1 Introduction
The main function of the grinding process is grinding and grading minerals, and separating the mineral from the gangue, which creates situations for subsequent sorting tasks, and which makes it one of the most fundamental parts in beneficiation production. With its slow producing process, complicated mechanism and external distorting factors, and large hysteresis, the grinding process is unstable. Additionally, the majority of the mining environment is remote and difficult to attract professional and technical personnel. The current fault diagnosis techniques develop slowly. An efficient fault diagnosis that analyzing data automatically draws wide attention because of the rapid growth of innovative techniques such as the internet, IoT, data-mining, and deep learning. As a result of the grinding process is relevant to the time series, this paper applies recurrent neural network (RNN) algorithm to analyze the relation between fault and grinding information, and predicts faults. The recurrent neural network is a form of neural network, which holds the same structure as the normal multi-layer perceptron, and feeds state or output to input, conveying state and information through the network, by which the RNN is able to handle time series problems. However, there are several difficulties encountered by the RNN such as gradient vanishing, weight exponentially exploding or disappearing, and these difficulties make it hard to acquire long-time association and train valuable network. This situation continued until the emergence of long short term memory (LSTM) [1], which conquered the difficulty of keeping the long-term memory, makes RNN be able to keep either long-term and short-term memory, and this makes RNN practical and productive. Currently, the LSTM becomes an efficient method analyzing time series, and it succeeded in handwriting recognition[2] and voice recognition[3] and so on, and played the kernel role in natural language processing.