💻 computer science

Cross device milling force prediction method based on self supervised pre training and domain adaptation

This paper proposes a cross-device milling force prediction method that combines self-supervised pretraining, domain adaptation, and few-label fine-tuning to effectively transfer vibration-force mapping across devices despite dynamic response differences and limited target-domain labels, achieving high accuracy with minimal labeled data.

Original authors: Zhigang Ma

Published 2026-09-01

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Zhigang Ma

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 high-stakes world of manufacturing, where aerospace components and precision molds are carved from solid metal, the cutting tool is constantly engaged in a violent, rhythmic struggle against the workpiece. This process, known as milling, is not a smooth glide but a series of rapid, intermittent impacts. As the rotating cutter bites into the material, it generates a force that dictates the quality of the final product, the lifespan of the tool, and the safety of the machine. If this force becomes too great or fluctuates unpredictably, the tool can wear out prematurely, the machine can vibrate dangerously, or the part can be ruined. For decades, engineers have relied on heavy, expensive sensors called dynamometers to measure this force directly. However, these devices are difficult to install, can alter the stiffness of the machine, and are often impractical for long-term use in a busy factory. A more attractive alternative exists: listening to the machine's vibrations. Just as a doctor might infer a heart condition from the rhythm of a pulse, engineers know that the vibration of a machine tool contains a hidden code that reflects the cutting force. The challenge, however, is that every machine is different. A vibration pattern that signals a specific force on one machine might look completely different on another due to differences in how the metal is built, where the sensors are placed, or how the signal travels through the structure. This makes it incredibly difficult to train a computer model on one machine and expect it to work accurately on another, especially when the second machine lacks the expensive force sensors needed to teach the model what to look for.

Researchers at Xinxiang Vocational and Technical College have developed a new approach to solve this problem, creating a method that allows a computer model to learn how to predict cutting forces on a new machine using very little data. Instead of trying to force a model trained on one device to work on another, they built a system that learns the fundamental language of vibration first, then adapts to the specific machine. The process begins by feeding the computer vast amounts of raw vibration data from both the original machine and the new target machine. Crucially, this initial data does not need to be labeled with force measurements. The computer is asked to play a game of "fill in the blanks," where parts of the vibration signal are hidden, and the model must reconstruct the missing pieces based on the surrounding patterns. This self-supervised training allows the model to learn the intrinsic rhythm and structure of milling vibrations without needing to know the exact force values yet. It learns what a stable cut looks like, how the signal changes when the tool hits the material, and how different frequencies interact, effectively building a universal dictionary of milling sounds.

Once the model has learned this general language, it is ready to adapt to the specific machine. The researchers then introduce a small amount of labeled data from the target machine—data where the vibration is paired with the actual measured force. They use a technique called domain adaptation, which acts like a translator, aligning the features learned from the first machine with the unique characteristics of the second. This step ensures that the model understands that a specific vibration pattern on the new machine corresponds to the same physical force, even if the signal looks slightly different due to the machine's unique structure. Finally, the model undergoes a brief period of fine-tuning using the few labeled samples available. This calibrates the prediction, adjusting the output to match the specific amplitude and dynamic response of the new environment. The result is a system that can take the abundant, unlabeled vibration data available on a new machine and, with only a tiny fraction of labeled force data, predict the cutting force with high accuracy.

The effectiveness of this method was tested by comparing it against traditional approaches. When the model was trained only on the source machine and applied directly to the target machine without any adaptation, the predictions were poor, often missing the mark by a significant margin. Even simple adjustments failed to bridge the gap between the two different machines. However, the new method, combining the self-supervised pre-training with domain adaptation and fine-tuning, showed remarkable results. On the target machine, using just five percent of the available data as labeled examples, the model achieved a prediction accuracy that explained 95.7 percent of the variation in the actual force. The average error in the prediction was reduced to 17.8 newtons, a level of precision that rivals models trained on much larger datasets. Even when the amount of labeled data was reduced to just one percent, the method still outperformed other techniques, maintaining a high level of reliability. The researchers also found that the method remained robust even when the vibration signals were contaminated with noise, a common occurrence in real industrial environments. By focusing on the underlying patterns of the signal rather than memorizing specific data points, the model could ignore the static and focus on the true cutting force.

This work demonstrates that it is possible to transfer knowledge from one machine to another without needing to re-measure every single cut with expensive sensors. The study explicitly rules out the idea that simply making a neural network deeper or more complex is enough to solve the problem of different machines; without the specific steps of self-supervised learning and domain alignment, the model tends to get confused by the differences in signal structure. The findings suggest that the key to success lies in teaching the model to understand the general physics of vibration before asking it to predict specific forces. By using the abundant unlabeled data that factories already have, and supplementing it with a minimal amount of labeled data, this approach offers a practical path toward intelligent, real-time monitoring of machining processes. It moves the industry closer to a future where machines can monitor their own health and performance without the burden of complex, intrusive sensor setups, making advanced manufacturing more efficient and accessible.

Technical Summary: Cross-Device Milling Force Prediction via Self-Supervised Pretraining and Domain Adaptation

1. Problem Statement

Milling force is a critical indicator for tool wear, surface integrity, and machining safety. While direct measurement using dynamometers is accurate, it is often impractical for industrial deployment due to high costs, installation constraints, and potential alterations to system stiffness. Consequently, indirect prediction using vibration signals is preferred. However, existing data-driven models face a significant bottleneck in cross-device applications:

Distribution Discrepancy: Different machine tools possess unique structural stiffness, spindle systems, and sensor installation positions. This causes the same cutting force to generate vibration signals with different amplitude scales, frequency distributions, and temporal characteristics across devices.
Label Scarcity: Acquiring labeled milling force data on a target device requires expensive equipment and synchronization, resulting in a scenario where the target domain has abundant unlabeled vibration data but very few labeled force samples.
Model Failure: Models trained on a source device often fail when applied directly to a target device because they learn device-specific artifacts rather than stable force-related dynamics.

2. Methodology

The paper proposes a unified framework integrating Self-Supervised Pretraining (SSL), Domain Adaptation (DA), and Few-Label Fine-Tuning to transfer vibration-force mapping capabilities from a source device to a target device.

2.1 Data Preprocessing and Task Formulation

Signal Processing: Raw vibration signals undergo stable cutting segment extraction, min-max normalization (to address amplitude differences), and temporal alignment (using cross-correlation to synchronize vibration and force signals).
Windowing: Continuous signals are segmented into overlapping sliding windows to capture local dynamic patterns and periodic tooth impacts.
Data Partitioning: The framework utilizes:
- $D_s$ : Labeled source domain (vibration + force).
- $D_t^u$ : Unlabeled target domain (vibration only).
- $D_t^l$ : A small subset of labeled target domain samples (few-label).

2.2 Self-Supervised Pretraining (Masked Vibration Modeling)

To leverage abundant unlabeled data from both source and target devices, the author employs a Masked Vibration Modeling task:

Mechanism: Input vibration windows are corrupted by masking random time-frequency segments. A decoder reconstructs the missing values based on the visible context.
Objective: This forces the feature extractor (encoder) to learn intrinsic temporal structures, channel correlations, and local dynamic patterns without relying on force labels.
Benefit: The encoder learns transferable representations that are less biased toward specific device characteristics before supervised training begins.

2.3 Domain-Adaptive Feature Learning

Following pretraining, the model undergoes joint optimization to align source and target distributions:

Supervised Loss ( $L_{pred}^s$ ): Uses labeled source data to ensure the features remain sensitive to milling force variations.
Adversarial Domain Adaptation: A domain discriminator is introduced to distinguish between source and target features. The encoder is trained to fool the discriminator (via a gradient reversal layer), encouraging the learning of domain-invariant features.
Loss Function: The objective combines supervised regression, adversarial domain loss, and self-supervised reconstruction loss.

2.4 Few-Label Fine-Tuning

To correct residual biases specific to the target device's force amplitude and dynamic response:

A small number of labeled target samples ( $D_t^l$ ) are used to calibrate the regression head and upper layers of the feature extractor.
Lower layers of the encoder are often frozen or updated with a smaller learning rate to preserve the general representations learned during pretraining.

2.5 Model Architecture

The prediction network is a hybrid feature fusion network:

Deep Branch: Uses residual convolutional blocks for local pattern extraction, bidirectional recurrent layers (Bi-LSTM/GRU) for temporal dependencies, and multi-head attention mechanisms to weigh important vibration fragments.
Physical Branch: Explicitly extracts statistical, spectral, and time-frequency features (e.g., energy, frequency distribution) to provide stable physical descriptions.
Fusion: Deep and physical features are concatenated and normalized before passing through a fully connected regression head.

3. Key Contributions

Unified Framework: Established a cross-device prediction framework that jointly utilizes labeled source data, unlabeled target data, and few labeled target samples, reducing dependence on large-scale target-domain labeling.
Self-Supervised Strategy: Designed a masked vibration modeling task to learn transferable temporal and spectral representations from unlabeled data, enabling the model to understand intrinsic vibration dynamics prior to force supervision.
Integrated Adaptation: Combined adversarial domain adaptation with few-label fine-tuning to simultaneously reduce device-induced distribution shifts and calibrate target-specific prediction biases.

4. Experimental Results

Experiments were conducted using data from two different vertical machining centers (Source: EM800A; Target: FANUC system) with Inconel 718 workpieces.

In-Device Performance: On the source device, the model achieved an average $R^2$ of 0.9923. On the target device (after fine-tuning), it maintained an average $R^2$ of 0.9908.
Cross-Device Performance (SSL + DA):
- 0% Target Labels: Achieved $R^2$ of 0.862 and RMSE of 47.3 N (outperforming direct transfer and DA-only baselines).
- 1% Target Labels: $R^2$ increased to 0.917, RMSE dropped to 35.2 N.
- 5% Target Labels: $R^2$ reached 0.957, RMSE 24.9 N, and MAE 17.8 N.
- 20% Target Labels: RMSE further decreased to 19.4 N, with $R^2$ of 0.973.
Feature Alignment: The Maximum Mean Discrepancy (MMD) distance between source and target features decreased from 1.00 (raw) to 0.22 after the full training process, indicating successful distribution alignment.
Robustness: Under strong noise (5 dB SNR), the proposed method maintained an $R^2$ of 0.794, significantly outperforming direct transfer ( $R^2$ 0.515).
Efficiency: The model uses 1.68M parameters and requires 0.058 ms per sample, offering a favorable balance between accuracy and inference speed compared to Transformer-based models.

5. Significance and Claims

The paper claims that this study provides a feasible solution for extending milling force prediction from single-device experimental validation to cross-device, few-label, and practical industrial deployment scenarios.

Addressing the Bottleneck: It directly addresses the key bottleneck of insufficient labeled target-domain data by demonstrating that high accuracy can be achieved with as little as 5% labeled target samples when combined with self-supervised pretraining and domain adaptation.
Transferability: The results verify that the proposed method effectively transfers vibration-force mapping capabilities across devices with different structural and dynamic characteristics, overcoming the limitations of traditional models that assume identical data distributions.
Industrial Viability: By reducing the need for expensive force sensors on every target machine and maintaining robustness against noise and parameter variations, the method supports non-invasive, real-time process monitoring and intelligent machining control.

The author concludes that the integration of self-supervised learning, domain adaptation, and few-label fine-tuning creates a robust framework capable of balancing prediction accuracy, transferability, and computational efficiency for industrial applications.