构造、物理与无损检测多源特征融合的华北落叶松木材力学性能机器学习预测

孙子上; 焦立超; 郑昌; 詹伟辉; 姜笑梅; 周海宾; 殷亚方

doi:10.12326/j.2096-9694.2026050

“古建筑木材科研基地”专栏 | 浏览量 : 0 下载量: CSCD: 0 CNKI被引量: 0 Crossref: 0 Wos: 0 Scopus: 0 更多指标

PDF
导出
分享
收藏
专辑

构造、物理与无损检测多源特征融合的华北落叶松木材力学性能机器学习预测
Machine Learning Prediction of Mechanical Properties of Larix gmelinii var. principis-rupprechtii Wood by Integrating Multi-Source Features of Anatomy, Physics and Non-Destructive Testing
1.中国林业科学研究院木材工业研究所，北京 100091
2.中国林业科学研究院木材标本馆，北京 100091
3.国家林业和草原局木材标本资源库，北京 100091
4.南京林业大学，江苏南京 210037
2026年40卷第3期页码：1-11
DOI： 10.12326/j.2096-9694.2026050
稿件说明：

移动端阅览

孙子上,焦立超,郑昌等.构造、物理与无损检测多源特征融合的华北落叶松木材力学性能机器学习预测[J].木材科学与技术,2026,40(03):1-11. DOI： 10.12326/j.2096-9694.2026050.

SUN Zishang,JIAO Lichao,ZHENG Chang,et al.Machine Learning Prediction of Mechanical Properties of Larix gmelinii var. principis-rupprechtii Wood by Integrating Multi-Source Features of Anatomy, Physics and Non-Destructive Testing[J].Chinese Journal of Wood Science and Technology,2026,40(03):1-11. DOI： 10.12326/j.2096-9694.2026050.

摘要

针对古建筑严重残损木构件替换选材时普遍存在的木材力学性能预测效率低、测试流程复杂等问题，构建木材力学性能快速预测方法并建立修缮用材适用性评价体系，是科学支撑古建筑替换选材的关键。以华北落叶松（

Larix gmelinii

var.

principis-rupprechtii

）为研究对象，测定生长轮宽度、晚材率两项构造特征与气干密度指标，并结合应力波波速、微钻阻力，以及二者耦合得到的波阻模量复合参数等无损检测参数，引入随机森林（random forest，RF）、决策树（decision tree，DT）、K近邻回归（K-nearest neighbors regression，KNNR）、支持向量回归（support vector regression，SVR）及表格先验数据拟合网络（tabular prior-data fitted network，TabPFN）5种机器学习算法，构建样本数量不足条件下木材抗弯强度、抗弯弹性模量、顺纹及横纹抗压强度等力学性能的预测模型。在各项指标中，气干密度、晚材率、微钻阻力值与力学性能的Pearson系数最高（均大于0.5），而生长轮宽度和应力波波速与力学性能相关性较弱。将气干密度、晚材率与微钻阻力值三类高相关多源特征融合构建的模型，其预测精度较单一使用构造与物理指标或无损检测参数分别提高23.26%和72.05%。5种模型中，基于预训练机制的TabPFN在力学性能预测中表现最优，预测抗弯强度的决定系数

²高达0.961 4。以沙普利加性解释（Shapley additive explanations，SHAP）值量化特征贡献度，进一步筛选出气干密度、晚材率与微钻阻力值为核心预测特征，其中气干密度贡献度最高（平均SHAP值为61.29%）。研究构建的融合构造、物理指标与无损检测参数的机器学习预测方法，实现对华北落叶松关键力学性能的快速、准确评估，为应县木塔等古木建筑修缮工程中严重残损构件的科学选材提供重要依据。

Abstract

To address the low efficiency of prediction and complex testing procedures for the mechanical properties of wood in selection of the wood materials for replacing severely damaged wooden components in ancient buildings

developing a rapid method for predicting wood mechanical properties and establishing a suitability evaluation system for restoration materials is of great importance. This study measured anatomical and physical indicators of

Larix gmelinii

var

. principis-rupprechtii

specimens

including growth ring width

latewood percentage

and air-dry density. Simultaneously

non-destructive testing (NDT) parameters were obtained

including stress wave velocity

micro-drilling resistance

and their coupled composite parameter (wave resistance modulus). Five machine learning algorithms including Random Forest (RF)

Decision Tree (DT)

K-Nearest Neighbors Regression (KNNR)

Support Vector Regression (SVR)

and Tabular Prior-Data Fitted Network (TabPFN)

were employed to construct prediction models for key mechanical properties

including modulus of rupture

modulus of elasticity

and compressive strength parallel and perpendicular to the grain

under limited sample conditions. Among all indicators

air-dry density

latewood percentage

and micro-drilling resistance exhibited the highest Pearson correlation coefficients with the mechanical properties (all

0.5)

whereas growth ring width and stress wave velocity showed relatively weak correlations. By integrating these three highly correlated multi-source features (air-dry density

latewood percentage

and micro-drilling resistance) for modeling

the prediction accuracy improved by 23.26% and 72.05%

respectively compared to using anatomical/physical indicators or NDT parameters alone. Among these five models

the pre-training-based TabPFN achieved the best prediction approach for the mechanical properties

yielding an

² of 0.961 4 for the modu

lus of rupture. Using Shapley additive explanations (SHAP) values

feature contribution analysis further identified air-dry density

latewood percentage

and micro-drilling resistance as the core predictive features

with air-dry density contributing the most (average SHAP value of 61.29%). The proposed machine learning-based prediction method

integrating growth features

physical

and NDT parameters

enables a rapid and accurate method for predicting the key mechanical properties of

Larix gmelinii

var.

principis-rupprechtii

wood

providing a critical scientific basis for selecting suitable materials for replacing severely damaged components in the restoration of ancient timber structures such as the Yingxian Wooden Pagoda.

关键词

Keywords

references

李铁英 , 秦慧敏 . 应县木塔现状结构残损分析及修缮探讨 [J ] . 工程力学 , 2005 , 22 ( S1 ): 199 - 212 .

LI T Y , QIN H M . Structural analysis and repair of Yingxian wooden tower [J ] . Engineering Mechanics , 2005 , 22 ( S1 ): 199 - 212 .

董梦妤 . 古建筑和出土饱水木材鉴别与细胞壁结构变化 [D ] . 北京 : 北京林业大学 , 2017 .

黄磊 , 王雨晨 , 韩旭 , 等 . 应县木塔普拍枋用材方法与残损状况分析 [J ] . 木材科学与技术 , 2024 , 38 ( 4 ): 8 - 19 .

HUANG L , WANG Y C , HAN X , et al . Material application method and component damage analysis of the Pu-Pai Fang in Yingxian wooden pagoda [J ] . Chinese Journal of Wood Science and Technology , 2024 , 38 ( 4 ): 8 - 19 .

邓扬 , 王琳溪 , 李雨航 , 等 . 应县木塔斗栱-倾斜柱的抗震性能和加固试验研究 [J ] . 建筑结构学报 , 2026 , 47 ( 1 ): 190 - 199 .

DENG Y , WANG L X , LI Y H , et al . Experimental study on seismic performance and reinforcement for inclined Dougong-column in Yingxian Wood Pagoda [J ] . Journal of Building Structures , 2026 , 47 ( 1 ): 190 - 199 .

周海宾 , 韩旭 , 黄磊 , 等 . 古建筑木构件开裂机制、评估与加固研究进展 [J ] . 木材科学与技术 , 2024 , 38 ( 1 ): 13 - 22 .

ZHOU H B , HAN X , HUANG L , et al . Research progress on cracking mechanism, evaluation and reinforcement of ancient building timber members [J ] . Chinese Journal of Wood Science and Technology , 2024 , 38 ( 1 ): 13 - 22 .

古建筑木结构维护与加固技术标准 GB/T 50165—2020 [S ] .

ICOMOS International Wood Committee . Principles for the conservation of wooden built heritage [Z ] . Delhi , India , 2017 .

Forest Products Laboratory (US) . Wood handbook: wood as an engineering material [M ] . The Laboratory , 1999 : 543 .

王忠铖 , 杨娜 , 李久林 . 无损检测指标与木材横纹局部抗压强度和密度的相关性 [J ] . 北京林业大学学报 , 2024 , 46 ( 5 ): 136 - 142 .

WANG Z C , YANG N , LI J L . Correlations between nondestructive testing indicators and local transverse compressive strength as well as wood density [J ] . Journal of Beijing Forestry University , 2024 , 46 ( 5 ): 136 - 142 .

杨伟 , 谢莉祥 , 吴晓静 , 等 . 基于无损检测技术的古建筑木构件力学性能测试研究 [J ] . 林产工业 , 2025 , 62 ( 7 ): 36 - 41 .

YANG W , XIE L X , WU X J , et al . Research on mechanical properties testing of wooden components of ancient buildings based on non-destructive testing technology [J ] . China Forest Products Industry , 2025 , 62 ( 7 ): 36 - 41 .

袁霄 , 钟慧娴 , 钟永 , 等 . 古建用落叶松木材物理力学性能预测及其影响因素 [J ] . 林业工程学报 , 2024 , 9 ( 2 ): 63 - 69 .

YUAN X , ZHONG H X , ZHONG Y , et al . Prediction of physical and mechanical properties and influencing factors of larch wood in historic buliding [J ] . Journal of Forestry Engineering , 2024 , 9 ( 2 ): 63 - 69 .

赵渊祥 , 吴明长 , 张俊 , 等 . Pilodyn在马尾松大径材活立木基本密度预测中的应用 [J ] . 西南林业大学学报 , 2025 , 45 ( 1 ): 213 - 219 .

ZHAO Y X , WU M C , ZHANG J , et al . Application of Pilodyn method in predicting the basic density of standing trees of Pinus massoniana large-diameter wood [J ] . Journal of Southwest Forestry University , 2025 , 45 ( 1 ): 213 - 219 .

王忠铖 . 藏青杨古建木材力学性质试验研究及预测方法 [D ] . 北京 : 北京交通大学 , 2022 .

CHEN Y , GUO W . Mechanical properties evaluation of two wood species of ancient timber structure with nondestructive testing methods [J ] . BioResources , 2016 , 11 ( 3 ): 6600 - 6612 .

高鑫 , 鲍咏泽 , 李珠 , 等 . 2个家系落叶松木材抗弯性能径向变化特征与预测 [J ] . 木材科学与技术 , 2025 , 39 ( 6 ): 18 - 25 .

GAO X , BAO Y Z , LI Z , et al . Radial variation and prediction of the bending properties of larch wood from two families [J ] . Chinese Journal of Wood Science and Technology , 2025 , 39 ( 6 ): 18 - 25 .

BREIMAN L . Random forests [J ] . Machine Learning , 2001 , 45 ( 1 ): 5 - 32 .

MIENYE I D , JERE N . A survey of decision trees: concepts, algorithms, and applications [J ] . IEEE Access , 2024 , 12 : 86716 - 86727 .

KANUNGO T , MOUNT D M , NETANYAHU N S , et al . An efficient k-means clustering algorithm: analysis and implementation [J ] . IEEE Transactions on Pattern Analysis and Machine Intelligence , 2002 , 24 ( 7 ): 881 - 892 .

CORTES C , VAPNIK V . Support-vector networks [J ] . Machine Learning , 1995 , 20 ( 3 ): 273 - 297 .

HOLLMANN N , MÜLLER S , PURUCKER L , et al . Accurate predictions on small data with a tabular foundation model [J ] . Nature , 2025 , 637 ( 8045 ): 319 - 326 .

无疵小试样木材物理力学性质试验方法第5部分：密度测定 GB/T 1927.5—2021 [S ] .

无疵小试样木材物理力学性质试验方法第3部分：木材生长轮与晚材率测定 GB/T 1927.3—2021 [S ] .

无疵小试样木材物理力学性质试验方法第9部分：抗弯强度测定 GB/T 1927.9—2021 [S ] .

无疵小试样木材物理力学性质试验方法第10部分：抗弯弹性模量测定 GB/T 1927.10—2021 [S ] .

无疵小试样木材物理力学性质试验方法第11部分：顺纹抗压强度测定 GB/T 1927.11—2021 [S ] .

无疵小试样木材物理力学性质试验方法第12部分：横纹抗压强度测定 GB/T 1927.12—2021 [S ] .

古建筑结构安全性鉴定技术规范第1部分: 木结构 DB11/T 1190.1—2015 [S ] .

BICHRI H , CHERGUI A , HAIN M . Investigating the impact of train/test split ratio on the performance of pre-trained models with custom datasets [J ] . International Journal of Advanced Computer Science and Applications , 2024 , 15 ( 2 ): 331 - 339 .

周贤武 , 张俊珍 , 周海宾 , 等 . 树龄对日本落叶松木材物理力学性质的影响 [J ] . 林业科技开发 , 2014 , 28 ( 4 ): 54 - 57 .

ZHOU X W , ZHANG J Z , ZHOU H B , et al . Tree age’s effects o n physical and mechanical properties of Larix kaempferi wood [J ] . China Forestry Science and Technology , 2014 , 28 ( 4 ): 54 - 57 .

黄鹤 , 杨晨 , 李珠 , 等 . 杂种落叶松早材和晚材拉伸弹性模量预测 [J ] . 木材科学与技术 , 2025 , 39 ( 6 ): 26 - 32 .

HUANG H , YANG C , LI Z , et al . Prediction of tensile modulus of elasticity of earlywood and latewood from hybrid larch [J ] . Chinese Journal of Wood Science and Technology , 2025 , 39 ( 6 ): 26 - 32 .

白晓彬 , 吴婧姝 , 杨娜 , 等 . 山西落叶松古旧木材材料性能无损检测技术研究 [J ] . 太原理工大学学报 , 2023 , 54 ( 6 ): 1101 - 1108 .

BAI X B , WU J S , YANG N , et al . Study on nondestructive testing technology of material properties for ancient larch wood in Shanxi [J ] . Journal of Taiyuan University of Technology , 2023 , 54 ( 6 ): 1101 - 1108 .

HOLLMANN N , MÜLLER S , EGGENSPERGER K , et al . TabPFN: a transformer that solves small tabular classification problems in a second [EB/OL ] . 2022 : arXiv : 2207 . 01848 . https://arxiv.org/abs/2207.01848 https://arxiv.org/abs/2207.01848 .

GRINSZTAJN L , OYALLON E , VAROQUAUX G . Why do tree-based models still outperform deep learning on typical tabular data? [C ] // Advances in Neural Information Processing Systems . New Orleans, Louisiana, USA. Neural Information Processing Systems Foundation, Inc. (NeurIPS) , 2022 : 507 - 520 .

VIERING T , LOOG M . The shape of learning curves: a review [J ] . IEEE Transactions on Pattern Analysis and Machine Intelligence , 2022 , 45 ( 6 ): 7799 - 7819 .

HOIEM D , GUPTA T , LI Z , et al . Learning curves for analysis of deep networks [C ] // International conference on machine learning . PMLR , 2021 : 4287 - 4296 .

VAN DEN BROECK G , LYKOV A , SCHLEICH M , et al . On the tractability of SHAP explanations [J ] . Journal of Artificial Intelligence Research , 2022 , 74 : 851 - 886 .

徐有明 , 徐姗姗 , 林汉 , 等 . 火炬松种源建筑材抗弯性质的变异及与树龄、晚材率、木材密度的关系 [J ] . 林业科学 , 2007 ( 2 ): 77 - 83 .

XU Y M , XU S S , LIN H , et al . Variation in wood bending properties of exotic loblolly pine provenances for building lumber and their relationships to tree age, tree growth and wood density [J ] . Scientia Silvae Sinicae , 2007 ( 2 ): 77 - 83 .