Research On In-situ Monitoring Of Cutters Wear And Worpiece Roughness Using Workpiece Texture Images

In recent years,with the promotion of the national strategy "Made in China 2025",the lean manufacturing concept featuring intelligence has been rapidly promoted and applied in the manufacturing industry.Establishing and improving the machining manufacturing industry’s condition monitoring and assessment system(e.g.,cutter degradation monitoring system)is a critical and basic requirement to enhance production efficiency and ensure product quality.In a typical metal-cutting manufacturing industry,cutter wear and the workpiece surface roughness are two important indicators in production,which are also important observations for cutter degradation assessment.Considering that it is difficult to achieve highfrequency in-situ and on-machine measurement of the above two indicators in the real mass production scenarios,this study takes end milling as the research object and analyzes the relationship between features extracted from the workpiece texture images and the two indicators(i.e.,cutter wear,specimen surface roughness)under multiple working conditions.The models of in-situ roughness evaluation and cutter wear identification are constructed.On this basis,the cutter health indicator and degradation evaluation model based on the two above-mentioned indicators are constructed,and the role of the machined surface texture image information in cutter degradation assessment based on multi-source information fusion is analyzed.The key and practical technology for in-situ monitoring and degradation evaluation of milling cutters using machined surface texture image analysis are studied.The main contributions are summarized as follows:(1)Focusing on the problem of accurate model construction of cutter wear monitoring with changing cutting parameters and the limited data of a typical working condition in real production scenarios,a method for monitoring data clustering and division under multiple working conditions is proposed.On this basis,a deep convolutional transfer learning neural network model is constructed,which is based on data adaptation of multi-source domain and target domain.And the mechanism of neural architecture transfer and weight-reuse under multiple working conditions is analyzed.A deep transfer learning model framework is proposed to fine-tune lightweight convolutional modules on the target domain for on-machine cutter wear recognition tasks at the edge computing devices.(2)To address the difficulty of accurate model construction of specimen roughness monitoring with changing cutting parameters and the insufficient model interpretability,an interpretable machine learning model for on-machine roughness monitoring is proposed.A topography simulation model considering spindle runout and cutting parameters is constructed,and the top-view projected texture patterns of the machined surface topography are obtained.A feature consistency check algorithm between the simulated texture images and the actual acquired texture images is designed.By verifying the correlation between the actual acquired texture images and the projected image features from topography simulation,the robustness of the roughness monitoring model is improved for cases under multiple working conditions.(3)Aiming at the problem that a single indicator(e.g.,cutter wear)is difficult to accurately and completely characterize the cutters’ degradation process,this study proposes a tool health indicator construction and degradation assessment method based on two-observed indicators(i.e.,cutter wear,workpiece surface roughness)using features extracted from texture images of the machined workpiece.Firstly,a multi-source feature extraction and fusion method of spindle vibration signal and workpiece texture images is constructed.Then,cutter degradation monitoring data with similar degradation patterns in historical monitoring data are extracted by the trajectory similarity computation.Finally,the role of texture image features in health indicator construction and degradation evaluation are analyzed,where models are built by multi-source feature fusion.(4)Aiming at the problem that general object detection algorithms are difficult to achieve satisfying recognition accuracy in real industrial monitoring tasks,a deep ensemble learning model and an optimized SSD model are constructed.For the key machined surface recognition task,a data distribution identification and weight assignment method based on clustering analysis are proposed to comprehensively utilize the performance advantages of the end-toend deep learning model and the traditional object detection model under different data scales and characteristics.In respect of the chip detection task,the model is optimized from the perspective of non-maximum threshold adjustment and cost-sensitive loss function construction,to enhance the model performance for small targets detection.(5)Considering the absence of a specific software system for on-machine condition monitoring and tool degradation assessment based on machined surface texture image analysis,some practical technology studies are carried out.The framework of machined texture image acquisition is built and the overall process of on-machine acquisition and macro focusing is designed.The front and back-end modules,structural framework,and interface of the software system are designed and implemented.