A tool wear monitoring approach based on triplet long short-term memory neural networks

Abstract

Advancements in artificial intelligence have significantly improved the monitoring of tool wear in machining processes, thereby enhancing the overall quality of machining. However, the scarcity of tool wear samples poses a challenge to the enhancement of model precision. This necessitates the exploration of monitoring techniques that are effective even with small sample sizes. A method involving a triplet long short-term memory (LSTM) neural network is introduced, which offers the potential for superior accuracy even with limited training data. During the machining process, spindle vibrations are captured using a triaxial accelerometer. The raw data is processed by a triplet network, which uses an LSTM as the base model, thereby facilitating the aggregation within classes and separation between classes. A soft-max classification layer is subsequently integrated into the model, which enables the precise determination of tool wear states. The base model is optimized using a Genetic Algorithm to ensure model efficiency and accuracy before it is expanded into a triplet network. Experimental results from a vertical machining center confirm that the triplet LSTM network offers superior accuracy compared to a standard LSTM network, even when the sample size is small.