⚡ electrical engineering

A Method for X-ray Computed Tomography Based Prediction of Local Cutting Power in Circular Sawing of Scots Pine

This study demonstrates that a machine learning model trained on high-resolution X-ray computed tomography data can accurately predict local cutting power during circular sawing of Scots pine by capturing the influence of specific anatomical features like fibre orientation, density, and knots, thereby outperforming traditional models based on global material properties.

Original authors: Yunbo Huang, Johannes Huber, Magnus Fredriksson, Julie Cool, Mikael Svensson

Published 2026-09-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yunbo Huang, Johannes Huber, Magnus Fredriksson, Julie Cool, Mikael Svensson

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

Imagine a world where the tools we use to build our homes and furniture could "see" the material they are about to cut before the blade even touches it. In the timber industry, this is not just a fantasy but a practical necessity. Wood is a living material, grown in complex patterns that vary from one tree to the next and even within a single board. When a circular saw slices through a piece of pine, the energy required to make that cut changes constantly. It spikes when the blade hits a hard knot, shifts when the grain runs at a sharp angle, and fluctuates with the density of the wood. For decades, machines have operated blindly, reacting to these changes only after they happen, often leading to uneven cuts, wasted energy, or damaged tools. The question researchers have long asked is whether it is possible to predict these energy demands in advance by looking inside the wood before the sawing begins.

A team of researchers from Sweden and Canada has taken a significant step toward answering this question by combining advanced imaging with modern computing. They focused on Scots pine, a common timber used in construction, and set out to see if they could predict exactly how much power a saw blade would need at every millimeter of its path. Instead of treating a piece of wood as a uniform block with average properties, they treated it as a detailed map of internal structures. Using a high-resolution X-ray scanner, similar to the medical CT scans used in hospitals but adapted for industrial use, they created a three-dimensional digital model of six pine boards. This scan revealed the wood's internal density, the orientation of its fibers, and the precise location of knots and other natural features.

The researchers then took these scanned boards to a specialized workshop where they cut 250 narrow grooves into the wood using computer-controlled saws. They tested different types of blades, cutting speeds, and directions to create a wide variety of cutting conditions. Crucially, they measured the electrical power consumed by the saw motor at a very high speed, capturing the exact energy fluctuations as the blade moved through the wood. The core of their work was aligning the digital map from the X-ray scan with the real-time power measurements. They matched every point of the cut to the specific wood structure the blade was encountering at that exact moment, creating a massive dataset of over 200,000 individual observations.

With this data in hand, the team trained several computer models to learn the relationship between the wood's internal features and the energy required to cut it. They tested different mathematical approaches, including linear models and more complex neural networks, to see which could best predict the power spikes and dips. The results were striking. The most successful model, a type of artificial intelligence known as a multilayer perceptron, was able to predict the local cutting power with remarkable accuracy. It achieved an error rate of just 5.3 watts, which is tiny compared to the total power of roughly 1,000 watts used during the cuts. More importantly, the model did not just guess an average value; it successfully reproduced the specific shape of the power curve, accurately predicting where the energy would rise and fall as the blade encountered knots or changes in grain direction.

The study revealed that the most important factor in predicting cutting power was simply the local density of the wood. Where the wood was denser, the saw needed more power. The angle of the wood fibers relative to the blade was the second most critical factor, followed by the number of teeth on the blade that were engaged in the cut at any given moment. The models showed that traditional methods, which rely on average properties of the entire board, miss these critical local details. By using the X-ray data, the researchers could account for the specific influence of growth rings, wood types like sapwood versus heartwood, and the disruptive effect of knots.

However, the researchers were careful to note the limits of their findings. While the models performed exceptionally well in the controlled laboratory environment, they found that the predictions were less perfect near knots and other abrupt changes in the wood structure. This is likely because the physical machine and the saw blade have their own inertia and dynamic responses that smooth out the sudden changes in resistance, effects that the current computer models do not yet fully capture. Additionally, the study was conducted on dry wood under specific conditions, meaning the approach would need further testing with wetter wood or frozen timber before it could be used in a real-world sawmill.

This work demonstrates that it is possible to move beyond average estimates and understand the cutting process at a very local level. By using X-ray imaging to see inside the wood and machine learning to interpret that data, the researchers have shown a path toward smarter, more efficient wood machining. In the future, this technology could allow sawmills to adjust their machines in real-time based on the specific internal structure of each log, reducing energy waste and improving the quality of the lumber produced. For now, the study stands as a proof of concept, showing that the invisible patterns inside a tree can be translated into precise predictions about the work required to shape it.

Technical Summary: A Method for X-ray Computed Tomography Based Prediction of Local Cutting Power in Circular Sawing of Scots Pine

Problem Statement
Circular sawing of wood is an energy-intensive process heavily influenced by the local anatomical variability of the material, including density fluctuations, grain orientation, and defects such as knots. While cutting power is a critical signal for monitoring tool wear, surface quality, and productivity, existing predictive models predominantly rely on global material properties or average piece-wise values. These conventional approaches fail to capture localized resistance variations caused by the three-dimensional fiber structure and the specific curved path of the circular blade. Consequently, local overloads and instability, which can degrade quality even when average power appears acceptable, remain difficult to predict. There is a gap in methods that align local descriptors of internal wood structure with tool engagement along the cut path and evaluate prediction under generalization settings that reflect real-world deployment rather than data leakage.

Methodology
The study assembled a dataset comprising approximately 212,000 localized cutting observations derived from 250 grooves cut in six Scots pine (Pinus sylvestris L.) boards. The experimental workflow integrated non-destructive testing with controlled machining:

CT Imaging and Feature Extraction: Six sawn and planed specimens were scanned using an industrial X-ray computed tomography (CT) system (Microtec Mito) at 0.3 mm spatial resolution. From the reconstructed 3D density volumes, the authors extracted local density, fiber orientation angles relative to the cutting, feed, and orthogonal directions, and structural masks for knots, sapwood/heartwood boundaries, and juvenile wood. Knot influence was modeled using a distance-dependent decay function from the knot pith.
Machining and Data Alignment: Specimens were machined on a 5-axis CNC sawing setup using two blade types (cross-cut and rip-cut) under varying spindle speeds, feed speeds, cutting depths, and feed orientations (climb and counter). Spindle motor current was recorded at 500 kHz to derive cutting power profiles.
Data Registration: A curvilinear coordinate transformation mapped the circular saw blade engagement path onto a straightened 1D local progress index ( $\xi \in [0, 1]$ ). This allowed the alignment of CT-derived features (density, angles, masks) with the measured power profiles at 759 uniformly spaced positions per groove.
Modeling and Validation: Four regression models were trained: Ridge regression, a Multilayer Perceptron (MLP), Extreme Gradient Boosting (XGB), and a spline-based model. To prevent optimistic bias from correlated observations within a single cut, the study employed group-aware validation strategies: stratified 10-fold cross-validation (keeping entire grooves within folds) and a groove-level leave-one-out (LOO) test. Feature importance was analyzed using SHapley Additive exPlanations (SHAP) on the XGB model.

Key Contributions

Local-Scale Prediction: The study demonstrates a data-driven approach that predicts cutting power on a millimeter scale, moving beyond global averages to resolve local variations driven by growth rings, wood type, and knots.
CT-Process Integration: It establishes a framework for registering high-resolution CT-derived anatomical features (specifically fiber orientation and density) to the exact curved trajectory of a circular saw blade.
Rigorous Validation: By utilizing groove-level LOO validation, the study ensures that models are tested on entirely unseen cutting paths, providing a more realistic assessment of generalization capability than standard random splitting.
Interpretability: The use of SHAP values provides a theoretically grounded attribution of how specific material properties (e.g., density, fiber angle) and process parameters influence local cutting resistance.

Results

Model Performance: The Multilayer Perceptron (MLP) achieved the highest predictive accuracy, with a mean Root-Mean-Square Error (RMSE) of 5.27 W and a coefficient of determination ( $R^2$ ) of 0.908 across 10-fold cross-validation. The XGB model performed second-best ( $RMSE \approx 6.54$ W, $R^2 \approx 0.887$ ). Linear models (Ridge and Spline) showed significantly higher errors and greater variability, suggesting the relationship between CT descriptors and cutting power is inherently nonlinear.
Profile Reproduction: In groove-level LOO tests, the models successfully reproduced the general shape of measured power profiles, particularly in clear-wood regions. The models accurately tracked power increases associated with knot engagement and density peaks.
Discrepancies: Larger prediction errors were observed near knots and in rip-cut scenarios. The authors attribute these deviations to dynamic system effects (e.g., spindle inertia smoothing rapid resistance changes) and machine-specific CT reconstruction artifacts (axial intensity shifts) that are not explicitly captured in the predictor set.
Feature Importance: SHAP analysis identified local CT-derived density as the dominant predictor, positively correlated with cutting power. The angle between the longitudinal fiber direction and the cutting direction was the second most influential feature. Machining parameters, such as the number of teeth in contact, also contributed but were partially mediated by local wood features.

Significance and Claims
The paper claims that CT-derived local anatomy and fiber orientation enable data-driven models to represent the influence of growth rings, wood type, and knots on local cutting resistance more accurately than models based on global properties. The framework demonstrates that local circular-sawing power can be predicted from CT data with high accuracy, supporting the potential for CT-informed process planning and intelligent monitoring at the log or board level.

However, the authors maintain a modest scope regarding immediate industrial application. They state that the method is currently most suitable for "relative and profile-level prediction under controlled laboratory conditions" rather than for direct industrial deployment. The study highlights that future maturation of the approach requires the incorporation of physically motivated descriptors of tooth engagement and machine dynamics, as well as validation on larger, more balanced datasets that include variables such as moisture content, temperature, and frozen wood states.

More like this