Active control of flexible rotors using deep reinforcement learning with application of multi-actor-critic deep deterministic policy gradient

Abstract

Vibration in rotating machinery is one of the main causes of machine failure. Active and passive control methods have been developed in order to reduce vibration levels and extend the operating speeds of machines. These controllers are either manually tuned or tuned during operation using adaptive control techniques and are tailored to a single vibration source. In the field of artificial intelligence, deep reinforcement learning has greatly impacted the field of continuous control, from mastering simple games to controlling multiple actuators in a robot doing complex tasks. Deep reinforcement learning agents are capable of finding optimal control policies without a model of the underlying system. This work proposes the multi-actor-critic deep deterministic policy gradient (MAC-DDPG) algorithm by integrating multiple criteria to train concurrent actors in a periodic system. Using the frequency footprint of each type of vibration, a cost function is designed to train the critics and actors. The proposed controller is evaluated on a test rig supported by two patented Smart Electro-Magnetic Actuator Journal Integrated Bearings (SEMAJIB). The proposed controller is capable of finding optimal control policies for reducing the synchronous vibration caused by the rotor’s unbalance and stabilizing a system with oil whip vibration. A derivative controller actor and a harmonic actor are used concurrently for controlling the vibrations. The proposed controller is able to reach unbalance vibration reduction up to 93%. In addition, the proposed controller is successful in completely eliminating oil whip instability with up to 99% reduction.