A Deep Learning Approach for Improving Detection Accuracy and Efficiency Based on a Mass-Position Sensing Scheme

A deep learning (DL) approach for improving detection accuracy and efficiency is conducted in this article based on a mass-position sensing scheme. In the scheme, masses and positions of multiple spheres can be determined using a length-adjustable cantilever with lower modes. Four DL networks, including two simple multilayer perceptron (MLP), an inverted triangle MLP, and a residual network, are constructed to process the datasets obtained by experimentally verified physics-based model. Comparing to iteration with the nonnegative linear least squares, the detection accuracy is increased by 80%, and the calculation efficiency is improved by more than 4000 times. The conducted DL approach does not rely on the modal shape functions of the cantilever which is essential for the iteration method. The size of the dataset has almost no impact on the predicted accuracy while more input dimensions can make significant improvement. If the principle of a physical sensor can be verified by simulation and experiment simultaneously, a dataset can be established with simulation and then the DL neural network can be trained to learn the relationship between input and output of the sensors. This is especially useful when it is difficult to reverse the input from the output of the sensor by traditional mathematical means. So, our approach that training DL networks with the dataset obtained by experimentally verified physics-based model is expected to be applicable to physical sensors besides resonant one.