The Theory And Application Of Deep Learning For Data-mechanism Fusion In Tool Condition Monitoring

With the development of technology,the smart machining has emerged as one of the most compelling research topics today.As an important component of the smart machining,tool condition monitoring provides significant value to advance product quality and processing efficiency.It helps to reduce human intervention,improve machining accuracy and reliability,and achieve more flexible process control.But so far,there are still some problems in related research.The tool condition monitoring is listed as one of the key technologies by developed countries.Deep learning is the most mature and outstanding technology in artificial intelligence.With its self-learning,complex nonlinear mapping and feature extraction capabilities,it occupies an irreplaceable position in modeling problems and provides a new way to realize tool condition monitoring.However,deep learning has not yet been independently applied to the tool condition monitoring due to the following reasons:1)The design of the deep network is completely based on experience which is not interpretable;2)Unable to provide sufficient samples covering complex working conditions,which restricts the performance of the model under unknown working conditions;3)The network is completely based on data,while there is no mechanical connection with the task.In order to solve these problems,this study is dedicated to shedding some light on the deep learning for data-mechanism fusion.Under the guidance of the machining mechanism,deep learning can predict the current tool condition according to the cutting signal.The main research contents and innovations of this study are summarized as follows:1)A pyramid network design method based on the spectrum characteristics of cutting signals is proposed,which can quickly and efficiently extract multi-scale features from the cutting signals containing a large number of sampling points.The learned patterns are restricted by the spectrum-based structure,which simplifies the monitoring task and reduces the model complexity.The efficiency of feature extraction is also greatly improved by reducing the number of units.The length of the cutting signal is also no longer limited by the memory capacity of LSTM.2)The dual-frequency attention and pyramid attention mechanisms are proposed,and a multi-scale pyramid network is constructed on this basis.Since the network structure is determined according to the periodic fluctuation of the cutting signal,the periodic characteristics of the cutting signal are effectively extracted and preserved.Therefore,the obtained attention distribution provides a basis for the interpretability of tool wear monitoring.On the basis of attention distribution,the periodic characteristics of interest of the model can also be further studied.3)Four practical data-mechanism fusion methods are proposed,which are suitable for seamlessly integrating different mechanistic information into data-driven models.The effects and characteristics of each data-mechanism fusion method are studied in detail,which helps researchers transform these methods to other fields.Based on these methods,a data-mechanism fusion model is specially designed for tool condition monitoring.The introduction of mechanism information limits the input space and feature space.Even with a small amount of data,the model is able to overcome the interference of machining conditions and environmental noise.The effectiveness and feasibility of the methods proposed in this study has been verified in high-speed milling experiments.The experimental results show that,compared with the data-driven model based only on cutting signals and the physics-based model based only on wear mechanism,the data-mechanism fusion model is more accurate and stable in tool condition monitoring under unknown machining conditions.