Tobacco leaf baking stage detection method based on electronic nose time-frequency MIC optimization characteristics

Abstract

The invention relates to a tobacco leaf curing stage detection method based on time-frequency MIC optimization characteristics of an electronic nose, in particular to a method for detecting a tobacco leaf curing stage by fusing time-domain characteristics and frequency-domain characteristics in response signals of the electronic nose and combining maximum mutual information coefficient (MIC) characteristic optimization and an identification model. Comprising the following steps: firstly, establishing a tobacco leaf curing stage detection model based on electronic nose time-frequency MIC optimization characteristics, and then detecting the tobacco leaf curing stage by using the detection model. According to the method, the time domain feature and the frequency domain feature in the electronic nose signal are fused to identify the tobacco leaf baking stage, so that the accuracy is high; in combination with an MIC method, feature optimization and data dimension reduction are realized; operation is simple, detection is rapid, the whole electronic nose collecting system and the matched device thereof are economical and practical, market application and popularization are facilitated, and technical reference can be provided for improving the intelligent level of tobacco leaf curing.

Description

Technical Field

[0001] The invention belongs to the technical field of tobacco leaf intelligent baking detection, and in particular relates to a tobacco leaf baking stage detection method based on an optimal electronic nose time-frequency characteristic, which can be used for detecting the baking stages of different types of tobacco leaves. Background Art

[0002] Baking is a key link to highlight the style and quality of tobacco leaves, and is an important step in the transformation and fixation of the internal chemical composition and aroma substances of tobacco leaves. However, the current baking state of tobacco leaves still relies on the subjective judgment of workers. The baking process is not targeted and has strong subjectivity. It is difficult to achieve accurate matching between the real-time state of tobacco leaves and the baking process, which affects the transformation and accumulation of aroma substances in tobacco leaves. Therefore, it is of great practical significance to intelligently distinguish the baking state of tobacco leaves. For the discrimination of the baking state of tobacco leaves, machine vision, near-infrared spectroscopy and other technologies are often used for detection. Compared with human eye observation, there has been some improvement, but there are still problems of limited observation area and poor representativeness. The electronic nose technology has the advantages of obtaining overall information of the sample and good representativeness. In addition, the current research on electronic noses usually uses the extracted maximum value, minimum value, average value, maximum first-order derivative, standard deviation and integrated area and other time domain features, as well as the frequency domain features extracted by wavelet transform combined with feature optimization methods for discriminant analysis, which has a good application prospect, but the research on the discrimination of tobacco leaf baking stage based on the optimal electronic nose time-frequency features is relatively rare. Summary of the invention

[0003] The purpose of the present invention is to provide a tobacco leaf baking stage detection method based on the time-frequency MIC optimization features of an electronic nose. The tobacco leaf odor characteristics are taken as the research object, and the electronic nose is used to collect tobacco leaf signals at different baking stages. The time domain features and frequency domain features in the signals are then fused and optimized. Then, the discrimination models of different baking stages are compared to obtain the optimal model for distinguishing the tobacco leaf baking stage, which is used for the detection of the tobacco leaf baking stage, aiming to solve the problems of limited observation area and poor representativeness of the existing tobacco leaf baking stage detection technology.

[0004] The technical solution adopted by the present invention is as follows: A tobacco leaf baking stage detection method based on electronic nose time-frequency MIC optimization features is characterized in that the time-frequency features in the electronic nose signal at different baking stages are obtained, and then the time-frequency features are optimized using MIC (maximum mutual information coefficient), and different tobacco leaf baking stages are discriminated in combination with a discriminant model, including the following steps: Step 1: Obtain gas samples from the baking room at different baking stages; Step 2: Under room temperature conditions, the electronic nose response signals of gas samples at different baking stages are obtained using the same acquisition parameters; Step 3: preprocessing the signal of the electronic nose sensor array; Step 4: extract the time domain features and frequency domain features in the electronic nose response signal respectively, fuse the extracted time and frequency features, and use MIC to optimize the time and frequency features; Step 5: Divide the optimized time-frequency features into a training set, a validation set, and a test set, and use the divided data sets to train and validate the model, select the best model, and determine the various parameters of the model; Step 6: If the effect of the established detection model meets the requirements, it means that the model is feasible; otherwise, expand the sample data set and repeat steps 1 to 5 until the requirements are met; The specific steps of using the model to detect the tobacco leaf baking stage are as follows: Step A, obtaining gas samples from the baking room at different baking stages; Step B, at room temperature, using the same acquisition parameters to obtain the electronic nose response signals of the gas samples at different baking stages; Step C, preprocessing the signal of the electronic nose sensor array; Step D, respectively extracting the time domain features and frequency domain features in the electronic nose response signal, fusing the extracted time-frequency features, and optimizing the time-frequency features using MIC; Step E: input the optimized time-frequency features into the established detection model, identify the baking room gas sample, and output the detection results.

[0005] When optimizing the time-frequency features of the electronic nose based on MIC, the MIC coefficients between the time domain features, the frequency domain features and the tobacco leaf baking stage are calculated respectively, and then the features are optimized using the threshold segmentation method. The threshold selection is based on the model accuracy improvement of less than 1% or no longer improving. The present invention selects the time-frequency features with a MIC coefficient greater than 0.36 as the preferred feature combination, wherein 18 features are extracted in the time domain and 32 features are extracted in the frequency domain, and the total number of features is reduced by 30.56%.

[0006] The above-mentioned electronic nose time-frequency characteristics, wherein the time domain characteristics include the maximum value, minimum value, average value, variance, integral value and maximum first-order derivative in the electronic nose response signal; the frequency domain characteristics refer to the mean, standard deviation and median value of the approximate coefficient and detail coefficient obtained by wavelet transform of the electronic nose response signal; the electronic nose time-frequency characteristics are arranged in the order of feature extraction, and then feature optimization is performed based on MIC; the tobacco leaf baking stages include the early yellowing stage, the late yellowing stage, the early color fixation stage, the late color fixation stage and the dry tendon stage.

[0007] In the above step 1 and step A, the gas samples of the baking room at different baking stages are obtained by sampling at representative positions, such as the upper, middle and lower points of the geometric center of the baking room.

[0008] In the above step 2 and step B, at room temperature, the acquisition parameters include a gas flow rate of 2000 ml / min and a reaction time of 2 min.

[0009] In the above step 3 and step C, when preprocessing the signal of the electronic nose sensor array, it is necessary to perform baseline correction and smoothing on the original signal, wherein the smoothing method includes but is not limited to the sliding average method, the exponential sliding average method and the SG filtering method.

[0010] In the above steps 6 and E, the discriminant model established has a network structure including 1 input layer, 2 convolutional layers, 2 batch normalization layers, 2 maximum pooling layers and 1 fully connected layer. The input layer uses a standardized method to normalize the data. The convolutional layer parameters are 2×1×16, the maximum pooling parameters are 2:1, the model learning rate is 0.001, and softmax is used as the mapping function.

[0011] The detection system used in the above method includes an electronic nose signal acquisition unit, a gas acquisition unit, a profiling gas chamber and a control and analysis unit; the electronic nose signal acquisition unit is composed of a gas sensor array, a conditioning circuit and a communication module, wherein the gas sensor array is located inside the profiling gas chamber, the conditioning circuit supplies power to the gas sensor and performs A / D conversion on the signal, and the communication module adopts USB communication; the gas acquisition unit includes an air pump, a speed regulator and a gas acquisition bag, the air pump is used to detect the flow of the gas to be tested and the reducing gas in the system, and the speed regulator is used to control the flow rate of the air pump; the profiling gas chamber is printed by resin material and is covered with an anti-stick coating and an airflow guide mechanism; the control and analysis unit includes a PC and a software interface, wherein the software interface is used for the real-time display, storage and discrimination analysis functions of the electronic nose signal.

[0012] Compared with the prior art, the present invention has the following advantages: First, the detection method of the present invention has the advantages of simple operation, economical and practical, and easy to promote; Second, the time-frequency characteristics of the electronic nose signal integrated in the detection method of the present invention can more comprehensively obtain the characteristic information in the signal and improve the detection accuracy of the tobacco leaf baking stage; Third, the present invention provides technical support and reference for the intelligent identification of tobacco leaf baking stages and other object baking stages, and also provides new ideas and methods for integrating time-frequency features in the field of agricultural product baking status detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 Schematic diagram of the electronic nose data collection system involved in Embodiment 1 and Embodiment 2 of the present invention Figure 2 The electronic nose response signal diagram at different baking stages involved in Example 2 of the present invention is Figure 3 This is the TFF-MIC-1DCNN network model architecture diagram involved in Example 2 of the present invention Figure 4 The importance coefficient of the electronic nose time-frequency characteristics calculated based on MIC in Example 2 of the present invention is a) time domain b) frequency domain Figure 5 The model training process diagram based on MIC optimization features involved in Example 2 of the present invention is Figure 6 The confusion matrix of the external test set of the TFF-MIC-1DCNN model involved in Example 2 of the present invention DETAILED DESCRIPTION

[0014] In order to make the purpose and advantages of the present invention more clear, the present invention is further described below in conjunction with specific embodiments.

[0015] Example 1: A detection system structure and use of a tobacco leaf baking stage detection method based on the time-frequency MIC optimization characteristics of an electronic nose Part a: Detection system structure of a tobacco leaf baking stage detection method based on the time-frequency MIC optimization characteristics of the electronic nose The tobacco leaf baking stage detection system includes an electronic nose signal acquisition unit, a gas acquisition unit, a profiling air chamber and a control and analysis unit; the electronic nose signal acquisition unit is composed of a gas sensor array, a conditioning circuit and a communication module, wherein the gas sensor array is located inside the profiling air chamber, the conditioning circuit supplies power to the gas sensor and performs A / D conversion on the signal, and the communication module adopts USB communication; the gas acquisition unit includes an air pump, a speed regulator, and a gas acquisition bag, the air pump is used to detect the flow of the gas to be tested and the reducing gas in the system, and the speed regulator is used to adjust the flow rate of the air pump; the profiling air chamber is printed with resin material and is covered with an anti-stick coating and an airflow guide mechanism; the control and analysis unit includes a PC and a software interface, wherein the software interface is used for the real-time display, storage and discrimination analysis functions of the electronic nose signal.

[0016] Part b: A method for using a detection system for tobacco leaf baking stage detection based on the time-frequency MIC optimization characteristics of an electronic nose b1. First, power on the system, preheat the sensor, and let the system run at no load. When the sensor response curve is stable, you can prepare for testing.

[0017] b2. Turn on the computer, open the electronic nose signal online acquisition software, and check whether the electronic nose device is successfully connected to the computer through the software interface.

[0018] b3. Set the gas flow rate to 2000 ml / min, the acquisition frequency to 1 Hz, and the acquisition time to 2 min.

[0019] b4. Connect the activated carbon tube to the electronic nose cleaning port.

[0020] b5. Connect the gas sampling bag to the inlet of the electronic nose and open the outlet valve of the gas sampling bag.

[0021] b6. Click the start button, and the software interface will display the collected electronic nose signal in real time and save it.

[0022] Example 2: A tobacco leaf baking stage detection method based on the electronic nose time-frequency MIC optimization feature Part C: Establishing a discriminant model for tobacco leaf baking stage detection c1. Sample acquisition Select mature leaves without pests and diseases and obvious mechanical damage, and braid them according to the local production process. After braiding, they are evenly loaded into the dense curing room. The curing process parameters are formulated with reference to the local process, and samples are taken 4 times at each temperature point. In order to improve the representativeness of the samples and enrich the diversity of the samples, samples are taken at the same time at the upper, middle and lower positions in the middle of the curing room. A total of 540 samples.

[0023] c2. Sample electronic nose signal collection The detection system includes an electronic nose signal acquisition unit, a gas collection unit, a contoured gas chamber, and a control and analysis unit. First, power on the system, preheat the sensor, and let the system run at no load. When the sensor response curve is stable, you can prepare for detection. Set the gas flow rate to 2000ml / min, the collection frequency to 1Hz, and the collection time to 2min. Then connect the activated carbon tube to the electronic nose cleaning port, connect the gas sampling bag to the electronic nose inlet, and open the gas sampling bag outlet valve. Click the start button, the software interface can display the collected electronic nose signal in real time and save it. The electronic nose response signal at different baking stages is as follows Figure 2 shown.

[0024] c3. Electronic nose signal preprocessing The electronic nose information obtained in the present invention is one-dimensional time series data with a duration of 2 minutes. In order to eliminate the influence of environmental gas on the sample response signal, it is necessary to perform baseline correction on the sample response signal; in order to reduce the influence of noise in the signal, the signal is smoothed using the sliding average method, exponential sliding average method and SG filtering method.

[0025] c4. Time-frequency feature extraction and optimization In order to characterize the information in the sensor response curve, six time domain features, including maximum value, minimum value, average value, variance, integral value and maximum first-order derivative, are extracted from the time domain dimension of the original response signal of the electronic nose. The maximum value and minimum value reflect the dynamic characteristics of the signal. The average value, variance and integral value reflect the overall level of the signal. The maximum first-order derivative reflects the change level of the signal. The frequency domain features extracted by wavelet transform can reveal the internal characteristics of the signal and are clearly distinguished from the time domain signal. In wavelet analysis, detail coefficients and approximate coefficients are used to describe the characteristic coefficients of the signal at different frequencies. The detail coefficient represents the detailed information of the signal in the high-frequency range, and the approximate coefficient represents the overall trend or slow change of the signal in the low-frequency range.

[0026] The maximum value indicates the maximum point of the sensor response value within 2 minutes, and the minimum value indicates the minimum point of the sensor response value within 2 minutes. The maximum and minimum values ​​of the i-th sensor are defined as follows:

[0027] in i is the sensor tag, x i1 , x i2 , ..., x i120 It is i The response value of a sensor.

[0028] The mean value represents the average value of the sensor response signal within 2 minutes. The variance represents the average of the square sum of the deviations of the sensor signal from the mean value within 2 minutes. The integral value represents the integral area of ​​the sensor response curve within 2 minutes and the time axis. The mean value, variance and integral value of the i-th sensor are defined as follows:

[0029] in i is the sensor tag, x i1 , x i2 , ..., x i120 It is i The response value of a sensor, μ is the average response of the sensor within 2 minutes, f(t) It is the expression of the sensor response curve within 2 minutes.

[0030] The maximum first-order derivative indicates the severity of the change in the sensor response signal within 2 minutes. The maximum first-order derivative of the i-th sensor is defined as follows: in iis the sensor tag, x i1 , x i2 , ..., x i120 It is i The response value of a sensor, Δt is the average response signal frequency of the sensor within 2 minutes. The time domain characteristics and sensor numbers are shown in Table 1.

[0031] Table 1 Correspondence between time domain characteristic parameters and sensors

[0032] Assume the received signal is x [ n ], then the detail coefficient and approximate coefficient obtained by DWT are expressed in mathematical formula as follows:

[0033] where c A [ j ][ k ] is the approximate coefficient, c D [ j ][ k ] is the detail coefficient, j represents the transformation series, k represents the subband index, represents the inverse transform of the wavelet basis function, Represents the inverse transformation of the scaling function. This paper selects the first wavelet of the Daubechies wavelet and performs a 1-layer decomposition. The present invention uses the mean, standard deviation and median of the detail coefficient and the approximate coefficient to represent the frequency domain features. The frequency domain features and sensor numbers are shown in Table 2.

[0034] Table 2 Correspondence between frequency domain characteristic parameters and sensors

[0035] The extracted time-frequency features are fused, and MIC is used for feature optimization. The maximum information coefficient (MIC) is a statistic that measures the relationship between two variables. Compared with traditional mutual information, MIC can capture a wider range of dependencies, including linear and nonlinear relationships. As a criterion for feature selection, MIC can effectively measure the strength of the association between features and target variables. The core idea of ​​MIC is to divide the data into different grid structures, then calculate the mutual information between variables under these grids, and finally select the grid segmentation method to maximize the mutual information. Specifically, MIC is effective for any two variables. X and YFor example, we will find the value that maximizes the mutual information under a certain grid division and standardize it to get the MIC value.

[0036] (1) Grid division The variable X and Y Assume that X Divided into m intervals, Y Divided into n interval, then form a m × n of grid.

[0037] (2) Calculate mutual information For each grid division, calculate X and Y Mutual information on this grid I ( X , Y ). Mutual information measures X and Y The difference between the joint probability distribution and the marginal probability distribution of :

[0038] in, p ( x,y )yes X and Y The joint probability distribution of p ( x )and p ( y )yes X and Y The marginal probability distribution of .

[0039] (3) Finding the maximum grid division By trying different m × n Grid combinations, find the mutual information I ( X,Y ) The largest grid division. For each m and n , calculate the maximum mutual information, and standardize its value:

[0040] in, M ( X , Y ) is the normalized mutual information, log 2 (min( m , n )) is the normalization factor.

[0041] (4) Calculation of MIC value Among all possible mesh partitions, MIC is defined as the normalized mutual information M ( X , Y ):

[0042] This value indicates X and Y The maximum correlation between them under the optimal grid partitioning is normalized to the range of [0,1].

[0043] The MIC coefficient of time-frequency characteristics is as follows Figure 4 As shown. Figure 4 In a, the top three feature variables in terms of importance are min6, av6, and pv6, with MIC values ​​of 0.8751, 0.8613, and 0.8583, respectively, indicating that the three features have a high correlation strength with the tobacco leaf baking stage at the time domain level. Figure 4 In b, the top three feature variables in importance are Dav6, Dme1, and Dme6, and their MIC values ​​are 0.7768, 0.6708, and 0.6702, respectively, indicating that the three features have a high correlation strength with the tobacco leaf baking stage at the frequency domain level. The present invention selects feature variables with a MIC value greater than 0.36 as the optimized feature combination, in which 18 features are extracted in the time domain and 32 features are extracted in the frequency domain, and the total number of features is reduced by 30.56%.

[0044] c5. Sample data set division Before building the model, all data were divided into training set, validation set and external test set. The model was trained with the training set, and then the model parameters were modulated according to the validation set to obtain a more accurate model. Finally, the stability of the model was tested with the external test set. When building the model, 10% of the tobacco gas samples were randomly selected as the external test set, 20% of the samples were selected as the validation set, and the remaining 70% of the samples were used as the training set. The number of samples in the training set, validation set and external test set for detecting the tobacco baking stage model were 378, 108 and 54, respectively.

[0045] c6. Final model determination of model optimization set MIC was used to optimize the features, and the optimized features were combined with SVM, 1DCNN, and LSTM models to build tobacco leaf baking stage discrimination models. The performance of the models is shown in Table 3. The training process is as follows: Figure 5As shown. When the classification model is trained using the optimized electronic nose time domain features, TF-MIC-1DCNN has the best effect, with the accuracy of the training set, validation set and test set being 92.59%, 90.74% and 85.19%, respectively, followed by TF-MIC-SVM and TF-MIC-LSTM. When the classification model is trained using the optimized electronic nose frequency domain features, FF-MIC-1DCNN has the best effect, with the accuracy of the training set, validation set and test set being 92.86%, 92.59% and 92.59%, respectively, followed by FF-MIC-LSTM and FF-MIC-SVM. When the classification model is trained by integrating the optimized time-frequency features, TFF-MIC-1DCNN has the best effect, with the accuracy of the training set, validation set and test set being 92.89%, 92.59% and 92.59%, respectively, followed by TFF-MIC-LSTM and TFF-MIC-SVM. Compared with single features, the accuracy of TFF-MIC-SVM, TFF-MIC-1DCNN and TFF-MIC-LSTM models trained by optimized fusion features has improved. Among them, TFF-MIC-LSTM has the highest improvement of 11.11%, followed by TFF-MIC-SVM with 7.41% and TFF-MIC-1DCNN with 5.55%. This shows that TFF-MIC-1DCNN has higher classification ability in the optimized feature matrix. The structure of the TFF-MIC-1DCNN model is as follows: Figure 3 shown.

[0046] Table 3 Performance of different discrimination models based on MIC optimization features

[0047] Part d: Using the recognition model to detect the tobacco leaf baking stage The detection of tobacco leaf curing stage using the optimal recognition model is performed according to step AE. The time-frequency features after optimization of the external test set are obtained, and the optimal time-frequency features are input into the optimal TFF-MIC-1DCNN discrimination model to obtain the discrimination results of tobacco leaf curing stage. By comparing the accuracy of the real stage and the predicted stage, the discrimination effect of the model is evaluated, and the discrimination accuracy is 92.59%. The confusion matrix is ​​as follows: Figure 6 shown.

[0048] The present invention uses two embodiments to explain the operating procedures of a tobacco leaf baking stage detection method based on the time-frequency MIC optimization features of an electronic nose from the perspectives of the structure and usage of a detection system, the establishment and optimization of a discrimination model, and the testing of the model discrimination effect. It can be seen from the results of the detection using the model that the present invention establishes and optimizes a discrimination model based on the MIC feature optimization method by making full use of the feature information of the time domain and frequency domain space of the electronic nose signal, thereby realizing the classification and discrimination of the tobacco leaf baking stage, and providing technical support and reference for the identification of the tobacco leaf baking stage and the baking stage of other objects.

[0049] The detection of the baking stage of other objects based on the electronic nose technology may be performed with reference to the detection method and detection process proposed in the present invention.

[0050] The above implementation modes are only used to illustrate the present invention and are not used to limit the present invention. Any modifications, equivalent substitutions and improvements made on the basis of the technical essence of the present invention within the spirit and principles of the present invention shall be included in the scope of the present invention. The patent protection scope of the present invention is defined by the claims.

Claims

1. A tobacco leaf baking stage detection method based on the time-frequency MIC optimization characteristics of an electronic nose, characterized in that: The time-frequency features of the electronic nose signals at different baking stages are obtained, and then the time-frequency features are optimized using MIC (maximum mutual information coefficient), and the different baking stages of tobacco leaves are discriminated in combination with the discrimination model, including the following steps: Step 1: Obtain gas samples from the baking room at different baking stages; Step 2: Under room temperature conditions, the electronic nose response signals of gas samples at different baking stages are obtained using the same acquisition parameters; Step 3: preprocessing the signal of the electronic nose sensor array; Step 4: extract the time domain features and frequency domain features in the electronic nose response signal respectively, fuse the extracted time and frequency features, and use MIC to optimize the time and frequency features; Step 5: Divide the optimized time-frequency features into a training set, a validation set, and a test set, and use the divided data sets to train and validate the model, select the best model, and determine the various parameters of the model; Step 6: If the effect of the established detection model meets the requirements, it means that the model is feasible; otherwise, expand the sample data set and repeat steps 1 to 5 until the requirements are met; The specific steps of using the model to detect the tobacco leaf baking stage are as follows: Step A, obtaining gas samples from the baking room at different baking stages; Step B, at room temperature, using the same acquisition parameters to obtain the electronic nose response signals of the gas samples at different baking stages; Step C, preprocessing the signal of the electronic nose sensor array; Step D, respectively extracting the time domain features and frequency domain features in the electronic nose response signal, fusing the extracted time-frequency features, and optimizing the time-frequency features using MIC; Step E: input the optimized time-frequency features into the established detection model, identify the baking room gas sample, and output the detection results.

2. A tobacco leaf baking stage detection method based on the time-frequency MIC optimization characteristics of an electronic nose according to claim 1, characterized in that: When optimizing the time-frequency features of the electronic nose based on MIC, the MIC coefficients between the time domain features, the frequency domain features and the tobacco leaf baking stage are calculated respectively, and then the features are optimized using the threshold segmentation method. The threshold selection is based on the model accuracy improvement of less than 1% or no longer improving. The present invention selects the time-frequency features with a MIC coefficient greater than 0.36 as the preferred feature combination, of which 18 features are extracted in the time domain and 32 features are extracted in the frequency domain, reducing the total number of features by 30.56%.

3. A tobacco leaf baking stage detection method based on the time-frequency MIC optimization feature of an electronic nose according to claim 1, characterized in that: The electronic nose time-frequency characteristics, wherein the time domain characteristics include the maximum value, minimum value, average value, variance, integral value and maximum first-order derivative in the electronic nose response signal; The frequency domain features refer to the mean, standard deviation and median of the approximate coefficients and detail coefficients obtained by wavelet transform of the electronic nose response signal; the electronic nose time-frequency features are arranged in the order of feature extraction, and then feature optimization is performed based on MIC; the tobacco leaf baking stages include the early yellowing stage, the late yellowing stage, the early color fixation stage, the late color fixation stage and the dry tendon stage.

4. A tobacco leaf baking stage detection method based on the time-frequency MIC optimization feature of an electronic nose according to claim 1, characterized in that: To obtain oven gas samples at different baking stages, samples should be drawn at representative locations, such as the top, middle, and bottom points of the oven's geometric center.

5. The method for detecting tobacco leaf baking stages based on the time-frequency MIC optimization characteristics of an electronic nose according to claim 1, characterized in that: The acquisition parameters included a gas flow rate of 2000 ml / min and a reaction time of 2 min at room temperature.

6. The method for detecting tobacco leaf baking stages based on the time-frequency MIC optimization characteristics of an electronic nose according to claim 1, characterized in that: When preprocessing the signal of the electronic nose sensor array, it is necessary to perform baseline correction and smoothing on the original signal, wherein the smoothing method includes but is not limited to the sliding average method, the exponential sliding average method and the SG filtering method.

7. The method for detecting tobacco leaf baking stages based on the time-frequency MIC optimization characteristics of an electronic nose according to claim 1, characterized in that: The established discriminant model has a network structure consisting of 1 input layer, 2 convolutional layers, 2 batch normalization layers, 2 maximum pooling layers and 1 fully connected layer. The input layer uses a standardized method to normalize the data. The convolutional layer parameters are 2×1×16, the maximum pooling parameters are 2:1, the model learning rate is 0.001, and softmax is used as the mapping function.

8. The method for detecting tobacco leaf baking stages based on the time-frequency MIC optimization characteristics of an electronic nose according to claim 1, characterized in that: The detection system used in the method comprises an electronic nose signal acquisition unit, a gas acquisition unit, a profiling gas chamber and a control and analysis unit; the electronic nose signal acquisition unit is composed of a gas sensor array, a conditioning circuit and a communication module, wherein the gas sensor array is located inside the profiling gas chamber, the conditioning circuit supplies power to the gas sensor and performs A / D conversion on the signal, and the communication module adopts USB communication; the gas acquisition unit comprises an air pump, a speed regulator and a gas acquisition bag, the air pump is used for detecting the flow of the gas to be tested and the reducing gas in the system, and the speed regulator is used for regulating the flow rate of the air pump; the profiling gas chamber is printed by a resin material and is covered with an anti-sticking coating and an airflow guide mechanism; the control and analysis unit comprises a PC and a software interface, wherein the software interface is used for the real-time display, storage and discrimination and analysis functions of the electronic nose signal.

Images