基于VMD-LightGBM特征融合的蛋壳强度敲击振动无损检测

Non-destructive Testing of Eggshell Strength Knocking Vibration Based on VMD-LightGBM Feature Fusion

摘要:

蛋壳强度是保障禽蛋耐储运性及加工品质的关键力学指标。针对传统破坏性检测造成样本不可逆损耗，以及现有声学无损检测在处理非平稳敲击信号时存在模态混叠、特征提取困难等问题，本文提出一种基于变分模态分解(Variational mode decomposition, VMD)与轻量化梯度提升机(Light gradient boosting machine, LightGBM)的蛋壳强度无损检测方法，旨在以准静态压缩试验作为精度验证基准，通过无损手段直接量化蛋壳力学性能。搭建了自动敲击振动采集系统，获取500枚鸡蛋瞬态振动响应信号。利用VMD算法的自适应频带分离特性，将复杂非平稳敲击信号分解为多个具有独立中心频率的固有模态分量(Intrinsic mode function，IMFs），有效抑制了信号模态混叠现象;构建了融合时域统计量、快速傅里叶变换频谱特征、小波时频系数及VMD模态参数37维多域混合特征集;建立LightGBM强度预测模型，采用10折嵌套交叉验证结合Bootstrap自助法，克服小样本数据带来的过拟合风险并量化泛化误差;引入沙普利加性解释(Shapley additive explanations,SHAP)方法揭示特征力学表征含义，并通过与破坏性实测值Fmax对比，验证了方法有效性。试验结果表明，构建的LightGBM无损检测模型预测值与破坏性实测值高度吻合，决定系数R2达到0.958，平均绝对误差和均方根误差分别低至1.19 N和1.50 N。研究结果为替代传统破坏性检测提供了可靠的理论依据与技术支撑。

Abstract:

Eggshell strength is a critical mechanical indicator determining the storage stability, transportability, and processing quality of poultry eggs. To address the irreversible sample damage caused by traditional destructive testing, as well as the challenges of mode mixing and difficult feature extraction in existing acoustic non-destructive testing (NDT) of non-stationary impact signals, a detection method was proposed based on variational mode decomposition (VMD) and light gradient boosting machine (LightGBM). Aiming to directly quantify eggshell mechanical properties via non-destructive means, utilizing quasi-static compression tests as the benchmark for accuracy validation, an automatic impact vibration acquisition system was constructed to capture transient vibration response signals from 500 eggs. Firstly, leveraging the adaptive frequency band separation capability of the VMD algorithm, the complex non-stationary impact signals were decomposed into multiple intrinsic mode functions (IMFs) with independent center frequencies, effectively suppressing mode mixing. Secondly, a 37-dimensional multi-domain hybrid feature set was established by integrating time-domain statistics, FFT spectral features, wavelet time-frequency coefficients, and VMD modal parameters. Finally, a LightGBM strength prediction model was developed. To mitigate the risk of overfitting associated with small sample sizes and to quantify generalization error, a 10-fold nested cross-validation combined with the Bootstrap method was employed. Furthermore, Shapley additive explanations (SHAP) were introduced to interpret the mechanical significance of the features, and the method's effectiveness was validated by comparing predictions with destructive measured values (Fmax). Experimental results demonstrated that the predicted values from the constructed LightGBM non-destructive detection model exhibited a high degree of agreement with the destructive measured values. The coefficient of determination (R2) reached 0.958, while the mean absolute error (MAE) and root mean square error (RMSE) were as low as 1.19 N and 1.50 N, respectively. The research result can provide a reliable theoretical basis and technical support for substituting traditional destructive testing.