A method for identifying the origin of American ginseng using an electronic nose based on sensor location embedding
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-10
- Publication Date
- 2026-08-11
AI Technical Summary
[0006]本发明的目的在于提供一种基于传感器位置嵌入的电子鼻西洋参产地识别方法，以解决现有电子鼻识别过程中未充分利用传感器在检测腔体中的空间分布信息、导致多传感器空间关联关系表征不足的问题，从而提高西洋参产地识别的准确性和鲁棒性
[0036]与现有技术相比，本发明具有如下有益效果：本发明通过获取传感器物理位置信息并构建位置描述向量，将空间位置先验嵌入识别模型，使各传感器不再被简单视为无结构排列的独立通道，从而增强了对多传感器空间结构信息的利用；同时，通过构建传感器关联结构和节点对几何关系特征，并将其引入图注意力建模过程，能够更充分地表征不同传感器之间的空间关联关系，提升融合特征的表达能力；此外，本发明利用几何偏置图注意力机制对邻域节点的信息传递权重进行自适应调节，使识别模型能够同时结合响应信号特征与空间几何先验信息，有利于提高西洋参等复杂挥发性样品的识别准确性和稳定性。
Smart Images
Figure CN122548231A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of intelligent sensing, electronic nose detection and pattern recognition technology, and in particular to a method for identifying the origin of American ginseng based on electronic nose embedded with sensor location. Background Technology
[0002] American ginseng is a common medicinal and edible plant raw material. Its volatile aroma information is closely related to factors such as origin, growing environment, cultivation conditions, processing methods, and storage conditions. American ginseng from different origins often has certain differences in aroma composition and dynamic release characteristics. Therefore, identifying the origin of American ginseng is of practical significance for quality evaluation, source tracing, authenticity identification, and market supervision.
[0003] Current methods for identifying the origin of American ginseng largely rely on human experience, physicochemical analysis, or chromatography-mass spectrometry. Human sensory judgment is highly subjective and has limited repeatability; while physicochemical detection and chromatography-mass spectrometry methods offer high precision, they typically suffer from long detection cycles, complex sample pretreatment, and high equipment costs, hindering rapid screening and on-site applications. Electronic nose systems, through sensor arrays that collect data on the volatile components of samples and combine this with pattern recognition algorithms for sample classification, offer advantages such as fast detection speed, low cost, and suitability for non-destructive analysis, showing promising application prospects in the identification of traditional Chinese medicine materials, agricultural products, and food.
[0004] However, existing electronic nose recognition methods typically treat multiple sensor channels as parallel inputs, focusing primarily on the temporal changes in response signals, while underutilizing the actual installation positions of the sensors within the detection chamber. In reality, the three-dimensional positions, layered arrangements, and relative spatial distances of different sensors within the detection chamber affect the local airflow distribution, diffusion conditions, and gas concentration field, resulting in sensors at different physical locations exhibiting different response characteristics. Ignoring this spatial prior information may lead to insufficient representation of the structural relationships between multiple sensors, impacting the recognition performance of samples with subtle differences.
[0005] Therefore, it is necessary to provide an electronic nose method for identifying the origin of American ginseng based on sensor location embedding. By acquiring the physical location information of the sensor and constructing a location description vector, spatial location priors are introduced into the identification model to improve the accuracy and stability of American ginseng origin identification. Summary of the Invention
[0006] The purpose of this invention is to provide an electronic nose method for identifying the origin of American ginseng based on sensor location embedding, in order to solve the problem that existing electronic nose identification processes do not fully utilize the spatial distribution information of sensors in the detection cavity, resulting in insufficient representation of the spatial correlation between multiple sensors, thereby improving the accuracy and robustness of American ginseng origin identification.
[0007] To achieve the above objectives, the present invention adopts the following technical solution:
[0008] This invention provides a method for identifying the origin of American ginseng using an electronic nose based on sensor location embedding, comprising the following steps:
[0009] S1: The electronic nose system is used to detect the American ginseng sample to be tested, and the timing response signals of multiple sensors during the detection process are collected to construct multi-sensor timing response data;
[0010] S2: Obtain the physical position information of the multiple sensors in the electronic nose detection cavity, and construct the position description vector of each sensor based on the physical position information;
[0011] S3: Construct a sensor association structure based on the relative spatial positional relationship between each sensor, and generate geometric relationship features between sensor node pairs;
[0012] S4: Based on the time-series response data, and embedding the position description vector and the geometric relationship features into the geometric bias graph attention modeling process, construct the geometric bias graph attention feature representation and obtain the fusion features;
[0013] S5: Identify the origin of the American ginseng sample to be tested based on the fusion features, and output the origin identification result.
[0014] Further, in step S2, the first The three-dimensional coordinates of the sensors in the electronic nose detection cavity are as follows:
[0015]
[0016] Using the central axis of the electronic nose detection cavity as a reference, the first The radial distance between the sensors is expressed as:
[0017]
[0018] Hierarchical identifier is denoted as Then the first The location description vector of each sensor is represented as:
[0019]
[0020] The hierarchical identifier is used to characterize the layer position, partition position, or functional area position of the corresponding sensor.
[0021] Further, in step S3, the sensor With sensors The Euclidean distance between them is expressed as:
[0022]
[0023] The connection weights between sensors are constructed based on the Euclidean distance, and the connection weights decrease as the Euclidean distance increases.
[0024] Furthermore, the connection weights are constructed using a distance kernel function, and the initial connection weights are expressed as follows:
[0025]
[0026] in, This is the distance to the kernel width parameter.
[0027] Furthermore, the initial association weights are corrected based on sensor hierarchy consistency to obtain the connection weights in the sensor association structure. The corrected connection weights are expressed as follows:
[0028]
[0029] in, This is the hierarchical enhancement coefficient. This is an indicator function.
[0030] Furthermore, in step S3, the sensor node pairs The geometric relationship characteristics may include one or more of the following: coordinate difference , , Euclidean distance ,sensor and sensors radial distance , and the corresponding hierarchical identifiers , .
[0031] Further, in step S4, for the first Time of the first Response characteristics of individual sensors Perform a linear mapping to obtain the node representation. The unnormalized attention score between node pairs is calculated based on node representation and geometric relationship features, and the unnormalized attention score is expressed as:
[0032]
[0033] in, For trainable attention parameter vectors, For node pairs Geometric relationship characteristics, This is a geometric relation mapping function used to map the geometric relation features into geometric bias terms.
[0034] Furthermore, the unnormalized attention score is normalized to obtain the node... The information transmission weights of its neighboring nodes are determined, and the neighboring nodes are then evaluated based on these information transmission weights. The nodes are aggregated to obtain the nodes. The fusion representation. Normalization can be achieved using the softmax function at the nodes. It is carried out within the neighborhood.
[0035] Furthermore, after step S1, the timing response signal can be preprocessed, including one or more of the following: denoising, smoothing, baseline correction, drift compensation, standardization, normalization, and first-order difference.
[0036] Compared with existing technologies, the present invention has the following advantages: By acquiring the physical location information of sensors and constructing location description vectors, the present invention embeds spatial location priors into the recognition model, so that each sensor is no longer simply regarded as an independent channel without structure, thereby enhancing the utilization of the spatial structure information of multiple sensors; at the same time, by constructing sensor association structure and node pair geometric relationship features and introducing them into the graph attention modeling process, the spatial association relationship between different sensors can be more fully represented, improving the expressive power of fused features; in addition, the present invention uses a geometric bias graph attention mechanism to adaptively adjust the information transmission weight of neighboring nodes, so that the recognition model can combine response signal features and spatial geometric prior information at the same time, which is beneficial to improving the recognition accuracy and stability of complex volatile samples such as ginseng. Attached Figure Description
[0037] Figure 1 This is a flowchart of an electronic nose method for identifying the origin of American ginseng based on sensor location embedding, according to the present invention.
[0038] Figure 2 This is a schematic diagram illustrating the construction of sensor physical location information and location description vector in this invention.
[0039] Figure 3 This is a schematic diagram of the geometric bias graph attention feature modeling process in this invention.
[0040] Figure 4 This is a schematic diagram of the confusion matrix of the one-fold test result, which is closest to the average accuracy of the five-fold cross-validation result in an embodiment of the present invention. Detailed Implementation
[0041] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.
[0042] Example 1
[0043] like Figures 1 to 3 As shown, this embodiment provides a method for identifying the origin of American ginseng using an electronic nose based on sensor location embedding, including the following steps:
[0044] S1: Acquire timing response signals from multiple sensors
[0045] An electronic nose system was used to test the American ginseng sample, and the timing response signals of multiple sensors during the testing process were collected to construct multi-sensor timing response data.
[0046] In this embodiment, the electronic nose system includes a detection chamber and multiple gas sensors arranged within the detection chamber. After the ginseng sample to be tested releases volatile components, the multiple sensors generate corresponding dynamic responses within a predetermined detection time. The response values of each sensor at each moment are arranged in chronological order to form multi-sensor time-series response data. Let the detection time sequence length be... The number of sensors is Then, the time-series response data of a single sample can be expressed as:
[0047]
[0048] When performing batch sample identification, the time-series response data of multiple samples can be represented as:
[0049]
[0050] in, Indicates the sample batch size. Indicates the number of time steps. Indicates the number of sensors.
[0051] In one embodiment, the timing response signal may be preprocessed after step S1. The preprocessing includes one or more of the following: denoising, smoothing, baseline correction, drift compensation, standardization, normalization, and first-order difference, in order to improve the stability and anti-interference ability in the subsequent recognition process.
[0052] S2: Obtain sensor physical location information and construct a location description vector
[0053] like Figure 2 As shown, the physical position information of multiple sensors in the electronic nose detection cavity is obtained, and the position description vector of each sensor is constructed based on the physical position information.
[0054] In this embodiment, the six sensors are arranged in two layers along the axial direction of the air chamber. The lower layer sensors G1-G3 have an axial height of 35 mm relative to the reference bottom surface of the detection chamber, while the upper layer sensors G4-G6 have an axial height of 65 mm relative to the reference bottom surface of the detection chamber. Each layer of sensors is distributed on a circle with a radius of 10 mm centered on the central axis of the air chamber. To facilitate position modeling, the three lower and three upper sensors can be considered as being located in the same circular plane, and a three-dimensional coordinate system is established based on their uniform distribution along the circumference. The origin of the coordinate system is the center of the reference bottom surface of the detection chamber, and the coordinate system is based on the central axis of the detection chamber. Establish a three-dimensional rectangular coordinate system along the axes, where, shaft and The axis is located in a plane perpendicular to the central axis of the detection cavity, and the coordinate unit is mm.
[0055] In one example, the spatial coordinates of the six sensors can be set as follows:
[0056] G1:(10.00,0.00,35.00)
[0057] G2:(-5.00,8.66,35.00)
[0058] G3:(-5.00,-8.66,35.00)
[0059] G4:(10.00,0.00,65.00)
[0060] G5:(-5.00,8.66,65.00)
[0061] G6:(-5.00,-8.66,65.00)
[0062] The hierarchy identifier of the lower-level sensor can be defined as follows: The hierarchical identifier of the upper-level sensor can be defined as follows: .
[0063] Let the first The three-dimensional coordinates of the sensors within the electronic nose detection cavity are:
[0064]
[0065] Using the central axis of the electronic nose detection cavity as a reference, the first The radial distance between the sensors is expressed as:
[0066]
[0067] Therefore, the first The location description vector of each sensor is represented as:
[0068]
[0069] The position description vectors of all sensors are combined to form a position description matrix. The position description vectors are used to characterize the spatial position attributes of each sensor in the detection cavity and provide prior position information for subsequent spatial relationship modeling.
[0070] The sensor array used in this embodiment is composed as shown in Table 1:
[0071] Table 1. Sensor Array Composition of the Electronic Nose System
[0072] G1 GM-402B CH4 G2 SMD1001 HCHO G3 GM-502B VOC G4 GM-702B CO G5 GM-602B H2S G6 GM-302B EtOH
[0073] S3: Constructing the sensor association structure and node pair geometric relationship features
[0074] The sensor association structure is constructed based on the relative spatial positional relationship between each sensor, and the geometric relationship features between sensor node pairs are generated.
[0075] Set sensor With sensors The three-dimensional spatial coordinates are respectively and The Euclidean distance between the two can be expressed as:
[0076]
[0077] An initial association weight is constructed based on the Euclidean distance, wherein the initial association weight decreases as the Euclidean distance increases. In this embodiment, the initial association weight is constructed using a distance kernel function, expressed as:
[0078]
[0079] in, This is the distance to the kernel width parameter.
[0080] Furthermore, the initial association weights are corrected based on sensor hierarchy consistency to obtain the connection weights in the sensor association structure. The corrected connection weights are expressed as follows:
[0081]
[0082] in, This is the hierarchical enhancement coefficient. This is an indicator function. When the sensor and sensors When they are at the same level, the corresponding connection weights are enhanced.
[0083] After obtaining the connection weights between each node, several nearest neighbors of each sensor node can be retained according to a preset nearest neighbor selection rule to obtain a sparse sensor association structure. The nearest neighbor selection rule can be to retain the nearest neighbors of each node with the largest weights. The method can be one adjacent node, or filtered according to a distance threshold; however, this invention does not limit the method.
[0084] In addition, sensor node pairs need to be constructed. The geometric relationship characteristics. These geometric relationship characteristics include one or more of the following: coordinate difference. , , Euclidean distance ,sensor and sensors their respective radial distances , and the corresponding hierarchical identifiers , In this embodiment, the geometric relationship features of node pairs can be represented as:
[0085]
[0086] The geometric relationship features are used to characterize the relative spatial relationship between node pairs, providing input for geometric bias graph attention feature modeling.
[0087] S4: Construct a geometric bias graph attention feature representation.
[0088] like Figure 3 As shown, based on the time-series response data, the location description vector, and the geometric relationship features, a geometric bias graph attention feature representation is constructed to obtain fusion features characterizing the spatial correlation of multiple sensors.
[0089] In this embodiment, for the first Time of the first Response characteristics of individual sensors Perform a linear transformation to obtain the node representation. Correspondingly, for the first Time of the first Response characteristics of individual sensors Perform a linear transformation to obtain the node representation. This process can be represented as:
[0090]
[0091]
[0092] in, These are trainable linear mapping parameters, learned from sample data during model training, used to map sensor response features to the target feature space. Each node represents the characteristic state of the corresponding sensor at the current moment and serves as input for subsequent node-to-node association modeling, attention weight calculation, and neighborhood information aggregation.
[0093] Furthermore, the unnormalized attention score between node pairs is calculated based on both node representation and geometric relationship features. The unnormalized attention score is expressed as:
[0094]
[0095] in, This is a trainable attention parameter vector used to calculate the feature correlation term between node pairs based on the node representation. For node pairs Geometric relationship characteristics, This is a geometric relationship mapping function used to map the geometric relationship features into geometric bias terms between node pairs. The geometric bias terms reflect the influence of the spatial relative positional relationship of node pairs on the attention score. In one embodiment, the... This can be implemented using linear mapping, nonlinear mapping, feedforward neural networks, or multilayer perceptrons. The first term in the above expression characterizes the correlation of node features, while the second term characterizes the influence of geometric bias, thus allowing the attention score between node pairs to be influenced by both response features and spatial geometric relationships.
[0096] The unnormalized attention score is normalized to obtain the first... Time Node Its neighboring nodes Information transmission weight The aforementioned Represents a node During the neighborhood feature aggregation process, nodes The level of attention. The normalization process uses the softmax function at the nodes. This is performed within the neighborhood of the target area, and is represented as:
[0097]
[0098] in, For nodes The set of neighboring nodes.
[0099] After obtaining the information transmission weight, the nodes Each neighboring node Node representation Perform weighted aggregation to obtain nodes The fusion representation:
[0100]
[0101] After performing the same weighted aggregation process on all nodes, the results of all sensor nodes at the [number]th node can be obtained. The fusion representation of time points is further used to form a fusion feature matrix:
[0102]
[0103] in, This represents the node feature dimension. The fused features are used to characterize the spatial correlation and temporal response information among multiple sensors.
[0104] In one implementation, before constructing the geometric bias graph attention feature representation, a step of fusing the location description vector with sensor response features is included. The fusion method includes one or more of splicing fusion, mapping fusion, embedding fusion, or gated fusion to further enhance the spatial location prior information contained in the initial node representation.
[0105] S5. Output the results of American ginseng origin identification.
[0106] The fused features obtained in step S4 are input into the classifier or discriminant module in the trained recognition model to determine the origin category of the American ginseng sample and output the American ginseng origin recognition result.
[0107] In this embodiment, the identification model is first trained using training samples with origin labels. The identification model includes a geometric bias graph attention feature modeling module and a classifier to determine the linear mapping parameters, attention parameters, parameters in the geometric relationship mapping function, and classifier parameters. In the identification stage, the fused features obtained in step S4 are input into the classifier or discrimination module in the trained identification model to output the origin category of the ginseng sample to be tested.
[0108] Example 2
[0109] This embodiment illustrates the specific application of the method of the present invention in the identification of the origin of American ginseng.
[0110] In this embodiment, American ginseng samples from five different origins were selected as the objects to be identified. These five origins include Changbai Mountain region in Jilin Province, China; Wendeng District in Shandong Province, China; Liuba County in Shaanxi Province, China; Ontario, Canada; and Wisconsin, USA. An electronic nose system was used to detect the American ginseng samples from these different origins, collecting multi-sensor time-series response signals, and modeling and identification were performed according to the method described in Embodiment 1.
[0111] In this embodiment, 120 samples were collected from each production area, resulting in a total of 600 samples from the five production areas, corresponding to the five categories. A five-fold cross-validation method was used for model training and performance evaluation of the dataset. This involved dividing all samples into five non-overlapping subsets, selecting four subsets as the training set each time, and using the remaining subset as the test set. This process was repeated five times, ensuring that each subset served as a test set once. Stratified sampling was used for each fold division to ensure that the proportion of samples from different production areas was approximately consistent across the subsets.
[0112] Table 2 shows the performance comparison of different models on the task of identifying the origin of American ginseng. The results in Table 2 are expressed as mean ± standard deviation. The results show that the method of the present invention achieves the best results in terms of accuracy, recall, and precision, reaching 94.44 ± 0.56, 94.44 ± 0.56, and 94.54 ± 0.52, respectively, which is better than the comparative models such as SVM, Random Forest, 1D-CNN, LSTM, ResNet18, and MobileNetV3. This indicates that the present invention, by explicitly introducing prior information on the physical location of the sensors and combining it with geometric bias graph attention modeling, can enhance the utilization of multi-sensor spatial structure information, thereby improving the performance of identifying the origin of American ginseng.
[0113] Table 2 Performance comparison of different models on the task of identifying the origin of American ginseng.
[0114] SVM 85.78 ± 2.10 85.78 ± 2.10 86.19 ± 2.12 Random Forest 89.78 ± 1.69 89.78 ± 1.69 90.29 ± 1.39 1D-CNN 88.00 ± 3.57 88.00 ± 3.57 88.65 ± 3.63 LSTM 86.00 ± 2.30 86.00 ± 2.30 87.44 ± 2.36 ResNet18 87.33 ± 4.72 87.33 ± 4.72 88.33 ± 4.04 MobileNetV3 89.67 ± 1.45 89.67 ± 1.45 89.90 ± 1.53 Method of the present invention 94.44 ± 0.56 94.44 ± 0.56 94.54 ± 0.52
[0115] The recognition results of the one-fold test set in five-fold cross-validation are as follows: Figure 4 As shown, the test set contains 120 samples. The confusion matrix indicates that the American ginseng samples from various origins in this test set generally have good distinguishing effects, demonstrating that the method of this invention can effectively utilize prior information about the physical location of the sensors and geometric relationship information between node pairs to improve the expressive ability of multi-sensor spatial structural features.
[0116] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. All equivalent structural changes, equivalent step substitutions, or other modifications and improvements that can be conceived by those skilled in the art without creative effort based on the description and drawings of the present invention should fall within the scope of protection of the present invention.
Claims
1. A method for electronic nose Panax quinquefolius origin identification based on sensor position embedding, characterized in that, Includes the following steps: S1: The electronic nose system is used to detect the American ginseng sample to be tested, and the timing response signals of multiple sensors during the detection process are collected to construct multi-sensor timing response data; S2: Obtain the physical position information of the multiple sensors in the electronic nose detection cavity, and construct the position description vector of each sensor based on the physical position information; S3: Construct a sensor association structure based on the relative spatial positional relationship between each sensor, and generate geometric relationship features between sensor node pairs; S4: Based on the time-series response data, and embedding the position description vector and the geometric relationship features into the geometric bias graph attention modeling process, construct the geometric bias graph attention feature representation and obtain the fusion features; S5. Identify the origin of the American ginseng sample to be tested based on the fusion features, and output the origin identification result.
2. The electronic nose ginseng origin identification method based on sensor position embedding according to claim 1, characterized in that, In step S2, the first The three-dimensional coordinates of the sensors in the electronic nose detection cavity are as follows: With the central axis of the electronic nose detection cavity as the reference, the first The radial distance between the sensors is expressed as: The hierarchy identifier is denoted as Then the first The position description vector of each sensor is represented as:
3. The hierarchical identifier is used to characterize the layer position, partition position, or functional area position of the corresponding sensor.
4. The electronic nose ginseng origin identification method based on sensor position embedding according to claim 2, characterized in that, The multiple sensors are arranged in a layered manner in the electronic nose detection cavity, including upper-layer sensors and lower-layer sensors. The layer identifier is used to distinguish between the upper-layer sensors and the lower-layer sensors.
5. The electronic nose ginseng origin identification method based on sensor position embedding according to claim 2, characterized in that, In step S3, the sensor With sensors The Euclidean distance between them is expressed as: An initial association weight is constructed based on the Euclidean distance, wherein the initial association weight decreases as the Euclidean distance increases.
6. The electronic nose American ginseng origin identification method based on sensor position embedding according to claim 4, characterized in that, The initial association weights are constructed using a distance kernel function, and are expressed as follows:
7. Among them, distance kernel width parameter.
8. The electronic nose ginseng origin identification method based on sensor position embedding according to claim 5, characterized in that, The corrected connection weight is expressed as:
9. Among them, This is the hierarchical enhancement coefficient. The indicator function is used; and, according to the preset nearest neighbor filtering rules, several nearest neighbor nodes of each sensor node are retained to obtain a sparse sensor association structure.
10. The electronic nose ginseng origin identification method based on sensor position embedding according to claim 2, characterized in that, In step S3, the sensor node pair The geometric relationship characteristics include one or more of the following: coordinate difference , , Euclidean distance ,sensor and sensors their respective radial distances , and the corresponding hierarchical identifiers , .
11. The method for identifying the origin of American ginseng based on an electronic nose using sensor location embedding according to claim 7, characterized in that, In step S4, for the first Time of the first Response characteristics of individual sensors Perform a linear mapping to obtain the node representation. The node representation is used to characterize the feature state of the corresponding sensor node at the current moment and serves as input for subsequent attention score calculation and neighborhood feature aggregation. The unnormalized attention score between node pairs is calculated based on the node representation and geometric relationship features, and the unnormalized attention score is expressed as:
12. Among them, For trainable attention parameter vectors, For node pairs Geometric relationship characteristics, This is a geometric relation mapping function used to map the geometric relation features into geometric bias terms.
13. The electronic nose ginseng origin identification method based on sensor position embedding according to claim 8, characterized in that, The unnormalized attention score is normalized to obtain the node. Its neighboring nodes Information transmission weight The Represents a node During the process of neighborhood information aggregation, nodes The degree of attention; and based on the information transmission weight, the neighboring nodes. The nodes are represented by weighted aggregation to obtain the nodes. The fusion representation; the normalization process uses the softmax function at the nodes. It is carried out within the neighborhood.
14. The electronic nose ginseng origin identification method based on sensor position embedding according to any one of claims 1 to 9, characterized in that, Before identifying the origin of the ginseng sample to be tested, the geometric bias map attention feature modeling module and classifier are trained based on training samples with origin labels to determine the linear mapping parameters, attention parameters, geometric relationship mapping parameters, and classifier parameters.
15. The method for identifying the origin of American ginseng based on an electronic nose using sensor location embedding according to claim 1, characterized in that, The process after step S1 includes a preprocessing step for the timing response signal, which includes one or more of the following: denoising, smoothing, baseline correction, drift compensation, standardization, normalization, and first-order difference.
16. The method for identifying the origin of American ginseng based on an electronic nose using sensor location embedding according to claim 1, characterized in that, In step S4, before constructing the geometric bias map attention feature representation, the step of embedding the position description vector into the sensor response feature representation is included. The embedding method includes one or more of splicing fusion, mapping fusion, embedding fusion, or gated fusion.
17. The electronic nose ginseng origin identification method based on sensor position embedding according to claim 1, characterized in that, In step S5, the origin identification result is used to indicate the preset origin category to which the American ginseng sample to be tested belongs.