Wear debris particle analysis is an equipment health monitoring technique used to identify possible failure modes in various machine components such as those found in engines. One of the first stages of analysis often involves the microscopic examination of particles from the component’s magnetic drain plugs or filters within the lubrication system. However, the diagnosis may not be consistent between technicians due to the subjectivity of their judgment. A software tool capable of automatically classifying the images of wear debris particles has been tested using a database of 800 images. It has shown that using automatic image analysis classification of wear debris particle images is more consistent, accurate and informative when compared to manual classifications performed by wear debris experts.
Attempts have been made to use image processing techniques to classify wear debris particle images in the past. The Computer-Aided Vision Engineering (CAVE) system is a good example of the application of standard image processing techniques which have been applied. Particle images are characterized by their size, shape, edge detail, edge curvature and surface texture. To obtain parameters that describe the shape of a particle, the Fourier technique is used. Edge detail is characterized using curvature analysis or fractals. Fractals are also used to determine the surface features of the particles. However, standard image processing techniques are challenged by marginal or poor-quality images and data noise.
In light of the difficulties encountered when using standard image processing techniques to automate the identification of wear debris particles, the development of computer-aided tools for particle classification has reverted to interactive systems such as the Systematic Classification of Oil-wetted Particles (SYCLOPS). With this software, an operator is required to select various characteristics of a particle using a set of stylized drawings. The software then selects the diagnosis consistent with the information supplied by the operator. However, the subjectivity of technicians’ judgements concerning the input to the system still indicates that a diagnosis may not be consistent between technicians.
This is because people are good at delivering complex qualitative descriptions, but they cannot make absolute measurements of their subjective impressions. In contrast, automatic image analysis offers the potential to collect statistical information associated with large numbers of particles and to create a reference database against which particles of unknown type and origin can be automatically classified. When successfully implemented, such a system could be used not only in the remote diagnosis of possible failure modes in oil-wetted components, but also in the study of archived time series of wear debris particle images to develop real-time prognostic capability.
Current work proposes using image processing techniques to remove subjectivity from the analysis of wear debris particle images. Image processing is a relatively young discipline and, as stated above, the results theoretically offered by standard techniques are often difficult or impossible to realize when these methods are applied to real image data. In response to this need, novel techniques have evolved that are able to cope with image data compromised by excessive data noise. In particular, natural shapes can be efficiently categorized using a technique that does not deal with exact representations, but rather has the ability to abstract generalizations in order to categorize both interclass similarities and the broad differences between classes of particles. This coincides with our own intuitive ideas concerning shape and texture.
Figure 1. Four Classifications of Wear Debris Particles
Such image analysis technology has been implemented and tested using the SYCLOPS database, an 800-image database compiled and classified by wear debris experts. The database is also used interactively to assist operators in the classification of unknown wear debris particle images. Four particle classifications are used: fatigue, severe sliding, mild sliding and cutting (Figure 1). To determine the parameters best suited to the discrimination of the four categories, the 800 images were segregated in the following way: 25 images from each of the four categories were selected as being highly representative of their classification. These 100 images were used as a training-image database, and the remaining images were used to test the resulting system.
Once the training image database had been selected, the various image-processing algorithms were implemented and assessed with respect to their power to discriminate between the various morphological and textural properties associated with the particles. It was concluded that five parameters were sufficient to distinguish between the four categories: a size parameter, two shape parameters and two surface texture parameters. A hybrid search-tree/multivariate- discriminant-analysis technique was developed to classify the unknown particle images. Figure 2 shows a schematic representation of the resulting system.
Figure 2. Wear Debris Classification
The order of classification was significant. In particular, it was important not to use size as a discriminator at the first level of the tree as this would exclude the detection of fine cutting particles. At the first level of the search tree, two morphological properties were used to discriminate between the long thin shape of the cutting particles and the other particles. Figure 3 shows these parameters used in distinguishing between the cutting particles shown in red and the remaining particles, shown in blue, green and cyan.
Figure 3. Use of Shape Parameters to Classify Cutting Particles
The next level of the tree concerned size discrimination. Figure 4 shows that the size of the mild sliding particles (green shapes) falls below 100 microns.
Figure 4. Thresholding on Size to Eliminate Mild Sliding (Benign) Particles
Finally, it remained to differentiate between fatigue and severe sliding particles. This was carried out using two parameters that characterize the different surface features of the particles. The fatigue particle has a granular surface texture made up of small cracks that do not exhibit any particular pattern. In contrast, the severe sliding particle has an association of local features that are correlated over the particle surface. This correlation is in the form of straight, parallel lines representing the striations associated with severe sliding. Figure 5 shows the processed images of the example fatigue particle and the severe sliding particle from which the granularity and striation parameters were extracted.
Figure 5. Surface Features of Fatigue and Severe Sliding Particles
Figure 6 shows the discriminating power of these two surface feature parameters. The results associated with the severe sliding particles are shown as blue shapes and those associated with the fatigue particles as cyan shapes.
Figure 6. Surface Feature Parameters Used to Discriminate
Between Severe Sliding and Fatigue Particles
The system was tested using the images remaining in the database once the training images were selected and their parameter distributions used to deduce the optimum classification routine. Each test particle image was automatically classified and the classification compared to that of the wear debris expert.
It soon became evident that the four categories of wear debris particles currently used are entirely notional. That is, they do not reflect the reality of the various gross physical differences between particles assigned to the same category by wear debris experts. The testing revealed that there are subcategories of particles, the study of which may provide an insight into wear processes that is not currently available. In particular, the size of the particles is not a continuum of values, but rather there are distinct distributions. Figure 7 shows the different categories of particles grouped according to size distribution where the results of processing have also revealed distinct statistical distributions with respect to the other four parameters.
Figure 7. Particle Groupings Suggested by Automatic Classification
What is clear is that the term “particle” is a misnomer for many of the images viewed, especially with respect to the fatigue and severe sliding particles in the size range of 100 microns to 1.0 mm where many are simply fragments as opposed to particles with a distinct shape. The size range above 1.0 mm is the most likely to be populated by particles as opposed to fragments of particles. Figure 7 provides examples of the groupings in this range. There are two distinct types of cutting particles: one is a long, thin, curved shape while the other is a much chunkier particle with a ruched (pleated or fluted) surface texture. Also in this size range is a type of severe sliding particle that exhibits a distinctive triangular shape such as that of a trowel.
The groupings in the size range 100 microns to 1.0 mm again show the two distinct types of cutting particles. With respect to the fatigue and severe sliding particles, there is considerable overlap between these two types where particles exhibit both radial cracks and striations. This reveals a further weakness in the current system of classification in that it limits the operator to either/or decisions. In particular, there is no way of representing where a particle shows evidence of both fatigue and severe sliding wear. Generally, it was found that the operator preferentially chose to label the particle as fatigue if it had radial cracks irrespective of the striated nature of the surface of the particle, effectively throwing away information.
Moreover, when the degree of fatigue and severity of striations are taken together, a transition zone is revealed where the wear debris experts’ opinions of classification became interchangeable. This makes such classifications ineffective and once again, significant information is lost. If an automatic classifier is used, then the evidence of both the degree of fatigue and severity of sliding can be included in the diagnosis. Figure 7 also shows two distinct shapes of severe sliding particles in the 100 micron to 1.0 mm range. One is flake-like and heavily striated, the other resembles a long thin splinter. It is clear that two different mechanisms have produced these two distinctive shapes. Use of the extra information provided by automatic classification may allow a more sophisticated diagnosis of wear mode.
Analysis of the test results revealed the extent to which wear debris experts are hampered by both the limitations of the human vision system and the optical systems used to visualize the particles. This is due to humans being poorly equipped to deal with extremes of lighting conditions and restricted depth of focus. Testing showed that some particles were wrongly classified by the wear debris experts due to either poor lighting of the specimen or blurred images. This was not the case for the automatic system, which can automatically compensate for both of these eventualities.
Acknowledgment The author would like to thank the QinetiQ Fuels and Lubricants Centre for its sponsorship of this work and for allowing access to the SYCLOPS image database. Special thanks to David Hodges and Tim Nowell for their advice and technical support throughout the project.
For additional information on techniques for automated particle classification, please review the following article: “Condition Monitoring and Predictive Analysis of Tribosystems by Wear Debris”, also found on the Web site.