Stacked sparse autoencoder base frameworks for tool condition monitoring

Abstract

Tool condition monitoring systems in High Speed Machining (HSM) are of great importance to maintain the quality of the products and diagnose the useful life of the tools. These systems are highly demanded for the suppliers of molds and dies in the aeronautic and automotive industry. Tool wear is a real problem for the industry, around 20% of the down-time is attributed to tool failure, resulting in reduced productivity and economic losses also, there seem to be numerous factors that can affect the surface roughness. Controllable process parameters include; feed, cutting speed, tool geometry and setup; but, there are other factors which are harder to control include machine vibrations, work piece and specially tool wear degradation. The complexity of the Tool Wear Monitoring (TWM) has led to a significant amount of literature related to this problem, various techniques have been proposed; but, TWM remains a real and complex problem to solve owing to its constant change in the process variables. Literature shows that the signal processing techniques are the most used for feature extraction. Extracting valuable features from raw data using signal processing techniques alone in most of the cases is not sufficient due to some of the important features are hidden in abstract features, which are not easy to access. For this task Deep Learning (DL) has the most notable advantage of a powerful complex learning ability. SSAE are a powerful classification tool of DL, the non-supervised training algorithm is a good option to extract very representative features of the signals. A new methodology based in SSAE is presented. Two approaches are taking on account, the first uses time domain signals as inputs, the second one is a hybrid methodology, the input sig- nals are pre-processed to obtain the Mel-Frecuency Cepstral Coefficients (MFCC) and use them as inputs for the Stacked sparse AutoEncoder (SSAE) neural network. The methodology evalu- ates different signals obtained from different sensors (accelerometer, dynamometer and acoustic emission), which were recorded during the machining of aluminum workpieces, with different hardness, tools and cutting trajectories. The methodology presents a fairly acceptable performance (99.8%) in the prediction of the tool wear condition, especially with the signals of the acoustic emission. SSAE neural network outperforms traditional neural network.