卷积神经网络用于近红外光谱古筝面板木材分级

Wood Quality of Chinese Zither Panels Based on Convolutional Neural Network and Near-Infrared Spectroscopy

目前,我国乐器制作行业在古筝面板用木材等级的筛选上主要依赖于技师主观评判,但此法缺少科学理论的依据,效率低,客观性及出材率的提高等方面受到限制,无法满足乐器市场的大量需求。实现古筝面板用木材快速、智能化的分级工作是一个急需解决的课题。近红外光谱非常适用于测量含氢的有机物质。古筝面板木材主要化学成分的化学键均由含氢基团组成,不同等级板材的化学成分存在差异,这些差异反映在近红外光谱中,为判断木材等级提供了可能。同时卷积神经网络对非线性数据具有较强的特征提取能力,所以提出一种应用卷积神经网络模型对光谱数据进行分析的方法,进而判别木材的等级。应用了Savitzky Golay一阶、二阶微分两种预处理方法和核主成分分析、连续投影算法两种数据压缩方法,通过所设计的卷积神经网络模型以样本识别准确率和模型构建过程中的损失值作为判定指标选出最佳预处理和数据压缩方法。为了提高模型提取分析光谱数据的能力和避免过拟合现象,应用了多通道卷积核、批量归一化和early stopping策略,将通过两层卷积层提取的特征信息送入全连接层,从而充分提取剩余信息,通过Softmax函数获得板材的最终预测等级,从而确定了最终模型。最终Savitzky Golay一阶微分和核主成分分析为最佳数据处理方法,同时得出用于区分不同等级的古筝面板用木材的主要关键谱带,分别为1 163~1 243, 1 346~1 375和1 525~1 584 nm。将该模型应用于测试集样本,古筝面板用木材的等级识别准确率为95.5%。实验结果表明所提出的方法可以高效地处理光谱数据,有效识别区分不同等级的古筝面板用木材的关键特征,从而为广阔的乐器市场提供一定的技术支持。

Currently, the instrument production industry relies mainly on the subjective judgment of instrumental technicians when selecting the wood for Chinese zither panels. However, this method lacks a summary of scientific theories and is inefficient, which limits the objectivity of the selection and the improvement of the yield. Moreover, the current model for judging the wood grade cannot satisfy the large demand of the musical instrument market. Therefore, achieving rapid and intelligent grading of wood for Chinese zither panels is an urgent problem to be solved. Near-infrared spectroscopy contains information about the molecular structure of an object and is very suitable for measuring organic substances containing hydrogen. The chemical bonds of the main chemical components of wood used in Chinese zither panels are composed of hydrogen-containing groups, and the chemical compositions of the panels of different grades are different. These differences are reflected in near-infrared spectral data by light, which makes it possible to judge the wood grade. Simultaneously, convolutional neural network(CNN) has a strong feature extraction ability for nonlinear data. Therefore, this paper proposes a method to analyze the spectral data by using the CNN model to determine the wood grade. In the experiment, this paper applied two spectral preprocessing methods, like the Savitzky Golay first-derivative and second-derivative preprocessing methods, and two data compression methods, like kernel principal component analysis(KPCA) and successive projections algorithm. Through the CNN model designed in the paper, the optimal preprocessing and data compression methods were selected by using the classification accuracy rate of samples and the loss value in the model construction process as the judgment indicators. In order to improve the ability of the experimental model to extract and analyze spectral data and avoid overfitting, this experiment applied multi-channel convolution kernel, batch normalization and early stopping strategies. Finally, the feature information extracted by the two convolution layers was sent into the fully connected layers to extract other residual features, and the prediction grade of the panel was obtained using the softmax function. Thus, the final experimental model was determined. Finally, Savitzky Golay first-derivative and KPCA were the optimal data processing methods. At the same time, the main key bands for distinguishing different wood grades were obtained, which were 1 163~1 243 and 1 346~1 375 and 1 525~1 584 nm, respectively. Applying the proposed model to the test set samples, the grade classification accuracy of the wood for Chinese zither panels was 95.5%. Experimental results revealed that the proposed method can efficiently process spectral data and identify the key features of different grades of wood for Chinese zither panels. Therefore, it can provide specific technical support for the broad instrument market.