有监督线性特征映射(SLFM)网络及刀具磨损量实时估计

A Supervised Linear Feature Mapping Network and Its Application in Tool Wear Estimation

提出了有监督线性特征映射网络（ＳＬＦＭ），并应用于刀具磨损量实时估计，研究了网络参数对学习速度和网络性能的影响，并与误差反馈式多层前向网络（ＢＰ网络）进行对比。研究表明，ＳＬＦＭ网络具有学习快、精度高的优点。

Recently sensor integration with artificial neural network for tool wear monitoring is attracting more attention. But the now widely used network is multilayered forward network with BP algorithm which requires excessive training time and only classifies the tool into worn or sharp group without giving an estimate of tool wear. In this paper, a Supervised Linear Feature Mapping (SLFM) network is proposed to estimate tool wear. The influences of network parameters on learning speed and performance are discussed. SLFM network is a variant or particular case of SelfOrganizing Feature Mapping (SOM) proposed by Kohonen(1982). The SLFM differs from SOM in two respects: (a) The output layer is a linear neuron set to provide a “discrete scale” estimate of tool wear. (b) The SLFM algorithm uses supervied learning law to adjust the weight vectors, instead of using competitive learning only. A schematic diagram of tool wear monitoring system for turning operation is presented. The system employs multiple sensors(acoustic emission, acceleration and motor power sensors). The output layer of the SLFM has 22 neurons corresponding to a range of 0～0.55 mm of flank wear. As comparison, a BP network with 7 inputs, 10 hidden neurons and one output neuron is trained using the same sample set as SLFM network. The results show that SLFM network has lower training iterations and higher accuracy than BP network.