基于LSTM的刀具数据异常检测方法的研究

Research on Cutlery Data Anomaly Detection Method Based on Long Short-Term Memory Network

刀具状态是机械加工过程中影响产品质量的因素之一,其中刀具异常数据的有效检测将有助于掌握刀具状态,针对这一问题,论文提出了一种基于LSTM的数据异常检测方法。该方法采用LSTM模型对正常刀具数据进行训练预测,预测之后使用正态分布建模方法求均值作为阈值,将实时采集到的刀具数据预测建模求得的值与阈值进行比较,得到异常数据。对比找到适合的神经网络层数和隐藏的神经元个数,最终结果显示,双层神经网络和128个隐藏神经元的结合,预测准确率提高50%;同时将LSTM算法与PCA降维算法在准确度方面进行了比较,准确度提高约20%,验证了LSTM的有效性。

The tool status is one of the factors that affects the product quality during machining,and the tool status can be grasped by effectively detecting the tool failure data. This document uses the LSTM-based method to detect data errors. This method uses a LSTM model to train and predict regular tooling data. After the prediction,the normal distribution modeling method is used to get the mean as a threshold,and the value obtained by predictive modeling of the tool data collected in real time with the threshold to get anomalous data is compared. The number of suitable neural network layers is compared with the number of hidden neurons.The final results show that the combination of a two-layer neural network and 128 hidden neurons improves the prediction accuracy by 50%. At the same time,the LSTM algorithm and the PCA dimension reduction algorithm are accurate. Compared to accuracy,accuracy is improved by about 20%,demonstrating the effectiveness of LSTM.