Feasibility study and development of an anti-pinch application with machine learning

An anti-pinch is a safety system that senses if a motor is running against an obstacle and prevents any injuries to people or damages to the obstacle or the motor itself. The anti-pinch sensing is usually performed by current monitoring, hall sensors or a combination of both. The state-of-the-art anti-pinch strategy is derivate based. It is simple to implement, and it check continuously the variation of the current and the variation of the velocity. Both are compared with a threshold and if they simultaneously report the presence of an obstacle, the motor is blocked. Although this task may look simple at first sight, many are the factor that add complexity and uncertainty to a standard algorithm: varying system conditions, mechanical and electrical uncertainties, mechanical wear of the motor system, etc. Moreover, the implementation of this algorithm requires a lot of run in order to setup the anti-pinch system correctly (e.g. for setting the thresholds). The scope of this project is to address these issues and to propose an innovative strategy: create an automated testbench and integrate a machine learning approach to automatically learn what is the best condition to trigger the anti-pinch. The project consists to build a motor bench that simulates a situation where a motor is moving a load and where, during the movement, a pinch can occur. In particular, the testbench could simulate an automatic back-seat system that works as follow: from an initial position, where the back-seat is parallel to the ground, the motor must move the back-seat until a certain position (expressed in degree). If there is an obstacle during the run, the motor must stop the movement of the back-seat. For the project, it is important to recognize the presence of an obstacle. Avoid false negative is more important than avoid false positive, according to the level of safety. The develop of this project is divided in different steps: Creation of a model of the system using Simulink. Mechanically build of the testbench. Design of an electronic circuit able to allow Arduino to control the two motors and to acquire measures from them. Develop the Arduino code in order to control the motors, to collect measures from sensors and to detect pinch. Design an HMI with Mathworks Matlab software, able to communicate with Arduino board, sending it test parameters and receiving the data Arduino measures. Apply machine learning training algorithm (logistic regression is used in this project) using Matlab. Implement on the Arduino board the machine learning algorithm to detect pinch. Finally, after several test, the machine learning anti-pinch algorithm and the classic anti-pinch algorithm are compared in order to evaluate if the innovative method is more convenient or not than the classic one by analyzing advantages and disadvantages and doing considerations about these methods.