): 345 354, DOI: 1.119/TSM.27.9767 [12] D.H. Pandya, S.H. Upadhyay, S.P. Harsha, Fault diagnosis of rolling element bearing with intrinsic mode function of acoustic emission data using APF-KNN, Expert Systems with Applications 4 (213) 4137 4145.
Table 1 K NN for Fault Detection and Classification Feature1 Feature2 Feature3 Feature4 Feature5 Feature6 Situation Rank Euclidean Distance Category.724.34.48.22.14.2 Unbalance 13.316 Unbalance.453.33.24.41.12.33 Misalignment 6.58 Misalignment.463.41.48.28.13.22 Health 7.64 Health.1129.28.39.27.13.24 Unbalance 11.719 Unbalance.44.52.22.34.2.14 Misalignment 5.44 Misalignment.461.41.45.25.9.22 Health 5.64 Health.1517.25.47.24.15.24 Unbalance 9.117 Unbalance.432.65.35.52.12.34 Misalignment 3.31 Misalignment.465.41.49.21.12.21 Health 5.68 Health.248.1.46.23.17.22 Unbalance 7.1638 Unbalance.429.61.48.53.17.45 Misalignment 2.3 Misalignment.465.41.51.32.15.21 Health 3.66 Health.2423.28.56.22.21.24 Unbalance 5.213 Unbalance.419.65.37.5.2.67 Misalignment 2.34 Misalignment.474.42.49.23.1.21 Health 2.75 Health.2734.47.66.35.18.26 Unbalance 3.2323 Unbalance.411.67.37.37.29.63 Misalignment 1.25 Misalignment.57.41.52.25.11.22 Health 1.14 Health New measured vibration data available, need to classify the operation condition..411.67.37.37.29.38? K Category Classification based on k = 3. 1 Misalignment Operation condition is Misalignment. 2 Misalignment 3 Misalignment