A compact self-learning intelligent system can be implemented in airplanes, drones, robots and assembly lines
Scientists of Samara National Research University (Samara University) have created an intelligent monitoring system capable of predicting technical faults, malfunctions and complex technical systems operation failures and warning about them in advance, even before they occur.
According to the developers’ plan, such solutions can be used to improve aviation safety in the future, however, the obtained characteristics already make it possible to implement them in industrial applications, in unmanned aerial vehicles and cars. They have created a prototype of the device monitoring production equipment, which is able to increase the reliability and safety of operating technological lines, and lessen the likelihood of downtime, preventing sudden equipment failures.
“We have developed a concept of the predictive system monitoring technological equipment and this concept has been implemented in practice – we have created a prototype of the system and its hardware modules, and formed a database of typical failures and malfunctions,” Albert Gareev, the head of Research Unit, an Associate Professor of Aircraft Maintenance Department, explained. “The most important thing here is the new principle we adopt: our development is based on a neural network, i.e. we use deep machine learning technology. As a result, we have created a unique self-learning software, which monitors equipment condition and notifies of impending component failure within a particular system.”
According to Assoc. Prof. Gareev, the uniqueness of the invented monitoring method lies in the software comparing so-called “dynamic images” of units and systems. A set of sensors records data on the real current equipment condition, and it is constantly compared to the ideal equipment condition, “ideal images”, put into the program database. The system detects deviations from this “ideal images” – e.g., changes in the oil system pressure, the difference in temperature levels or fuel consumption. The data is recorded at each section of a unit or system, and the neural network program determines the likelihood of a malfunction according to an algorithm developed through the self-learning.
To train this neural network program, the scientists created simulation models based on the German SimulationX software package, and then the system underwent additional training during experiments on a test bench. A Mi family helicopter hydraulic system assembled on the test bench was used as a test monitoring object. The test bench simulated a leakage of the working fluid and gas of the hydraulic system, changes in the pump speed, temperature and pressure rise, as well as various actions of the helicopter pilot. The results proved the ability of the neural network system to evolve, gradually learn and gain experience like a human being. According to the results of experiments, the accuracy of fault detection has reached 98%.
The scientists managed to keep the predictive monitoring system rather compact, cheap and energy efficient. The hardware platform of the system (without sensors) is based on a mobile neural processing unit with an energy consumption of 5-10 W and a cost of about 100 Euro. The processor board is comparable in size to a regular smartphone. These characteristics allow implementation of the “failure predictor” not only on the ground within industrial applications, but also in the air, e.g. in unmanned aerial vehicles. The system can also be useful for humanoid robots.
“The development enables us to provide any production facility with its own monitoring system reducing financial losses from potential downtime. Knowing of a pre-failure condition of a particular pump within an assembly line, you can switch to a spare line and not halt the production. Meanwhile the pump can be replaced or repaired, and the system will also provide you with a specific recommendation on which unit or element should be removed ,” Assoc. Prof. Gareev said. “Sure enough, our system was designed to be implemented in enterprises, especially in automotive and aviation industries, assembly lines, multi-axis machines, anthropomorphic devices, robots. However, the system can also be implemented in aviation – e.g., in unmanned aerial vehicles and airplanes.”
The university currently negotiates the possibility of the system implementation with a number of enterprises.
Find out more about Samara University.
