Electronic nose system for honey authenticity identification based on deep learning

Abstract

The invention relates to an electronic nose system for honey authenticity identification based on deep learning. The electronic nose system comprises a smell collection device, a sensor array module, a data processing module, an artificial neural network classification module and an SHAP interpretability analysis module. According to the system, the characteristics of volatile organic compounds (VOCs) of honey are collected through a multi-channel smell sensor, and the characteristics are preprocessed and then input into an artificial neural network model to realize true and false identification. And introducing an SHAP method to carry out global and local visual interpretation on model output, and realizing self-calibration and anomaly detection based on the contribution degree of the sensor. The system can be integrated and deployed on a honey filling production line, online monitoring and linkage control are achieved, and the system is suitable for various honey varieties and application scenes. Compared with an existing method, the method has the advantages of being high in detection accuracy, high in interpretability, flexible in deployment and high in automation degree.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of food authenticity detection, and particularly relates to an electronic nose system for identifying true and false honey based on deep learning. BACKGROUND

[0002] Honey is a natural nutrient product, which is favored by consumers due to its rich nutritional ingredients and health care effects, especially honey with geographical indications and specific nectar sources, such as Xinjiang Nileke black bee honey, which has a high market value. However, in order to maximize profits, honey adulteration behaviors in the market are rampant, which has become the third largest easily adulterated food in the world. Such adulteration not only affects the quality of honey and the health of consumers, but also seriously damages the brand credibility of honey and the economic value of regional characteristic products. Therefore, developing a scientific, accurate and generalizable method for identifying true and false honey has become the focus of the industry.

[0003] Existing honey true and false identification technologies mainly include sensory evaluation, physicochemical detection and instrument analysis. Among them, instrument analysis methods such as gas chromatography-mass spectrometry (GC-MS), high performance liquid chromatography-mass spectrometry (HPLC-MS / MS), nuclear magnetic resonance (NMR) and isotope ratio mass spectrometry have the advantages of high detection sensitivity and strong accuracy, but also have the disadvantages of complex operation, long detection period, high equipment cost and dependence on professional technical personnel, which are difficult to meet the needs of the honey industry for rapid, efficient and on-site deployment of true and false detection.

[0004] In recent years, the electronic nose (e-nose) system as an intelligent detection device simulating human olfaction has been widely used in food authenticity detection. It collects volatile organic compound information of sample gas through a sensor array and realizes pattern recognition and classification judgment combined with machine learning algorithms. Although the electronic nose has achieved good results in some food fields, due to the complexity and strong difference of honey odor components, the electronic nose system relying only on traditional pattern recognition algorithms still has certain limitations in processing multi-source information and achieving high-precision identification, especially in identifying true and false Xinjiang Nileke black bee multi-flower honey, the existing technology is difficult to achieve accurate identification.

[0005] In addition, some current research has tried to combine electronic nose with machine learning methods to improve recognition performance, but traditional machine learning models (such as support vector machine, random forest, etc.) have the characteristics of "black box", which makes it difficult to explain the decision basis of the model, reducing its credibility and acceptability in the food safety supervision scenario.

[0006] Therefore, there is an urgent need for a honey authenticity identification technology that combines electronic nose technology with an interpretable artificial intelligence method, which can maintain detection accuracy while improving model transparency and reliability, meeting the technical requirements of "high accuracy, low cost, interpretability, and deployability" in practical applications, especially for rapid screening and authenticity verification of honey categories with complex flavor characteristics such as Xinjiang Nileke black bee honey. SUMMARY

[0007] To solve the technical problems of complex operation, high detection cost, and uninterpretable model in existing honey authenticity detection methods, the present application proposes an electronic nose system and method for honey authenticity identification based on deep learning. The system, by constructing an intelligent architecture composed of a multi-channel gas sensor array, a preprocessing algorithm, an artificial neural network classifier, and a SHAP interpretability analysis module, can achieve rapid, accurate, and transparent discrimination of honey authenticity. The technical solution of the present application not only improves the classification performance of the model, but also enhances the visual understanding of the model decision logic, with good practicality and promotional value.

[0008] In one embodiment of the present application, an electronic nose system for honey authenticity identification based on deep learning is provided, comprising: a gas sensor array, a signal preprocessing module, an artificial neural network (ANN) classification module, an interpretability analysis module, and a hardware control unit. The gas sensor array is used to collect volatile organic compound (VOCs) signals of honey samples, composed of 30 different types of sensors, and 18 key sensors are selected by Monte Carlo simulation method for subsequent data analysis and classification modeling.

[0009] Further, the signal preprocessing module is used to standardize the collected sensor raw signal data, specifically including: identifying outliers using the interquartile range (IQR) method, replacing outliers by median interpolation, and normalizing the data to the [-1, 1] interval to improve the stability and robustness of subsequent model training.

[0010] Preferably, the artificial neural network classification module is constructed as a structure with two hidden layers, each containing 11 neurons, the learning rate of model training is 0.001, the maximum number of iterations is set to 1500, and the output sample is the probability value of true honey or adulterated honey.

[0011] In one embodiment of the present application, the interpretability analysis module is implemented based on the SHAP (Shapley Additive Explanations) algorithm, which provides global and local visual explanation results of the model, including feature contribution heat maps, 3D partial dependence plots, and local dependence relationship graphs, to reveal the influence of key sensor features in the model decision-making process.

[0012] Further, the hardware control unit is used to cooperate with the electronic nose device, and specifically controls the flow rate of gas sampling (400 mL / min), detection time (100 seconds) and sampling time (30 seconds), to ensure the consistency and stability of sensor signal acquisition.

[0013] In an embodiment of the present application, a method for identifying true and false honey is provided, comprising the following steps: S1, collecting 30-channel VOCs gas sensor signals of a honey sample; S2, screening out 18 key sensor features by using a Monte Carlo simulation method; S3, performing outlier detection and median interpolation on the sensor data, and standardizing to [-1, 1]; S4, inputting the preprocessed data into the ANN classification model, and outputting true and false probabilities; S5, analyzing the contribution degrees of two key sensors CH13 and CH28 by using a SHAP method; and S6, setting a discrimination threshold, and determining that the honey is true when the output probability is greater than or equal to 0.7, and determining that the honey is false otherwise.

[0014] Preferably, the SHAP contribution degree result shows that the average absolute contribution degrees of sensors CH13 and CH28 to the model output are not less than 42%, which are the most significant features in the model.

[0015] In an embodiment of the present application, the ANN model divides a sample set into a training set and a test set by using a KS algorithm, the ratio is 7:3, and the model is evaluated by using a ten-fold cross-validation method, and the classification performance of the test set reaches AUC≥0.85 and F1 score≥0.81.

[0016] Alternatively, to adapt to the characteristics of different honey varieties, the classification probability threshold can be dynamically adjusted according to the honey variety: the threshold is set to 0.65 for multi-flower honey, and the threshold is set to 0.75 for single-flower honey.

[0017] Further, the system can be integrated and deployed on a honey filling production line, when the classification probabilities of two consecutive samples are less than 0.6, the system automatically triggers an alarm device, and the filling process is paused, to realize automatic quality control.

[0018] Preferably, to reduce the hardware cost, the gas sensor array can be selected to be a combination structure containing only five sensors CH13, CH28, CH6, CH10 and CH22, and in this configuration, the system recognition accuracy can still reach more than 70%.

[0019] In an embodiment of the present application, when the SHAP analysis shows that the contribution degrees of sensors CH13 or CH28 are abnormal (less than 5% or higher than 60%), the system can automatically trigger a sensor calibration program, to improve the stability and long-term reliability of the device.

[0020] Alternatively, the training samples of the ANN model are derived from Xinjiang Nileke black bee honey, including 53 real honey samples and 55 fake honey samples, to ensure the representativeness and scientificity of the model training data.

[0021] Based on the above technical scheme, the honey true and false identification electronic nose system based on the interpretable artificial intelligence can realize efficient, accurate and interpretable identification of the authenticity of the honey sample by constructing a multi-level collaborative architecture integrating a gas sensor array, a signal preprocessing module, an artificial neural network classification module, an interpretability analysis module and a hardware control unit.

[0022] Specifically, the Monte Carlo simulation method is applied to the electronic nose sensor feature selection for the first time, effectively screening out 18 key gas sensors that most contribute to the performance of the model, thereby reducing the equipment cost and the data dimension while improving the generalization ability and stability of the model. The IQR method is used in the signal preprocessing module to identify outliers and perform median interpolation, effectively enhancing the robustness of the model and avoiding performance fluctuations caused by extreme values. In addition, the normalization strategy of standardizing to the [-1, 1] interval makes the input data stable in neural network training, improving the model convergence efficiency.

[0023] The artificial neural network (ANN) classifier in the application has a double-hidden layer structure, which realizes high-accuracy classification and judgment of the authenticity of honey by adjusting the learning rate and the maximum number of iterations. The SHAP method is used to perform global and local visual analysis of the decision-making process of the ANN model, which not only improves the transparency of the prediction results, but also significantly enhances the acceptability and credibility of the model in the food authenticity detection scene. Through multi-dimensional analysis methods such as feature contribution heat map, local dependence graph and 3D interactive graph, the influence mechanism of key sensors such as CH13 and CH28 on the classification results can be intuitively revealed, meeting the compliance and auditability requirements of the industry for "interpretable AI".

[0024] In addition, the application also provides a low-cost sensor combination scheme, which can simplify the sensor array to five channels of CH13, CH28, CH6, CH10 and CH22, reduce the hardware deployment cost under the premise of ensuring an accuracy rate of not less than 70%, and is suitable for resource-limited or on-site rapid screening scenarios. The system can also be deployed on a honey filling production line, realizes automatic alarm and shutdown control by monitoring the sample classification probability in real time and combining a dynamic threshold judgment mechanism, and thus builds a closed-loop quality supervision system.

[0025] In conclusion, the application provides an intelligent honey authenticity detection system with high accuracy, strong interpretability and good practicability, effectively solves the technical problems of high cost, low efficiency and uninterpretability of existing instrumental analysis technology, and insufficient recognition accuracy and transparency of traditional electronic nose, and has a wide application prospect, and is especially suitable for quality control and brand protection of regional characteristic foods such as Xinjiang Nileke black bee honey. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 It is a whole structure schematic diagram of the electronic nose system for identifying true and false honey based on deep learning, which shows the main component modules of the system and the functional connection relationship therebetween.

[0027] Figure 2 It is a structure schematic diagram of the artificial neural network (ANN) classification module, which shows the neuron structure and connection path of the input layer, the hidden layer and the output layer.

[0028] Figure 3 It is a cross-validation score trend chart when the Monte Carlo simulation method is used to evaluate different number of feature channel combinations in the sensor feature optimization stage, which is used to determine the optimal sensor number.

[0029] Figure 4 It is a schematic diagram of global interpretability analysis of the model.

[0030] The figure includes: (A) SHAP feature contribution column chart and (B) SHAP summary bee colony chart, which are used to show the importance ranking of each gas sensor channel to the output result of the artificial neural network model. The results show that the contribution of CH13 and CH28 sensors in the model decision is the highest, and the overall contribution is not less than 42%, which provides the most significant discrimination information basis for the model.

[0031] Figure 5 It is a SHAP heat map, which shows the SHAP contribution distribution of all samples on all sensor channels, and is used to reveal the consistency and difference of different channels in the whole data set.

[0032] Figure 6 It is a schematic diagram of local interpretability analysis of the model.

[0033] The figure includes: (A) SHAP waterfall chart and (B) SHAP gravity chart, which are used to show the item-by-item contribution of each odor sensor to the classification result in a specific honey sample. Among them, Figure 6 A presents the feature contribution path of sample 1 in the form of waterfall chart, which shows the positive and negative effects of CH6, CH28 and other channels on the judgment of “adulterated honey”; Figure 6B is the force plot of sample 1 and sample 9, which reveals how SHAP values push the model output towards "true" or "false" direction, and helps to understand the visualization process of model decision-making basis.

[0034] Figure 7 The SHAP dependence plot (single variable dependence plot) shows the trend of the influence of a single sensor channel response value (such as CH13, CH28) on the SHAP value, reflecting its nonlinear dependence on the model output.

[0035] Figure 8 The SHAP interaction dependence plot (two-channel interaction plot) shows how the interaction response between two sensor channels (such as CH28 and CH13) jointly affects the model prediction result.

[0036] Figure 9 The three-dimensional partial dependence plot (3DPDP) shows the influence surface of two input features (such as CH13 and CH28) on the prediction result under different numerical combinations, which is used to reveal the high-dimensional decision boundary of the model.

[0037] Figure 10 The SHAP explainability analysis process diagram shows the closed-loop processing flow of the module from model prediction to anomaly detection in the system, which is the key path diagram of realizing model transparency and self-calibration logic.

[0038] Figure 11 The deployment structure diagram of the honey production line integrated detection system shows the integration mode of the system with the filling production line, which supports online quality control and automatic response. DETAILED DESCRIPTION

[0039] In order to more clearly understand the technical solutions of the present application and its beneficial effects, the specific embodiments of the present application will be described in detail below with reference to the accompanying drawings. The electronic nose system for identifying true and false honey based on deep learning provided by the present application is developed around the key links of honey sample odor collection, feature extraction, intelligent identification, result interpretation and application deployment, and the system structure is composed of multiple functional modules. The data-driven and feedback-controlled closed-loop architecture is formed between the modules. For the sake of understanding, this part will introduce the system overall structure, sensor array and feature extraction module, artificial neural network classification module, explainability analysis module, classification decision logic and dynamic threshold, self-calibration mechanism and production line deployment mode and other contents in turn.

[0040] It should be understood that the embodiments of the present application are not limited to the following specific description, and the transformations or substitutions made by those skilled in the art without departing from the core idea of the present application should also be covered within the protection scope of the present application.

[0041] I. Description of the overall structure of the system In one possible implementation, as shown in Figure 1 The present application provides an electronic nose system for identifying true and false honey based on deep learning, which has the overall structure including a gas sensor array, a signal preprocessing module, an artificial neural network (ANN) classification module, an explainability analysis module, and a hardware control unit. Each module works together to realize the extraction of volatile organic compounds (VOCs) characteristics of honey samples, data preprocessing, true and false classification judgment, and result visualization explanation, and can be integrated into a honey production line for online quality monitoring.

[0042] Specifically: The gas sensor array is used to collect the gas volatile characteristic signals of the honey sample. The array is composed of 30 different types of gas sensors, covering multiple common VOCs response channels such as oxygen, ethanol, aldehyde ketone, ester, alcohol, etc. Among them, 18 key sensors are selected for subsequent analysis by the Monte Carlo simulation method, and the preferred sensor channels include CH3, CH6, CH9-CH11, CH13-CH16, CH19, CH22-CH25, CH27-CH28, etc.

[0043] The signal preprocessing module is connected with the sensor array and is used for data cleaning and standardization processing of the collected original sensor data. This module is configured to identify outliers using the interquartile range (IQR) method, and to replace the abnormal data with the median interpolation strategy, thereby improving the data quality and model stability. Then, the processed data is normalized to the interval [-1, 1] to ensure the numerical consistency of the neural network input features.

[0044] The artificial neural network (ANN) classification module is used for non-linear mapping and classification learning of the preprocessed sensor feature data. This module adopts a two-layer hidden layer structure, each layer containing 11 neurons, the activation function being ReLU, the learning rate being set to 0.001, and the maximum number of iterations being 1500. The model output is the classification probability value of true and false honey.

[0045] The explainability analysis module is used for visualizing the decision-making process of the ANN model. The module is constructed based on the SHAP (Shapley Additive Explanations) algorithm and can output global feature importance ranking graphs (such as SHAP bar charts, summary swarm plots) and sample-level local dependence graphs (such as waterfall charts, force diagrams, etc.), further assisting in understanding the influence mechanism of key sensors such as CH13 and CH28 in the model discrimination process.

[0046] The hardware control unit is responsible for coordinating the sampling and running process of the system hardware, including setting the flow rate of gas sampling to 400 mL / min, the detection time to 100 seconds, the sampling time to 30 seconds, and controlling the working state of the JSMB-30G analyzer to ensure the stability and repeatability of the data acquisition process.

[0047] The above modules are connected through a data communication interface to form a closed-loop system, which can automatically extract odor signals from honey samples, perform intelligent discrimination and interpretable analysis, and use the results for rapid screening, alarm prompting or production line control of honey authenticity, with good on-site deployment adaptability and scalability.

[0048] II. Sensor array design and optimization In one possible implementation, as shown in Figure 1 One of the core components of the honey authenticity identification electronic nose system described in the present application is a gas sensor array. The array is initially composed of 30 different types of gas sensors, each sensor channel has response capability to a specific class of volatile organic compounds (VOCs), including but not limited to oxygen, ethanol, aldehyde ketones, alcohols, acids, esters, hydrocarbons and sulfur compounds, etc. To improve the discrimination accuracy and operation efficiency of the system, and reduce the noise interference of redundant channels to the model, the present application introduces a Monte Carlo simulation method to screen and optimize the sensor characteristics.

[0049] As shown in Figure 3 The system completes the optimization and screening of sensor characteristics through the following steps: 1. Feature importance ranking: First, the importance of all 30 sensor channels in the model is evaluated and ranked, and an initial ranking list is generated according to their contribution to the honey authenticity classification task.

[0050] 2. Incremental feature evaluation (step-by-step addition): Based on the importance ranking, features are gradually introduced into the model in sequence. In each step, the cross-validation method is used to calculate the accuracy of the model to evaluate the performance of the current feature combination.

[0051] 3. Cross-validation score analysis: Through Monte Carlo simulation, a large number of random samples are recorded, and the cross-validation accuracy of different sensor combinations in each round of training is recorded, and the variation curve between the number of sensors and the performance of the model is drawn.

[0052] 4. Optimal feature subset determination: Figure 3 The curve in

[0053] According to the present application, the optimal sensor subset is determined to have a total of 18 channels, which are CH3, CH6, CH9, CH10, CH11, CH13, CH14, CH15, CH16, CH19, CH22, CH23, CH24, CH25, CH27, CH28, and two groups of important interaction feature channels. This optimal combination not only retains the key VOCs response capability, but also greatly reduces the model input dimension, improves the training efficiency and model generalization performance.

[0054] It should be noted that the two sensor channels CH13 and CH28 are repeatedly verified in multiple experiments as the channels with the highest contribution to the model, corresponding to the sensitive response to characteristic VOCs such as oxygen and ethanol, and become the key indicators in the subsequent SHAP explainability analysis.

[0055] In summary, by using the Monte Carlo feature selection method, combining the ANN model performance curve and the cross-validation results, the present application successfully constructs an optimal sensor subarray, effectively improving the accuracy, stability and hardware deployability of the electronic nose system in honey true and false identification.

[0056] III. Signal acquisition and preprocessing details In one possible implementation, in order to efficiently capture and process the characteristics of volatile organic compounds (VOCs) in honey samples, the present application proposes a standardized signal acquisition process and preprocessing strategy. This process covers sample pretreatment, electronic nose detection parameter setting, and outlier processing and standardization of signal data, ensuring that the subsequent artificial neural network model can obtain high-quality, clearly structured data input.

[0057] 3.1 Honey sample preparation and signal acquisition In the specific implementation process, each honey sample (8 grams in mass) is placed in a 20 mL glass container with a lid, and the container opening is sealed with a sealing film to prevent gas leakage. Then, the sample is placed in a room temperature condition overnight to allow the VOCs in the honey to be fully released and reach a gas dynamic equilibrium state, ensuring stable odor concentration during the detection process.

[0058] Signal acquisition is completed by a JSMB-30G intelligent olfactory analyzer, as shown in Figure 1 The device performs gas injection, sample injection and acquisition operations through an automatic control system, and the specific parameter settings are as follows: Gas flow rate: 400 mL / min; Detection time: 100 seconds; Sample injection time: 30 seconds; Number of repeated detections: 3 independent acquisitions are performed for each sample, and the average value is taken as the final input.

[0059] The analyzer outputs the sensor's electrical signal response to VOCs as a raw data matrix, containing response curves of 30 channels, reflecting the odor characteristics of the honey sample during the entire detection period.

[0060] 3.2 Data preprocessing method To improve the stability of the data and the generalization ability of the model, after obtaining the original sensor data, the system performs outlier rejection, missing value filling and normalization operation on the data through a signal preprocessing module. The specific steps are as follows: Outlier detection: The interquartile range (IQR) method is used to identify outliers. The Q1, Q3 values and IQR range of each sensor channel data are calculated. If a data point is less than Q1-1.5xIQR or greater than Q3+1.5xIQR, it is considered an outlier.

[0061] Median interpolation: Instead of deleting the detected outliers, the median of the corresponding channel is used for interpolation replacement. This method preserves the overall structure of the sample while effectively avoiding model bias caused by extreme values.

[0062] Data standardization: In order to input the data of different channels into the artificial neural network model on the same scale, all sensor channel data are normalized to the interval [-1, 1]. This standardization method can improve the training stability and convergence efficiency of the model, and prevent feature value differences from adversely affecting model performance.

[0063] After the above processing, the original response data output by the electronic nose is converted into a structured and balanced feature vector, which is used for subsequent true or false classification tasks by the artificial neural network model.

[0064] In summary, the present embodiment ensures the quality consistency and robustness of the input data through standardized collection and systematic preprocessing of volatile gases from honey samples, providing a solid data foundation for high-precision honey true or false identification models.

[0065] Four, artificial neural network classification module In one possible implementation, as shown in Figure 2 The present application adopts an artificial neural network (ANN) classification module as the core judgment mechanism for honey true or false identification. This module is responsible for feature learning and nonlinear mapping of preprocessed sensor data, outputting the probability value of the honey sample being "true honey" or "fake honey", and achieving high-accuracy classification.

[0066] 4.1 Model structure As shown in Figure 2 The ANN model consists of the following hierarchical structure: Input layer: receives the feature values of the 18 key gas sensors channels after preprocessing, with a dimension of 18.

[0067] First hidden layer: contains 11 neurons, with ReLU activation function, used to extract the preliminary nonlinear relationship of the data.

[0068] Second hidden layer: also set to 11 neurons, to further deepen the abstract expression of the feature space.

[0069] Output layer: uses Sigmoid activation function, with output values between [0, 1] as probabilities, representing the predicted probability of the honey sample being "real honey".

[0070] The entire model parameters, including weight matrix and bias terms, are optimized through the training process to minimize the loss function.

[0071] 4.2 Model training process To ensure the training quality and generalization ability of the model, the training process of the ANN module includes the following key steps: Training set division: use the KS algorithm to divide the total of 108 samples into training set and test set according to the ratio of 7:3, to ensure the representativeness and uniformity of the samples.

[0072] Training strategy: The learning rate is set to 0.001 to control the rate of parameter update; The maximum number of iterations is set to 1500 to ensure the convergence of the model; The loss function is selected as Binary Cross-Entropy; The optimization algorithm uses Adam optimizer to improve the training efficiency.

[0073] Cross-validation: through ten-fold cross-validation to evaluate the training process to prevent overfitting of the model.

[0074] 4.3 Model performance The ANN classification model constructed in this invention has excellent performance on the test set, as follows: Accuracy: reaches 79%; F1 score: 0.81; AUC value under the ROC curve: as high as 0.85, indicating that the model has strong discriminant ability and generalization ability in distinguishing true and false honey; Confusion matrix analysis: shows that the true positive rate (TPR) and true negative rate (TNR) are at a good level, indicating that the model can effectively identify fake and genuine samples.

[0075] The above results show that the ANN structure used can fully extract the discriminative features in the electronic nose signals and achieve accurate classification of honey authenticity through multi-layer nonlinear mapping, providing a stable and reliable basis for judgment for the system.

[0076] In summary, the artificial neural network classification module serves as the core intelligent discrimination engine in the system, not only achieving deep learning of high-dimensional odor feature data, but also improving the overall intelligence and practicality of the system through structure optimization and training strategies.

[0077] V. Explainability analysis module In one possible implementation, the honey authenticity identification system described in the present application further includes an explainability analysis module, which is constructed based on the SHAP (Shapley Additive Explanations) method and provides global and local explanations for the classification results of the artificial neural network (ANN) model. By quantifying the marginal contribution of each odor sensor channel in the model output, this module can intuitively reveal the positive and negative effects of features on classification decisions, effectively improving the credibility, transparency, and auditability of the model, and providing technical traceability for honey authenticity identification results.

[0078] Figures 4 to 9 The SHAP explainability analysis module is shown in the core visualization output in the system, covering multiple dimensions from overall feature importance ranking, typical sample personalized explanation, to response dependence and interaction effects between odor sensors, which helps users of the system to fully understand the discrimination logic and feature driving mechanism of the model.

[0079] 5.1 Global feature importance analysis (combined with Figure 4 ) To identify the odor sensor channels that have the greatest impact on the classification decision of the model, the system performs global feature contribution analysis on all training samples based on the SHAP (Shapley Additive Explanations) method. Figure 4 The typical global explainability visualization results generated by the SHAP analysis module are shown, which intuitively reflect the average influence of each sensor channel in the overall prediction of the model.

[0080] As Figure 4 shown, the system generates the following two key graphics through the SHAP global explanation function: Figure 4 A: SHAP feature contribution column chart The importance ranking of all sensor channels to the model output is shown. The results show that the average absolute SHAP value of channels CH13 and CH28 is significantly higher than that of other channels, and the combined contribution is not less than 42%, providing the most core discriminant information for the model. Among them, CH13 mainly responds to oxygen-containing components, and CH28 is sensitive to ethanol and some organic solvents. The composition of these two types of volatile substances is significantly different between real honey and adulterated honey, which reflects the consistency of the feature selection learned by the model and the chemical basis.

[0081] Figure 4 B: SHAP summary bee colony chart The chart is based on the SHAP values of all samples in the training set, and the contribution trend between feature values and model output is shown in the form of density distribution. The color depth in the figure represents the high and low of the feature value, and the distribution direction of the point reflects the positive and negative contribution relationship. The results show that the low response value of CH13 and CH28 is usually positively correlated with the determination result of real honey, while the high response value usually appears in adulterated honey samples, revealing the corresponding relationship between the nonlinear discrimination mechanism learned by the model for these channels and the actual chemical law.

[0082] This section analysis provides a theoretical basis and feature selection support for subsequent local explanation and feature interaction mechanism research.

[0083] 5.2 Local sample explanation and visualization (combined with Figure 5 and Figure 6 ) In practical application scenarios, the system not only needs to output the true or false determination result of honey, but also needs to have the ability to trace and explain the classification basis of individual samples, in order to enhance the transparency and verifiability of the model prediction results. Based on the local explanation function of SHAP, this subsection presents the feature contribution of different samples in each odor sensor channel through heat map, waterfall chart and gravity chart, and reveals the classification logic of the model for specific samples.

[0084] Figure 5 The heat map based on SHAP value is shown, which is used to analyze the contribution distribution of multiple samples at the feature level. The horizontal axis of the figure is the sample number, the vertical axis is the 30 odor sensor channels, and the color depth represents the positive and negative size and absolute value intensity of SHAP value. From the figure, it can be seen that CH13 and CH28 show significant positive or negative contribution in multiple samples, indicating that these two sensors play a key role in the model classification process, further verifying their importance in global feature analysis.

[0085] Figure 6 Then from the local sample level, the detailed explanation process of SHAP model for individual samples is shown. Among them: Figure 6A (SHAP waterfall plot) shows the reasoning path of the model for sample 1. CH6, CH28, CH13 provide positive support, pushing the output closer to "adulterated honey"; while CH9, CH25, etc. produce negative pull. After integrating the contributions, the model finally outputs a low probability value, judging that the sample is "adulterated honey".

[0086] Figure 6 B (SHAP force plot) shows the feature contribution distribution of sample 1 and sample 9, respectively, simulating how each sensor feature cooperates to affect the classification result. Sample 9 is a true honey example, and the high response values of CH13 and CH28 produce significant positive SHAP values, with the model output probability f(x) ≈ 0.86, judging it to be "true honey".

[0087] Through the above local explanation plots, the system not only explains "why it is so", but also helps quality supervision personnel to trace the specific feature pattern of adulteration behavior, thereby improving the practical value and acceptability of the model in the field of honey detection.

[0088] 5.3 Feature dependence and interaction analysis (combined with Figure 7 、 Figure 8 、 Figure 9 ) To further analyze the decision-making logic of artificial neural network models in determining the authenticity of honey, this section uses the dependence analysis and interaction visualization tools in the SHAP method to explore the nonlinear influence of different odor sensor channel response patterns on the model output and the synergistic effect between channels. This type of analysis not only enhances the explainability of the model, but also provides data support for optimizing sensor array configuration and parameter selection.

[0089] Figure 7 The single-factor dependence graph shows how the response value of a particular sensor channel affects its SHAP contribution value in the model output. Taking CH13 as an example, when its response value is below about 573.12, the SHAP value is generally positive, and the model tends to judge it as "true honey"; when the response value exceeds this threshold, the SHAP value gradually turns negative, indicating that its high response is actually a "fake" signal. This type of single-variable dependence curve reveals the complex decision boundaries learned by the model in different feature value intervals.

[0090] Figure 8 The two-factor interaction graph visualizes the relationship between the response combination of two sensors (such as CH13 and CH28) and the model output, showing the nonlinear interaction between features. For example, when CH13 and CH28 are both in a low response state, the model clearly tends to predict "true honey"; when CH28's response value exceeds 800, its contribution changes from positive to negative, inhibiting the model's judgment result, reflecting the existence of threshold sensitivity and synergistic turning points between features.

[0091] Further, Figure 9 A three-dimensional partial dependence plot (3DPDP) based on SHAP is demonstrated to visualize the influence surface of CH13 and CH28 channels on the model output under different response value combinations. Such three-dimensional graphics help researchers understand the model's discrimination strategy in a multivariate input scenario, and also provide targeted optimization basis for sensor array design.

[0092] In summary, by introducing the dependency and interaction analysis function of SHAP, the invention effectively breaks through the "black box" characteristics of traditional neural networks, and gives the honey true and false detection system interpretability, traceability and data transparency. This module not only enhances the user's trust in the model decision results, but also provides a visual basis for food quality supervision and enterprise product responsibility, opening up a new path of "intelligence + credibility" for honey and other agricultural product authenticity detection.

[0093] 5.4 SHAP interpretability analysis process In one possible implementation, as Figure 10 shown, the interpretability analysis module of the invention builds a complete analysis process based on the SHAP method, which is visual, auditable and triggerable, to enhance the transparency and credibility of the honey true and false discrimination model.

[0094] The process includes the following four main stages: 1. Feature contribution calculation: the system receives the prediction results of the artificial neural network model and the corresponding sample feature input, calls the SHAP algorithm to estimate the marginal influence value of each sensor channel, and generates a SHAP value matrix for each sample; 2. Global and local visualization generation: based on SHAP output, automatically draw global feature importance bar chart, summary bee colony chart, and local sample waterfall chart, force chart and dependence chart, to reveal the model's decision mechanism at multiple levels; 3. Abnormal contribution detection: the system monitors whether the SHAP contribution of key sensors (such as CH13, CH28) exceeds the set threshold (such as less than 5% or more than 60%), if there is abnormal fluctuation, the sensor self-calibration mechanism is automatically triggered; 4. Result output and archiving: the final explanation result is presented to the user through the graphical interface, and can be stored as a log record or report for traceability audit, model optimization or quality control.

[0095] Figure 10 The logical connection between the above steps is intuitively presented in the form of a flowchart, emphasizing that the SHAP method not only serves as an "interpretation tool", but also is embedded in the system closed-loop control logic, supporting multiple functions of data intelligent judgment, sensor self-checking and result traceability.

[0096] In summary, the SHAP explainability analysis process module of the present application not only provides visual explanation of the basis for model prediction, but also realizes system-level fault perception and dynamic response, significantly improving the reliability and acceptability of honey authenticity detection technology in practical application.

[0097] VI. Classification decision logic and dynamic threshold In one possible implementation, to realize automatic true or false discrimination of honey samples, the present application proposes a probability decision logic mechanism based on artificial neural network (ANN) classification output, and further introduces a dynamic threshold adjustment strategy related to honey varieties, to adapt to the differences in odor characteristics of diversified honey types, and improve the practicality and adaptability of the model.

[0098] 6.1 Classification probability decision logic After the artificial neural network model performs inference calculation on the preprocessed 18-channel odor characteristics, it outputs a probability value in the range of [0, 1], representing the prediction confidence that the sample is "real honey" (denoted as P). The system performs binary classification based on the comparison result of the probability value and the set threshold value: If P≥0.7, it is determined to be "real honey"; If P<0.7, it is determined to be "fake honey".

[0099] This standard is applicable to true or false identification in general cases, and is suitable for deployment in standard scenarios where honey varieties are not distinguished by default.

[0100] 6.2 Dynamic threshold adjustment strategy Due to the natural differences in VOCs composition of different flower species of honey, directly using a fixed threshold value may lead to misjudgment of some samples. Therefore, in one possible implementation of the present application, a dynamic threshold mechanism is introduced to flexibly set the decision criteria according to the honey variety, as follows: Multi-flower honey (such as Xinjiang Nileke black bee honey): the decision threshold is set to 0.65 to enhance the identification sensitivity of honey samples with complex characteristics but clear authenticity; Single-flower honey (such as acacia honey, jujube flower honey, etc.): the decision threshold is set to 0.75 to enhance the model's ability to distinguish weak fake signals in relatively simple component varieties.

[0101] The system can manually select the honey variety when the sample is loaded, or automatically identify the honey variety through an external information system (such as a database interface or a two-dimensional code traceability system), thereby dynamically adjusting the discrimination threshold to further improve the accuracy and scene adaptability of classification.

[0102] 6.3 Configurable parameter interface To meet the flexibility requirements of practical applications, the system software platform reserves threshold configuration interfaces, allowing users to perform the following operations according to specific detection requirements: Manually set the determination threshold; import the threshold configuration table (supporting multiple flower species mapping); automatically switch the threshold parameters at the system level; view the historical records of the threshold and classification results.

[0103] In summary, the present application significantly improves the determination flexibility and practical application accuracy of the honey authenticity identification system by constructing a classification logic system centered on probability output and setting a dynamic threshold mechanism based on honey flower characteristics, and is particularly suitable for rapid screening and quality verification in diversified honey scenarios.

[0104] Seven, system deployment scenarios In one possible implementation, as shown in Figure 11 The honey true-false identification electronic nose system based on explainable artificial intelligence according to the present application is not only suitable for laboratory detection environment, but also can be integrated and deployed in the filling production line of honey production workshop, realizing online, automatic and closed-loop quality control function.

[0105] 7.1 Production line deployment structure and process In the production line deployment scenario, the system is connected with the existing honey filling production line as an embedded module, and the overall deployment process is as follows: 1. Sample gas sampling unit: automatically extract samples from the honey batch before filling, and guide them to the sensor inlet through a sealed small container.

[0106] 2. Electronic nose detection module: the system starts detection according to the set process (flow rate 400 mL / min, detection time 100 seconds, sampling time 30 seconds), and automatically collects VOCs feature data.

[0107] 3. Data preprocessing and ANN classification: after the collected data are processed by the preprocessing module, they are sent to the ANN model, and the true-false probability results are output.

[0108] 4. SHAP explainability analysis (optional): the system can automatically generate local explanation graphs for this sample in the background, such as SHAP waterfall chart and key channel contribution chart, for subsequent review by supervisors.

[0109] 5. Production line linkage control and alarm response: if the detection result meets the abnormal trigger condition, the interface signal feedback controls the production line action (such as pause, alarm, etc.).

[0110] 7.2 Alarm and shutdown trigger logic To improve the food safety risk response capability, the system has multiple trigger mechanisms built in, which are used to identify potential quality problems in continuous detection and control the production line operation in time. The specific logic is as follows: When the classification probability of two or more consecutive samples is below 0.6, the system immediately triggers an alarm signal; At the same time, the system sends a "pause filling" instruction to the filling control unit to block the material flow, and waits for the operator to intervene; The system records the sample data and prediction results when the alarm is triggered, and generates a local audit log.

[0111] This mechanism ensures that a closed-loop quality assurance can be achieved during honey production, significantly reducing the risk of adulterated honey entering the market.

[0112] 7.3 System integration interface To ensure seamless integration with various production line systems, the system supports multiple interface protocols, including: RS-232 / RS-485 serial communication; TCP / IP network communication; programmable logic controller (PLC) trigger interface; SCADA or MES system data synchronization module (optional).

[0113] In addition, the system also supports remote access to historical detection records and real-time alarm information through Web services or industrial cloud platforms, enabling honey quality traceability and centralized supervision at the workshop and group levels.

[0114] In summary, the honey true and false identification system provided by the present application has good on-site deployment capability and can efficiently integrate into the honey production automation process, forming a data-driven online detection + visual interpretation + real-time control closed-loop system, significantly improving the food quality and safety protection capability of the honey industry chain.

[0115] Eight, hardware simplification and low-cost configuration scheme In one possible implementation, considering that different application scenarios have different requirements for the cost, size, and energy consumption of detection equipment, to improve the adaptability of the system in basic detection, mobile detection, or resource-constrained scenarios, the present application proposes a sensor array simplification configuration scheme to achieve low-cost deployment without significantly reducing identification performance.

[0116] 8.1 Simplified sensor array configuration In the original system, 18 key gas sensor channels were selected based on the Monte Carlo simulation method, but in resource-sensitive scenarios, to further reduce the number of sensors and manufacturing costs, the present application optimizes the top 5 core sensor channels with the highest contribution to the model to construct a simplified sensor array. Specifically, it includes: CH13: sensitive to oxygen, with the highest contribution; CH28: sensitive to ethanol and other organic solvents, with high contribution; CH6, CH10, CH22: good response to typical honey volatile substances such as ketones, alcohols, and esters.

[0117] This five-channel combination provides the most representative odor fingerprint information for the discriminant performance of the model, achieving a good balance between precision and complexity.

[0118] 8.2 Performance evaluation under simplified configuration Through experimental verification, the ANN model trained using the above-mentioned 5-channel sensor array can still achieve a classification accuracy of ≥70%, and has good recognition ability under non-extreme sample conditions. The specific performance is as follows: It can effectively capture the dominant VOCs characteristics of honey samples; it still maintains high sensitivity when identifying high-concentration adulterated samples; the stability of the model is controlled by the state of the main sensor, which is easy to maintain.

[0119] This configuration is particularly suitable for the following scenarios: Quick screening equipment for the circulation link of the honey market; mobile detection terminals for grassroots law enforcement departments; simple quality control equipment for bee farmers or small and medium-sized processing plants; temporary detection occasions where complete systems cannot be deployed in remote areas.

[0120] 8.3 Cost control and deployment advantages Compared with a full 30-channel system or an 18-channel optimal configuration, the five-channel scheme has the following advantages: Hardware costs are reduced by more than 50%; circuit board area and power consumption are significantly reduced; system response time is shortened, and model inference speed is improved by about 20%; it can be integrated into portable devices or handheld terminals, making it easy to use outdoors.

[0121] In addition, to further improve the applicability of the equipment, the present application also supports a pluggable design of the sensor module, and users can dynamically adjust the sensor configuration according to the detection task, taking into account flexibility and economy.

[0122] In summary, the simplified hardware configuration scheme provided by the present application selects high-contribution degree sensor channels to achieve a significant reduction in cost and portable design while maintaining core detection capabilities, significantly expanding the coverage of the system in actual honey detection applications, and meeting the diverse needs of users at different levels for detection cost, efficiency, and deployment methods.

[0123] Nine, sensor self-calibration mechanism In one possible implementation, to ensure the stability of the system during long-term operation and the consistency of the discrimination results, the present application constructs a sensor self-calibration mechanism based on SHAP analysis results. This mechanism combines model interpretability output and device control strategies, and can automatically trigger the calibration program when key sensors fluctuate abnormally, thereby prolonging the service life of the device and reducing the risk of misjudgment.

[0124] 9.1 Calibration trigger logic As mentioned before, CH13 and CH28 sensors are the two channels with the highest average contribution in the model, playing a leading role in honey authenticity judgment. To monitor their running state and output stability in real time, the present application sets the following calibration trigger conditions: When the SHAP contribution value is lower than 5%, or When the SHAP contribution value is higher than 60%, it is considered that the channel has abnormal fluctuations or drift trends in the current sample, and the system automatically starts the sensor self-calibration process.

[0125] The trigger logic is not only based on a single judgment, but also can be set to execute when abnormal contribution occurs in continuous multiple samples to avoid accidental noise triggering.

[0126] 9.2 Calibration process overview When the trigger condition is met, the system automatically performs the following calibration steps: 1. Sensor state check: the system reads the baseline output, response time and drift trend of CH13, CH28 and other sensors through the internal diagnostic module; 2. Reference gas cleaning (if applicable): if the device has a cleaning module, reference gas (such as high-purity nitrogen) can be introduced to zero the sensor; 3. Sensitivity recovery and initialization: the system performs bias correction or sensitivity factor reset for the channel; 4. SHAP contribution monitoring after calibration: the sample is re-detected and the contribution of the channel is analyzed to see if it returns to the normal range, forming a closed-loop verification mechanism; 5. Abnormal record and operation prompt: if calibration fails or the contribution is continuously abnormal, the system records logs and prompts the user to check or replace the corresponding sensor module.

[0127] 9.3 Calibration strategy configurability To adapt to different scenarios for device stability requirements, the system allows users to customize the following parameters through the configuration interface: Contribution upper and lower threshold (default 5% / 60%, adjustable); whether to enable the continuous sample trigger mechanism; calibration frequency, maximum calibration attempt number; enable / disable automatic calibration function; push alarm information to the operation and maintenance platform (such as connecting to SCADA or MES system).

[0128] 9.4 Implementation effect and advantage The SHAP-based contribution monitoring has the following advantages: Using the intrinsic discriminant of the model to judge the change of sensor performance is more targeted than traditional "signal amplitude + noise" monitoring; it realizes intelligent diagnosis and self-recovery capability within the system, reducing the frequency of manual maintenance; it improves the stability and reliability of the system in long-term operation or multi-batch detection process.

[0129] In summary, the application realizes continuous monitoring and automatic correction of the state of the core gas sensor by constructing a sensor self-calibration mechanism with SHAP abnormal contribution degree as the core trigger condition, significantly improving the durability, accuracy and intelligent maintenance capability of the honey authenticity detection system in industrial scenarios.

[0130] Ten, experimental sample and data set source In one possible implementation, in order to construct a honey authenticity identification model with high representativeness, high discriminability and good generalization ability, the application relies on the Xinjiang Nileke black bee honey producing area, collects and constructs a structured honey sample database, and completes the whole process of electronic nose signal collection, feature extraction and model training on this basis.

[0131] 10.1 Sample collection description A total of 108 honey samples were collected in the experiment, and the specific composition was: 53 genuine honey samples, all from the authentic Nileke black bee honey in the Nileke County Black Bee Reserve in Yili State, Xinjiang, confirmed by a third party as a non-fraudulent product; 55 fake honey samples were constructed by the following methods: Dilute with syrup (such as high fructose syrup, corn syrup, etc.) and real honey in different proportions; Mix with different geographical origins, different bee species or non-Nileke black bee honey samples; Fake way of artificially changing volatile components during processing.

[0132] Each sample was collected strictly in accordance with the operation specifications of sample sealing, label numbering, data control, etc., to ensure the authenticity and reviewability of the data source.

[0133] 10.2 Data collection and preprocessing standards All samples were sampled three times independently using a JSMB-30G intelligent olfactory analyzer, with the following parameters set: Gas flow rate: 400 mL / min; detection time: 100 seconds; sampling time: 30 seconds.

[0134] The raw data obtained were uniformly processed by the following preprocessing procedures: 1. Outlier detection and interpolation: IQR method was used to identify extreme values, and median was used for replacement; 2. Normalization: standardize each channel data to the interval [-1, 1] to ensure that the model input features are on a uniform scale; 3. Feature selection: apply Monte Carlo simulation method to select the 18 highest contribution sensors; 4. Sample set division: The data is divided into training set and test set by the Kennard-Stone algorithm in the ratio of 7:3, and is used for model training and cross-validation.

[0135] 10.3 Data set characteristics and reusability The experimental data has the following characteristics: High representativeness: covering local unique bee species in Xinjiang and typical adulteration methods, consistent with real market scenarios; Balanced structure: the number of true and false samples is basically balanced, which is conducive to avoiding model bias; High signal quality: stable sampling period, low data curve noise, convenient for subsequent modeling analysis; Strong reusability: all raw data, processing procedures and label information can be archived, supporting subsequent iteration and retraining of the system.

[0136] In addition, the data set constitutes the main basis for the training and evaluation of the system, and is the core support for its performance verification, and can be used as a reference basis for future large-scale honey true and false detection sample standards.

[0137] The present application provides a solid training and verification basis for the electronic nose system and artificial neural network model by constructing a data set with clear source, standardized structure and covering real adulteration behavior, ensuring the reliability and scientificity of the honey true and false identification system in various application scenarios.

[0138] In summary, the present application proposes an electronic nose system for identifying true and false honey based on interpretable artificial intelligence, which combines a multi-channel gas sensor array, a high-precision artificial neural network model and a SHAP interpretability analysis method, realizes intelligent identification and transparent judgment of honey volatile odor information, and has the advantages of high identification accuracy, strong model interpretability, flexible system deployment, wide adaptability, etc. The system supports online deployment, dynamic threshold adjustment, self-calibration control and production line linkage, and can be widely applied to multiple links of honey production, circulation, supervision, etc., significantly improving the efficiency and credibility of honey quality identification, and has good practical application prospect and popularization value.

[0139] It should be understood that the above description is only a preferred embodiment of the present application and is not intended to limit the protection scope of the present application. For those skilled in the art, various equivalent transformations or replacements of the above technical solutions can be made without departing from the principles and essence of the present application, which should be covered within the protection scope of the present application.

Claims

1. An electronic nose system for identifying genuine and fake honey based on deep learning, characterized in that, include: A gas sensor array was used to collect volatile organic compound (VOC) signals from honey samples. The array consisted of 30 gas sensors, and 18 key sensors were selected for subsequent analysis using Monte Carlo simulation. The signal preprocessing module is used to perform outlier detection, median interpolation, and standardization on the gas sensor data. The outlier detection adopts the interquartile range (IQR) method, and the standardization process normalizes the data to the [-1,1] interval. The Artificial Neural Network (ANN) classification module is used to classify the preprocessed sensor data and output the probability of honey authenticity. The ANN model includes two hidden layers, each containing 11 neurons, with a learning rate of 0.001 and a maximum number of iterations of 1500. The interpretability analysis module is used to interpret the ANN model using the SHAP method, generating feature contribution maps, 3D partial dependency maps, and local dependency maps. The hardware control unit is used to control the analyzer's operating parameters, including a gas flow rate of 400 mL / min, a detection time of 100 seconds, and an injection time of 30 seconds.

2. A method for identifying genuine and fake honey, characterized in that, Includes the following steps: S1: Collects signals from a 30-channel gas sensor on a honey sample; S2: 18 key sensor features were selected based on Monte Carlo simulation; S3: Perform outlier processing and median interpolation on the sensor data using the IQR method, and then standardize it to [-1,1]; S4: Input the processed data into a double-hidden-layer ANN model to obtain the true / false classification probability; S5: Obtain the feature contribution of key sensors CH13 and CH28 using the SHAP algorithm; S6: Set a judgment threshold. When the classification probability is ≥0.7, it is judged as real honey, and when it is <0.7, it is adulterated honey.

3. The system according to claim 1, characterized in that: The SHAP analysis showed that the combined average absolute contribution of sensors CH13 and CH28 to the model output was no less than 42%.

4. The system or method according to claim 1 or 2, characterized in that: The ANN model training uses the KS algorithm to divide the training set and test set into a 7:3 ratio, and uses 10-fold cross-validation. The AUC value of the test set is not less than 0.85, and the F1 score is not less than 0.

81.

5. The method according to claim 2, characterized in that: The classification threshold is dynamically adjusted based on the type of honey flower. The threshold is set at 0.65 for multi-flower honey and 0.75 for single-flower honey.

6. The system according to claim 1, characterized in that: The system is integrated into the honey bottling production line. When the predicted probability of two consecutive samples is lower than 0.6, the system automatically triggers an alarm and suspends the production line.

7. The system according to claim 1, characterized in that: The gas sensor array can be replaced with a low-cost combination containing only five sensors: CH13, CH28, CH6, CH10, and CH22. This combination has an accuracy rate of no less than 70% in determining the authenticity of honey.

8. The system according to claim 1, characterized in that: When the SHAP analysis results show that the contribution of the CH13 or CH28 sensor is less than 5% or more than 60%, the system automatically triggers the sensor calibration procedure.

9. The system or method according to claim 1 or 2, characterized in that: The training samples for the ANN model include 53 genuine samples and 55 adulterated samples of Xinjiang Nileke black bee honey.

Images