基于数据融合与机器学习的奇楠沉香真伪多模态识别策略

Highlights
•Authentic Qi-nan was classified by machine learning algorithms using odor-signals.
•The accuracy of authentic Qi-nan was enhanced through the decision-level fusion method.
•A new rapid analysis technique for detecting the authenticity of Qi-nan was developed.

亮点
• 利用气味信号，通过机器学习算法对真品奇楠进行了分类。
• 通过决策级融合方法，提升了真品奇楠识别的准确率。
• 开发了一种用于检测奇楠真伪的快速分析新技术。

Abstract
Owing to its distinctive aroma and high economic value, Qi-nan agarwood is in high demand. However, its limited availability has resulted in rampant forgery, highlighting the urgent necessity for rapid, nondestructive, and intelligent authentication techniques. This study proposes a multimodal recognition framework employing an intermediate data fusion strategy for the accurate discrimination between genuine and counterfeit specimens. Initially, texture features were extracted from sample images utilizing discrete wavelet transform (DWT). Subsequently, odor signals were captured using an electronic nose (E-nose). In unimodal classification endeavors, multiple machine learning algorithms were applied to the E-nose data, achieving classification accuracies surpassing 90 %, with logistic regression (LR) demonstrating the optimal performance. Additionally, a decision-level data fusion model amalgamating E-nose odor features and DWT-derived texture features was established, achieving a classification accuracy of 100 %, thus outperforming the performance of all unimodal models. Further, the volatile compounds were validated by headspace solid-phase microextraction coupled with gas chromatography-mass spectrometry (HS-SPME-GC-MS) to verify the differences in the odor components of genuine and counterfeit samples. This study presents a novel, efficient, and practical methodology for the reliable authentication of Qi-nan agarwood, offering extensive potential applications in quality assurance and market regulation.

由于其独特的香气和高经济价值，奇楠沉香市场需求旺盛。然而，其供应量有限导致了猖獗的伪造行为，突显了对快速、无损、智能化真伪鉴别技术的迫切需求。本研究提出了一种采用中级数据融合策略的多模态识别框架，用于准确鉴别真伪样本。首先，利用离散小波变换从样本图像中提取纹理特征。随后，使用电子鼻捕获气味信号。在单模态分类尝试中，将多种机器学习算法应用于电子鼻数据，分类准确率超过90%，其中逻辑回归表现出最优性能。此外，建立了一个融合了电子鼻气味特征和DWT纹理特征的决策级数据融合模型，达到了100%的分类准确率，从而超越了所有单模态模型的性能。进一步地，通过顶空固相微萃取-气相色谱-质谱联用技术对挥发性化合物进行了验证，以确认真伪样品气味成分的差异。本研究为奇楠沉香的可靠鉴别提供了一种新颖、高效且实用的方法，在质量保证和市场监督方面具有广泛的应用潜力。

Introduction
Qi-nan agarwood, named Kyara or Kinam, represents an exceptionally rare and precious variety of agarwood. It is formed through the slow accumulation of resin over many years or even decades in Aquilaria spp. (family Thymelaeaceae) under stress conditions such as mechanical injury, fungal infection, or chemical attack [1], [2], [3]. Compared with conventional agarwood, Qi-nan agarwood is characterized by a higher yield and resin content, more complex and abundant aromatic constituents [4]. Renowned as the “King of Incense,” it holds significant cultural and economic value in traditional Chinese medicine, religious rituals, and premium fragrance markets [5], [6].

奇楠沉香，又称伽罗或奇楠，是一种极为稀有珍贵的沉香品种。它是由瑞香科沉香属植物在机械损伤、真菌感染或化学攻击等胁迫条件下，经过多年甚至数十年的树脂缓慢积累而形成的[1,2,3]。与常规沉香相比，奇楠沉香具有更高的产香量和树脂含量，以及更复杂丰富的芳香成分[4]。被誉为”众香之王”，它在传统中药、宗教仪式和高端香料市场中具有重要的文化和经济价值[5,6]。

However, the rarity and depletion of wild resources have created a supply-demand imbalance, thereby stimulating the rapid expansion of cultivated agarwood. Major agarwood-producing regions in China have introduced policies to promote cultivated agarwood, with the cultivated area now exceeding 67000 ha and continuing to expand. Among them, cultivated Qi-Nan agarwood is a newly developed premium agarwood variety in recent years. Compared with traditional agarwood, cultivated Qi-Nan offers several advantages, including a shorter resin-formation cycle, higher yield, and the ability to emit fragrance at ambient temperature, thereby becoming a major focus of industrial development in recent years [7], [8]. In terms of utilization, cultivated Qi-Nan agarwood is primarily processed into handicrafts such as bracelets, as well as essential oils and incense products, among which bracelets exhibit the greatest market advantage. Owing to their relatively affordable price and rich aromatic profile, Qi-Nan bracelets account for more than 70 % of sales through both online and offline channels, making them the most economically valuable product in the agarwood industry.

然而，野生资源的稀有和枯竭造成了供需失衡，从而刺激了人工种植沉香的快速发展。中国主要沉香产区已出台政策推广种植沉香，目前种植面积超过67000公顷，并持续扩大。其中，人工种植奇楠是近年来开发出的一个高端沉香新品种。与传统沉香相比，人工种植奇楠具有结香周期更短、产量更高、能在常温下发香等优势，因此成为近年来产业发展的重点[7,8]。在利用方面，人工种植奇楠沉香主要加工成手串等工艺品，以及精油和线香等产品，其中手串展现出最大的市场优势。由于其相对亲民的价格和丰富的香韵，奇楠手串通过线上线下渠道的销售额占70%以上，成为沉香产业中经济价值最高的产品。

Despite the significant reduction in price compared with wild agarwood, cultivated Qi-nan still commands a much higher market value and profit margin than most other wooden handicrafts. Coupled with its rapidly expanding consumer market, the supply of raw materials remains insufficient, which has led to the proliferation of various counterfeiting practices. These counterfeit goods not only cause economic losses to consumers due to price and quality discrepancies, but also pose health risks of skin allergies due to the possible chemical additives they may contain. Counterfeiting methods can generally be divided into low-grade methods and sophisticated methods. Low-grade counterfeiting typically involves carbonizing inferior agarwood at high temperature to imitate the visual appearance of high quality Qi-nan. More sophisticated approaches, however, differ from traditional falsification using non-agarwood materials, synthetic fragrances, or chemical additives. Instead, counterfeiters often employ supercritical extracts obtained directly from cultivated Qi-nan agarwood, which are either coated onto the surface of bracelets or injected into the wood matrix [9]. Since the chemical composition of such extracts originates from genuine Qi-nan, detection by a single analytical method is particularly challenging. Nevertheless, the extraction process inevitably causes the loss or imbalance of certain volatile constituents, leading to perceptible alterations in aroma even when the extracts are reintroduced into the wood or applied to its surface.

尽管与野生沉香相比价格大幅降低，但人工种植奇楠的市场价值和利润率仍远高于大多数其他木质工艺品。加上其快速扩张的消费市场，原材料供应仍然不足，这导致了各种造假行为的泛滥。这些假冒商品不仅因其价格和质量差异给消费者造成经济损失，还可能因其可能含有的化学添加剂而带来皮肤过敏的健康风险。造假方法大致可分为低级方法和高级方法。低级造假通常涉及将劣质沉香高温碳化以模仿高品质奇楠的外观。然而，更高级的方法与使用非沉香材料、合成香料或化学添加剂的传统伪造不同。造假者往往采用直接从人工种植奇楠沉香中获得的超临界萃取物，将其涂覆在手串表面或注入木材基质中[9]。由于此类萃取物的化学成分来源于真正的奇楠，通过单一分析方法检测尤其具有挑战性。尽管如此，萃取过程不可避免地会导致某些挥发性成分的损失或失衡，即使将萃取物重新注入木材或涂覆在其表面，也会引起香气的可察觉变化。

To address authenticity issues, conventional analytical methods such as thin-layer chromatography (TLC), mass spectrometry (MS) and fourier transform infrared spectroscopy (FTIR) are commonly employed [10], [11], [12]. Although these techniques provide detailed chemical compositional information with high sensitivity, they are generally time-consuming, complex in operation, and often destructive in nature—thus limiting their applicability in rapid, convenient, and non-destructive identification scenarios. More importantly, these methods have obvious drawbacks in terms of technology and application: they are difficult to effectively distinguish between agarwood counterfeit samples coated or injected with extracts. Although the extracts used in these counterfeit samples have the same chemical composition as genuine agarwood and can to some extent avoid the detection of traditional techniques, their aroma characteristics are significantly different from genuine agarwood due to the loss or imbalance of volatile components during the extraction process, which cannot be accurately evaluated by the above methods.

为了解决真伪问题，通常采用薄层色谱、质谱和傅里叶变换红外光谱等常规分析方法[10,11,12]。尽管这些技术能提供高灵敏度的详细化学成分信息，但它们通常耗时、操作复杂，并且往往具有破坏性——从而限制了它们在快速、便捷、无损鉴别场景中的应用。更重要的是，这些方法在技术和应用上存在明显缺陷：难以有效区分涂覆或注入萃取物的沉香假冒样品。尽管这些假冒样品中使用的萃取物与真品沉香具有相同的化学成分，能在一定程度上规避传统技术的检测，但由于萃取过程中挥发性成分的损失或失衡，其香气特征与真品沉香存在显著差异，而上述方法无法对此进行准确评估。

In recent years, electronic nose (E-nose) technology has gained widespread attention as a non-destructive detection tool that mimics the human olfactory system, owing to its advantages such as rapid response, simple operation, and high-throughput capability. In different application contexts, this technology is sometimes termed Instrumental Odour Monitoring Systems (IOMS), commonly employed for real-time and continuous monitoring of environmental odours, such as in industrial zones and waste treatment facilities [13]. In the fields of agricultural products and food, E-nose is primarily applied for aroma analysis and quality discrimination [14], [15], [16]. Particularly with the advancement of artificial intelligence, machine learning algorithms such as principal component analysis (PCA), orthogonal partial least squares-discriminant analysis (OPLS-DA), support vector machine (SVM), and random forest (RF) have been extensively applied to authenticity identification and geographical origin tracing studies based on E-nose data [17], [18]. These data-driven approaches demonstrate remarkable advantages in improving classification accuracy, model generalization capability, and high-dimensional data processing efficiency, thereby providing a novel technological pathway for the efficient, objective, and intelligent identification of Qi-Nan agarwood.

近年来，电子鼻技术作为一种模仿人类嗅觉系统的无损检测工具，因其响应快速、操作简单、高通量等优势而受到广泛关注。在不同的应用背景下，该技术有时被称为仪器气味监测系统，通常用于工业区和废物处理设施等环境气味的实时连续监测[13]。在农产品和食品领域，电子鼻主要用于香气分析和品质鉴别[14,15,16]。特别是随着人工智能的进步，主成分分析、正交偏最小二乘判别分析、支持向量机和随机森林等机器学习算法已广泛应用于基于电子鼻数据的真伪鉴别和地理产地溯源研究[17,18]。这些数据驱动的方法在提高分类准确率、模型泛化能力和高维数据处理效率方面展现出显著优势，从而为奇楠沉香的高效、客观、智能化识别提供了新的技术途径。

However, E-nose technology still faces certain limitations in the recognition of complex samples, especially in cases where aroma features are similar or mixed. To further improve recognition accuracy and model robustness, researchers have begun to attempt to fuse E-nose data with multimodal information such as images, mass spectrometry, and spectra [19], [20], [21]. Data fusion technology integrates complementary information from different types of data, which helps to comprehensively characterize sample features and compensate for the shortcomings of a single sensor system. Currently, fusion strategies mainly include feature level fusion, model level fusion, and decision level fusion [22], [23], [24]. Different levels of fusion methods have their own advantages and can be flexibly selected and combined according to specific research needs and data characteristics. It is worth emphasizing that the fast and non-destructive identification technology has not been reported in the field of identifying agarwood bracelets. A systematic application of multiple data fusion strategies will fill the research gap in this field.

然而，电子鼻技术在复杂样本的识别方面仍面临一定局限，尤其是在香气特征相似或混合的情况下。为了进一步提高识别准确率和模型稳健性，研究者开始尝试将电子鼻数据与图像、质谱、光谱等多模态信息进行融合[19,20,21]。数据融合技术整合了不同类型数据的互补信息，有助于全面表征样本特征，弥补单一传感器系统的不足。目前，融合策略主要包括特征级融合、模型级融合和决策级融合[22,23,24]。不同层级的融合方法各有优势，可根据具体研究需求和数据特征灵活选择和组合。值得强调的是，在手串鉴别领域，快速无损识别技术尚未见报道。多种数据融合策略的系统应用将填补该领域的研究空白。

In this study, an E-nose system combining a gas sensor array with machine learning algorithms was employed to distinguish genuine and counterfeit Qi-nan agarwood samples. Multiple classification algorithms, including Orthogonal Partial Least Squares Discriminant Analysis (OPLS-DA), Random Forest (RF), Logistic Regression (LR), Extreme Gradient Boosting (XGBoost), AdaBoost, and Support Vector Machine (SVM), were used to analyze the samples. To further evaluate the detection capability of the E-nose and identify volatile organic compounds (VOCs) potentially involved in sample discrimination, solid-phase microextraction coupled with headspace gas chromatography-mass spectrometry (SPME-HS-GC-MS) was conducted to provide supporting chemical evidence. In addition, to enhance the classification accuracy and robustness of the model, image-based information was introduced as a complementary modality. Discrete Wavelet Transform (DWT) was applied to extract texture features from the surface of Qi-nan agarwood samples. A decision-level data fusion strategy was then employed, in which key features were independently extracted from both the texture data and E-nose responses and subsequently integrated into a unified classification model. This strategy leverages the complementary strengths of multimodal data, offering a more precise and intelligent approach for the authentication of Qi-nan agarwood.

在本研究中，采用结合气体传感器阵列和机器学习算法的电子鼻系统来区分真伪奇楠沉香样本。使用了多种分类算法，包括正交偏最小二乘判别分析、随机森林、逻辑回归、极限梯度提升、AdaBoost和支持向量机来分析样本。为了进一步评估电子鼻的检测能力并识别可能参与样本区分的挥发性有机化合物，进行了顶空固相微萃取-气相色谱-质谱联用分析，以提供支持的化学证据。此外，为了提高模型的分类准确率和稳健性，引入了基于图像的信息作为补充模态。应用离散小波变换从奇楠沉香样本表面提取纹理特征。随后采用了一种决策级数据融合策略，其中从纹理数据和电子鼻响应中独立提取关键特征，然后整合到一个统一的分类模型中。该策略利用了多模态数据的互补优势，为奇楠沉香的真伪鉴别提供了一种更精确、更智能的方法。