Key feature in-situ recognition method for electronic nose smell recognition

Abstract

The invention discloses a key feature in-situ recognition method for electronic nose smell recognition, which comprises the following steps of: acquiring a complex smell data sample for in-situ recognition, and ensuring high precision and high resolution of smell information through a multi-level and multi-dimensional data processing algorithm-multi-dimensional layered mapping technology; the in-situ identification is more accurate and rapid; a dynamic smell information extraction self-adaptive updating mechanism is adopted, specifically, deep microscopic characteristic extraction is carried out on a smell sample, comprehensive overall characteristic construction is carried out to represent smell information, and a smell recognition model is adjusted and optimized in a self-adaptive mode; visual representation based on smell evolution information is adopted; and an intelligent in-situ identification system of the electronic nose system from an information end to a decision end is established. According to the method disclosed by the invention, the key smell characteristics can be quickly identified and positioned in the in-situ identification process by constructing an efficient smell information visualization method, so that the high efficiency and accuracy of classification are ensured.

Description

Technical Field

[0001] This invention relates to electronic nose odor recognition technology, and more specifically to a method for in-situ recognition of key features for electronic nose odor recognition. Background Technology

[0002] Currently, with the advancement of technology and the increasing demands for product quality, various industries are gradually increasing their need for odor detection and analysis. Traditional odor detection methods mostly rely on physical and chemical analysis techniques, such as mass spectrometry and gas chromatography. These methods are not only complex to operate, time-consuming, and costly, but also require destructive sampling, making it difficult to meet the needs for rapid and non-destructive detection.

[0003] Electronic nose technology, as an emerging odor detection method, has been widely used in various fields due to its advantages of high sensitivity, rapid response, and non-destructive detection. Electronic nose systems simulate the human olfactory system, using multiple sensors to collect odor information and analyzing the data through pattern recognition technology, enabling rapid and accurate detection and differentiation of different odors. However, in practical applications, electronic nose technology still faces challenges such as complex odor data, difficulties in feature extraction, and poor model interpretability. In particular, it struggles to accurately identify and characterize critical regions essential for classification performance among numerous odors. These challenges significantly limit the effectiveness and reliability of electronic nose technology in complex environments.

[0004] Deep learning technology has made significant progress in pattern recognition in recent years, and its powerful feature extraction and pattern recognition capabilities have provided new solutions for the application of electronic nose technology. However, deep learning models are often considered "black box" models, and their decision-making processes are difficult to interpret, leading to reduced user trust in the model results. Therefore, in-situ identification plays an important role in gas information classification, the core of which lies in the real-time identification and analysis of key locations and features affecting classification performance. By directly identifying key odor features during sampling and detection, in-situ identification can significantly improve the accuracy and efficiency of classification. Odor information visualization technology, as an important tool for realizing in-situ identification, has gradually received widespread attention. By presenting odor information in a visual form, users can intuitively observe the characteristics of odor data and the decision-making process of deep learning models. This visualization method not only improves the interpretability of the model but also enhances user trust in the model results. In in-situ identification scenarios, real-time detection and analysis of odor information are particularly critical. Odor information visualization can monitor changes in odor characteristics in real time, quickly identify abnormal odors, and provide timely warnings and responses, thereby ensuring the efficiency and reliability of the classification process. Summary of the Invention

[0005] The main objective of this invention is to provide an in-situ identification method for key features in electronic nose odor recognition. By combining electronic nose odor data with the decision-making process of a deep learning model, real-time monitoring and analysis of odor features are achieved, thereby improving classification performance and rapidly identifying and locating key odor features. In terms of theoretical prospects, this invention aims to discover and characterize deep key regions and features that affect classification performance, and to gain a deeper understanding of the complex structure and dynamic changes of odor information.

[0006] The technical solution adopted in this invention is: a method for in-situ recognition of key features for electronic nose odor recognition, comprising: S1. Acquire complex odor data samples for in-situ identification, and ensure high accuracy and high resolution of odor information through multi-level and multi-dimensional data processing algorithms - multi-dimensional hierarchical mapping technology, making in-situ identification more accurate and faster; S2. A dynamic odor information extraction and adaptive update mechanism is adopted, including deep microscopic feature extraction of odor samples, comprehensive overall feature construction to characterize odor information, and adaptive adjustment and optimization of the odor recognition model. S3. Employ a visual representation based on odor evolution information; S4. Establish an intelligent in-situ recognition system for the electronic nose system, from the information end to the decision-making end.

[0007] Further, in step S1, a complex odor data sample for in-situ identification is acquired, and the sample is quantitatively sealed in a sampling bottle and left to stand for a period of time to ensure that the odor is fully volatilized. A dynamic headspace sampling method is used to introduce the sample gas into the system at a constant flow rate for a period of time to acquire gas information. The data from the sensor array during the sampling period are integrated to form a data matrix. Through multi-dimensional hierarchical mapping technology, the original odor information is mapped from its initial representation to a richer and deeper expression space. This hierarchical mapping is achieved by stacking multiple layers of nonlinear transformation. Each layer can be regarded as a new abstract representation of the input features. The expressive power of the features is enriched and expanded by introducing nonlinear transformations to correct the linear unit ReLU, the Sigmoid function, the hyperbolic tangent function Tanh, Leaky ReLU, and increasing the number of feature channels. (1).

[0008] Furthermore, in step S2, the adaptive update mechanism for dynamic odor information extraction and the deep microscopic feature extraction extract odor features from the temporal dimension and the sensor chain dimension, respectively.

[0009] Furthermore, in step S2, a time-series feature extraction method is employed. This method uses multi-scale convolution to extract temporal features, designing the receptive field to be 1×1, 1×3, or 1×5. Padding terms are used to adjust and implement multi-branch kernel calculations at different kernel sizes, capturing the dynamic characteristics of odor information changing over time. The fused calculation result is then formed by summing elements together. ; (2) (3) (4) (5).

[0010] Furthermore, in step S2, the sensor chain feature extraction method is used to extract temporal features through multi-scale convolution. The receptive field is designed as 1×1, 3×1, and 5×1, and the padding term is used to adjust the multi-branch kernel calculation under different convolution kernel sizes. The interaction information of different sensor response signals is extracted, and the fused calculation result is formed by summing the elements. ; (6) (7) (8) (9).

[0011] Furthermore, in step S2, by capturing the spatial layout correlation and multi-dimensional channel dynamic change pattern of odor signals, the comprehensive and overall characteristics of odor information are effectively extracted. A lightweight, comprehensive feature extraction module is adopted, which includes two parts: spatial obfuscation and channel obfuscation. The spatial obfuscation module uses spatial attention to focus on odor information. The importance of elements in electronic nose data The dimension that characterizes the sensor. Characterize the time series in each sensor; Input data Perform max pooling and average pooling operations separately, and then concatenate the two to form a single value. As shown in the following formula: (10) right Perform convolution operations, and then... Finally, with the input data Multiply; (11) The channel obfuscation module employs a channel attention mechanism; For input data Max pooling becomes In the form of, and then proceed Finally and Multiply.

[0012] (12).

[0013] Furthermore, in step S3, a visual representation based on odor evolution information is used to intuitively display the information flow of odor information in the network. Through this odor information flow, the feature regions that the deep learning model focuses on when analyzing odor information are intuitively displayed. The specific steps for visualizing odor evolution information are as follows: Generating Odor Information Visualizations: Odor information visualizations are generated using odor evolution information. For the output category, its gradient with respect to the input odor map is calculated. These gradients reflect the importance of the output category relative to each pixel in the input odor map.

[0014] (13) in, It is a category relative to the input odor spectrum Medium pixel The odor gradient.

[0015] Global average pooling: Perform global average pooling on these gradients to obtain the weights for each channel, which represent the importance of that channel to the target class.

[0016] (14) in, It is the first The weight of each channel, It is a normalization factor, which is the time-series dimension of the odor feature map multiplied by the sensor dimension. It is the first Each channel is located in The odor gradient value.

[0017] Weighted summation of feature maps: Multiply these channel weights by the corresponding convolutional layer output feature maps, and sum the results to generate the final odor information visualization map.

[0018] (15) in, It is the first The convolutional layer outputs feature maps for each channel. Is the target category A visual representation of odor information.

[0019] Furthermore, in step S4, the specific calculation process of the intelligent in-situ identification system from the information end to the decision-making end is as follows: The electronic nose response value over 90 seconds was optimized into a 10×90×1 tensor; Temporal dimension feature extraction and sensor chain feature extraction methods are employed to capture the dynamic and multidimensional features of odor data; By spatially and through channel obfuscating odor information, a comprehensive and holistic characteristic can be constructed. Visualizing odor evolution information can intuitively show the key areas that the model focuses on when processing odor data; By utilizing fully connected layers combined with nonlinear mapping techniques and applying the SoftMax function, the feature space is effectively mapped to the label space, thereby achieving high-precision in-situ recognition.

[0020] Furthermore, in step S4, an adaptive analysis system for the electronic nose system is established from the information end to the decision end to ensure the robustness and generalization ability of the model under different samples, environments and conditions. When training and testing the model, the batch size of the training set and the test set is 32, the Adam optimizer is used, the number of iterations is 500, the learning rate is 0.001, and the weight decay coefficient is 0.0001. When identifying odors, the trained classification model is used for testing on the test set, and the test results after 10 repetitions of training are averaged.

[0021] Advantages of this invention: Enhancing the interpretability of deep learning models: By constructing a method for visualizing odor information, the interpretability of deep learning models is effectively improved, making key features and decision-making processes more transparent, and helping users better understand and trust the model's judgments.

[0022] Achieving accurate in-situ identification of odor features: Based on the odor information identification paradigm, advanced odor information visualization technology is used to monitor and analyze odor features in real time, ensuring accurate in-situ identification of odors and significantly improving the accuracy and stability of the model.

[0023] Broad application prospects: This method has broad application potential, including multiple fields such as agricultural product origin identification, food safety testing, and environmental monitoring. By rapidly identifying and locating key odor features, it ensures the efficiency and accuracy of classification, providing reliable technical support for related industries.

[0024] Promoting the development of odor recognition technology: By deeply studying and characterizing the deep key regions and features that affect classification performance, this invention not only enriches the connotation of odor recognition technology in theory, but also provides important reference for the development and application of electronic nose technology in practice.

[0025] In terms of application prospects, the method of this invention, by constructing an efficient odor information visualization method, enables the rapid identification and location of key odor features during in-situ identification, ensuring the efficiency and accuracy of classification, and has broad application potential, including multiple fields such as agricultural product origin identification, food safety testing, and environmental monitoring.

[0026] In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The invention will now be described in further detail with reference to the figures. Attached Figure Description

[0027] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0028] Figure 1 This is a flowchart of the in-situ recognition method for key features of electronic nose odor recognition according to an embodiment of the present invention; Figure 2 This is an odor response diagram shown in an embodiment of the present invention, using a certain category of a peanut dataset as an example; Figure 3 This is a radar image showing the odor information of seven origin categories of peanut dataset in this embodiment of the invention; Figure 4 This is a visualization of the odor information of peanut and rice datasets in an embodiment of the present invention; Figure 5 This is an intelligent in-situ identification and recognition system from the information end to the decision end in this embodiment of the invention; Figure 6 This is a diagram showing the odor recognition results in an embodiment of the present invention. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0030] refer to Figures 1 to 6 A method for in-situ identification of key features for electronic nose odor recognition, comprising: S1. Acquire complex odor data samples for in-situ identification, and ensure high accuracy and high resolution of odor information through multi-level and multi-dimensional data processing algorithms - multi-dimensional hierarchical mapping technology, making in-situ identification more accurate and faster; S2. A dynamic odor information extraction and adaptive update mechanism is adopted, including deep microscopic feature extraction of odor samples, comprehensive overall feature construction to characterize odor information, adaptive adjustment and optimization of odor recognition model, thereby improving the accuracy and reliability of in-situ recognition. S3. The use of visual representation based on odor evolution information not only improves the comprehensiveness and intuitiveness of odor recognition, but also provides detailed odor feature analysis and dynamic change trends, providing rich data support for in-situ recognition. S4. Establish an intelligent in-situ recognition system for the electronic nose system from the information end to the decision-making end to ensure the robustness and generalization ability of the model under different samples, environments and conditions, and to ensure the effectiveness and universality of the analysis system.

[0031] In step S1, complex odor data samples for in-situ identification are acquired, and the samples are quantitatively sealed in sampling bottles and left to stand for a period of time to ensure that the odors are fully volatilized. A dynamic headspace sampling method is used to introduce the sample gas into the system at a constant flow rate for a period of time to acquire gas information. The data from the sensor array during the sampling period are integrated to form a data matrix. Through multi-dimensional hierarchical mapping technology, the original odor information is mapped from its initial representation to a richer and deeper expression space. This hierarchical mapping is achieved by stacking multiple layers of nonlinear transformation. Each layer can be regarded as a new abstract representation of the input features. The expressive power of the features is enriched and expanded by introducing nonlinear transformations to correct the linear unit ReLU, the Sigmoid function, the hyperbolic tangent function Tanh, Leaky ReLU, and increasing the number of feature channels. (1).

[0032] In step S2, the adaptive update mechanism for dynamic odor information extraction and the deep microscopic feature extraction extract odor features in both the temporal dimension and the sensor chain dimension.

[0033] In step S2, a time-series feature extraction method is employed. This method uses multi-scale convolution to extract temporal features, designing the receptive field to 1×1, 1×3, and 1×5. Padding terms are used to adjust the kernel size and implement multi-branch kernel calculations, capturing the dynamic characteristics of odor information over time and improving sensitivity and recognition accuracy. To adaptively adjust the temporal receptive field and control the flow of multi-scale information, element-wise summation is used to form the fused calculation result. .

[0034] (2) (3) (4) (5).

[0035] In step S2, the sensor chain feature extraction method is used, similarly extracting temporal features through multi-scale convolution. The receptive field is designed to be 1×1, 3×1, and 5×1, and the padding term is used to adjust the multi-branch kernel calculation under different convolution kernel sizes, extracting the interaction information of different sensor response signals, and enhancing the model's ability to understand and express complex odor features. To adaptively adjust the sensor chain feature dimensions and receptive field and control the flow of multi-scale information, the fused calculation result is formed by element-wise summation. .

[0036] (6) (7) (8) (9).

[0037] In step S2, by capturing the spatial layout correlation and multi-dimensional channel dynamic change patterns of odor signals, the comprehensive and overall characteristics of odor information are effectively extracted, thereby realizing the calculation of key features and improving the model's ability to recognize complex odor signals and its robustness.

[0038] The construction of comprehensive and holistic characteristics can more fully represent odor information and facilitate the specific interpretation of odor information visualization. Considering that the dataset for validating in-situ recognition is a small dataset, large models or deep network structures are generally unsuitable. Therefore, a proposed... A lightweight, comprehensive feature extraction module is adopted, which includes two parts: spatial obfuscation and channel obfuscation. The spatial obfuscation module uses spatial attention to focus on odor information. The importance of elements in electronic nose data The dimension that characterizes the sensor. By characterizing the time series in each sensor, the cross-sensitivity and temporal characteristics of odor information can be fully considered through spatial location information.

[0039] Input data Perform max pooling and average pooling operations separately, and then concatenate the two to form a single value. As shown in the following formula: (10) right Perform convolution operations, and then... Finally, with the input data Multiply; (11) The channel obfuscation module employs a channel attention mechanism, aiming to comprehensively capture the complexity and integrated characteristics of deep information expression within the gas information channel dimensions. Through the rational combination of these two modules, it first focuses on the spatial dimensional characteristics of odor, then considers the relationships between channel dimensions, thereby comprehensively constructing global information. This design can consider the characteristics of odor and spectral data from multiple angles and levels.

[0040] For input data Max pooling becomes In the form of, and then proceed Finally and Multiply.

[0041] (12).

[0042] In step S3, a visual representation based on odor evolution information is used to intuitively show the flow of odor information in the network. Through this flow of odor information, the feature regions that the deep learning model focuses on when analyzing odor information are intuitively shown, thereby improving the interpretability of the model and the user's trust in the model's decision-making process.

[0043] The specific steps for visualizing odor evolution information are as follows: Generating Odor Information Visualizations: Odor information visualizations are generated using odor evolution information. For the output category, its gradient with respect to the input odor map is calculated. These gradients reflect the importance of the output category relative to each pixel in the input odor map.

[0044] (13) in, It is a category relative to the input odor spectrum Medium pixel The odor gradient.

[0045] Global average pooling: Perform global average pooling on these gradients to obtain the weights for each channel, which represent the importance of that channel to the target class.

[0046] (14) in, It is the first The weight of each channel, It is a normalization factor, which is the time-series dimension of the odor feature map multiplied by the sensor dimension. It is the first Each channel is located in The odor gradient value.

[0047] Weighted summation of feature maps: Multiply these channel weights by the corresponding convolutional layer output feature maps, and sum the results to generate the final odor information visualization map.

[0048] (15) in, It is the first The convolutional layer outputs feature maps for each channel. Is the target category A visual representation of odor information.

[0049] In step S4, the specific calculation process of the intelligent in-situ identification system from the information end to the decision-making end is as follows: The electronic nose response value over 90 seconds was optimized into a 10×90×1 tensor. In this way, we not only increased the dimensionality of the data but also provided richer input information for the convolutional neural network. Then, we used multidimensional hierarchical mapping technology to map the odor information to a high-dimensional space, which not only captured odor features more effectively but also made in-situ recognition more accurate and faster. Temporal feature extraction and sensor chain feature extraction methods are employed to capture the dynamic and multidimensional features of odor data. This strategy greatly enhances the understanding and analysis of odor information, thereby improving the accuracy and robustness of in-situ identification. By spatially and channel-wise obfuscating odor information, a comprehensive and holistic feature is constructed. This method further improves the model's ability to perceive odors in complex environments, not only optimizing the data processing flow but also ensuring the reliability and universality of the model under various conditions, further proving the reliability of in-situ identification. Visualizing odor evolution information allows for a clear view of the key regions that the model focuses on when processing odor data. This intuitive feedback mechanism enables in-situ identification of key odor features, further optimizing the model's learning and reasoning process and enhancing its interpretability and transparency. By utilizing fully connected layers combined with nonlinear mapping techniques and applying the SoftMax function, the feature space is effectively mapped to the label space, thereby achieving high-precision in-situ recognition. This process not only improves the recognition accuracy but also ensures the reliability and feasibility of the model in practical applications.

[0050] In step S4, an adaptive analysis system for the electronic nose system is established from the information end to the decision-making end to ensure the robustness and generalization ability of the model under different samples, environments and conditions. When training and testing the model, the batch size of the training set and the test set is 32, the Adam optimizer is used, the number of iterations is 500, the learning rate is 0.001, and the weight decay coefficient is 0.0001. When identifying odors, the trained classification model is used for testing on the test set, and the test results after 10 repetitions of training are averaged. Specific implementation examples: A method for in-situ identification of key features for electronic nose odor recognition includes: S1. Construction of an in-situ odor identification dataset: To demonstrate the effectiveness and universality of the proposed in-situ odor information visualization identification method, the origin identification of two agricultural products was used for verification: peanuts from seven different origins and rice from seven different farms. The experimental samples were directly obtained from Jilin Visual Agriculture Technology Development Co., Ltd., and all were newly produced agricultural products to ensure the authenticity and consistency of the samples. Peanut samples came from Tongren, Guizhou; Yantai, Shandong; Nanyang, Henan; Harbin, Heilongjiang; Jinzhou, Liaoning; Songyuan, Jilin; and Cangzhou, Hebei. Rice samples were all from the Daohuaxiang brand from seven different farms in Jilin Province. All experimental samples were stored uniformly under room temperature, dry, and shaded conditions.

[0052] During the electronic nose experiment, the ambient temperature was 25±1℃, and the process of acquiring gas information was as follows: First, the electronic nose system is powered on and preheated for 1 hour to ensure the sensor array is within its normal operating range and outputs a stable signal. Second, the sample is placed in a 500 mL sample vial and sealed for a period of time to ensure sufficient headspace air. Before each experiment, clean air filtered through activated carbon is introduced into the system at a rate of 300 mL / min for 30 seconds to purge the gas path and sensor chamber. Using a dynamic headspace sampling method, the sample gas is introduced into the system at a flow rate of 300 mL / min for 90 seconds to acquire gas information, such as... Figure 2 The image shows the odor response diagram for a specific category of peanuts. Finally, parallel samples were obtained by repeating the above steps to obtain parallel gas samples. 50 samples were obtained from each peanut origin, totaling 350 samples from 7 origins. 30 samples were obtained from each rice origin, totaling 210 samples from 7 origins. Figure 3 As shown, using the peanut dataset as a class, a radar chart displaying the aroma information of peanuts from different origins is presented.

[0053] Secondly, a multi-dimensional hierarchical mapping technique is used to process the odor information. This technique allows us to map the raw odor information from its initial representation to a richer and deeper expression space through a series of nonlinear transformations. Specifically, we can achieve this hierarchical mapping by stacking multiple layers of nonlinear transformations. Each layer can be viewed as a new abstract representation of the input features, enriching and expanding the expressive power of the features by introducing new nonlinear transformations and increasing the number of feature channels.

[0054] (1).

[0055] in, It is a weight matrix. It is a bias term. It is a non-linear activation function, which may employ the modified linear unit (ReLU), sigmoid function, hyperbolic tangent function (Tanh), Leaky ReLU, etc.

[0056] S2. Dynamic odor information extraction and adaptive update mechanism: Multidimensional hierarchical mapping maps odor information to a high-dimensional space to facilitate subsequent extraction of deep microscopic characteristics. Then, odor microscopic characteristics are extracted separately along the temporal and sensor chain dimensions.

[0057] A time-series feature extraction method is employed, utilizing multi-scale convolution to extract temporal features. The receptive field is designed to be 1×1, 1×3, and 1×5, and padding is used to adjust the kernel size to achieve multi-branch kernel computation, capturing the dynamic features of odor information changing over time and improving sensitivity and recognition accuracy. To adaptively adjust the temporal receptive field and control the flow of multi-scale information, element-wise summation is used to form the fused computational result. .

[0058] (2) (3) (4) (5).

[0059] Then, using a sensor chain feature extraction method, temporal features are extracted through multi-scale convolution. The receptive field is designed to be 1×1, 3×1, and 5×1, and the padding term is used to adjust the multi-branch kernel calculation under different convolution kernel sizes, extracting the interaction information of different sensor response signals and enhancing the model's ability to understand and express complex odor features. To adaptively adjust the sensor chain feature dimensions and receptive field and control the flow of multi-scale information, the fused calculation result is formed by element-wise summation. .

[0060] (6) (7) (8) (9).

[0061] S3 Comprehensive Overall Feature Building: Constructing comprehensive holistic features can more fully represent odor information and facilitate the visualization and interpretation of odor information. Considering that agricultural product origin identification involves small datasets, large models or deep network structures are generally unsuitable. Therefore, a lightweight comprehensive holistic feature extraction module is proposed, which includes spatial obfuscation and channel obfuscation. The spatial obfuscation module uses spatial attention to focus on odor information. The importance of elements in electronic nose data The dimension that characterizes the sensor. By characterizing the time series in each sensor, the cross-sensitivity and temporal characteristics of odor information can be fully considered through spatial location information.

[0062] Input data Perform max pooling and average pooling operations separately, and then concatenate the two to form a single value. As shown in the following formula: (10) right Perform convolution operations, and then... Finally, with the input data Multiply; (11) Secondly, the channel obfuscation module employs a channel attention mechanism, aiming to comprehensively capture the complexity and integrated characteristics of deep information expression within the gas information channel dimensions. Through the rational combination of these two modules, it first focuses on the spatial dimensional characteristics of odor, then considers the relationships between channel dimensions, thereby comprehensively constructing global information. This design can consider the characteristics of odor and spectral data from multiple angles and levels.

[0063] For input data Max pooling becomes In the form of, and then proceed Finally and Multiply.

[0064] (12).

[0065] S4 Visual representation based on odor evolution information: Visualization techniques are used to intuitively display the flow of odor information within a network, showcasing the feature regions that deep learning models focus on when analyzing odor information. This enhances the interpretability of the model and increases user trust in its decision-making process. The specific steps for visualizing odor evolution information are as follows: Generating Odor Information Visualizations: Odor information visualizations are generated using odor evolution information. For the output category, its gradient with respect to the input odor map is calculated. These gradients reflect the importance of the output category relative to each pixel in the input odor map.

[0066] (13) in, It is a category relative to the input odor spectrum Medium pixel The odor gradient.

[0067] Global average pooling: Perform global average pooling on these gradients to obtain the weights for each channel, which represent the importance of that channel to the target class.

[0068] (14) in, It is the first The weight of each channel, It is a normalization factor, which is the time-series dimension of the odor feature map multiplied by the sensor dimension. It is the first Each channel is located in The odor gradient value.

[0069] Weighted summation of feature maps: Multiply these channel weights by the corresponding convolutional layer output feature maps, and sum the results to generate the final odor information visualization map.

[0070] (15) in, It is the first The convolutional layer outputs feature maps for each channel. Is the target category A visual representation of odor information.

[0071] These steps help to understand the focus and decision-making basis of deep learning models when classifying specific categories. For example... Figure 4 As shown, an odor information visualization map is generated by odor evolution information, where (a) is a visualization map of a certain type of odor of peanuts from different origins, and (b) is a visualization map of a certain type of odor of rice from different origins.

[0072] This approach not only improves the interpretability of in-situ identification models but also enhances user trust in the model's decision-making process. Through intuitive odor information visualizations, users can clearly see the areas of focus and features the model focuses on when analyzing odor information, thus better understanding and trusting the model's classification results.

[0073] S5 Intelligent In-situ Recognition System from Information Terminal to Decision Terminal: To meet the requirements of deep learning, the number of experiments needs to be increased to ensure that the task can be correctly identified. 75% of the samples are used as the training set, and the remainder as the test set. Considering the small dataset, a deep network is not suitable; therefore, a shallow network decision framework is established. The network structure is shown in Table 1.

[0074] Table 1

[0075] The specific calculation process is as follows: (1) First, the electronic nose response value for 90 seconds is optimized into a 10×90×1 tensor. In this way, we not only improve the dimensionality of the data, but also provide richer input information for the convolutional neural network. Then, we use multidimensional hierarchical mapping technology to map the odor information to a high-dimensional space, which not only captures odor features more effectively, but also makes in-situ recognition more accurate and faster. (2) Temporal dimension feature extraction and sensor chain feature extraction methods are used to capture the dynamic and multidimensional features of odor data. This strategy greatly enhances the understanding and analysis of odor information, thereby improving the accuracy and robustness of in-situ identification; (3) By spatially and channel-wise obfuscating odor information, a comprehensive and holistic feature is constructed. This method further improves the model's ability to perceive odors in complex environments, not only optimizing the data processing flow but also ensuring the reliability and universality of the model under various conditions, further proving the reliability of in-situ identification; (4) Visualizing odor evolution information allows for an intuitive display of the key areas the model focuses on when processing odor data. This intuitive feedback mechanism enables in-situ identification of key odor features, further optimizing the model's learning and reasoning process and enhancing its interpretability and transparency; (5) Finally, by using fully connected layers combined with nonlinear mapping techniques and applying the SoftMax function, the feature space is effectively mapped to the label space, thereby achieving high-precision in-situ recognition. This process not only improves the accuracy of recognition but also ensures the reliability and feasibility of the model in practical applications.

[0076] like Figure 5As shown, this is the overall network framework in the intelligent in-situ identification and recognition system from information to the decision-making end. During the training process of the pre-trained network, the training hyperparameters of the network are designed after parameter pre-adjustment as follows: the number of iterations is 500, the batch training size is set to 32, the network learning rate is 0.001, the weight decay coefficient is 0.0001, the batch normalization method is applied after the convolution calculation to achieve a uniform distribution of parameters, the ELU activation parameter is applied after the batch normalization calculation to enhance the nonlinear mapping capability of features, and the cross-entropy loss function with L2 regularization term is applied to realize the loss calculation of the network (regularization coefficient is 0.01).

[0077] Figure 6 The curves showing the changes in accuracy and loss value on the test set in the pre-experiment are displayed respectively. It can be seen that as the number of iterations increases, the accuracy shows a continuous upward trend, while the loss value shows a steady downward trend.

[0078] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for in-situ identification of key features for odor recognition using an electronic nose, characterized in that, include: S1. Acquire complex odor data samples for in-situ identification, and ensure high accuracy and high resolution of odor information through multi-level and multi-dimensional data processing algorithms - multi-dimensional hierarchical mapping technology, making in-situ identification more accurate and faster; S2. A dynamic odor information extraction and adaptive update mechanism is adopted, including deep microscopic feature extraction of odor samples, comprehensive overall feature construction to characterize odor information, and adaptive adjustment and optimization of the odor recognition model. S3. Employ a visual representation based on odor evolution information; S4. Establish an intelligent in-situ recognition system for the electronic nose system, from the information end to the decision-making end.

2. The in-situ identification method for key features of electronic nose odor recognition according to claim 1, characterized in that, In step S1, complex odor data samples for in-situ identification are acquired, and the samples are quantitatively sealed in sampling bottles and left to stand for a period of time to ensure that the odors are fully volatilized. A dynamic headspace sampling method is used to introduce the sample gas into the system at a constant flow rate for a period of time to acquire gas information. The data from the sensor array during the sampling period are integrated to form a data matrix. Through multi-dimensional hierarchical mapping technology, the original odor information is mapped from its initial representation to a richer and deeper expression space. This hierarchical mapping is achieved by stacking multiple layers of nonlinear transformation. Each layer can be regarded as a new abstract representation of the input features. The expressive power of the features is enriched and expanded by introducing nonlinear transformations to correct the linear unit ReLU, the Sigmoid function, the hyperbolic tangent function Tanh, Leaky ReLU, and increasing the number of feature channels. （1）。 3. The in-situ identification method for key features of electronic nose odor recognition according to claim 1, characterized in that, In step S2, the adaptive update mechanism for dynamic odor information extraction and the deep microscopic feature extraction extract odor features in both the temporal dimension and the sensor chain dimension.

4. The in-situ identification method for key features of electronic nose odor recognition according to claim 1, characterized in that, In step S2, a time-series feature extraction method is used. Temporal features are extracted through multi-scale convolution. The receptive field is designed to be 1×1, 1×3, or 1×5, and padding is used to adjust the multi-branch kernel calculation under different kernel sizes. This captures the dynamic features of odor information changing over time. The fused calculation result is formed by summing elements. ; （2） （3） （4） （5）。 5. The method for in-situ identification of key features for electronic nose odor recognition according to claim 1, characterized in that, In step S2, the sensor chain feature extraction method is used, and temporal features are extracted through multi-scale convolution. The receptive field is designed to be 1×1, 3×1, and 5×1, and the padding term is used to adjust the multi-branch kernel calculation under different convolution kernel sizes. The interaction information of the response signals of different sensors is extracted, and the fused calculation result is formed by summing the elements. ; （6） （7） （8） （9）。 6. The method for in-situ identification of key features for electronic nose odor recognition according to claim 1, characterized in that, In step S2, by capturing the spatial layout correlation and multi-dimensional channel dynamic change patterns of odor signals, the comprehensive and overall characteristics of odor information are effectively extracted. A lightweight comprehensive and overall characteristic feature extraction module is adopted, which includes two parts: spatial obfuscation and channel obfuscation. The spatial obfuscation module uses spatial attention to focus on odor information. The importance of elements in electronic nose data The dimension that characterizes the sensor. Characterizing the time series in each sensor; input data Perform max pooling and average pooling operations separately, and then concatenate the two to form a single value. As shown in the following formula: (10) right Perform convolution operations, and then... Finally, with the input data Multiply; (11) The channel obfuscation module employs a channel attention mechanism; For input data Max pooling becomes In the form of, and then proceed Finally and Multiplication: (12)。 7. The method for in-situ identification of key features for electronic nose odor recognition according to claim 1, characterized in that, In step S3, a visual representation based on odor evolution information is used to intuitively display the information flow of odor information in the network. Through this odor information flow, the feature regions that the deep learning model focuses on when analyzing odor information are intuitively displayed. The specific steps for visualizing odor evolution information are as follows: Generating Odor Information Visualizations: Odor information visualizations are generated using odor evolution information. For the output category, its gradient with respect to the input odor map is calculated. These gradients reflect the importance of the output category relative to each pixel in the input odor map. (13) in, It is a category relative to the input odor spectrum medium pixel The odor gradient; Global average pooling: These gradients are subjected to global average pooling to obtain the weights for each channel. These weights represent the importance of that channel to the target class. (14) in, It is the first The weight of each channel, It is a normalization factor, which is the time-series dimension of the odor feature map multiplied by the sensor dimension. It is the first Each channel is located in odor gradient values; Weighted summation of feature maps: Multiply these channel weights by the corresponding convolutional layer output feature maps, and sum the results to generate the final odor information visualization. (15) in, It is the first The convolutional layer outputs feature maps for each channel. Is the target category A visual representation of odor information.

8. The method for in-situ identification of key features for electronic nose odor recognition according to claim 1, characterized in that, In step S4, the specific calculation process of the intelligent in-situ identification system from the information end to the decision-making end is as follows: The electronic nose response value over 90 seconds was optimized into a 10×90×1 tensor; Temporal dimension feature extraction and sensor chain feature extraction methods are employed to capture the dynamic and multidimensional features of odor data; By spatially and through channel obfuscating odor information, a comprehensive and holistic characteristic can be constructed. Visualizing odor evolution information allows for a direct demonstration of how the model processes odor data. The key areas of focus; By utilizing fully connected layers combined with nonlinear mapping techniques and applying the SoftMax function, the feature space is effectively mapped to the label space, thereby achieving high-precision in-situ recognition.

9. The method for in-situ identification of key features for electronic nose odor recognition according to claim 1, characterized in that, In step S4, an adaptive analysis system for the electronic nose system is established from the information end to the decision-making end to ensure the robustness and generalization ability of the model under different samples, environments and conditions. When training and testing the model, the batch size of the training set and the test set is 32, the Adam optimizer is used, the number of iterations is 500, the learning rate is 0.001, and the weight decay coefficient is 0.0001. When identifying odors, the trained classification model is used for testing on the test set, and the test results after 10 repetitions of training are averaged.

Images