💻 computer science

A Tool Wear Prediction Method Based on Multi-Scale Temporal Feature Fusion

This paper proposes GLDformer, a novel tool wear prediction model that integrates multi-scale temporal feature fusion, global-local feature modeling, and trend-detail decomposition to effectively capture complex signal dynamics, achieving superior accuracy and generalization in cross-tool validation compared to existing baselines.

Original authors: Chen Jinguo, Sun Da, He Genghuang, Luo Manting, Chen Youxing, Li Miao, Zhang Qinkun, Chen Linhan

Published 2026-10-01

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Chen Jinguo, Sun Da, He Genghuang, Luo Manting, Chen Youxing, Li Miao, Zhang Qinkun, Chen Linhan

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the world of modern manufacturing, the cutting tool is the unsung hero of precision. Whether shaping the components of an aircraft engine or crafting a medical implant, a sharp tool ensures a smooth surface and exact dimensions. But tools do not last forever. As they slice through metal, they slowly dull, a process known as wear. This gradual degradation is dangerous; a tool that is too worn can snap, ruining the expensive part it is cutting and potentially damaging the machine itself. To prevent this, engineers have long tried to listen to the machine. By attaching sensors to the equipment, they can "hear" the subtle changes in vibration, force, and sound that occur as the tool dulls. These signals are like a continuous conversation between the tool and the workpiece, but the language is complex. The signals contain a mix of slow, steady changes that show the tool getting older, alongside sudden, sharp spikes caused by the roughness of the cut. For years, computer programs designed to read these signals struggled to understand both the slow story and the sudden events at the same time, often missing the warning signs until it was too late.

A team of researchers from universities in China has developed a new way to listen to this conversation, one that pays attention to every nuance of the signal. They created a system called GLDformer, a sophisticated software model designed to predict exactly how much a tool has worn down just by looking at the data streaming from the machine. Instead of trying to force the data into a single, simple pattern, their approach treats the signal as a rich tapestry of different time scales. It looks for the slow, steady drift of the tool's life while simultaneously tracking the quick, chaotic jitters of the cutting process. By separating these different rhythms and then weaving them back together, the model can see the full picture of the tool's health with remarkable clarity.

The researchers tested their idea using a well-known collection of data from a public challenge involving milling machines. They focused on three specific tools, each with a complete record of its life from the moment it was sharp until it was finally discarded. To make the test truly rigorous, they used a method called "leave-one-tool-out." This means they trained the computer on two of the tools and then asked it to predict the wear on the third tool, a tool it had never seen before. This is a critical test because in a real factory, a new tool might behave slightly differently than the ones the computer learned from. If the system can predict the wear on an unknown tool, it proves the method is robust enough for real-world use.

The results were striking. When the new model looked at the unseen tool, it predicted the amount of wear with an average error of just 6.07. To put that in perspective, the model was off by less than one-tenth the width of a hair. This level of precision was significantly better than the best existing methods, which missed the mark by a much larger margin. The model also showed a very high degree of accuracy in matching the overall trend of the tool's life, correctly identifying when the tool was healthy and when it was rapidly approaching the end of its useful life.

What makes this approach so effective is how it handles the different types of information hidden in the sensor data. The researchers realized that standard computer models often get confused when a signal contains both long-term trends and short-term noise. Their solution was to build a system with two distinct "brains" working in parallel. One part of the system focuses on the big picture, looking at the entire sequence of data to understand the long-term relationship between the tool's age and its performance. The other part zooms in, examining small, local chunks of the signal to catch the immediate, sharp fluctuations that happen during a single cut. A smart switching mechanism then decides how much weight to give to each of these views at any given moment, blending them into a single, coherent understanding.

Furthermore, the model includes a special step that separates the slow, steady decline of the tool from the fast, erratic vibrations. Imagine trying to hear a single instrument in a loud orchestra; if you focus only on the volume, you might miss the melody. Similarly, this model isolates the slow "melody" of the tool wearing down from the "noise" of the cutting process. By analyzing these two components separately and then recombining them, the system avoids the common mistake of smoothing out important details or getting distracted by temporary spikes. This separation allows the model to remain calm and accurate even when the cutting process becomes chaotic.

The researchers also tested what would happen if they removed these special features one by one. When they took away the part that looked at local details, the model's accuracy dropped significantly, proving that the small, sharp fluctuations are essential for a correct prediction. When they removed the part that looked at the long-term trends, the model failed to understand the overall life of the tool. And when they removed the step that separated the slow trend from the fast noise, the model made larger errors, particularly when the tool was wearing out quickly. These tests confirmed that every part of the system plays a vital role and that the combination of these different perspectives is what leads to such high performance.

This work demonstrates that by respecting the complexity of the data—acknowledging that a tool's life is a mix of slow changes and fast events—we can build better tools for monitoring industrial machines. The model does not just guess; it learns to distinguish between the signal and the noise, the trend and the event. While the study was conducted under specific, controlled conditions, the success of predicting the wear on a completely unknown tool suggests that this method could be a powerful addition to the toolkit of modern manufacturing. It offers a way to keep machines running smoothly, prevent unexpected breakdowns, and ensure that the products we rely on are made with the highest possible precision.

Technical Summary: A Tool Wear Prediction Method Based on Multi-Scale Temporal Feature Fusion

Problem Statement
Standard Transformer architectures often struggle to simultaneously characterize the complex nature of multi-sensor cutting signals, which contain multi-scale local fluctuations, long-range dependencies, and slowly varying degradation components. Existing indirect monitoring methods, whether based on handcrafted features or deep learning, often face limitations: traditional models rely heavily on domain expertise and specific signal processing strategies, while deep learning approaches (CNNs, RNNs) may suffer from constrained receptive fields or recursive computation bottlenecks. Furthermore, there is a lack of unified frameworks that explicitly separate and recombine trend and detail components while maintaining robust generalization across different tools under strict data isolation protocols.

Methodology: GLDformer
The authors propose GLDformer, a tool wear prediction model designed to integrate global-local feature modeling with multi-scale trend-detail decomposition. The methodology proceeds through the following stages:

Data Preprocessing & Feature Construction:
- Raw signals from the PHM Society 2010 dataset (three-axis cutting forces, three-axis vibrations, and acoustic emission) are processed by discarding the initial and terminal 2.5% to retain only the stable cutting phase.
- Signals are divided into 100 consecutive time windows. For each window, three statistical features (Maximum, Mean, Variance) are extracted per sensor channel, resulting in a 21-dimensional feature vector per window.
- Features are normalized using Min-Max scaling based strictly on the training set to prevent information leakage.
Multi-Scale Convolutional Embedding:
- Instead of standard linear projection, the model employs a Multi-Scale Conv Tokenizer. Three parallel 1D convolutional branches with kernel sizes of 3, 7, and 15 capture local patterns at varying receptive fields (fine-grained, medium-range, and broad context).
- These outputs are concatenated, normalized, and passed through GELU activation and Dropout to form the initial sequence representation.
Global-Local Dual-Branch Gated Fusion:
- The model utilizes a dual-branch architecture to process the embedded features:
  - Global Branch: Consists of Transformer encoder layers to model long-range dependencies across the 100 time windows.
  - Local Branch: Comprises stacked LocalConvBlocks (using depthwise and pointwise convolutions) to capture neighborhood dynamics.
- An Adaptive Gating Mechanism fuses these branches. A gating matrix, generated via a Sigmoid activation on concatenated global and local features, dynamically controls the balance between global and local information for each time step and feature dimension.
Multi-Scale Trend-Detail Adaptive Modeling:
- To address the mutual masking of trends and fluctuations, a dedicated module decomposes the fused features.
- Trend Extraction: Three depthwise convolutions (kernels 5, 11, 21) extract slowly varying trend components.
- Detail Extraction: The detail component is computed as the residual ( $D = H_F - T$ ).
- Adaptive Recombination: A gating mechanism dynamically weights the trend and detail components based on context before recombining them via residual addition.
Prediction Head:
- Temporal Attention Pooling: Focuses on informative time windows rather than simple averaging.
- Residual Regression: The final prediction combines the main branch output (attended features) with an auxiliary branch output (global mean pooling). The auxiliary branch, weighted at 0.1, provides an inductive bias toward the overall wear level to prevent excessive deviation.

Key Contributions

Unified Framework: The paper introduces a framework that explicitly coordinates multi-scale local fluctuations, long-range dependencies, and slowly varying components within a single Transformer-based architecture.
Novel Architectural Components:
- A Multi-Scale Conv Tokenizer replaces linear projection to better encode local context.
- A Gated Fusion Mechanism allows for content-dependent selection between global and local representations.
- A Trend-Detail Decomposition Module explicitly separates and recombines degradation trends and local perturbations.
Rigorous Validation Strategy: The study employs a Leave-One-Tool-Out (LOTO) cross-validation strategy on the PHM2010 dataset (using tools C1, C4, and C6). This ensures that training, validation, and test data are mutually isolated, providing a strict evaluation of generalization capability to unseen tools.

Experimental Results
The model was evaluated against baselines including LSTM, BiLSTM-Attention, CNN, GRU, and standard Transformers.

Performance Metrics: GLDformer achieved an average MAE of 6.07 μm, RMSE of 8.25 μm, and R² of 0.94.
Improvement: Compared to the best baselines, GLDformer reduced MAE by 30.9% and RMSE by 28.5%, with an R² improvement of approximately 0.05.
Ablation Studies:
- Removing the Local Convolutional Branch resulted in the most significant performance drop (MAE +50.6%), highlighting the necessity of capturing fine-grained local fluctuations.
- Removing the Transformer Branch caused a 44.7% increase in MAE, confirming the importance of long-range dependency modeling.
- Removing the Trend Decomposition Module increased RMSE by 32.4%, indicating its role in suppressing large prediction errors.
- Replacing the multi-scale tokenizer with linear projection degraded performance, validating the benefit of multi-scale feature extraction at the embedding stage.

Significance and Claims
The paper claims that GLDformer effectively improves the accuracy and generalization capability for predicting tool wear on unknown tools, addressing the limitations of single-scale or single-temporal modeling approaches. The authors assert that the complementary roles of local features, long-range correlations, and explicit trend-detail separation are critical for characterizing tool wear evolution.

The study acknowledges limitations, noting that experimental validation was restricted to three tools under fixed operating conditions and that the model tends to underestimate the accelerated wear phase in the end-stage of certain tools (e.g., Tool C4). The authors suggest that future work should investigate model applicability under varying tools and processing conditions. The work is presented as a viable approach for multi-sensor tool condition monitoring, offering a robust method for handling the complex temporal dynamics inherent in industrial cutting signals.

Technical Summary: A Tool Wear Prediction Method Based on Multi-Scale Temporal Feature Fusion

More like this