Machine vision takes the quality control of biofuels to the next level


Quality control is more important than ever


The internal load quality variation in a fuel type load often causes problems in sampling and continuous quality control. Fuel that was thought to be of uniform quality can, in fact, be a mixture of this and that and this is not easy to discover by quick visual assessment. You can see this for yourself by taking a look at recycled wood or waste components before and after processing. Raw material that can be highly heterogeneous is crushed to the desired size and at the same time, recyclable metals and foreign substances that are harmful in combustion are extracted.

The crushed fuel is transported to power plants and received there with little else in the way of information than these: supplier-fuel code-applied load. With recycled wood, the code (Statistics Finland) used in reception does not even indicate its quality even though an ABCD-type quality classification exists for recycled wood. Even this classification is advisory in nature, not a set of values that can be continuously measured. Traditionally, quality assurance involves assessing the general look of the load when unloading or taking samples from the load either manually or by using an automated sampler. At this point, however, it is impossible to tell what the goods originally looked like and whether the processed fuel yielded representative samples that reflect the whole load’s quality.

The situation was equally poor for us too when we verified the functionality of our equipment by comparing the online measurement data with the reference samples. The samples are, indeed, taken from falling flow at regular intervals. Yet they tend to get taken every 1-2 minutes, as it is neither sensible nor very pleasant to monitor the whole delivery lot that can take anything between 10 minutes to 36 hours to unload depending on the facility. When sampling, we try to visually verify that the taken samples look similar to what mostly passes by on the conveyor. This phase is not as easy as one might think, especially when dealing with a closed, dark scraper conveyor. Successful sampling is not 100 % guaranteed even after the oven drying results are nicely in line with our measuring device’s continuous data!


The role of machine vision in quality control


We have identified the problem and our shortcomings in forecasting success on a Gaussian curve when dealing with challenging fuel components such as recycled wood. To tackle this issue, our FuelControl systems are equipped with a computer vision system that takes snapshots of the fuel flow at regular intervals. At first, we evaluated load quality with spot checks from photographs, but this also proved quite slow and painstaking. That is why we automated fuel type recognition so that the system classifies the photos according to the fuel types used at the facility. 

With recycled wood, we took things even further and developed a tool that utilizes neural networks – the Self-Organizing-Map (SOM). As input, the SOM utilizes data produced by the X-ray machine, computer vision system, and the fuel supply system, processing the delivery lot with the accuracy of each X-ray picture. Starting from the classification of individual X-ray pictures, we move on to calculate a database subclass – that the recycled wood in question represents – with the accuracy of 30-60 seconds. Now this information helps us to assess the whole load’s quality on a general level. We can verify whether the reference samples we took represent the load’s main quality with high probability or whether they deviate from it.


Small system, big benefits


Using SOMs to analyze recycled wood loads in more detail also yields valuable information to both the facility burning them and to the fuel supplier. If the SOM starts detecting abnormal deviation in the incoming material flow, it is usually caused by lower quality that can, at worst, result in costly stoppages for the boiler. Another reason for deviations can be caused by recycled wood that is drier than expected, which in turn can increase the risk of explosion and fire due to dust formation. The SOM can also indicate changes in the fuel’s manufacturing process, and with this information readily available, we can locate the reason, fix the problem, and potentially even prevent it from occurring. The easiest fix can involve adjusting the crushing line speed, after which collecting metals using magnets and eddy current separators is improved and the fuel supplier can then benefit from this added value.

In the future, Inray software can be used to search for and compile statistics on color deviations in received solid fuel. Color is added to the pictures by, for example, so-called green chips and painted wood. Burning large quantities of green chips adds chlorine to the boiler, and painted recycled wood is not a popular material in many boilers because of the impurities it contains. In the past, lead was added to some paints, and these compounds are highly soiling and corrosive. 

The computer vision system enables analysing biofuels even more comprehensively than before. The series of pictures below provides an example of how a camera can be utilised at a power plant. The same soaking wet and indefinable material was delivered to the facility under three different fuel codes. The computer vision system is in itself a highly efficient solution and combining it with the company’s FuelControl system enables an extensive, real-time quality control system for biofuels.