Novel fractional-order convolutional neural network based chatter diagnosis approach in turning process with chaos error mapping

The chatter not only brings about poor surface quality of the workpiece but also causes the tool wear and then leads to the increase in production cost over time. For this reason, it would be imperative that the chatter signal should be checked in an accurate manner whenever required. Because the chatter belongs to the nonlinear vibration phenomenon during the machining process, varied chatter characteristics will be presented under the different material, cutting speed and depth cutting conditions. Therefore, the machining learning method is used by many research programs by combining the database in order to analyze the vigorously changed data. To the chatter signal, thousands and even tens of thousands lots of data should be collected for use as training data and it would be extremely difficult for ordinary manufacturers and laboratories. It is because that not only will the tool be consumed but massive materials and time will also be required as far as the data collection is concerned. In this research, the chaotic error map is employed to accelerate the data processing in that 94.8% accuracy and 99.62% precision can be achieved simply with 60 lots of data only. Through the attractor properties of the chaotic system in the 3D space the input signal will be allowed to move along the attractor. Through such process, it transfers the highly variated chatter data to the consistent output result between the same classes. Further, this research is also the first program that proposes the fractional-order (FO) convolutional neural network (hereafter briefed as FOCNN) for chatter detection. Through the computation of fractional-order, it reduces 42.3% of trainable parameters when compared with the CNN having approximate training conditions while enhancing 3.8% of accuracy. Accordingly, our technique is also practical in use for the machining process.

Automated summary

Chatter not only affects the surface quality of the workpiece and tool wear, but also increases production costs. Accurate detection of chatter signals is therefore necessary. Due to the nonlinear vibration nature of chatter during machining, different chatter characteristics are observed under different conditions. This research uses a machining learning method combined with a database and employs chaotic error mapping to accelerate data processing. With only 60 data points, an accuracy of 94.8% and precision of 99.62% can be achieved. Additionally, this research introduces the fractional-order (FO) convolutional neural network (FOCNN) for chatter detection, reducing trainable parameters by 42.3% compared to approximate training conditions while improving accuracy by 3.8%.