key: cord-102238-g6dsnhmm authors: Wescoat, Ethan; Mears, Laine; Goodnough, Josh; Sims, John title: Frequency Energy Analysis in Detecting Rolling Bearing Faults date: 2020-12-31 journal: Procedia Manufacturing DOI: 10.1016/j.promfg.2020.05.137 sha: doc_id: 102238 cord_uid: g6dsnhmm Abstract Component failure analysis is sometimes difficult to directly detect due to the complexity of an operating system configuration. Raw time series data is not enough in some cases to understand the type of fault or how it is progressing. The conversion of data from the time domain to the frequency domain assists researchers in making a more discernible difference for detecting failures, but depending on the manufacturing equipment type and complexity, there is still a possibility for inaccurate results. This research explores a method of classifying rolling bearing faults utilizing the total energy gathered from the Power Spectral Density (PSD) of a Fast Fourier Transform (FFT). Using a spectrogram over an entire process cycle, the PSD is swept through time and the total energy is computed and plotted over the periodic machine cycle. Comparing with a baseline set of data, classification patterns emerge, giving an indication of the type of fault, when a fault begins and how the fault progresses. There is a separable difference in each type of fault and a measurable change in the distribution of accumulated damage over time. A roller bearing is used as a validating component, due to the known types of faults and their classifications. Traditional methods are used for comparison and the method verified using experimental and industrial applications. Future application is justified for more complex and not so well-understood systems. As manufacturing products become more complex, the machinery that makes the products has increased in complexity. With an increase in machinery and equipment, the maintenance required takes more time and costs more to a company. This in turn makes it harder to identify failures on equipment before they occur. Untimely equipment failure has one of the biggest impacts on the operating costs for a manufacturer with costs in some cases exceeding 100k€ [1] . All aspects of the manufacturing line are susceptible to failure, whether human or machine. It only takes one component to fail for a stoppage to occur in production. To avoid this, manufacturers and maintenance staff are developing and using analysis tools and methods to predict when component and equipment failures will occur. By knowing when failures will occur, maintenance is scheduled to repair or tend to equipment before untimely failure occurs. Different tools of analysis have been proposed as cost effective for manufacturing maintenance and effective in determining the remaining lifetime of equipment components. For these different tools, the forms of analysis require different types of data to inform the user on subsequent needs. As an example, data scientists and researchers have used data features such as mean and variance of raw time series data to justify equipment analysis and maintenance scheduling. These methods perform well with an isolated component, but when incorporated into manufacturing equipment with multiple different components, the signal is susceptible to misinterpretation and outside interference from other components. System level predictive maintenance and modeling is extending beyond the capability of such conventional and customary analyses. Time series data is convertible to frequency data using a variety of different types of frequency transforms, notably the Fourier Transform and Wavelet Transform. From the frequency spectrum, different features can be extracted such as power and phase with respect to a range of frequencies. For components in a system, there are corresponding nominal characteristic frequencies [2] . When there are deviations from these expected frequencies, then a possible defect to the As manufacturing products become more complex, the machinery that makes the products has increased in complexity. With an increase in machinery and equipment, the maintenance required takes more time and costs more to a company. This in turn makes it harder to identify failures on equipment before they occur. Untimely equipment failure has one of the biggest impacts on the operating costs for a manufacturer with costs in some cases exceeding 100k€ [1] . All aspects of the manufacturing line are susceptible to failure, whether human or machine. It only takes one component to fail for a stoppage to occur in production. To avoid this, manufacturers and maintenance staff are developing and using analysis tools and methods to predict when component and equipment failures will occur. By knowing when failures will occur, maintenance is scheduled to repair or tend to equipment before untimely failure occurs. Different tools of analysis have been proposed as cost effective for manufacturing maintenance and effective in determining the remaining lifetime of equipment components. For these different tools, the forms of analysis require different types of data to inform the user on subsequent needs. As an example, data scientists and researchers have used data features such as mean and variance of raw time series data to justify equipment analysis and maintenance scheduling. These methods perform well with an isolated component, but when incorporated into manufacturing equipment with multiple different components, the signal is susceptible to misinterpretation and outside interference from other components. System level predictive maintenance and modeling is extending beyond the capability of such conventional and customary analyses. Time series data is convertible to frequency data using a variety of different types of frequency transforms, notably the Fourier Transform and Wavelet Transform. From the frequency spectrum, different features can be extracted such as power and phase with respect to a range of frequencies. For components in a system, there are corresponding nominal characteristic frequencies [2] . When there are deviations from these expected frequencies, then a possible defect to the As manufacturing products become more complex, the machinery that makes the products has increased in complexity. With an increase in machinery and equipment, the maintenance required takes more time and costs more to a company. This in turn makes it harder to identify failures on equipment before they occur. Untimely equipment failure has one of the biggest impacts on the operating costs for a manufacturer with costs in some cases exceeding 100k€ [1] . All aspects of the manufacturing line are susceptible to failure, whether human or machine. It only takes one component to fail for a stoppage to occur in production. To avoid this, manufacturers and maintenance staff are developing and using analysis tools and methods to predict when component and equipment failures will occur. By knowing when failures will occur, maintenance is scheduled to repair or tend to equipment before untimely failure occurs. Different tools of analysis have been proposed as cost effective for manufacturing maintenance and effective in determining the remaining lifetime of equipment components. For these different tools, the forms of analysis require different types of data to inform the user on subsequent needs. As an example, data scientists and researchers have used data features such as mean and variance of raw time series data to justify equipment analysis and maintenance scheduling. These methods perform well with an isolated component, but when incorporated into manufacturing equipment with multiple different components, the signal is susceptible to misinterpretation and outside interference from other components. System level predictive maintenance and modeling is extending beyond the capability of such conventional and customary analyses. Time series data is convertible to frequency data using a variety of different types of frequency transforms, notably the Fourier Transform and Wavelet Transform. From the frequency spectrum, different features can be extracted such as power and phase with respect to a range of frequencies. For components in a system, there are corresponding nominal characteristic frequencies [2] . When there are deviations from these expected frequencies, then a possible defect to the 48th SME North American Manufacturing Research Conference, NAMRC 48 (Cancelled due to component may have been identified. However, there is still a high degree of variation in the frequency spectrum associated with observing an FFT output, in terms of deviation from the expected frequencies and power values for similar data. Wavelet transforms can offer results that incorporate both steady-state and transient information, but there are inherent restrictions in application of this method in systems with bursts in data or non-stationarity [3] . However, the authors have proposed a new method of interpreting frequency domain results using the energy content of the signal over time. Two different applications are used for demonstration. The first is a bearing test station for observing the energy changes with different types of induced faults. The second application is a vertical vehicle lift at an automotive manufacturing facility. The authors' proposed method involves calculating the energy content from the PSD over the course of a defined test process, as gathered from a spectrogram at each time step of a discrete test process. For the experimental validation, baseline data and fault data are generated at a component test stand. For the vertical lift applications, a system with one good bearing and one faulty bearing was chosen. The energy content over time is assessed to evaluate changes and their associated patterns. For further validation, a distribution of points is created to help visualize potential changes to the data for the bearing station. In the remainder of this paper, the literature review first covers a brief overview of predictive maintenance and analysis methods relating to vibration and FFTs. The next section covers the FFT and different types of bearing failures. The methodology is described as an overview in terms of how data is gathered and a more in-depth description of how the analysis is performed. The results cover the initial findings employing the method, and conclusions over the larger impacts as well as future work are presented. As manufacturing equipment complexity increases, maintenance teams move from corrective to predictive maintenance. Figure 1 below shows a representation of the timeline of the different types of maintenance and when they were introduced [4] . Predictive maintenance changes the state of maintenance from anticipation of faults with preventive maintenance to knowing when machine faults will occur. For predictive maintenance to occur for machines, sources of data are gathered from equipment and test stands. From this data, different data features are created to classify potential faults. In 1985, J.T. Renwick et al. wrote on how vibration data is a proven technique for predictive maintenance due to ease of implementation and cost [5] . In 2000, M. Lebold created a review of different data features in vibration data for classification such as kurtosis [6] . These terms also appear in other vibration literature. Liu et al. made a list of data features that he used in his analysis and noted that as the amount features increased so did the analysis [7] . Further reviews, such as those written by Jardine et al. and Carden et al., deal with the application of these features and techniques to machinery and structural engineering applications, respectively [8] , [9] . However, there is a tipping point referenced when analysis became too computational costly for the value of identifying additional features [10] . Focusing on bearings, D. Dyer et al. proposed a detection method based on kurtosis in 1978 in select frequency ranges as a simple and cheap technique moving away from traditional trend analysis [11] . In 1984 and 1985, P.D. McFadden et al. proposed two models for detecting defects in a rolling element bearing [12] , [13] . In McFadden's analysis, envelope analyses of various frequency spectra are observed and tested with determining identification of fault based on bearing load, geometry and speed. Ian Howard wrote an extensive review in 1995 in rolling element bearing vibration, highlighting envelope analysis, Fourier analysis and wavelet transform. He provided several use cases for frequency techniques as well as highlighting the use of a spectrogram for changing speed applications. He also makes mention of using envelope analysis and recognizes the difficulty in choosing the correct frequency bands for envelope analysis [14] . 2006, Wei Zhou et al. reviewed the different methods for monitoring bearings in electric machines, determining vibration and current analysis as the more popular methods [15] . In 2008, Blodt et al. using current analysis, developed a model to determine bearing damage based on magnetomotive force and observing the PSD [16] . In 2010 and 2012, Lau et al. and Chen et al. used wavelet package transform on current analysis and vibration data, respectively [17] , [18] . For both papers, failure and predictive models are created using the wavelet transform for their respective data sources. As machinery complexity increases, the need to filter out noise increases [19] . If there is too much outside interference in the analysis, the interference can disrupt the analysis by either indicating a fault when it has not occurred or a "hide" a fault until it is too late. Taking machine bearings as an example, the bearings operate under several known frequencies' dependent on process factors: size, speed of rotation and rolling element dimensions. Researchers use a frequency transform for better visualization of bearing frequencies and to note if any differentiations in the data start to happen. Neural networks and Naïve Bayes classifiers are a common form of pattern recognition utilized across many different industries [20] , [21] . In 1992, Liu et al. proposed a neural network for classing failure features in bearings, with a success rate of 100% versus other know techniques [7] . In 2004, Samanta et al. used artificial neural networks and genetic algorithms to detect bearing failures. The classification is based on data features, such as kurtosis and skewness in the time domain [22] . While the algorithm worked well for identifying single failures, the authors ran into problems with regards to multiple failure classification. In 2013, Prieto et al. presented a neural network classification based on manifold learning techniques versus statistical time techniques [23] . Prieto was able to ascertain his classification within 95% and validate his methodology against statistical time techniques. Bayesian Classifiers are commonly found in feature classification and fault detection. In 2015, Mehta et al. developed a Condition-Based Maintenance (CBM) system architecture that investigated spindle damage using Bayesian classification and sensor fusion [24] . From the multiple sources of sensor data, they were able to identify the rising trends of impending failure in spindles beyond single-source classification methods. In 2015 as well, Sharma et al. used sound signal data to compare a Naïve Bayes and a Bayes net classifier on the detection of a faults within a roller bearing. The results were both able to accurately determine the fault and showed comparison of strength in early or late detection of fault. In 2017, Islam et al. used a reliable bearing fault diagnosis involving a combination of Bayesian and multi-class support vector machines for bearing fault classification. This study incorporated the classification of frequency events in addition to time evets [25] . In each of the following papers, one of the main distinctions is the training set of data to classify the results and the separation of data for this training. Neural Networks and Naïve Bayesian Classifiers are very susceptible to misclassification if the training data set is not carefully defined; this motivates offline learning, defined as not connected to the production line [26] . Setting up an experimental stand to generate training data is a means of both validation and training a model in an offline environment. Conversely, online learning takes place directly on the production line; neural networks and genetic algorithms are examples of methods capable of self-classification. However, without parameters carefully set, then there is a chance for misclassification that would require a reset of the model [27] . A positive to both offline and online methods however is the continuation of learning as more data is added to better build the prediction models [17] . Lau et al. made the distinction in his training of offline and online learning for their experiments. Neural Networks and Bayesian Classifiers are adaptable based on new training data. One common item to note from the following papers is the isolation of one type of bearing fault for each classification and model. A continued area of research is the propagation of a multi-class fault classifier, which is a common problem in bearings as wear propagates [2] . As mentioned earlier, frequency ranges and frequency space are a key area in failure classification stretching back to 1978 [11] . In this regard certain transforms are heavily cited in literature in failure classification, notably: Fourier Transform [28] , Hilbert Huang Transform [29] and Wavelet Analysis [30] . Of these techniques, the most cited technique is the wavelet transform. Many papers compare this technique directly with the FFT. Focusing on bearings, Tse et al. compare the two techniques in analyzing bearing faults [31] , finding that both are capable of classification of each bearing fault. Yet, the FFT required more data for the same accuracy faults. Jingling Chen et al. writes about the application of fault diagnosis in wind turbines using the wavelet transform, finding their method valid even in a heavily noisy environment during an application test [32] . Ziwei proposes a method of using wavelets based on the signal-to-noise ratio and mean square error for a roller bearing [33] . The method shows high accuracy and efficiency, as well as accounting for a noisy environment in their model. While good classification results are common from all the papers, a couple of aspects arise; the first is the need for a specific definition of features common across each category. One paper raised this being a potential issue with machine operators, who may not appreciate the frequent interruption to their equipment to make adjustments to sensors or changing the sensors [31] . A need for a more general case of identifying wavelet parameters was also called for. In a review of wavelet transforms by M. Gomez et al., the resolution of the wavelet was raised as a concern [34] , citing better visualization required higher computational resources. In a production setting, directing computational resources may be difficult when analyzing a system where operators or maintenance staff need to make real-time line decisions. An additional advantage of wavelets is the dissemination of time data of when frequency events occur. However, in response to that the authors use a spectrogram to register when time events occur and the corresponding frequencies as raised by Howard [14] . As a means to add further time information to an FFT, Mehla et al. introduced the concept of a windowed FFT or a Short Time Fourier Transform to add time data [35] . The Windowed Short Time Fourier Transform combined aspects of both wavelet analysis and Fourier transform analysis in achieving mixed results compared to either independently. Returning to the FFT, one key problem stems from its inability to accurately predict events over a variable speed or load [36] , [37] , an issue that wavelet analysis is capable of eliminating. The FFT is heavily dependent on the periodic samples of a system, which are in turn affected by the characteristic frequencies of the system. When changing the speed of the shaft in a rotodynamic system or the amount of load mid process, it could misinform the analysis. As a means of investigating this issue for manufacturing, the authors explored other applications of the FFT. FFTs are used in seismology [38] - [40] . One area the FFT has been used is to detect frequency events related to earthquake timing versus other natural events. One paper was interested in understanding the separation of tensile fault and shear fault propagation before an earthquake [35] . The approach they took involved calculating the energy related to the elastic wave triggered by one of the propagated faults. In doing so they can derive indicators and factors to fault displacement of tectonic plates. For medical devices, Karim et al. were involved with studying the change in energy related to the PSD for medical data classification when epilepsy occurs [41] . This application looks primarily at ensuring quick data extraction and speeding up computational time. In the results of the paper, Karim referenced that using energy spectral density for classification did speed up computational time and had higher classification results when coupled with Support Vector Machines over conventional data features. Based on research in seismology and biomedical devices, there is clear indication that the energy spectral density is a valid feature for data classification. One element that has not been broadly explored is the change in energy over time from the FFT. The authors believe this to be an important aspect for manufacturing equipment in seeing how the energy of the signal changes over the course of a process with respect to the PSD. Another aspect is the clear use of the change in the distribution of energy points over time. Haskell made mention to the change in distribution regarding spectral space for calculating the displacement but did not care for the mapping the probability density function line to measure faults [35] . One of the first components tested are bearings due to their well-documented failure modes and the existing literature on bearing failures and how they initiate. Companies such as SKF, Timken and Barden have all released documents detailing the many different types of failures and defects for bearings [42] - [44] . The common one that each document mentions is normal fatigue life or wear, considered unavoidable and once detected will progressively increase until the bearing failure is complete. Another common failure mode is spalling, a stress-induced delamination defect that will spread around the entire bearing surface once initiated. However, what is of more interest for early detection are the causes of premature failure and how they present in the data domain. In the document from Barden [42] , the primary causes of early failure are the load on the bearing, the environmental conditions and geometrical bearing defects. Excessive and reverse loading can also cause rolling elements and rings to deform. Some environmental defects are overheating, lubrication failure and corrosion. Each of these defects can cause lubrication breakdown, which in turn can cause excessive wear on the bearings. Geometrical failures involve tight and loose fits. Either of these defects could cause excessive heat and wear leading to bearing deformation and destruction. Figure 2 contains some of the common causes for bearing failures and faults. Figure 2a comes from reverse loading of the bearing ball. Figure 2b shows the effect of corrosion on the bearing. Figure 2c is caused by a loose fit of the bearing when affixed in a pillow block. Figure 2d shows the effects of inadequate lubrication on a bearing. The causes of failure are identified usually from the location, type and size of the defect on the bearing. The geometry of the application plays particular role in the identification of the failure as it is typically a problem of the system that causes bearing failure to appear. In the SKF document [43] , other bearing failures mentioned include smearing and more explicitly Brinelling indentations. This causes damage to the rings of the bearings and can in turn lead to spalling. While the list of defects and failures for bearings go on, it is important to be able to classify each major type of bearing defect to identify and fix the issue causing damage. In the Timken document [44] , the authors include the previous defects listed as well as defects that could cause the roller cage damage, which can lead to ball damage and deformation. Figure 3 contains a representation of a bearing with the various labeling of parts referenced in equations 1 through 3. With reference to the bearing frequencies mentioned earlier, equations 1-3 refer to how to calculate the bearing characteristic frequencies. The IRF is the inner race frequency, Figure 2a describes reverse loading, when the ball is acting in the opposite direction it is meant too [42] . Figure 2b is the effect of corrosion due to improper sealing or maintenance [42] . Figure 2c is caused by a loose fit to a pillow block [42] . Figure 2d shows wear brought about due to improper lubrication [42] . Most of the failures come from the application or improper maintenance. the ORF is the outer race frequency and RF is the roller frequency. Nb is the number of rolling elements, and S is the rotational speed. Bd is the ball diameter, Pd is the pitch diameter, and ∅ is the contact angle. Bearing defect data is created at a bearing station comprising an electric motor, a steel shaft, a linear table and a pillow block bearing assembly (NP-205), as shown in Figure 4 . The bearing in the motor next to the shaft is incrementally damaged on purpose to gather failure data from different defects. For the tests, the outer ring of the bearing is held fixed while the inner ring rotates with the shaft. The shaft connects the motor to the pillow block to introduce an "application". The linear table induces misalignment, allowing for loading on the bearing. This loading causes the bearing vibration to increase, making it easier to identify the defect frequencies. Bearings themselves are chosen as the validating component due to the widespread understanding and application in manufacturing equipment. The bearings used are 608ZZ bearings, commonly found in roller skates, fidget spinners, fans and small hand tools. These were chosen due to easy access and the ability to use them in bulk for the application. Figure 5 shows the type of damage for an inner ring defect in the initial state. Different damage tests are designed for each bearing to isolate the onset of failure for each test and to then see how the damage spreads to other components or increases along the individual components. Different bearing faults were tested at varying degrees of misalignment. The minimum degree misalignment was 0degree, while the maximum misalignment was assessed just at 1-degree. Typically, the maximum misalignment for a 608 bearing is 1 degree. Equation 4 shows how misalignment was calculated regarding the 608Z bearing. Δf refers to the allowable misalignment. Δa refers to the axial play of the bearing. Dp refers to the pitch diameter. The damage and failures were induced to replicate the progression of failure leading to a destroyed bearing found in an automotive manufacturing facility (shown in Figure 6 ). The 608ZZ is not the exact bearing used in the manufacturing plant, however the experiment designed involves the same type of bearing and application use. What is of interest to the authors is primarily the failure signal. There were three different bearings tested. One bearing was a default 608ZZ bearing tested from the original motor. The second bearing tested was a bearing with an induced inner race defect. The third bearing tested was a bearing with an induced outer race defect introduced. For the inner race defect bearing, the bearing is destroyed to the final stage of failure. The final stage is meant to simulate a fully damaged bearing as seen in Figure 6 . Table 1 shows the different tests and bearings described in the data gathering phase. The defects were caused by a Dremel tool with a fine tip. This bit was used to "flake" and "roughen" the surface of the bearing to simulate each defect. This additional damage affects the possibility of isolating failure signals to faults. The smallest defect on the race surface was 1 mm x 1 mm and 0.1 mm deep. For Tests 4-6 and 13-15, this was the initial size for each defect on the damaged bearings. For Tests 7-9, the damage grew to 2 x 2 x 0.2 mm as this was the next size of defect on the bearing. The amount of defect on the bearing also exponentially increased. Tests 10-12 were of the maximum defect damage state, again exponentially increasing with the entire surface marked up with at least an initial defect size of 1 x 1 x .1 mm. This mimicking of exponential destruction is similar as well to the increase of destruction found in most bearings according to literature. For testing and making defects on the bearings, the shape of each defect is circular. This was to ensure precision of making the defect on the bearing naturally versus a square defect area. For the experimental data, since a ball bearing is used, the defect begins in the center of the race as this is where the rolling element makes most contact with the inner ring. 5 individual "process" conditions are run for each test. Each test is run on the same electric motor. The motor is given time to warmup of an initial 10 seconds. This gives the signal time to stabilize, then data collection can begin for the process. A process constitutes turning on the motor, leaving it on to gather at least 20 seconds of data and then turning the motor off. This is repeated until the 5 processes of the data are collected. The sensor and data capture software come from ifm Efector Inc. The sensor is an accelerometer sensor that samples at 50 kHz. At least 400 MB worth of data are gathered following this method per process. This is done to simulate a lift process that is seen in the second application of this method. For the second application, vibration data is collected using an accelerometer from a vertical lift application at BMW. There are two motors in each lift application. The lift is used to move cars between the different levels of the manufacturing facility. The bed of the lift is connected to a belt that is then attached to a bearing and motor. Only one motor and bearing are coupled at any time to the belt that moves the lift. The other motor is held in reserve if maintenance must be scheduled on the running motor. The idle motor can quickly slot in and lift operations can continue as normal until the former motor is repaired. The data collection system is like the one used in the experimental data collection. The same accelerometer and data acquisition system is used. The same sampling rate is used as is the same overall amount of data collected as in the experimental data setup. Two cases of data were collected. One was from a lift with a known bearing fault. The other case of data was collected from lift that was considered "healthy". Each data test was saved into one file of five processes. For the analysis, each file was then split into five separate files, one for each process. Each file is made up of the raw vibration data. From this vibration data, a spectrogram of the process is created. A spectrogram is a representation of frequencies of a signal as it varies with time. The primary reason for selection of the FFT was ease of implementing possible different filters and the relatively low requirement of needing to choose parameters as denoted by the Wavelet transform and highlighted in Section 2.3. The spectrogram usually appear with time along the x axis, frequencies along the y axis and then power represented as different colorations in the graphs. Using the spectrogram, energy is calculated over time from the FFTs. This energy value comes from integrating the power spectral density with regards to the frequency range of the FFT. This converts the multiple power peaks of the FFT into one value of energy represented in time. This energy calculation closely corresponds with the motions of the equipment. The firs method the authors consider is as a distribution of energy points with respect to the probability density function. The baseline case is expected to have a single peak in terms of the distribution and should represent a normal distribution of energy. This is an established expectation from the Vibration Handbook by Chris Mechesfke [2] based on existing healthy and faulty data values from machinery. Any deviations from this would then appear as a potential fault. The second application used classifiers to compare healthy processes versus processes of fault data. This was done to asses how possible it was for the data to be incorporated in different types of analysis. In this case a Support Vector Machine and a Density Based Spatial Clustering method was used. This was also meant to represent the difference of using supervised versus unsupervised learning in the analysis. These methods are further explained in the results. The FFT had the following characteristics. The window for the FFT is the Hamming window, due to windows use in literature. The size of each block of the FFT is 2048 data points. This increased the detail of the FFT and increased the computational time. The overlap point between the blocks was 128 points. The sampling frequency was around 50000 Hz. For the faulty bearings, the expected energy values are expected to have a wider variance. When fault begins to appear in raw values, the variance and amplitude of the raw data will increase and cause a wider variation in the signal rather than having a single peak [2] . The probability density function is expected to be lower due to a fact the energy values have a wider range to cover. This is then determined as representative of the process and determining between healthy or an unhealthy machine. Figure 7 and 8 show FFT response from Test 3 and 6. Test 3 was the baseline case at a misalignment of 1 degree, while Test 6 was the inner race defect case at a misalignment of 1 degree. The conventional case for the FFT is comparing differences with regards to the response of established baseline data. The calculated BPFO from equation 2 was 117 Hz. The calculated BPFO was 203 Hz. The calculated rolling element frequency was 157 Hz. The misalignment of the shaft is attributed to the deviation of the BPFI. The introduced defect to the inner race is causes the sidebands seen in Figure 8 . Sidebands are present in all three frequencies seen here. With the roller frequency from Figure 7 and Figure 8 , there is a noticeable difference in terms of the PSD. Figure 7 has singular high peaks at each characteristic frequency, while Figure 8 has noticeable sidebands at each response. The difference in peak location and peak amplitude for the BPFO and RF is noticeable, with each peak location dropping by around 2 to 3 Hz. The peak amplitude for the outer and inner rings is also reduced by 2 W/Hz in both cases. While the differences are noticeable to humans, a computer program may not be able to register the difference in the signal seen. Algorithms such as Naive Bayesian Classifiers and neural networks require strict training data sets for determinable analysis on whether the data is faulty or healthy. The closeness of the signal difference could lead to multiple misclassification of data points. There are ways filter out the signal digitally and through the experimental setup, but that will also require more software and hardware, then some manufacturers can or are willing to provide. With the energy method proposed, the differences in those FFTs will become pronounced and determine not only healthy or fault, but also the type of defect associated with the different bearing tests. Figure 9 and Figure 10 show the probability density function with respect to the energy points. The probability density function is based on the occurrence of the energy point over the length of the process. Figure 9 shows the baseline case data Figure 8 : FFT response from Bearing Test 6. This is from Bearing Test 6 of the inner race defect case. On the x axis are the different frequencies. The y axis is the power components of the signal. Single Sided Power (W/Hz) Frequency (Hz) Figure 9 : Baseline Data. This is the plotted data taken from Bearing Test 1 shown in Table 1 . The x axis has the energy values with units of J/Hz for energy spectral density. The y axis is the probability density function. from Test 1, while Figure 10 shows the initial induced damaged inner race bearing from Test 4. Each figure has three processes plotted. The different processes are denoted by different dashed lines. The first peak of the process is centered around 2 J/Hz. This is indicative of the start of the process and represents when the process is not in motion. The normal operation of the process begins at 175 J/m 2 /s and then ranges up to 300 J/Hz. Only one baseline has a clear normal distribution shown as the third process with the dotted line. However, the range of energy values is similar for each individual process. For Figure 10 , the process data plotted appears to follow a similar distribution as seen with the baseline data, which was not expected. The peak of the energy data occurred at a lower energy value than was expected. The peak location in terms of energy values is around 15 J/Hz. The amplitude of the pdf is around 0.11. There is a slight peak around 0 J/Hz for each process, like the baseline data. A higher distribution in energy values was expected with a lower pdf value. This may be due to the initial size of the defect compared to the baseline data. The highest concentration of energy values was more than the baseline data, which was unexpected. Figure 11 shows the change in misalignment for the baseline bearing. Interesting enough the baseline bearing changes from a normal distribution and then begins to form two peaks in the intermediate stage at a wider distribution of energy points before centralizing at two peaks at a narrower and lower energy content. From Figure 11 , at the extreme case of misalignment a wider distribution of energy values was expected denoted by the dotted dashed line. However, in that case, the distribution began to centralize around two peaks at 50 J/Hz and 170 J/Hz. Figure 12 shows the baseline misalignment with the average distribution of the tested bearings for the single inner race defect, the single outer race defect and the baseline case, all at 1 degree of misalignment. From this Figure 12 , the classification of each bearing defect from the baseline is not possible. At the maximum misalignment, the noise generated by that state drowns out the possibility of classification based on the defect. The information that Figure 12 does provide is that misalignment is detectable from the 1-degree misalignment state, irrespective of defect. Misalignment is considered a bearing defect and one that if neglected could lead to warpage and damage to the bearing, motor and shaft. Figure 13 shows the difference in extreme damage on the inner race bearing with a comparison to the baseline data. As noted in Figure 8 , the initial bearing damage started off very low and culminated in a high peak at low energy values. In Figure 13 , the average bearing energy signature appears to have stabilized close to a middle value between 25 J/Hz and 100 J/Hz This is between the initial damage reading shown and the baseline damage process lines shown in Figures 9 and 10 . This could relate to the possibility of the appearance of more sidebands in the FFT and higher peak amplitudes in the PSD. The average values are lower for the maximum defect data peaking at a location around 48 J/Hz and 80 J/Hz for the defect data in the dotted line. The average process data for test 1, 2 and 3 is shown above. The difference in each of these tests is the angle of misalignment for the shaft and bearing. The x and y axis remain the same as the previous plots. Table 1 is plotted over each other. On the x axis, the spectral energy density and on the y axis is the probability density function In the Figures 9-13 , there are several different conclusions drawn. The first is that fault data is discernible from baseline data on the process. This is based both on the shape of the distributions as well as the location of where the peaks are occurring. The second conclusion is that misalignment interferes with additional bearing defect classification. In this case, the authors meant to use it as a means of loading the bearing for excitation by the motor. However, when comparing the bearings in the extreme misalignment, it was too difficult to recognize which defect was under observation. Another observation is the subsequent appearance of additional peaks in more defect data. This could be due to the appearance of additional peaks in the FFT from the defect. This next section shows a case study from BMW on one of their vertical lift applications. The vehicle car lifts are a high priority in ensuring they stay healthy as if one were to fail unexpectedly it can shut down the line for multiple shifts. The lifts have IFM vibration sensors similarly to the ones used in the experimental testing. Data is taken from two sources. The baseline comes from a lift with similar characteristics as to the defect lift. The maintenance staff had deemed it healthy. The defect data comes from a lift with a known bearing defect. The detection of the bearing defect stemmed from calculating the bearing fault frequencies originally. In this event, the bearing defect is a ball defect that has scraped against the inner race surface, causing a small surface of the bearing ball to appear. The data collected goes through the same configuration, however the distributions are not shown. A Support vector machine (SVM) classifier and a density based spatial clustering method were used to compare and see if there was a possible difference in the classification. Figure 14 shows the process events with respect to the energy content versus the corresponding FFT for the healthy lift. For every 1000 sampled FFTs, that represents approximately a minute of process time. For lift operations, there is the up motion and the down motion. In this case the down motion is the events without the initial spike. The energy spike seen at the beginning of each process correspond then to the motor initially ramping up before settling into the operation. The end spike is the brake engaging, bringing the motor to a sudden halt. For the process, the average energy content remains around 1 J/Hz except for the spike. Figure 15 shows the process events with respect to the energy content versus the corresponding FFT for the unhealthy events. In the case for this data, the brake events and the initial motor spikes are not discernible. The average energy is also higher closer to 1.5 J/Hz versus the healthy 1 J/Hz. This is different from what is seen in the experimental data. In Figure 12 and 13, the distribution of the data shows the healthy data as having a higher energy content as oppose to the defect data. A possible reason could be due to the increase of load of holding a car. Based on Figure 14 and 15, it is possible to discern the difference between the average energy level as well as near elimination of the brake events in the data. These are a few of the features with a measurable difference that are useful in terms of classification. From the data presented, it shows that tracking of process events is also possible regarding using the energy spectral method. There is a 0.5 J/Hz difference about the energy content between the healthy lift and damaged lift. Figure 16 shows a comparison side by side of a process of each data. Figure 17 shows the classification of the data using a one class support vector machine. The kernel used was the rbf kernel, which denotes a Gaussian distribution. The tolerance for the stopping criterion was set to 0.1. The degree of the polynomial was set at 4 to account for the spikes seen in Figure 4 and 5. The classification was based on 2 features. The The healthy data was taken after the bearing change, while the unhealthy data was taken just prior to the bearing change. difference in the maximum peak positions of the FFT within a 10 Hz tolerance and the difference in the energy of the sample. From the classification, the classifier had 98 % success rate in classifying the fault failure data from the baseline data. Figure 17 shows the classification results for each point with regards to the fault data energy and top maximum frequency positions. In the first graph, Class 1 is considered a health data point, while class 2 is an unhealthy data point. There is still some misclassification from using this method stemming around the brake events. However, the authors believe those errors are removeable through better SVM parameters. A density based spatial clustering method (DBSC) was used to see if it was possible to for unsupervised learning in clustering the data. The DBSC was used from the machine learning library, sklearn. Here the command is termed as DBSCAN. The Euclidean distance of points from one another was a metric with a minimum of 20 points to become a cluster. Figure 18 shows the initial results of comparing the area and position of the maximum FFT peak position. Figure 17 shows a comparison of this points in an unsupervised learning method. Three clusters emerged around three different positions. The first cluster corresponds to the lower energy values typically found in the baseline data as seen in the dotted and dashed circle. The dashed circle holds values associated with the brake event, while finally the furthest cluster is failure events. This was more of a proof of concept to see if the data is cluster able in unsupervised method. Future use of the method could lead to detection of different faults if different clusters begin to emerge or if one cluster grows larger than another. A method for using the energy spectral density to detect component faults was presented. Bearings were used to validate the method. Data from a vertical lift application was used to show the use in a production environment. Two different applications were represented with an experimental and an industrial application. Fifteen different cases were presented shown in Table 1 . Each case is represented by a different location of the defect, the size of the defect, the location of the defect, the amount of misalignment and the relative amount of damage compared to the overall surface of the bearing. Based on the results, the energy content from the power spectral density is a valid method for determining bearing faults from healthy data over the course of a process. From the experimental test stand, each bearing defect had a different distribution in comparison to each defect and the baseline data. When misalignment was introduced, however the faults became more difficult to discern from one another and the baseline misalignment case. The distribution of energy values was used to validate the detection of faults in the experimental data. For the industrial data, the energy content with respect to the corresponding FFT and by extension the corresponding time was used to show the difference regarding a healthy lift versus a damaged lift. From Figures 14 -18 , the data is separable between healthy and unhealthy processes and classification is possible. However, there is still work required before completely validating the tested method on an industrial application. A more rigorous experimental plan should be developed for offline learning with a specific emphasis on multi failure classification. An experimental test was considered for ball defects to validate the data seen in the corresponding lift application. Another issue is the dependence of FFT levels on speed and load. For the industrial data, simply taking data from a corresponding and similar lift could invite potential error, if the speed and load are different between different lift applications. However, a relationship between the speed and load can be defined using the characteristic bearing equations and modeling of the applications for each case. Another instance is the amount of noise generated in the bearing vibration test stand. For some of the test cases, the variation exceeded 50% and did not match the baseline or any other know defect state. Therefore, for Figures 9 and 10 , only three test processes are plotted instead of five test processes. A new bearing station has been designed as a more rigid configuration to allow for better and more rigorous testing. This will eliminate much of the expected environmental noise. Another possible elimination of ambient noise is the use of filtering. A bandpass filter added into the model could focus in on the selected frequencies and mitigate the amount of possible error. Fine tuning the parameters in the classifications will lead to more robust results and reduce the number of misclassifications. This is a problem seen with wavelet transforms as well. A more general case will need to be developed for applications to other components. Another application is the extension of incorporating wavelets into this method. Wavelet analysis is the accepted method for bearing fault diagnosis and offers the ability to show the time occurrence of when frequency events occur. Incorporating the occurrence of energy events from the frequency over time could increase accuracy for multi class failures as additional features. Future work will also seek to extend this application to other equipment and see if it is possible to detect faults with them. Bearings were used to see if this was a valid technique for classification and the data features were truly separable. The objective in the future is to test the method on different industrial applications, for continued validation. Operations having more significant process variation will be selected, in order to provide a more robust test of the classification approach. Possible equipment and components involve vehicle platforms and valves from non-Newtonian sealant applicators. Current status of machine prognostics in condition-based maintenance: A review Machine condition monitoring and fault diagnostics Strengths and Limitations of the Wavelet Spectrum Method in the Analysis of Internet Traffic Overview of maintenance strategy, acceptable maintenance standard and resources from a building maintenance operation perspective Vibration Analysis -a Proven Technique As a Predictive Maintenance Tool Review of vibration analysis methods for gearbox diagnostics and prognostics Intelligent monitoring of ball bearing conditions Vibration based condition monitoring: A review A review on machinery diagnostics and prognostics implementing condition-based maintenance Optimal number of features as a function of sample size for various classification rules Detection of rolling element bearing damage by statistical vibration analysis Vibration monitoring of rolling element bearings by the high-frequency resonance technique -a review The vibration produced by multiple point defects in a rolling element bearing A review of Rolling Ellement Bearing Vibration 'Detection, Diagnosis and Prognosis Bearing condition monitoring methods for electric machines: A general review Models for bearing damage detection in induction motors using stator current monitoring Detection of motor bearing outer raceway defect by wavelet packet transformed motor current signature analysis Fault feature extraction of gearbox by using overcomplete rational dilation discrete wavelet transform on signals measured from vibration sensors An Overview on Vibration Analysis Techniques for the Diagnosis of Rolling Element Bearing Faults Pattern Recognition Statistical, Structural and Neural Approaches A review of neural networks for statistical process control Bearing Fault Detection Using Artificial Neural Networks and Genetic Algorithm Bearing fault detection by a novel condition-monitoring scheme based on statistical-time features and neural networks Condition based maintenancesystems integration and intelligence using Bayesian classification and sensor fusion Reliable bearing fault diagnosis using Bayesian inference-based multi-class support vector machines Online Learning versus Offline Learning A Baseline for Detecting Misclassified and Out-Of-Distribution Examples in Neural Networks Health monitoring, fault diagnosis and failure prognosis techniques for brushless permanent magnet machines Bearing fault diagnosis using FFT of intrinsic mode functions in Hilbert-Huang transform Wavelets for fault diagnosis of rotary machines: A review with applications Wavelet analysis and envelope detection for rolling element bearing fault diagnosis-their effectiveness and flexibilities Generator bearing fault diagnosis for wind turbine via empirical wavelet transform using measured vibration signals Fault Diagnosis of a Rolling Bearing Using Wavelet Packet Denoising and Random Forests Review of recent advances in the application of the wavelet transform to diagnose cracked rotors A comparative study of FFT, STFT and wavelet techniques for induction machine fault diagnostic analysis Processing for improved spectral analysis Fault detection in induction machines using power spectral density in wavelet decomposition Spectral Approach to Geophysical Inversion by Lorentz, Fourier, and Radon Transforms Time-frequency analysis of earthquake records Total energy and energy spectral density of elastic wave radiation from propagating faults: Part II. A statistical source model A new framework using deep auto-encoder and energy spectral density for medical waveform data classification and processing Bearing failure: Causes and cures Bearing Failures and Their Causes Timken Bearing Damage Analysis with Lubrication Reference Guide