Tool Wear Prediction Based on Adaptive Feature and Temporal Attention with Long Short-Term Memory Model

Wanzhen Wang, Sze Song Ngu, Miaomiao Xin, Rong Liu, Qian Wang, Man Qiu, Shengqun ZhangInternational Journal of Engineering and Technology Innovation 2024 · DOI

TL;DR

Researchers have developed a hybrid deep learning model, FA-LSTM-TA, that significantly improves tool wear prediction in CNC machining. By combining Long Short-Term Memory (LSTM) networks with dual attention mechanisms—one for adaptive feature weighting and one for temporal importance—the model outperforms traditional methods by up to 41%, reducing the need for manual feature engineering.

What Problem Does This Paper Address?

In high-precision manufacturing, knowing exactly when a cutting tool will fail is a multi-million dollar question. Tool failure accounts for roughly 20% of machine downtime. While we can't easily measure "wear" while a machine is running (direct measurement), we can listen to its "symptoms" through vibration and force sensors (indirect measurement).

However, traditional machine learning requires experts to manually pick which signal features matter most—a tedious process called feature engineering. Furthermore, standard LSTM models, while great at handling sequences, often "forget" critical information or struggle to capture global dependencies when dealing with the long, noisy data streams generated during machining.

Methods

The authors propose a model called FA-LSTM-TA (Feature Attention - LSTM - Temporal Attention). The workflow consists of three main stages:

  1. Multi-Domain Feature Extraction: Instead of feeding raw vibration signals into the model, they extract 24 specific features across the time domain (e.g., RMS, Skewness), frequency domain (via Fast Fourier Transform), and time-frequency domain (via Discrete Wavelet Transform).
  2. Feature Attention Mechanism: This is the "adaptive" part. A self-attention layer automatically weights these 24 features, amplifying those most related to tool degradation and suppressing irrelevant noise. This eliminates the need for manual feature selection.
  3. Temporal Attention LSTM: The weighted features pass through two LSTM layers to capture sequential patterns. A second attention layer then looks across different time steps to identify which moments in the machining process are most indicative of wear.
  4. Optimization: The team used a Genetic Algorithm (GA) to automatically find the best hyperparameters (like the number of hidden units and batch sizes), rather than relying on trial and error.

Datasets

The study utilized the PHM2010 competition dataset, a gold standard in the field of Prognostics and Health Management.

  • Source: Milling data from a CNC machine center.
  • Details: It includes force, vibration (x, y, z), and acoustic emission (AE) signals.
  • Scope: The authors focused on the x-direction vibration signals from three specific tools (C1, C4, and C6) that had labeled wear measurements.

Key Insights and Ideas

  • The "Goldilocks" Window: The authors found that the number of sub-sequences (how you slice the data) is critical. If you slice the data too thinly, the model loses the "big picture" of temporal dependence; if the slices are too large, the training process becomes inefficient.
  • Attention as a Filter: The feature attention mechanism acts as an automated expert. It learns through backpropagation which of the 24 extracted features actually correlate with a dulling blade, effectively doing the job of a data scientist.
  • Global vs. Local: While LSTMs are good at local "step-by-step" memory, the temporal attention layer allows the model to look at the entire machining cycle at once to find the most relevant "wear signatures."

Key Results

The FA-LSTM-TA model was compared against Support Vector Regression (SVR) and a standard LSTM. The results showed a clear victory for the attention-based approach:

  • Accuracy: The proposed model achieved significantly lower error rates. For tool C6, the RMSE (Root Mean Square Error) improved by 35% and the MAE (Mean Absolute Error) improved by 41.6% compared to a standard LSTM.
  • Goodness of Fit: The $R^2$ (coefficient of determination) reached as high as 0.944, indicating the model explains nearly 95% of the variance in tool wear.
  • Consistency: The model consistently outperformed SVR, which struggled with the non-linear, non-smooth nature of machining data.

Limitations and Future Work

Despite the high accuracy, the authors noted a few hurdles:

  • The Boundary Problem: The model is less accurate during the very beginning (initial wear) and the very end (rapid failure phase) of a tool's life. In these stages, wear happens too fast and signals become too unstable for the current model to track perfectly.
  • Generalization: The model was trained and tested on data from the same machining conditions. In a real factory, speeds, feeds, and materials change constantly.
  • Future Directions: The authors suggest exploring transfer learning and domain adaptation to help the model stay accurate even when the machining environment changes. They also see potential in "Digital Twins" that combine these data-driven models with physical laws of physics.

Summary updated

Abstract

Effective monitoring of tool wear status can improve productivity and reduce losses. In previous studies, extensive feature selection was required when using the traditional machine learning method. The gating mechanism in the traditional long short-term memory (LSTM) model may incur information loss and a weaker representation of global sequential dependencies in handling long sequences. This paper aims to enhance the performance of the LSTM model in tool wear prediction by combining feature and temporal attention. Firstly, the original vibration signal is divided into sub-sequences and related features extracted. Secondly, the ability to capture global sequential dependencies using the LSTM model is improved by feature and temporal attention. Finally, a fully connected layer is used to predict tool wear values. Compared to traditional LSTM, the proposed method performs best in three evaluation metrics, RMSE, MAE, and the coefficient of determination.
BibTeX
@article{wang2024tool,
  author = {Wanzhen Wang and Sze Song Ngu and Miaomiao Xin and Rong Liu and Qian Wang and Man Qiu and Shengqun Zhang},
  title = {{Tool Wear Prediction Based on Adaptive Feature and Temporal Attention with Long Short-Term Memory Model}},
  journal = {International Journal of Engineering and Technology Innovation},
  year = {2024},
  doi = {10.46604/ijeti.2024.13387},
}
Figure