Convolutional neural network with Huffman pooling for handling data with insufficient categories: A novel method for anomaly detection and fault diagnosis

Abstract

The rotating component is an important part of the modern mechanical equipment, and its health status has a great impact on whether the equipment can safely operate. In recent years, convolutional neural network has been widely used to identify the health status of the rotor system. Previous studies are mostly based on the premise that training set and testing set have the same categories. However, because the actual operating conditions of mechanical equipment are complex and changeable, the real diagnostic tasks usually have greater diversity than the pre-acquired datasets. The inconsistency of the categories of training set and testing set makes it easy for convolutional neural network to identify the unknown fault data as normal data, which is very fatal to equipment health management. To overcome the above problem, this article proposes a new method, Huffman-convolutional neural network, to improve the generalization ability of the model in detection task with various operating conditions. First, a new Huffman pooling kernel is designed according to the Huffman coding principle and the Huffman pooling layer structure is introduced in the convolutional neural network to enhance the model's ability to extract common features of data under different conditions. Second, a new objective function is proposed based on softmax loss, intra-class loss, and inter-class loss to improve the Huffman-convolutional neural network's ability to distinguish different classes of data and aggregate the same class of data. Third, the proposed method is tested on three different datasets to verify the generalization ability of the Huffman-convolutional neural network in diagnosis tasks with multi-operating conditions. Compared with other traditional methods, the proposed method has better performance and greater potential in multi-condition fault diagnosis and anomaly detection tasks with inconsistent class spaces.