Not on schedule, but in accordance with the prediction
The solution we are presenting here today is exact Machine Learning that is being talked about so much. For all 11 years of existence of the company we are engaged in automation of the industrial domestic and foreign enterprises, that is we introduce "classics" software — SAPR, CAM, PDM, PLM systems. At the same time, we always use modern innovative technologies in our projects, including augmented reality, virtual reality, BIG DATA, Machine Learning, and then BIM. This is our constant innovative component.
One of our most well-known and desirable product is equipment maintenance and repair system. It provides automation of equipment inspections and scheduled preventive maintenance. We have been actively developing and improving this system for more than ten years, using emerging technological innovations in it. For example, a few years ago we realized that it is possible to use BIG DATA and Machine Learning technologies to analyze the data stored in our systems. This was the impetus for the creation of our machine learning system capable of predicting equipment failure. A well-known problem of many of our companies – the equipment is repaired in accordance with the plan, schedule, and not on the principle of predictive maintenance. According to various statistics, up to 50% of equipment repairs are carried out not when it is necessary, but much earlier. For enterprises, this eventually turns into a significant loss of time and money. In this case, we offer our customers the tools to save both. With the help of machine learning systems (artificial intelligence, predictive models), you can achieve different effects. For example, reducing or increasing the operating time of equipment, reducing repair costs, optimizing maintenance costs. Our product provides reduction of repair cycles, allows you to build the optimal modes of operation of the equipment in terms of different parameters. It is possible to obtain an optimal regime by developing a resource, and it is possible-by reducing costs. This can be achieved through mathematical models (artificial learning).
How does this system start? The first stage collects the data necessary to start its work. They are then processed so that you can use them to create a model. That is, the data is preprocessed, analyzed, and distributed into the desired groups. Then there is a predictive analysis, including the detection of anomalous states, the construction of various models, the identification of complex modes of operation, training of the system (the model itself). And this cycle: data collection — predictive analysis — training — is repeated until the desired result is obtained, that is, the accuracy of predicting equipment failure will be at least 98%.
Of course, sometimes also happens that in some projects the desired figures are unattainable, or they can be achieved at such a price that the use of such approaches ceases to be justified economically. The price is always a fundamental point. And here, the most important thing is, first of all to take into account the goals of the project and specific tasks that are supposed to be solved through the implementation of our system. Often in these projects, the accuracy of predicting can vary in prices. For example, to achieve accuracy 90% accurate — this one costs, and 91 percent- another. The choice is for the customer.
When the targets for accuracy are reached, we move on to the last stage — visualization. This is providing information directly to the employees who will work with it.
I will give two cases for example-these are implemented projects where the results are already visible.
The first example is the solution of the problem of reducing the cost of repair and maintenance of tank cars at the enterprise engaged in transportation of liquefied gas. The customer has its own fleet of such tank cars, which is about five thousand units. For this project, data were obtained for the period of five years on the maintenance and operation of tank cars. On their basis, a mathematical model was built, 150 parameters were analyzed at the entrance, on which the condition of tank cars can depend. As a result, we found that only fifteen parameters have a direct impact on the condition of the tank car, five of them are crucial. Then a system was developed for the customer that works very simply and efficiently. The operator in CRM system makes shipment, it means sets a task to load one of tanks with a certain fuel and to send to the customer. And when the "shell" is formed, the system gives information that the tank with a probability of 64% will not reach, in the way it will break the wheel set. Statistically proved the effectiveness of this system, primarily due to it prevents breakdowns in the way, the most difficult for the customer.
Another example relates to the operation of gas turbines. We had at our disposal historical telemetry data for a long period of time collected by the MES system. We analyzed the data on how this system works, what can be predicted from it. In fact, the enormous economic effect of the fact that we have learned to predict the failure of several key turbine units has been confirmed. But there is a very interesting point related to machine learning. At the enterprise two gas turbines, one year of release, from one producer, one brand, at one time were put into operation. When the project started, we made a mathematical model for one gas turbine, implemented it, and then made sure that the result satisfied the customer. When we began to implement this model for the second turbine, the result was much worse. This is quite obvious for machine learning, but it is not obvious for the customer. Turbines during these 20 years worked differently, they were differently serviced and had different production cycles. As a result, their state is significantly different and therefore the mathematical model for one turbine cannot be applied to the same other turbine. It needs to be adapted. And this approach applies to all equipment. In current projects, we confirm this theory — it is necessary for each type of equipment to make its own mathematical model that would describe it and help us solve the problems of operation
This product, as well as a number of others that we produce, is widely known both in domestic and foreign markets. Moreover, the maim part of orders — foreign, they account for about 70% of the turnover of Rubius. And in the Russian market it is more difficult to work. We earn mainly abroad and invest in Russia. As for marketplaces and other forms of work with customers: now we are working closely with Skolkovo Technopark, its residents, we consider Skolkovo as a kind of showcase for finding customers. We attract foreign customers only as a result of our trips abroad. In the United States, we organize a business trip, during which we attend conferences, in which the necessary business contacts are established. In developed economies, recommendation systems also work effectively when your customers recommend your company to their partners, but you must first develop a stable reputation as a reliable manufacturer. Our best projects abroad began with trust, with personal communication with the owner, and not even about the tasks of their business — we talked about our philosophy in business. Talking about how the work in the company is organized, how we raise our programmers, how much we invest in it — we talk about our values. And if the customer says: well, at the level of values you are right for us, then let's work - we start with a small project. In Russia, for the first time last year we had such a dialogue with a potential customer at the level of values, but in general this approach does not work in Russia, here the rules of the game are different.