Open Cut » Maintenance & Equipment
Microscopic examination of wear debris for machine condition monitoring was first introduced in the 70's It is reliable but, for the human analyst, can become a tedious and time consuming task. This has limited its widespread usage. In 1994, CMTE received an ACARP grant to demonstrate the feasibility of developing a computer based tool to automate the process
The aim of this project is to develop a computer-based pattern recognition procedure to properly identify contaminants and wear particles in lubricant oil, and to indicate the type and severity of wear. Its success will enable us to develop a system whereby such analysis can be done automatically
Equipment maintenance is 30-50% of the operation costs in the mining industry or $10-15b/year for the nation. The final aim is to create a new analytical instrument for machine condition monitoring through used oil analysis, with applications beyond the mining industry. A market size of 100+ instruments would be a conservative estimate for Australia
This was originally a two year project. It was a joint project between CMTE and ACIRL. Oil samples were collected from two 784 coal haulers at BHPAC's Goonyella mine and filtergrams were prepared by ACIRL and forwarded to a wear consultant, Mr Peter Ball of Machine Reliability Services.
A written report was made from a visual examination of the filtergram and returned to the Goonyella mine. The filtergrams were then passed on to CMTE for image analysis. Due to the slow start of the project, and some difficulties with the filtergrams, an extension of one year was given. In this extension, no further oil samples were collected and most of the work was conducted at CMTE.
Objectives
During this century there has been a gradual shift from the traditional maintenance schemes of breakdown, to preventative maintenance. Today this shift continues towards condition maintenance. It is based upon the principle that maintenance need only be performed if the machine is about to fail. This predictive ability can only be achieved with regular monitoring of the machine condition.
There are three primary components of such monitoring: performance, vibration and lubrication. In this work, we are primarily concerned with the third, and in particular, the analysis of wear and contaminant particles within the lubricating oil. Wear particles contain valuable information about the wear processes that occur within the machine.
Unfortunately, assessing this information must be done visually by highly trained operators. This makes the analysis both expensive and time consuming. To extend the benefits of such analysis to industry it is essential that it becomes automated
This project aimed to develop a computer-based pattern recognition procedure to properly identify the contaminants and wear particles and to indicate the type and severity of associated wear. Success would enable the development of a system whereby such analysis could be done automatically.
Results
In the first year of this project a literature review was conducted to establish the nature and location of various research centres around the world which might be working on the problem. Although there are a number of groups trying to characterise wear debris, none of the work is directly related to the mining industry and little success has been achieved in regard to automation
In the same year a prototype vision system was built to acquire images of particles on a filtergram. It consisted of a standard TV camera fitted to an optical microscope and connected to a PC frame grabber. The images were acquired and processed with a number of different commercial and public domain image processing systems. Since none of these systems was able to calculate the shape and textural features needed for the classification of the wear debris, it became necessary to write our own image processing software.
In the second year of the project, over eighty oil samples, were collected from two haul trucks at Goonyella. From these samples, images were acquired and processed. In previous work, there appeared to be a clear difference between the fractal dimension of road dust and wear, but in this work, this feature alone was insufficient to classify the different types of wear debris found in used oil.
Although preliminary analysis of a number of shape and textural features showed some promise (eg the cutting particle could be easily identified from the other wear modes) there were a number of significant failures. In these failures, the software was unable to correctly segment, or interpret, the particle from the background.
In hindsight, this came as no surprise because the images were acquired with little prejudice, ie the particles were chosen randomly to duplicate the behaviour of an automated system. This meant that the particles could be poorly lit, partially out of focus, or obstructed by agglomeration. Since all of these problems will occur in an automated system they need to be resolved.
In the third year of this project, strategies for the success of the project were re-evaluated. Firstly, a high-resolution digital camera was purchased to improve the quality of the images and a further 180 images were acquired. Secondly, our focus of activity shifted from classification to segmentation.
The most significant impediment to the classification of wear debris, is not the need for more sophisticated classification rules, but the ability to segment the particle cleanly and accurately from the background. When a wear particle is generated in a machine, the texture, colour and shape of the particle reveals the wear mode that generated the particle. It is these features that identify the wear debris.
Unfortunately, we do not often get to see this particle because it has been attacked whilst it is in the lubricant. What we see on the computer screen is the result of attack whilst the particle is in the lubricant, and distortions generated by the microscope and vision system used to acquire the image. To solve this problem we need to establish the nature of any attack or distortion upon the particle. One way that this can be achieved is with a technique of over segmentation and reconstruction.
The particle is first broken up into a number of small segments. This process can be based upon the colour, texture and shape of different parts of the image. The particle is then rebuilt, starting at the central segment, based upon the relative properties of each segment. The shape and surface features of the particle are then derived from all (or part) of the rebuilt particle.
Any knowledge gained during the reconstruction is used to influence the level of confidence in the classification.
This new technique is able to double the segmentation success rate from 39% to 79% of the sample.
Achievements
What has been achieved in this project is the development of a sophisticated vision system capable of analysing filtergrams found in the mining industry. It consists of a high resolution digital camera and more than 16,000 lines of image processing software. The software is based upon a technique of over segmentation and reconstruction, which is both unique and robust. This allows the software to work under a wide variety of lighting conditions, and more importantly, it enables the software to have a level of confidence in the validity of the features that it calculates. It means that the software is able to reject an image if it is too difficult to segment and lowers the chance of a false classification. Unfortunately, the delay involved in improving the software did not allow us time to develop the classification rules necessary to identify the different wear modes. Therefore, given the original aim of this project:
" to develop a computer-based pattern recognition procedure to properly identify the containments and wear particles present in lubricating oil, and to indicate the type and severity of associated wear."
then this project has not succeeded, but if we take a wider perspective and examine the overall, or commercial aim of this project:
" to develop a piece of analytical equipment that can automatically classify wear particles on a filtergram."
Then we have made a significant stride towards this goal.