Development of a Novel Wear Detection System for ... - IEEE Xplore

IEEE SENSORS JOURNAL, VOL. 14, NO. 2, FEBRUARY 2014

465

Development of a Novel Wear Detection System for Wind Turbine Gearboxes Andrew Hamilton, Alison Cleary, and Francis Quail

Abstract— This paper presents a low-cost, inline, gearbox lubrication monitoring sensor. The purpose of the research was to develop a sensor that can analyze wear particles suspended in gearbox lubricant systems. Current inline sensor systems rely on methods that prevent significant morphological classification. The size and shape of the particles are often indicative of the type of wear that is occurring and is therefore significant in assessing the gearbox state. A demonstration sensor consisting of a webcam that uses an active pixel sensor combined with a rectangular cross section optically transparent acrylic pipe was developed. A rig that simulates a gearbox lubrication system was used to test the sensor. Images of wear particles suspended in the lubricant were captured in real time. Image analysis was then performed to distinguish particles from the lubricant medium. Object characteristics, such as area and major axis length, were used to determine shape parameters. It was found that the sensor could detect particles down to a major axis length of 125 µm. Classification was also demonstrated for four basic shapes: square, circular, rectangular and ellipsoidal. ellipsoidal, was also demonstrated. Index Terms— Condition monitoring, gearbox, sensor, wind turbine, lubrication.

I. I NTRODUCTION

W

IND turbine farms are subject to variable load regimes compared to conventional power stations. Since the wind loads are highly stochastic and can cause transient loads in excess of 2MN within the turbine drivetrain, a high number of faults can occur to multiple components. The gearbox has been identified as one of the most critical components due to the cost and time involved in repairing or replacing damaged modules [1]. In addition, a relatively large period of wind turbine down time is required to conduct repairs, resulting in further loss of revenue [2]. Fig. 1 shows the failure frequency and downtime for a range of wind turbine components. Whilst Manuscript received July 16, 2013; revised September 12, 2013; accepted October 2, 2013. Date of publication October 9, 2013; date of current version December 5, 2013. This work was supported by EPSRC under Project Grant EP/G037728/1 and was conducted within the University of Strathclyde between the Wind Energy Systems Centre for Doctoral Training and the Centre for Advanced Condition Monitoring. The associate editor coordinating the review of this paper and approving it for publication was Prof. Boris Stoeber. A. Hamilton is with the Wind Energy Systems Centre for Doctoral Training, University of Strathclyde, Glasgow G11XW, U.K. (e-mail: [email protected]). A. Cleary is with the Centre for Intelligent Dynamic Communications, University of Strathclyde, Glasgow G11XW, U.K. (e-mail: [email protected]). F. Quail is with the Centre for Advanced Condition Monitoring, University of Strathclyde, Glasgow G11XW, U.K. (e-mail: [email protected]). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/JSEN.2013.2284821

Fig. 1. Failure/turbine/year and downtime results from European wind turbines [1].

the failure rate is lower than most components, the downtime is far greater. The problem is compounded when wind farms are located offshore and cannot be conveniently or inexpensively accessed. To optimise fault maintenance intervention and enable fault detection, condition monitoring techniques can be applied where various parameters can be measured and tracked that correlate strongly to the components’ operational state [3]. A variety of techniques exist that range from analysing data from wind turbine Supervisory Control and Data Acquisition (SCADA) units to deploying sensors that measure features such as atypical mechanical vibration [4]. Monitoring of wear particles suspended in gearbox lubricant offers particular advantages in predicting faults. The number, size and shape of the particles are often indicative of the type of wear that is occurring and so are significant in assessing the overall health of the gearbox [5]. Particle size and liberation rates have also been linked to useful remaining life and specific component failures [6]. In Fig. 2 an example of inline gearbox wear particle trending is shown. Tracking the cumulative detection of different size classes of particles and using suitable analysis algorithms allows wind turbine operators to assess the current state of the gearbox and specific components. This paper presents a novel inline sensor system that categorises wear particles by size and shape. It is envisioned that the system may be used to complement other wear sensors. II. L UBRICATION -BASED C ONDITION M ONITORING Currently, there are commercially available inline or bench top sensors that monitor wear in wind turbine gearboxes.

1530-437X © 2013 IEEE

466

Fig. 2.


1.5 MW onshore wind turbine gearbox cumulative particle count.

However, there are limitations in deployment of these systems including high cost, the requirement of human intervention for data analysis and the limited value of data that is captured [7]. A. Current Sensors Optical methods use an LED or laser beam combined with a photodetector to generate a signal that varies with the amount of incident light. When a particle partially occludes the laser beam, the decrease in incident light is registered as a change in voltage or current, and an equivalent particle diameter is calculated. In some systems, diffraction or scattering of light is measured instead. A variety of algorithms can be used to correlate the signal change to a particle size [8]. Dielectric constant and inductance based electrical sensors rely on measuring the change in electromagnetic properties due to the presence of metallic particles [9]. These optical systems can detect particles with a diameter of 3 μm, allowing component wear to be tracked at a very early stage [12]. This conforms to the lowest particle size used by ISO oil cleanliness code [10]. The presence of particles less than 10 μm is associated with the earliest stages of machine wear. Whilst determining the source of particles cannot be achieved, the overall health of the machine may be accurately assessed. However, in both of these existing systems, particle morphology is not determined due to the nature of the signals generated. In optical systems, the signals are dependent on light scattering or occlusion which is affected by particle orientation. Similarly in electrical systems, the changes in measured parameters such as inductance are mainly related to the general size and mass of material as opposed to shape [11]. In both techniques, it is assumed that the particles are circular with the largest detected dimension setting the diameter. As particle morphology is not captured, it cannot be used to correlate the type of wear that is occurring [12]. B. Particle Morphology Wear particle morphology has a history of use in determining the type of damage that is or has been occurring in machinery [13].

Rolling contact fatigue, associated with bearing failures, generates spherical particles in the order of 5 μm that can increase in size as the problem worsens. They can also be a result of cavitation erosion and abrasive wear, requiring more detailed analysis [7]. Sliding contact surface wear generally creates elliptical shapes type particles (in the order of 15 μm in length and 1 μm in thickness) which can indicate abnormal loading regimes and lubricant viscosity reduction [14]. Cutting wear occurs when one component surface penetrates another indicating misalignment deformation and can create particles in the range of 25 to 100 μm. Advanced bearing faults may increase wear rates of elongated particles with a major to minor axis length ratio ranging from 10:1 to 30:1 [7]. A wide variety of laboratory and bench top based equipment can be used to fully categorise particle shape. This is achieved using different techniques, such as optical and scanning electron microscopy and laser scattering [15]. Developing an inline sensor for wind turbine gearboxes that can classify particles by morphology offers more detailed fault identification. This could reduce failure rates by allowing wind turbine operators to take remedial action and optimise their maintenance schedule. While using imaging to monitor particles in a number of industrial processes is not novel, we believe that this is the first demonstration of such a system applied specifically to on/inline gearbox oil monitoring [7]. C. Active Pixel Sensors Advances in digital video technology have lowered the cost of high resolution active pixel sensors (APS’s). Webcams are available with 5.0 mega pixel (MP) video resolution and high frame rates, allowing a high level of detail to be captured at a low cost. These APS’s may be capable of providing inline, real time images of particle suspending in wind turbine gearbox lubricant [16]. Once images are captured, software must be used that can distinguish wear particles from the surrounding lubricant, nonwear contaminants and the enclosing pipe. Particle characteristics such as area, eccentricity and various axis lengths must also be determined from the wear images to allow correct morphological classification. Currently, there are a number of digital video and image processing techniques that would allow wear particles to be distinguished as individual objects. These range from basic binary conversion combined with total pixel count to full color histogram analysis with morphological filtering [17]. To assess the principle, a low-cost webcam was used for particle capture and an image processing script was developed using MATLAB [18]. III. S ENSOR D ESIGN AND T EST R IG The sensor was developed to integrate with a purposebuilt fluid circulation rig that simulates the operation of a gearbox lubrication system shown Fig. 3. Different lubricant parameters could be altered such as temperature and flow rate. Wear regimes can be simulated by introducing particles into the lubricant circulation using a T-piece control line and 2 double check valves. The arrangement allows particles to

HAMILTON et al.: DEVELOPMENT OF A NOVEL WEAR DETECTION SYSTEM FOR WIND TURBINE GEARBOXES

Fig. 5.

467

Total FOV for webcam and AOI used in image acquisition.

Particles close to being 5 μm potentially could be detected but could only be interpreted as a single pixel. Fig. 3.

Gearbox lubricant simulation rig used to test sensor.

Fig. 4. Upper - Sensor installed on test rig with oil flow connected with light shield removed. Lower - Diagram of sensor construction.

enter the rig whilst the lubricant is flowing. Particles are then removed from the system using two standard industrial filters of different size class (10 and 400 μm). The entire imaging system was controlled by a computer using MATLAB. Mobil Gear XMP320 gearbox lubricant was selected as the test fluid due to its widespread use in wind turbines [19]. The sensor consisted of a commercially available webcam APS (Bush model number VP-71VP2) combined with a clear acrylic pipe, shown in Fig. 4. A rectangular cross section of the pipe was selected to allow a clearer image of the lubricant with minimal visual distortion. Piping with a circular cross section magnified the image non-uniformly. The field of view (FOV) of the camera was chosen to be slightly larger than the usable transparent area of the pipe. Light variations were determined to be a significant factor in the overall accuracy of the APS due to deviations that were observed whilst running basic video processing programs. An LED array with an integral optical diffuser was used to give a constant and uniform light source to illuminate the particles and show changes in the lubricant color. The sensor was encased in an opaque, non-reflective shield to prevent external light from being detected. The webcam was capable of producing 60 frames per second (FPS) with 1.9 MP video output and 7.8 MP image output. It was selected due to its manual focussing option, suitable FOV specifications for the purpose and its low cost (4 GBP). A manual focus was preferred over auto or digital focus, allowing easier adjustments once the webcam was installed in place. The webcam pixels were 5 μm in diameter which meant particles smaller than this could not be detected.

IV. S ENSOR P ROGRAMMING A. Image Acquisition An image acquisition program was developed using MATLAB with the Image Processing Toolbox to convert the raw data stream into an appropriate format. The program automatically detects the webcam characteristics such as frame rate and number of pixels. The user can set the capture area and frame rate appropriate for the type of testing. The capture area used here corresponded to the inner dimensions of the acrylic pipe of 6 × 6 mm. The webcam had a relatively large FOV that was unnecessary due to the small width of the pipe. This was reduced to capture a 6 × 6 mm by setting an area of interest (AOI) as shown in Fig. 5. The number of captured pixels was reduced resulting in faster acquisition. The centre camera was aligned with the midpoint of the pipe. Setting the appropriate frame rate was important as it could lead to two types of error, and needed to be matched to flow rates of the oil test rig. A high frame rate could potentially capture the same particle multiple times as it traverses the field of view. A low frame rate could allow particles to pass undetected, particularly at high flow rates. In addition, particles did not traverse the pipe at uniform speeds due to wall surface roughness within the pipe. Larger, heavier particles tended to flow along the bottom of the pipe at a slower rate compared to light particles that were suspended in the middle. As the system presented here was intended to demonstrate the principle of operation it was decided to use a low frame rate with the risk of some particles being undetected. It was calculated that if the flow rate was 0.216 ml s−1 , a particle would pass the 6 × 6 mm field of view in approximately 0.5 s, requiring a 2 FPS capture rate. This would mean a frame is captured every 0.5 s, matching the expected period for a particle to cross the FOV. The images were initially recorded as an m × n × 3 array, where m and n are corresponding to the image height and width and 3 is the number of color channels (red, green, blue or RGB). In this format, the color perceived by the webcam pixel is represented as combination of 3 colors. The data class used was uint8 which meant that RGB values varied between 0 and 255 [20]. B. Object Identification Object detection relies on identifying features that are unique from the surrounding image using a variety of methods

468


Fig. 7.

Left - 4 connectivity, Right - 8 connectivity for center pixel.

Fig. 6. 2× 750 μm particles. Left - Image from APS, Right - Conversion of image using Otsu’s method into bi-modal.

in combination; histogram averaging, edge detection and texture differentiation [21]. Typically, images will be processed in several operations, depending on the features that need to be extracted and the degree of accuracy required [22]. Wear particles are often characterised by such methods in ferrography and other wear particle analysis systems due to the relative complexity of the objects and the advantages in robust categorisation [23], [24]. For the purposes of the sensor demonstration, a relatively simple program that distinguished particles as discrete objects was developed that utilised bimodal image conversion and edge detection. The RGB image was converted into grey scale by averaging the individual ui nt8 values of red green and blue for each pixel G S pi xel = (R pi xel + G pi xel + B pi xel )/3

(1)

where G S pi xel is the grey scale image consisting of an m × n matrix populated with uint8 grey intensity classes (0-black, 255-white) [25]. Wear particles appeared dark when compared to the surrounding lubricant whilst other contaminants such as air bubbles appeared light. To enhance this feature, a histogram based threshold was applied which utilised Otsu’s method of minimising weighted within-class variance [26]. This method assumes that the image is bi-modal consisting of background pixels and foreground pixels and computes the optimal grey scale threshold value, t. Pixels equal to or greater than t will have value 1, otherwise the pixel value will be 0. The result was that particles within the image appeared black and maintained borders whilst the lubricant became white. No spatial coherence is used to define the object structure and uniform illumination is assumed so the class variance is determined only by object appearance. The method first requires a histogram to be calculated on the grey scale image that holds the total number of pixels in each grey intensity class. The between class variance, σ B2 (t), for t[1] can then be evaluated and is defined as σ B2 (t) = ωb (t)ω f (t)[μb (t) − μ f (t)]2

(2)

where weights ωb (t) and ω f (t) are the probabilities of a pixel be longing to the background and foreground classes when separated by, t, the threshold. μb (t) and μ f (t) are the background and foreground class means when the probability of each histogram bin value occurring within the given class is evaluated for threshold t. The maximum between class variance is then evaluated and the corresponding value of t is selected as the threshold. Within the MATLAB code,

recursion was incorporated to evaluate the threshold to minimise processing time [27]. The progression of the processing techniques can be seen in Fig. 6. Some particle clarity was lost in this process, especially with particles less than a major dimension of 400 μm. The overall conversion did not require significant processing power and could be maintained at the highest camera frame rate of 60 FPS. C. Particle Counting Objects were identified by counting the number of boundaries that existed between black and white pixels in each image [28]. The program scanned each line of pixels until a change value occurred: 0 values (white) were considered as background so that no objects consisting of these pixels were counted. Pixel values of 1 were assessed for connectivity to determine the outline for individual objects. An example of different types of connectivity is shown in Fig. 7. 4-face connectivity allowed faster image processing but reduced detail. However at 50FPS, it was found that 8-face connectivity was possible to determine a reasonably accurate outline of particles in the range of 300 to 500 μm. At lower sizes, outlines became less accurate until particle morphology was significantly reduced. Objects were considered to have no internal boundaries by not recognising any pixel change once object connectivity had been established. This prevented small light discrepancies on particles from manifesting as object structure. One problem that can be experienced when using Otsu’s method of bi-modal image thresholding is that when the threshold, t, is applied, some pixels can be above this limit but separated from the main object by a number of pixels whose value is below t. The result is small objects consisting of a few pixels clustered around the edges of larger objects that represent particles. This distorts the total particle count and the wear profile estimated later in the program. To prevent this, a morphological structuring element that removed objects under 10 pixels was applied. D. Size Analysis Particle size was determined by measuring the overall area and the major axis length of the objects identified in each image. Area was established by summing the total pixel count for each object. During sensor design, the camera offset that produced clear images was measured and this remained constant during the testing phase. Images of 2000 graded particles (125–150 and 450–500 μm) were captured and the individual areas measured. The pixel to area ratio was found to be 1:260 (measured in μm2 ) with an error bound of ±15%.


Fig. 8.

469

Example images of regular objects used to test shape parameters.

The percentage error was found by binning objects into area classes of 20 pixel increments for the two sets of graded particles. The average pixel to area ratio was found for each bin, with the highest and lowest ratios used to estimate error bounds. The process was repeated using the binned ratios to find the average pixel to area ratio with the error bounds carried through to weigh the total error bounds. A similar process was used to develop a pixel to length ratio which was found to be 1:16.1 with an error bound of ±17%.

Fig. 9.

Object shape parameters for square and rectangular test objects.

Fig. 10.

Object shape parameters for circular and ellipsoidal test objects.

E. Shape Analysis Particle shape was estimated by using object area, eccentricity, major axis length and minor axis length to create 4 parameters. The first parameter, PC , estimated how circular the object was, and is defined as Pc = (I − Mecc )/(L min /L ma j )

(3)

where Mecc is the object eccentricity, L min is the object minor axis length and L ma j is the object major axis length, both in μm. If PC is approximately 1, the object will display circular characteristics. The second parameter, PS , estimated how square the object was, defined as PS = ((L min /L ma j ) + Mext )/2

(4)

where Mext is the object extent, the ratio of the object area to the bounding box area. The bounding box is the smallest possible rectangle that may fit inside the object. If PS is approximately 1, the object will display square characteristics. The third parameter, PE , estimated how ellipsoidal the object was, defined as PE = 1 − (L min /L ma j )

(5)

If PE is approximately 1, the object will display ellipsoidal characteristics. The last parameter, PR , estimated how rectangular the object was, defined as PR = ((1 − L min /L ma j ) + Mext )/2

It can be seen that the thin rectangular object best fit PR , closely followed by PE . The rectangular object also best fit PR and was followed by PS . The square object best fit PS but was instead followed by PC which indicates that the square fits the profile of a circle better than a rectangle due to the similar axis lengths. In Fig. 10 the results for the circular and square objects are expressed as a percentage correspondence. It can be seen that the thin ellipsoidal object best fit PE closely followed by PR due to the close resemblance to the thin rectangle. Similarly the circle best matched PC and PS due to the close resemblance of these shapes in terms of the measured characteristics. Additional characteristics may be applied in the future to improve shape categorisation accuracy. However the system was intended to demonstrate the principle. More characteristics requires increased processing capabilities so basic scalar parameters were used.

(6)

If PR is approximately 1, the object will display rectangular characteristics. The 4 parameters allowed a shape profile to be developed for each object and allowed non-regular shaped objects to be categorised. In Fig. 8, regular test objects with the same formatting as the images produced by the program are shown to demonstrate how objects are characterised. Using the described categorisation process on the test objects, the degree to which objects fit each parameter could be measured. In Fig. 9 the results for the square and rectangular objects are expressed as a percentage correspondence.

V. T ESTING M ETHODOLOGY The sensor was integrated in the lubricant test rig and the light shield was attached. The camera was checked for correct position, orientation and focus length. The rig was run for approximately 30 min at 15 ml s−1 to remove large air bubbles from the piping and also to heat the lubricant to the correct operating temperature. Air bubbles did continue to move through the oil, however they rarely presented as particles. The frequency of oil bubbles manifesting was approximately 5%. Since air bubbles tended to be of similar color and shade to the surrounding oil,

470


TABLE II T OTAL N UMBER OF PARTICLES FOR 50 mg S AMPLES

Fig. 11. Left - Original image of air bubble in oil flow (approx. 400 μm length). Right - processed image of oil flow with no detected objects. TABLE I I NDIVIDUAL PARTICLE M ASS FOR D IAMETER L IMITS

Fig. 12. Basalt spheres particles suspended in oil under microscope inspection. Left - 476 μm, 490 μm and 498 μm particles. Center - 483 μm particle. Right-various particles ranging from 127–145 μm.

they were classed as white due to the grey scale threshold that was selected in the initial image processing stage of the program as shown in Fig. 11. A. Calibration and Lowest Particle Size Testing Basalt spherical particles were used as the first test material to determine experimentally the smallest distinguishable object possible. The particles were graded by sieve fractions with the largest being 475–500 μm and the smallest being 125–150 μm. Particles within these grades could have a diameter value anywhere between the two limits. Basalt spheres were used for calibration as they could be sourced with known diameter distributions and uniform shape. Graded metallic particles with similar characteristics could not be sourced as readily. As the sensor operates purely on visual properties, the dark color of the basalt particles was considered an appropriate match. The particle size ranges were selected for a variety of reasons. The camera was initially tested using a series of printed dots on plain white paper. The same analysis program, light source and camera positioning were used. It was found that particles larger than 500 μm appeared very clearly on the output images. Particles less than 100 μm only presented rarely as objects. Therefore it was decided to limit initial testing to this range. The mass of particles was selected to determine the number of particles in a given sample size. Using spherical volume and a basalt density of 2900 kg/m3, the individual particle masses for each grade limit was found, shown in Table 1. 50 mg test samples were used for each grade. A significant degree of error existed within each sample group as each had an upper and lower limit corresponding to the sieve fractions. To compensate, a geometric mean, Dave, was used to calculate an average diameter for use in volume-based calculations [29], defined as (7) Dave = (D L DU )

where D L is the lower sieve limit and DU is the upper sieve limit, both measured in microns. Using the geometric mean the total number of particles per 50 mg was estimated and is presented in Table 2. These values were used to calculate the detection rate. To verify the size and shape of the basalt particles a microscope at 20 × zoom was used to capture images. From visual inspection, it was found that the general shape was circular. It was also found that for each grade, a variety of particle sizes could be found within the limits. This was confirmed by running the sensor program and finding the major axis length. An example of the images obtained by microscope and their processed form can be seen in Fig. 12. Samples were pre-mixed with the same lubricant in a tube that could be connected to the test rig. This minimised clustering and also allowed the particles to enter the flow at a more uniform rate. The flow rate used was 5 ml s−1 flow rate, requiring a 23 FPS image capture rate. B. Shape Analysis Testing Shape detection was tested by introducing steel particles of varying sizes into the oil flow. In the absence of oil samples from gearboxes with known wear characteristics the particles were obtained from a machine tool workshop as they had similar sizes and shapes of those found in gearbox lubrication systems. The same wear mechanisms occurring are also similar [7]. The total mass of particles was 200 mg and consisted of various particle shapes and sizes (validated through microscopy) ranging from approximately 100 to 500 μm. The sensor was not assessed for detection rate due to the unknown sample composition. Instead the sensor was tested in its intended operation format: inline collection and sorting of data in real time. Objects were initially sorted into 50 μm size bins, ranging from 100 to 500 μm and above. The shape with the highest


471

TABLE III PARTICLE D ETECTION R ATES FOR BASALT S PHERES

Fig. 13. Left–480 μm particle in original captured image. Center - converted image with color assigned according to object number. Right-converted binary image used in particle counting and classification stage of program.

percentage parameter would then determine what shape class they were binned. Within each object bin, the objects were further sorted by their highest shape correspondence. For example, an object with a major axis length of 265 μm would be counted in the 250–300 μm bin. If the shape parameter with the highest percentage was PC , the object would be classed as square within the 250–300 μm bin. To assess the effectiveness of the shape classification system, all captured images were stored with the corresponding object area, major axis length, minor axis length and eccentricity. This allowed a comparison between the measured parameters and the actual particles. Images with a large number of particles and a variety of shape classifications were sampled to verify the accuracy of the 4 measured parameters.

Fig. 14. Left-Group of 125–150 μm particles in original captured image. Center - converted image with color assigned according to object number. Right - converted binary image used in particle counting and classification stage of program. TABLE IV PARTICLE S HAPE D ETECTION R ATE AND S IZE B INNING

VI. R ESULTS The results of the detection rates for the 4 rounds of lowest particle sizing tests are presented in Table 3. A number of detected objects during testing were found to be greater than the expected size of the test objects. Since they may be a result of particle clustering, they were recorded separately. Clustering of particles may have occurred due to laminar flow within the lubricant and the uniform mass of particles. This caused particles to move in close proximity and periodically they would appear clustered together when traversing the AOI. It was found that the overall detection rate for the 475–500 μm particles was 92%. The particles appeared clearly on the original images and also maintained the circular shape features after processing. An example of the returned image can be seen in Fig. 13. The centre image is the intermediate stage of image processing where different objects are assigned different numerical values. This is represented by different colors for discrete objects. The overall detection rate for the 125–150 μm range was 68.8% as particles out with the range were not included despite being a significant number. The majority of particles appeared clearly on the original images. Some of the clarity was lost during processing, rendering some of the circular particles as

square or rectangular. An example of the returned image can be seen in Fig. 14. The results for the shape and size binning test can be seen in Table 4. The system captured and analyzed over 4500 individual particles by size and dominant shape. To check the accuracy of classification, images that included a variety of particles were selected and analysed, an example of which is shown in Fig. 15. It can be seen that the largest parameter does correspond to the overall shape for the objects that are detected within the image. VII. D ISCUSSION The detection rate for particles of grade 125–150 μm was lower than expected for a number of reasons. From analysing sample images, it was found that the vast majority of particles were captured by the sensor. However, clustered particles manifested as much larger particles and so were not included in detection rate for that grade. Further signal processing

472

Fig. 15. Upper - Captured image with processing stages (different colors in middle image denotes discrete object). Lower - Parameter graph for sample image frame with bar color matching discrete object.

would be required to eliminate this problem. Particles at this grade had much less clarity. This meant that a large number of grey scale levels existed within each object after processing. When Otsu’s method is then applied some grey levels were considered as black or white incorrectly. This resulted in inaccurate object representation of the particle size, preventing their inclusion in the detection rate. The lubricant color also exacerbated both problems. As the lubricant had a slight yellow shade, when processed it would manifest as mid to lower grey scale. This prevented Otsu’s method from being as effective. During the operation of the gearbox, the lubricant changes color due to chemical changes and the presence of unfiltered particles. In addition, XMP 320 samples obtained from operational gearboxes displayed significantly darker shades which would affect the sensor systems ability to distinguish particles from the surrounding lubricant. The refractive index of the lubricant (approximately 1.5) does not alter significantly during this change. Further image processing correction would be required to compensate for background color changes. Compared to optical particle counters, the smallest detectable particle was 2 orders of magnitude greater. This would prevent early wear trending offered by current systems. However the detectable particle size range is still relevant for failure diagnosis. 150–400 μm particles give strong indications of the type and source of wear when morphology is known, and this size range corresponds to that used by several commercial metallic particle counters. Therefore there is potential for its inclusion in a multi-sensory lubricant condition monitoring system that also included an optical particle sensor [12]. The accurate detection of optical sensors could be used to verify data obtained from an improved APS sensor.


Improving the detection rate may be achieved by a variety of methods. Using an APS sensor with smaller pixel size would increase the clarity of images, especially for small particles. An enhanced object detection algorithm that analyses the greys cale levels in greater detail may allow greater discrimination for particles in close proximity. The shape classification routine of the sensor programme was able to bin particles by dominant size and basic shape. However, the classification process requires improvement sin multiple areas to ensure accurate identification. In size ranges less than 150 μm, greater distinction between circular and square particles is required to prevent false classification. The cause of this error is a combination of clarity loss and basic shape parameters based on a small number of objects characteristics. This error in identification may account for the particularly large number of particles classes as square for range 100–150 μm in Table 4. Improvements may be made by recording a larger number of object characteristics such as solidity, equivalent diameter, total perimeter and convex area. In addition, further algorithm development would allow classification improvements. The intention of analysing particle morphology was to determine the type of damage and wear affecting gearbox components. The wear regime introduced to test shape classification capabilities of the sensor system was for test purposes only not intended to represent a particular fault type of failure as real samples with known faults were not available. However,if a large number of 150 μm particles classed as square were in reality circular, the wear regime may indicate cutting bearing wear. The high number of small circular particles could indicate bearing wear whilst the large number of small ellipsoidal and rectangular particles may indicate cutting upon the bearing surface.The larger number of particle classed as square for particle size groups >250 μm indicates that an advanced gear tooth fault has developed which is associated with large angular particles. Future tests should aim to simulate exact failure types using data obtained from operational wind turbine gearboxes. VIII. C ONCLUSION The principle of using a low cost APS sensor to detect and categorise wear particles within lubricant flow has been demonstrated. The sensor was able to distinguish between the surrounding lubricant, background pipe, air bubbles and wear particles to determine what constituted an object. Particles above 125 μm could be identified with a reasonably high detection rate. In addition basic particle shape could be estimated which can aid fault diagnosis and source identification. ACKNOWLEDGMENT The authors would like to thank Prof. S. Marshall, Deputy Head of the Electronics and Electrical Engineering Department, University of Strathclyde, for useful discussions on shape analysis and G. Brown, who is a Researcher with the Centre for Microsystems and Photonics, University of Strathclyde, for his help in obtaining microscopic images of


test particles. The authors would also like to thank Y. Wang, Ph.D. student with the Energy Technologies Institute, University of Strathclyde, for providing data on wear particle trends from wind turbine gearboxes. R EFERENCES [1] C. J. Crabtree, Y. Feng, and P. J. Tavner, “Detecting incipient wind turbine gearbox failure: A signal analysis method for on-line condition monitoring,” in Proc. Eur. Wind Energy Conf., Warsaw, Poland, 2010, pp. 154–156. [2] F. Spinato, P. J. Tavner, G. J. W. van Bussel, and E. Koutoulakos, “Reliability of wind turbine subassemblies,” IET Renew. Power Gen., vol. 3, no. 4, pp. 387–401, Dec. 2009. [3] T. Burton, N. Jenkins, D. Sharpe, and E. Bossanyi, “Offshore wind turbines and wind farms,” in Wind Energy Handbook, 2nd ed. Chichester, U.K.: Wiley, 2011, ch. 11, sec. 5, pp. 665–666. [4] P. J. Tavner, “Monitoring wind turbines,” in Offshore Wind Turbines: Reliability, Availability and Maintenance. Stevenage, U.K.: IET, 2012, ch. 7, sec. 2, pp. 116–117. [5] L. A. Toms and A. M. Toms, “Condition monitoring tests,” in Machinery Oil Analysis—Methods, Automation & Benefits, 3rd ed. Park Ridge, IL, USA: STLE, 2008, ch. 8, sec. 2, pp. 388–389. [6] R. Dupuis, “Application of oil debris monitoring for wind turbine gearbox prognostics and health management,” presented at the Annual Conf. of the Prognostics and Health Management Society, Portland, OR, USA, 2010. [7] A. Hamilton and F. Quail. (2011, Oct.). “Detailed state of the art review for the different on-line/in-line oil analysis techniques in context of wind turbine gearboxes,” J. Tribol. [Online], 133(4), pp. 44001–44018. Available: http://tribology.asmedigitalcollection. asme.org/article.aspx?articleid=1468840 [8] M. Heim, B. J. Mullins, H. Umhauer, and G. Kasper, “Performance evaluation of 3 optical particle counters with an efficient ‘multimodal’ calibration method,” J. Aerosol. Sci., vol. 39, no. 12, pp. 1019–1031, Dec. 2008. [9] D. Yang, X. Zhang, Z. Hu, and Y. Yang, “Oil contamination monitoring based on dielectric constant measurement,” in Proc. Int. Conf. Measuring Technol. Mech., Apr. 2009, pp. 249–252. [10] ISO Hydraulic Fluid Power—Fluids—Method for Coding the Level of Contamination by Solid Particles, ISO Standard 4406:1999, 1999. [11] L. Du, J. Zhe, J. E. Carletta, and R. J. Veillette. (2010, Mar.). “Inductive Coulter counting: Detection and differentiation of metal wear particles in lubricant,” Smart Mater. Struct. [Online]. 19(5), pp. 057001–057007. Available: http://stacks.iop.org/SMS/19/ 057001 [12] P. O. Vähäoja and H. V. S. Pikkarainen, “Trends in industrial oil analysis—A review,” Int. J. Condit. Monitoring, vol. 1, no 1, pp. 4–11, May 2010. [13] D. Anderson, “Wear particle atlas (revised),” NAEC Advanced Technology Office, Lakehurst, NJ, USA, Tech. Rep. NAEC-92-163, Jul. 1982. [14] S. Ghosh, B. Sarkar, and J. Saha, “Wear characterization by fractal mathematics for quality improvement of machine,” J. Qual. Maintenance Eng., vol. 11, no. 4, pp. 318–332, Feb. 2005. [15] S. Raadnui, “Wear particle analysis—Utilization of quantitative computer image analysis: A review,” Tribol. Int., vol. 38, no. 10, pp. 871–878, Oct. 2005. [16] T. H. Wu, J. H. Mao, J. T. Wang, J. Y. Wu, and Y. B. Xie, “A new on-line visual ferrograph,” Tribol. Trans., vol. 52, no. 5, pp. 623–631. Jul. 2009. [17] A. C. Bovik, “Morphological operators for feature detection,” in The Essential Guide to Image Processing. London, U.K.: Elsevier, 2009, ch. 13, sec. 5, pp. 309–317. [18] MATLAB 2012, MATLAB Version 7.14.0.739 (R2012a), The MathWorks Inc., Natick, MA, USA, 2012. [19] K. Williamson. (2010, Nov. 2). “Smooth operator—Wind turbine gearbox lubrication,” Renew. Energy Focus [Online]. Available: http://www.renewableenergyfocus.com/view/13637/smooth-operatorwind-turbine-gearbox-lubrication/ [20] M. Petrou and C. Petrou, “Image segmentation and edge detection,” in Image Processing: The Fundamentals. 2nd ed. Chichester, U.K.: Wiley, 2010, ch. 3, sec. 2, p. 37.

473

[21] B. Jähne, “Feature extraction,” in Digital Image Processing, 6th ed. Berlin, Germany: Springer-Verlag, 2005, ch. 11, sec. 1, pp. 299–330. [22] T. Wu, J. Wang, Y. Penga, and Y. Zhanga, “Description of wear debris from on-line ferrograph images by their statistical color,” Tribol. Trans., vol. 55, no. 5, pp. 606–614, Sep. 2012. [23] Z. Lu, X. Yan, C. Sheng, and C. Yuan, “Wear trend analysis based on ferrograph cover area rates compares with other methods,” in Proc. IEEE Veh. Power Propuls. Conf., Dearborn, MI, USA, Sep. 2009, pp. 1792–1797. [24] M. S. Laghari, F. Ahmed, and J. Aziz, “Wear particle shape and edge detail analysis,” in Proc. Int. Conf. Comput. Autom. Eng., Singapore, 2010, pp. 122–125. [25] L. Kirkup, “Dealing with uncertainties,” in Experimental Methods. Milton, Australia: Wiley, 1994, ch. 4, sec. 3, p. 58. [26] N. Otsu, “A threshold selection method from gray-level histograms,” IEEE Trans. Syst. Man Cybern., vol. 9, no. 1, pp. 62–66, Jan. 1979. [27] O. Marques, “Color image procesing,” in Practical Image and Video Processing Using MATLAB. Hoboken, NJ, USA: Wiley, 2010, ch. 16, sec. 5, pp. 414–415. [28] M. Petrou and C. Petrou, “Image segmentation and edge detection,” in Image Processing: The Fundamentals. 2nd ed. Chichester, U.K.: Wiley, 2010, ch. 6, sec. 1, pp. 542–544. [29] J. T. Carsetensen and M. Dali, “Particle size distribution in mesh cuts and microscopically estimated volumetric shape factors,” Drug Develop. Ind. Pharmacy, vol. 25, no. 3, pp. 347–352, Mar. 1999.

Andrew Hamilton received the M.S. degree in civil engineering from the University of Strathclyde, in 2009. He is currently working toward the Ph.D degree at the Wind Energy Centre for Doctoral Training, University of Strathclyde, Glasgow, U.K. His research interests include machine condition monitoring and lubrication analysis. He is an associate member of the IMechE.

Alison Cleary is a member of the Centre for Intelligent Dynamic Communications. She received the M.Eng. degree in electronic engineering with physics in 2000 and the Ph.D. degree in optical sensors in 2005, both from the Department of Electronics and Electrical Engineering, University of Glasgow. She has interest in sensor and characterization techniques, and has worked on optical sensors, optical generation/detection of ultrasound, spectroscopy and condition/process monitoring at the Universities of Glasgow and Strathclyde. Dr. Clearyis is a member of the IET.

Francis Quail is the Director of the University of Strathclyde’s Centre for Advanced Condition Monitoring and Director of the Centre for Intelligent Asset Management. He received the Ph.D. degree in mechanical engineering from the University of Strathclyde, Glasgow. His research interests include gearbox condition monitoring, asset management and predictive maintenance of turbo machinery. Dr. Quail is a Fellow of the IMechE and a Fellow of the CMI.