Robot milling tool wear monitoring based on variational autoencoder with multidomain depth features

Due to the open-chain, multilink structure and weak rigidity characteristics of the industrial robots, the stability of their milling process is poor, significantly exacerbating tool wear and breakage. Therefore, this article proposes a tool wear condition monitoring method based on variational autoencoders (VAE) and deep learning neural networks. This method combines unsupervised and supervised learning to analyze the mapping relationship between multidomain features of monitoring signals and tool wear, achieving the monitoring and identification of tool wear conditions. Firstly, a multidomain feature extraction method is proposed to obtain wear characteristics from milling signals. Then, based on the proposed method, VAE is combined with long short-term memory neural networks (LSTM), Transformer encoders integrated with LSTM layers, and bidirectional LSTM neural networks (BiLSTM) to construct three tool wear prediction models: VAE-LSTM, VAE-LSTM-Transformer, and VAE-BiLSTM. These models achieve deep integration of multidomain features and map them to tool wear values. Finally, experiments were conducted to validate the tool wear conditions for both robot milling and traditional machine tool milling. The results show that this method has high prediction accuracy, strong adaptability to changes in milling parameters, and strong compatibility with different milling equipment.