It is quite normal nowadays for modern manufacturing processes to implement comprehensive control and monitoring systems. One of advantages is that they can provide huge volumes of data in a timely and usable manner: including process flows, pressures, temperatures, qualities and energy use.
These parameters may be recorded on a minute by minute basis and databases that contain tens, even hundreds, of thousands of records of operating conditions can be made available.
The work has aimed to identify new opportunities for reducing energy costs, and improving output and yield, whilst still maintaining product quality.
Data Mining – An Overview
Data mining is a relatively new approach, which aims to extract knowledge from huge volumes of data. It can be defined as “the semi-automated process of discovering non-trivial, implicit, previously unknown, potentially useful and understandable information from large, historical and disparate data sets.”
Data mining techniques can broadly be divided into three categories:
- Manual identification of patterns
- Semi-automatic identification
- Automatic identification (which has been applied in this case).
The tree shown in Figure 1 shows part of a decision tree, which relates to the particle size of a powder in a production plant.
Figure 1. Example Decision Tree
In-Line Particle Size Analysis – Overview
The use of long-established laser diffraction techniques within a fully automatic optical system for the determination of particle size and distribution in dry powders offers new possibilities for process monitoring and for improved and consistent product quality, together with reduced wastage and rationalisation of resources.
The sampling technique makes sure that real-time measurement under isokinetic conditions is possible and hence a full and accurate characterisation of the particulate.
Operations at Nexpress include pre-blending, compounding, extrusion, pelletising and granulation, followed by pulverisation, classification and sieving. The product is toner, a fine powder product with a tight particle size distribution. A schematic of the pulverising, classifying and sieving operations is shown in Figure 2.
Figure 2. Process Flow Schematic Diagram
Toner is vacuum-conveyed from bins to the filter receiver and from there, via a series of valves, to the pulveriser feed hopper.
A screw conveyor transfers the toner to the base of the pulveriser, which reduces the particle size to below 12 microns using a fluidised bed opposed jet mill with internal air classification.
The pulverised product is then fed to a dedicated air classifier, in which a fine fraction is removed. The coarse product from the classifier is then passed through an air-jet sieve to give the finished toner.
Analysis of the Process Data
Operating data was collected at 30s intervals from the process monitoring system. This data included all the key process parameters. In total, more than 29,000 records representing in excess of one week of operation were obtained. Figure 3 shows the product particle size for a typical day.
Figure 3. Typical Variation in Particle Size Distribution
Figure 4 shows the frequency distribution of process electricity costs over the entire data set:
Figure 4. Electricity Cost Figure Variation
The data was analysed using data mining techniques, to identify patterns, which explain the variability in the key process performance parameters. Figure 5 below shows part of the decision tree where the ‘outcome’ (or objective) is particle size.
Figure 5. Part of Decision Tree with Particle Size as the Outcome
The tree identified the pulveriser bed pressure as the most significant factor affecting particle size. A second factor is the speed of the pulveriser rotors – higher speeds reducing the average particle size.
The decision trees for each of the process lines were examined in relation to each other and it was expected that good correlation would be seen between them. However, this was not the case – in particular with the outcome set as “yield”.
Investigation of this by using the re-heat valve output as the outcome showed that the classifier wheel gap airflows were the most important factors, and by tracing this route through it was found that the volume of air through these gaps was far to high.
From this discovery it became apparent why the yield of product from this process line was significantly inferior to that of the other lines: it had higher airflows, and airflow directly affects the classification cut point. That is, higher airflow = coarser cut point.
The larger air volume being utilised on this system gave rise to higher energy usage due to the additional power required to compress this extra volume of air, and in running classifier motors and exhausters at higher speeds to compensate for the coarser cut point.
The data mining exercise on these production lines has shown that some of the parameters that are thought not to be so important can have a significant effect on the overall plant performance and energy usage.
Many of the process variables show long and short-term cyclic variations, whilst the long term ones are relatively easy to investigate and improve; it is the shorter term ones which contain the greatest improvements in product quality.
Applying Data Mining
The discussions above illustrate that the analysis of data using data mining and in-line particle size analysis techniques can identify new opportunities to improve process performance, and also quantify the impact of those new opportunities.
The toolkit contains not only the facility to mine new data easily, but also ‘expert system’ capabilities that can be used to apply the rules to process data, as it becomes available, and provides the process operators with online advice.
The work carried out has illustrated that data mining can be used effectively to improve toner-manufacturing performance. At Nexpress, improvements to yield, quality, energy use and throughput have been identified, including:
- Reduction of electricity usage by adopting the identified lower usage scheme
- Reduction in out of specification material produced during start-up and shutdown
- Improvement in yield by monitoring air flowrates
- Improved product quality by better understanding of process step changes
- Fault diagnosis of utilities and major plant items
- Tighter particle size distribution by less frequent and more informed adjustments
- Improved plant output through smoothing of cyclic changes
- Process investigation by comparing optimum schemes against actual operations
This information has been sourced, reviewed and adapted from materials provided by Hosokawa Micron.
For more information on this source, please visit Hosokawa Micron.