Predictive and Standalone Fault Diagnosis System for Induction Motors

Sudden faults created in induction motors result in catastrophic failures and loss of production. Therefore, the industry is in need of a predictive based system that can identify developing faults in advance. Condition monitoring is used as the general method of identifying faults and taking measures before the dreadful situation. However, there is limited work done on the predictive methodologies based on the trend analysis. The study presented in this paper proposes a novel method that identifies trend variation of critical harmonics of the vibration spectrum with increasing fault severity for frequent mechanical faults; structural looseness, misalignment, bearing eccentricity and bearing inner race fault. Faults were artificially induced on a three-phase induction motor and vibration data obtained was analysed with a MATLAB based algorithm.


Introduction
Induction motors are the most commonly encountered electrical machines in the industry due to their self-starting features, simplicity, reliability, and cost-effectiveness. In failure, it can lead to severe losses due to both, costs of repair or replacement of motor and loss of production. Hence condition monitoring is a critical process for the industry. The wellestablished approaches for condition monitoring can be classified based on the diagnostic signal such as vibration [1], [2], acoustic [3], thermal [4], motor current [5], [6]. Among them, the vibration signature is the mostly implemented methodology for observing motor malfunctions as the unique vibration patterns assist to identify the root cause of each fault.
Vibration based analysis is the primary technique for predictive maintenance in the industry [7]. Vibration fault simulation systems were implemented to observe vibration patterns before applying the concept in real life [8]. Obtaining vibration RMS (Root Mean Square) is the most traditional way of identifying faults in an induction motor [9]. As a more advanced method, routine tests for motor vibration based on vibration time wave, frequency spectrum, and phase analysis are performed once in a few months. Apart from this, deep learning based vibration analysis techniques have been developed and implemented in the industry for online monitoring [10]. Yet, there is a need for increasing the accuracy of fault diagnosis and correct prediction of them with an alternative and advanced monitoring system such as trend analysis. When considering vibration, only a few developed techniques can be found on continuous monitoring with trend analysis which is imperative for an industrial perspective [11]. Some comprehensive researches have already revealed that the RMS value of the acquired motor vibration signal can be trended throughout the operating time and early-stage fault detection can be implemented [12]. Considering the easiness of integration with existing machinery and technologies, IoT (Internet of Things) techniques are used for condition monitoring based on both scheduled intervals [13] and real time condition monitoring [14], taking account of the RMS value of the vibration spectrums of induction motors. In addition, there exists wireless vibration monitoring systems instead of wired systems for real time measurements of motors [15].
Moving forward, in this research we have identified the trend variations of each critical harmonics of the vibration spectrum for artificially induced faults while scaling up the fault severity. Through that, a novel method can be developed to identify and predict induction motor-based faults.
Here, vibration data is analysed, and trends of variations of harmonics of vibration spectrums are plotted for following faults: 1. Structural looseness 2. Angular misalignment 3. Bearing eccentricity

Bearing inner race faults
Experiments for the first three faults were carried out by inducing faults on the developed lab setups and, for bearing inner race faults, a publicly available set of data is used.

Proposed Monitoring System
The principle proposed in this research is to capture the growing trend of a motor fault through measuring and online processing of the monitored information based on RMS and critical harmonics values in order to achieve fault identification and prognosis. It includes those particular faults occurring due to the electrical and mechanical stresses which are detectable with this concept. The accuracy and consistent behaviour of this monitoring concept is accomplished through the integration of software and low cost-highly accurate sensors.

Figure 1-Data Acquisition System
Data acquisition was done through a Raspberry-Pi type microcontroller. For vibration data, ADXL345 accelerometer was used at a sample rate of 3200Hz and a sensitivity of 4mg/LSB. In the setup, two vibration sensors were used for obtaining vibration in axial and radial directions compared with the motor shaft. Further, as a supportive measurement for harmonic identification, motor speed was measured with KY-032 IR sensor. In order to evaluate whether a trend variation can be observed for supply current harmonics, ACS712 Hall-effect current sensor was used with an ADC. Here, Python was used as the programming language for preparing the data acquisition system. The block diagram of data acquisition system is shown in Figure 1. Figure 2 shows the placement of the sensors in the proposed monitoring system.

3.2
Data Analysing Process It is found through previous works that, for different ranges of expected fault frequencies, vibration measurement technique varies [7]. If we expect to see a spectral line below 10Hz the most effective way is to take the vibration in displacement (m) before taking the frequency spectrum. For most of the induction motor faults, these fault frequencies lie between 10-1000Hz where the velocity spectrum is much more effective.
In this data acquisition system, an accelerometer sensor was used for vibration measurement. Thus, it is essential to take the velocity by integrating it before applying fast Fourier transform. Further, the signal is passed through a low pass filter and a high pass filter. As the amplitudes of harmonics obtained from the vibration spectrum at different occasions are compared, it is essential to eliminate spectral leakage applying to them. So, a windowing function (Hanning Window) is also used. Figure 3 illustrates the block diagram for signal analysis process of the proposed monitoring system. Data analysing process was done in the micro-controller, programmed with MATLAB. In making the standalone system, the algorithm was converted to a standalone executable and was deployed on the Raspberry Pi microcontroller.

Fault Analysis
As mentioned earlier, for the four different faults that are simulated in this experiment procedure, different experiment setups and analysing processes were used to create a fault and increase its severity. For each fault, the methodology followed, and the results obtained are discussed below.

Structural Looseness
Having bolted joints at the base of a motor causes unintentional loosening as a result of its operational conditions of industries. Therefore, this phenomenon should be identified before it gets worse which can lead to several other severe faults. In the frequency spectrum, looseness can take various forms that are created internally or externally [16]. The most common form of looseness could start with an increase in the amplitude of harmonics of vibration. Figure 4 shows the three-phase induction motor used to create structural looseness by loosening the above highlighted four bolts which are tightening the motor to the base. For all these scenarios, the frequency spectrum was obtained while increasing the looseness at an equal phase. Data were obtained in the radial direction at both the drive end and fan end of the motor. Then the amplitudes of the first three harmonics of motor running speeds were plotted against the degree of looseness. Figure 5 indicates the vibration velocity time signal and frequency spectrums for healthy condition and the worst conditions of looseness respectively, at 1500rpm. A clear increase is observed in the first five harmonics of the vibration spectrum for the faulty motor. As previously mentioned, in order to critically evaluate the looseness, four scenarios were considered as above and the variations of the

Misalignment
Induction motors which involve industrial processes typically experience misalignment due to various factors such as improper assembly of machines, thermal distortion of machines, asymmetry in applied loads, unequal settlement of the foundation, etc. [17]. For some reason, ultimately, there is an incorrect positioning along the shaft centreline between bearings when two machines are coupled. Types and degrees of misalignment basically depend on coupling type, shaft size, and speed, etc.

Experimental Setup
Characteristics of angular misalignment was observed using following experimental framework ( Figure 6).

6.2
Methodology Experiment tests for misalignment fault were conducted in a laboratory environment using a two-pole, 5.5kW, 1500 rpm three phase induction motor. Initially, the shaft of the motor was aligned with the load. Then the shaft was misaligned by 0.15 mm stepwise and axial vibration was measured. The speed of the motor was kept at 1500 rpm and the coupling type was three-jaw coupling.

6.3
Results and Discussion Frequency time-domain signals for velocity and acceleration for aligned and misaligned conditions are shown in Figure 7. For correctly aligned condition, frequency velocity spectrum only shows the fundamental harmonic. However, 1x, 2x and 3x harmonics become dominant when the fault is dominant and gives the highest contribution to the vibration of the motor. Figures 21, 22 and 23 show the patterns of variation for each 1x, 2x and 3x harmonics, respectively. Although all three are increasing, the rates of growth for 1 st and 2 nd are decelerating and 3 rd is accelerating.

Bearing Looseness
Bearing eccentricity is a common occurrence due to bearing wear and shaft looseness. It results in an unbalanced force distribution around the bearing. Not attending to this can lead the motor to complete bearing failure or even a winding failure.

Experimental Setup
In order to simulate a bearing looseness, a separate setup (Figure 8) was developed to exert a force on the shaft which has an induced lump to simulate the bearing wear.
As mentioned, in a bearing looseness, the force is not distributed equally around the shaft. By exerting an external force on the lump created, the fault can be simulated. The setup was developed in a way that it has a rotating element (rubber roller with a bearing) which engages the motor shaft, with the other end

Figure 7-Velocity and Acceleration Vibration for Healthy Condition and Misaligned Condition
fixed to a shock absorber and hence a gradual increase of the force can be applied to the lump. A 5.5 kW rated motor was used here. Figure 24 shows the velocity vibration of time domain and frequency spectrum for bearing looseness, respectively.

Results and Discussion
An increase in the first harmonic of the spectrum was expected and it was observed. But an unintentional increase in the 2 nd and 3 rd was also observed. It was observed while the bearing looseness is being simulated, due to the push exerted at one side from the coupling point, a parallel misalignment was induced. And the features of a parallel misalignment are the increase in 2 nd and 3 rd harmonics. Figures 25, 26 and 27 illustrate 1 st , 2 nd, and 3 rd harmonic variation, respectively, while exerted force is increasing stepwise.

Inner Race Bearing Fault
Bearing failure is considered one of the most common problems for machinery failure. A bearing can fail due to a fault induced at the inner race, outer race and the ball [18].
For bearing inner race faults the critical frequency which shows the spectral line can be calculated from equation (1). Although velocity spectrums are used for looseness and misalignment detection, as BPFI is a considerably large frequency, the spectrum is obtained by directly applying fast fourier transform for acceleration data.

Methodology
To evaluate how the bearing fault affects the trends of harmonics of the vibration spectrum, external data obtained from a wind turbine a developing inner race fault is used. It is a 2 MW wind turbine and data are obtained for 50 consecutive days. It has a nominal speed of 1800 RPM fluctuating over the days and these fluctuations are also considered in harmonic detection.
The calculated BPFI value for this specific case is 280Hz. Further, in the case of inner race faults, it is a common thing to observe two sidebands around the BPFI frequency at running speed. Hence, two sidebands can be observed at 250Hz and at 310Hz. Here, in order to identify how the trend varies with increasing severity, BPFI and its sideband amplitudes are plotted against time.

Results and Observations
The following plots show the two different vibration spectrums at day 1 ( Figure 28) and day 50 (Figure 29). It can be clearly identified that there is a huge spike at BPFI and considerably large sidebands around them. 2*BPFI and 3*BPFI are available as well which indicates high severity in Inner race fault.
The trend of variation for BPFI and its sidebands are plotted in Figure 30 for the 50day period.

Conclusions
This research mainly focuses on developing the methodology for condition monitoring based trend analysis. It was found that the proposed method has worked successfully when faults were experimentally tested on actual machines. For all the faults considered in this study, experimental results show that there are specific trend variations of critical harmonics and RMS values of vibration when the severity of faults is increasing. Current harmonic spectrums did not give a clear outcome of a trend at a tested level of the developing each fault. Thus, with continuous monitoring, this trend variation of vibration can be interpreted as slow development of the fault and be alarmed accordingly.
In this way it is possible to create a standalone system powered by a microcontroller to monitor the sensor outputs and identify the trends of variations of harmonics. Then these trends can be used to predict a time of failure in advance. In the future, the proposed concept can be implemented for more fault conditions in induction motors by measuring current and temperature together with vibration.