Adaptive multi-stable stochastic resonance assisted by neural network and physical supervision

Abstract Stochastic resonance can utilize the energy of noise to enhance weak frequency characteristic. This paper proposes an adaptive multi-stable stochastic resonance method assisted by the neural network (NN) and physics supervision (directly numerical simulation of the physical system). Different from traditional adaptive algorithm, the evaluation of the objective function (i.e., fitness function) in iteration process of adaptive algorithm is through a trained neural network instead of the numerical simulation. It will bring a dramatically reduction in computation time. Considering predictive bias from the neural network, a secondary correction procedure is introduced to the reevaluate the top performers and then resort them in iteration process through physics supervision. Though it may increase the computing cost, the accuracy will be enhanced. Two examples are given to illustrate the proposed method. For a classical multi-stable stochastic resonance system, the results show that the proposed method not only amplifies weak signals effectively but also significantly reduces computing time. For the detection of weak signal from outer ring in bearings, by introducing a variable scale coefficient, the proposed method can also give a satisfactory result, and the characteristic frequency of the fault signal can be extracted correctly.