Research On Tool Wear Condition Recognition And Prediction Based On Deep Learning Models

Tool wear condition directly affects the surface quality of work piece,production cost and manufacturing efficiency,which is of great significant to monitoring tool wear conditions in production line timely.Enormous manufacturing big data generates during the production process,which is capable of reflect the tool wear degradation trending thoroughly.However,it requires big data analytical methods and deep learning models for signal processing and condition monitoring.Meanwhile,the signals are heavily polluted during the manufacturing process,which is more difficult to extract features.In this paper,a demand research on machining accuracy and computing resource under different processing scenarios has been conducted.Then,target signals and tool condition monitoring models have been chosen.In the end,a tool condition monitoring prototype system has been developed to meet the requirements under different processing scenarios.The contributions of this thesis is concluded as follows:A deep residual bidirectional long-short term memory(DRBLSTM)model is investigated for tool wear condition recognition.The developed model combines the bidirectional long-short term memory cell with residual layer to alleviate the gradient problems of the model effectively and promote model convergence.At the same time,time-average pooling layer is proposed to build a direct connection between features over channels and the labels,and avoid model overfitting by model parameters reduction.The following experiments validated the proposed model is competent to recognize current tool wear status.Aiming at the problems of model recognition accuracy improvement,an improved deep residual neural network is proposed.Short-time Fourier Transform and bilinear interpolation are applied to process vibration signals in sequence,and the results are treated as the model input.Res Net block is further optimized,and 9 variants are proposed,and the optimal is decised by experiment.The optimal model structure are stacked to extract more sensitive features.Final comparison experiments prove that the optimized model is adequate to predict tool wear value accurately,and the accuracy is 0.9449.In order to improve model prediction accuracy and efficiency,multi frequency-bands deep convolutional neural network(MFB-DCNN)is introduced to prediction tool wear under strong interference situation in this thesis.Wavelet package coefficients from different frequency bands are reshaped into two-dimensional matrices.These matrices are learned by corresponding convolutional neural network blocks.And the extracted features are fused for further analyzation.Further performance experiments validate the brilliant model performance of the proposed model and the testing time cuts down up to 25.181%.In order to improve model robustness,an attention-based multi frequency-bands recursive convolutional neural network(AB-MFBRCNN)is constructed.This model is built on the basis of MFB-DCNN,and recursive convolutional neural network structure is applied to comprehensive analyze features under multiple frequency band without any parameters addition.Moreover,attention mechanism is introduced for further model robustness and recognition accuracy improvement.Experiments have been conducted to prove the effectiveness of the two proposed structures,and the proposed AB-MFBRCNN model overperforms the previous MFB-DCNN model with heavily polluted vibration signals.Given the complex manufacturing requirements under 3C industry,a tool condition monitoring prototype system is developed.Micro-service framework is adapted for greater service scalability and flexibility.Then,low-value signal filter algorithm is proposed for fast signal selection.In the end,tool condition monitoring models proposed in this thesis is also applied for reliable and stable tool condition monitoring.In conclusion,the main contributions of the thesis are summarized,and the research directions in the future are forecasted.