“…In the large-scale data statistics, the digit recognition of the meter is a very important part of the whole intelligent "Internet of meters", so it has become a hot research topic for many years [1,2]. Digit recognition generally uses the traditional methods of feature matching and feature discrimination, but the recognition rate of these methods is not high [3][4][5].With the rapid development of neural network technology, it itself has a high degree of parallelism, strong self-organizing ability and fault tolerance, and better noise interference suppression ability to open up a new route for rapid and accurate digit recognition [6,7]. However, neural network has a lot of local extreme points because of highly complex nonlinear structure and its training process is easy to fall into local extremum under large amounts of data.…”