Get best of the week

Accelerating the implementation of AI projects in the Segezha forest holding



An interesting case from Dmitry Bocharov, vice president of internal control and audit at Segezha Group, was heard at our RAIF artificial intelligence forum . Dmitry told how machine learning tools are used in the largest woodworking holding in Russia and how obstacles to implementation are overcome. We give him the floor.

First, a few words about the company


Segezha Group is one of the largest vertically integrated woodworking forest holdings in the country.



I am sure that many of you have heard about our company. In the end, if you saw a paper bag in IKEA, "ABC of Taste" or "Auchan", then it was produced, including, by our company.



Now I want to convey, on the one hand, the value of artificial intelligence in solving specific business problems, and on the other, tell about our experience and even the design pain we encountered when we were dealing with this case.

Harvesting process


To begin with, a little about how logging is done:



Forest is cut using special equipment - harvesters. Then, the workpieces are transported to the warehouses by timber carriers with manipulators, so that from there by rail or by road to the mills where pulp, paper, plywood, lumber and other paper products are delivered.

Timber measuring mechanism


One of the key problems not even of Segezha, but of the entire industry is the process of measuring this very forest product, or rather, logs.



How is this happening now?



With the help of a special ruler, the height, length and width of the stack are measured, which is multiplied by various coefficients, prescribed even in the USSR in various state standards and industry standards. The most basic coefficient is the “full wood coefficient”, that is, the indicator of the actual number of cubes in a stack minus the gaps between the logs. This is where the problem of the human factor arises - if the employee is inexperienced, he is likely to measure inaccurately.

However, the biggest difficulties from the point of view of the audit are deliberate violations, since the total salary of the employees delivering the forest to us is less than the cost of timber in a timber truck (one cubic meter costs 4-5 thousand rubles). A little mathematics - and here you have the opportunity for various conspiracies, abuses, manipulations .... Then it’s impossible to understand how much “forest” there really was. There is a car, there is even an act with the number of logs fixed in it, but if there were just so many of them there are no confirmations, except for those that they measured with a ruler. And here the problem is not even that we do not trust all our employees or the employees of our contractors. There is simply a critical lack of clarity in this process, first of all, real documentary evidence that something was really measured.

Modern approach


We have developed a special algorithm that, based on a photograph, using a neural network, not only determines the number of logs and the diameter of each log (also an important indicator for us) and considers the same coefficient of full wood, but most importantly, it takes it not from some GOST, but corrects it for specific stack of forest products.

These photos are tied to the geolocation of the car and are stored in a special database. Therefore, after we can always take and verify: was this forest really and how much was it. The plans for the next couple of months are to train the system so that it can automatically compare departing and arriving cars by heuristic search. First, the system photographs the car when it leaves the plot from the forest, then the second time when it already arrives at the plant. Further, it automatically checks the photos and fixes whether some of the logs were removed from above and whether they were replaced. Such automatic control is based on artificial intelligence. This greatly simplifies the work of, for example, the security service, because we can’t run through all the forests of Russia (and we have a cutting area of ​​nearly eight million hectares!), Just like we can’t control every lumberjack,because it is expensive and inefficient.

When we tried to implement the system together with the company that made the pilot project, we started with the Telegram bot to demonstrate the capabilities of this algorithm.



By the way, this Telegram bot is still there.

The main problems and their solution


We faced the basic problems that all companies that implement artificial intelligence or related projects face. Firstly, the budget issue is where to get the money from. Secondly, cost justification issues . Thirdly, the largest block of problems is procurement procedures and tenders .

For ourselves, we solved this problem as follows: Segezha Group has the so-called “Pilot Projects” in the procurement procedures. If we want to introduce something new and small, besides previously undescribed, there is no need to invent TK. We don’t yet know how this will work, therefore, writing up the appropriate TK is just a waste of time. There is a certain budget for such projects, and by decision of the procurement commission it is absolutely officially possible to choose one of the contractors. Thus, our company works in the spirit of a startup. We are ready to lose this money, but we can try to solve a specific problem.

My colleague, Segezha vice president of IT, at one of the forums talked about one of our projects like this: it cost several million rubles, but it could bring about three hundred million. We took a chance, made a “pilot” and as a result it paid off many times - maybe not a hundred, but at least ten times for sure. Obviously, such experiments suffer losses, but you can and should try, because any implemented case is a very valuable experience. The use of technologies developed in specific business problems is bearing fruit. But here it is necessary to know the measure as well - artificial intelligence and machine learning should not be implemented everywhere, if only to be implemented.

Another internal life hack: we agreed with colleagues (with financiers, purchasers and company management) that we will reinvest part of the money that similar projects bring to us in the future - that is, we will constantly invest the money saved in new technologies and most promote similar stories in Segezha.



Now we are just finishing piloting the “wood” case. To make it clear on economic effects: the error in the method of measurement with a standard ruler according to GOST is 5%, but in fact it is much larger. Segezha Group annually harvests and buys timber for 15 billion rubles. Even if you take 1% of this amount, this is a significant loss. And such projects, which at the same time do not cost billions or even hundreds of millions of rubles, allow these risk zones to be closed. Maybe there is no direct economic effect (that is, we will not earn more or we will not have new production), but from the point of view of preventing possible losses in logging, rather high efficiency is obvious.

I think many people are interested in the timing of the manufacture of such prototypes and I want more specific numbers. I can’t name the numbers for obvious reasons, but I’ll designate the most problematic point - getting relevant data. For example, what photo or what data to take for training a neural network? We cannot use the results of manual measurements (the same measurements with a ruler), because then the algorithm will work on incorrect data. We have to take each timber carrier and make the so-called "point": the logs are completely unloaded from the timber truck and each is measured in diameter and length. Thus, it is possible to determine the reliable volume of all wood with minimal error. Another feature is that the wood is different: pine, spruce, larch ... Accordingly, each has its own characteristics in the dimension. To measure everythingmy employees had to travel to different regions - to Kirov, Arkhangelsk region, Krasnoyarsk, Karelia - and measure every timber carrier there. So the main time (about two weeks) was spent on collecting a sufficiently representative sample for training the model.

Posted by Dmitry Bocharov, Vice President, Internal Control and Audit, Segezha Group

More articles:


All Articles