Tea quality identification method, system and device based on gas

Abstract

The invention discloses a tea quality identification method, system and device based on gas, and relates to the technical field of gas identification, and the method comprises the following specific steps: obtaining sensor array data; inputting the sensor array data into a pre-trained classification model to obtain a classification result; the classification model is constructed by multilayer convolution, and a grouping convolution mixed attention mechanism is introduced. According to the method, odor molecules released by the tea leaves are directly detected through the gas sensor, physical or chemical treatment on the tea leaves is not needed, and non-destructive evaluation on the quality of the tea leaves is achieved; according to the invention, based on the tea gas classification network obtained by combining GCHAM, depth classifiable convolution and elastic network regularization, the precision and reliability of tea quality identification are significantly improved.

Description

Technical Field

[0001] The present invention relates to the field of gas identification technology, and more particularly to a method, system and device for identifying tea quality based on gas. Background Art

[0002] Tea, as an ancient and popular drink, enjoys a high reputation around the world. The quality of tea not only affects consumers' drinking experience, but is also closely related to its inherent organic compound components. These compounds include but are not limited to tea polyphenols, caffeine, amino acids and trace elements, which play a key role in the taste, aroma and potential health benefits of tea.

[0003] However, in the tea market, some unscrupulous merchants, in pursuit of high profits, pass inferior tea as good tea, or even mix other substances to impersonate high-quality tea. This behavior seriously infringes on the legitimate rights and interests of consumers and disrupts the normal order of the market. Therefore, it is particularly urgent to develop methods and technologies that can quickly and accurately identify the quality of tea.

[0004] In recent years, electronic nose technology and intelligent pattern recognition algorithms have been widely used in the field of food quality testing, especially for the evaluation of tea quality. The electronic nose simulates the working principle of the human olfactory organ and uses a series of metal oxide semiconductor (MOS) gas sensors to capture and analyze the odor molecules released by the sample. Then, the collected data is processed and classified through specific pattern recognition algorithms, such as principal component analysis (PCA), linear discriminant analysis (LDA) or artificial neural network (ANN), so as to achieve non-destructive evaluation of tea quality.

[0005] Despite this, existing electronic nose systems still have many limitations. First, such devices are often bulky and costly, requiring professional technicians to operate and maintain, which limits their application in daily consumer scenarios. Secondly, traditional electronic nose systems mostly use dynamic headspace analysis, a process that involves complex mechanical components (such as micropumps, mass flow controllers, etc.), resulting in poor overall portability of the equipment, which is not conducive to field testing. In addition, large-scale gas measuring instruments usually generate a large amount of raw data, which requires that subsequent data processing must rely on advanced algorithm support, further increasing the technical threshold and economic burden of use. Therefore, how to develop a portable and fast method, system and device for identifying tea quality is a problem that technicians in this field urgently need to solve. Summary of the invention

[0006] In view of this, the present invention provides a gas-based tea quality identification method, system and device to overcome the above-mentioned defects.

[0007] In order to achieve the above object, the present invention adopts the following technical solution:

[0008] A method for identifying tea quality based on gas, the specific steps are:

[0009] Acquire sensor array data;

[0010] The sensor array data is input into a pre-trained classification model to obtain a classification result; the classification model is constructed by multi-layer convolution and a grouped convolution hybrid attention mechanism is introduced.

[0011] Optionally, the step of acquiring the sensor array data is:

[0012] Collect gas data of samples to be identified based on multiple sensors to obtain raw data;

[0013] The raw data is intercepted according to a preset threshold to obtain the sensor array data.

[0014] Optionally, the classification model includes a layer of two-dimensional convolution and multiple layers of depth-separable convolution.

[0015] Optionally, the grouped convolutional hybrid attention mechanism includes a channel attention branch, a spatial attention branch and a feature fusion module, the channel attention branch generates a channel attention map based on multiple convolution modules; the spatial attention branch generates a spatial attention map based on the grouping module and the spatial attention module; the feature fusion module performs feature fusion on the channel attention map and the spatial attention map to obtain a fused feature.

[0016] Optionally, the expression of the grouped convolutional hybrid attention mechanism is:

[0017] F'=(F×M C (F)+F×M S (F)) / 2;

[0018] Where F is the given feature map; M C (F) is the channel attention map; M S (F) is the spatial attention map.

[0019] Optionally, the expression of the channel attention map is:

[0020] M C (F) = σ(Conv1d(F,k));

[0021] Where M C (F) is the channel attention map; σ is the sigmoid function; Conv1d is a one-dimensional convolution; F is the given feature map; k is the convolution kernel size.

[0022] Optionally, the expression of the spatial attention map is:

[0023] M S (F) = σ(Conv2d(GAP C (F),k);

[0024] Where M S (F) is the spatial attention map; σ is the sigmoid function; Conv2d is the two-dimensional convolution; GAP C is the global average pooling; F is the given feature map; k is the convolution kernel size.

[0025] A tea quality identification system based on gas, comprising:

[0026] A data acquisition module, used to obtain sensor array data;

[0027] The classification module is used to input the sensor array data into a pre-trained classification model to obtain a classification result; the classification model is constructed by multi-layer convolution and introduces a grouped convolution hybrid attention mechanism.

[0028] Optionally, the data acquisition module includes a plurality of tin dioxide series sensors.

[0029] A gas-based tea quality identification device comprises: a collection end and a processing end, wherein the collection end is integrated with a plurality of sensors for collecting gas data of a sample to be identified; and the processing end stores a classification model for performing grade classification according to the gas data using the classification model.

[0030] It can be seen from the above technical solutions that the present invention provides a method, system and device for identifying tea quality based on gas, which have the following beneficial effects compared with the prior art:

[0031] 1. Realize fast non-destructive testing: The gas sensor can directly detect the odor molecules released by tea leaves, without the need for physical or chemical treatment of the tea leaves, thus realizing non-destructive evaluation of tea quality. This can not only maintain the integrity and original characteristics of the tea leaves, but also quickly conduct quality testing in various links such as picking, processing, storage and sales, thus improving detection efficiency and accuracy.

[0032] 2. Portability and low cost: The device is compact and portable, which reduces the use of complex mechanical components in traditional electronic nose systems, reducing overall cost and power consumption. At the same time, by optimizing the sensor array and data processing algorithm, the amount of raw data is reduced, and the demand for advanced algorithms and computing resources is reduced, making the system easier to use in daily consumer scenarios and field test environments.

[0033] 3. Improve recognition accuracy and reliability: The tea gas classification network obtained by combining GCHAM, deep classifiable convolution and elastic network regularization in the present invention significantly improves the accuracy and reliability of tea quality recognition;

[0034] 4. Improve user experience and satisfaction: By quickly and accurately identifying the quality of tea, consumers can more easily choose tea products that suit their tastes and health needs, thereby improving consumer experience and satisfaction; in addition, it can also provide tea manufacturers with scientific quality control methods to help them improve product quality and market competitiveness. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0036] Figure 1 A schematic diagram of the structure of the detection rod provided by the present invention;

[0037] Figure 2 A schematic diagram of the method provided by the present invention;

[0038] Figure 3 A schematic diagram of the flow chart of the method for acquiring sensor array data provided by the present invention;

[0039] FIG4(a) is a schematic diagram of the response curve of the gas sensor provided by the present invention when detecting third-grade tea leaves; FIG4(b) is a schematic diagram of the response curve of the gas sensor provided by the present invention when detecting second-grade tea leaves; FIG4(c) is a schematic diagram of the response curve of the gas sensor provided by the present invention when detecting first-grade tea leaves; FIG4(d) is a schematic diagram of the response curve of the gas sensor provided by the present invention when detecting premium-grade tea leaves;

[0040] Figure 5 A schematic diagram of the tea quality grades provided by the present invention;

[0041] Figure 6 A schematic diagram of the structure of the grouped convolutional hybrid attention mechanism provided by the present invention;

[0042] Figure 7 A schematic diagram of the structure of the spatial attention branch provided by the present invention;

[0043] In the figure, 1 is the rod body; 2 is the base. DETAILED DESCRIPTION

[0044] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0045] On the one hand, this embodiment discloses a gas-based tea quality identification device, a collection end and a processing end. The collection end is integrated with multiple sensors for collecting gas data of samples to be identified; the processing end stores a classification model for grade classification based on the gas data using the classification model.

[0046] Furthermore, in this embodiment, the collection end adopts a detection rod structure, the structure is as follows Figure 1 As shown, the rod body 1 and the base 2 are fixedly connected. The rod body 1 is a hollow cylinder, and a plurality of holes are arranged at the end away from the base 2 for arranging sensors. The base 2 is used for arranging a microcontroller. Six MQ series tin dioxide sensors (MQ2, MQ3, MQ4, MQ5, MQ7, and MQ9) are connected inside the detection rod through a development board to detect methane, propane, butane, ethanol, isobutane, carbon monoxide, combustible gas, etc. The specific parameters are shown in Table 1, and an Arduino Uno is used. The R3 microcontroller system controls the sensor to collect data (obtain gas concentration data from the sensor, perform algorithm analysis to detect gas leakage or abnormal concentration, and trigger the alarm system or display the results). The six sensor analog output pins are respectively connected to the six analog input pins of the microcontroller, and are fixed on the inner wall of the rod body 1 in two rows. The connection between the detection rod and the processing end can be wired, wireless (Bluetooth connection, wifi connection) and other methods. It is connected to the processing end to collect and detect the gas volatile concentration of tea leaves. The pre-processed data is detected using a classification model. The classification model is a high-performance tea gas classification network proposed by combining GCHAM with deep classifiable convolution and elastic network regularization to identify the quality of tea leaves. In this embodiment, the processing end is a PC end, using an NVIDIA GeForce RTX 4060Laptop GPU processor with 8gb of memory, and the program runs in the pycharm software under the PyTorch2.0.1 underlying framework. Python3.9 is the programming language.

[0047] Table 1 Parameters of MQ sensor

[0048]

[0049] Furthermore, the working principle is as follows: (1) Gas collection: Obtain gas samples, which can be odor molecules in the air, or gases from chemical reactions, biological processes or other sources. (2) Sensor detection: The collected gas samples enter the sensor unit. (3) Result output: The identified odor or gas information is output in the form of a digital signal or other form for the user to further analyze, display or record. These output results can be information such as the type, concentration, relative intensity of the odor. The detection rod structure design used in this embodiment is lightweight and compact, easy to carry and operate, not limited by space, and usually has certain customization functions. Users can choose specific sensor models or configure different sensor combinations according to their own needs to meet specific detection needs. In addition, the "probe rod" type gas sensor has multiple gas detection functions and can monitor multiple different types of gases at the same time, thereby improving the applicability and flexibility of the sensor.

[0050] On the other hand, this embodiment discloses a method for identifying tea quality based on gas. Figure 2 As shown, the specific steps are:

[0051] Step 1: Obtain sensor array data;

[0052] Step 2: Input the sensor array data into the pre-trained classification model to obtain the classification result; the classification model is constructed by multi-layer convolution, and the grouped convolution hybrid attention mechanism is introduced.

[0053] In one embodiment, the step of acquiring sensor array data is as follows: Figure 3 As shown, specifically:

[0054] Step 11: Collect gas data of the sample to be identified based on multiple sensors to obtain raw data;

[0055] Step 12: intercept the original data according to the preset threshold to obtain sensor array data.

[0056] Furthermore, during the collection, the probe rod is placed into the bulk tea leaves, and the depth of placement should be sufficient for the tea leaves to completely cover all sensors of the probe rod. The collected raw data is saved in a csv file format; 50s-600s of data are intercepted as input data for the model.

[0057] In one embodiment, the sensor is powered on for one hour for aging and preheating before data collection.

[0058] In one embodiment, the classification model includes one layer of two-dimensional convolution and multiple layers of depthwise separable convolution.

[0059] Furthermore, the classification model is a deep neural network with five layers of convolution, including one layer of ordinary two-dimensional convolution and four layers of depth-separable convolution, to reduce the number of model parameters. The specific parameters are shown in Table 2. Then, the GCHAM attention mechanism is used to enhance the effective feature extraction capability of the model, specifically:

[0060] Sample collection:

[0061] Taking green tea as an example object, the sensor array data steps are as follows: 1. Power on the sensor for one hour for aging and preheating before collecting data. 2. In order to prevent the sensor from directly contacting the tea sample and affecting the sensor performance, a nylon filter with a filtration accuracy of 250 mesh is used. It can effectively filter the dust in the tea sample while not affecting the volatile gas in the tea. Contact with the sensor to generate a response signal. 3. Take about 100g of tea leaves and place it in a 1L container. Insert the detection rod into the tea sample until the sensor is completely buried in the tea. The sampling time is 15 minutes and the sampling frequency is 1hz. After each sampling is completed, take out the sensor and let it stand in the air for 15 minutes to allow the sensor response to return to the baseline level. The response curves of the gas sensors are respectively Figure 4(a)-4(d) As shown in the figure, the data collection process lasted for 25 days, 50 sets of data were collected for each grade of tea, and a total of 200 tea gas data samples were collected. 4. In terms of data processing, since the data collected by the gas sensor in the early stage contained a lot of noise, the data from 50s to 600s was intercepted as the input data of the classification model.

[0062] In this embodiment, tea leaves are divided into four grades, namely, special grade, first grade, second grade and third grade. The specific information of the samples is shown in Table 3 and Figure 5 As shown, in Figure 5 In the figure, from left to right they are grade three, grade two, grade one and special grade.

[0063] Quality identification:

[0064] First, the sample data is reshaped into a (1,6,550) tensor, where 6 is the number of sensors and 550 is the number of sampling points for each sensor. Ordinary two-dimensional convolution is used to map the original feature vector to a higher-dimensional feature space. Then, a four-layer depth-separable convolution is used to extract the deep features of gas information, reduce parameters and calculations, and improve the efficiency and performance of the network. GCHAM is used to suppress the influence of noise and irrelevant information, helping the model to dynamically adjust the features and regions of interest, thereby enhancing the representation of effective features. Then, an average pooling layer is used to compress the output features and reduce FC parameters. Finally, FC is used to achieve nonlinear mapping and classification.

[0065] Table 2 Gas classification model parameters

[0066]

[0067] Table 3 Sensory evaluation of four different grades of green tea

[0068]

[0069] The classification model uses cross entropy as the loss function and adds elastic network regularization, where the L1 and L2 norm regularization coefficients and λ1 and λ2 are set to 0.0001 and 0.001 respectively after multiple attempts. The AdamW algorithm is selected for gradient training, and the learning rate decay strategy is adopted. The learning rate decays to 1 / 2 of the original every 50 epochs. The initial learning rate is set to 0.0001, and the maximum number of training rounds is set to 200. The total number of samples is 200, and the training set and test set are divided into a ratio of 7:3. The accuracy, precision, recall rate and F1 score are introduced as evaluation indicators of network performance, which are expressed as:

[0070]

[0071] In the formula, Accuracy is the correct rate; Precision is the accuracy; Recall is the recall rate; TP is the true positive; FP is the false positive; TN is the true negative; FN is the false negative.

[0072] To avoid randomness, the model was run 10 times, and the accuracy, precision, recall, and F1-score of these 10 results were averaged as the final classification result to evaluate the model performance, and the standard deviation was calculated to evaluate the model stability.

[0073] Furthermore, the experimental results show that the method proposed in this embodiment has better recognition performance, with a recognition accuracy of 95.33%, a precision of 95.47%, a recall rate of 95.92%, and an F1 score of 95.52%, which proves the effectiveness of the attention mechanism and model proposed in this embodiment. Combining the portable tea odor detection stick system proposed in this embodiment with the tea gas classification network can be used as a convenient and efficient non-destructive testing technology for tea quality evaluation.

[0074] In one embodiment, a grouped convolutional hybrid attention mechanism (i.e., a GCHAM attention mechanism) includes a channel attention branch, a spatial attention branch, and a feature fusion module. The channel attention branch generates a channel attention map based on multiple convolution modules; the spatial attention branch generates a spatial attention map based on the grouping module and the spatial attention module; the feature fusion module performs feature fusion on the channel attention map and the spatial attention map to obtain a fused feature.

[0075] Furthermore, the gas sensor array data is collected by multiple sensors at the same time. These sensors may have different response characteristics and noise levels to different gases, which results in the gas sensor array outputting deep gas information, which contains effective information, redundant information, and interference information. Therefore, an effective method is needed to reduce the impact of redundant information and interference information on model performance. This embodiment proposes a grouped convolution hybrid attention mechanism that combines channel attention and spatial attention to help the model dynamically adjust the features and regions of interest, effectively suppress noise and irrelevant information, and enhance the model's feature representation capabilities. Its structure is as follows: Figure 6 shown.

[0076] In one embodiment, the channel attention branch compresses the channel features of tea gas information into one value through a global average pooling operation without considering spatial information. Then, the local cross-channel interaction of gas information is realized through effective interactive convolution. Effective interactive convolution is a one-dimensional convolution method, which is specifically expressed as:

[0077] M C (F) = σ(Conv1d(F,k)) (5);

[0078] Where M C (F)∈R 1×1×C is the generated channel attention map; σ is the sigmoid function; Conv1d is a one-dimensional convolution; F∈R C×H×W is a given feature map; k is the convolution kernel size, which is proportional to the channel dimension C, indicating that when capturing local cross-channel interaction information, only the relationship between each channel and its k adjacent channels is considered. The value of k is determined by the following formula:

[0079]

[0080] Wherein, B represents the number of channels, γ and b are hyperparameters, which are set to 2 and 1 respectively in this embodiment, so the value of k is 3.

[0081] In one embodiment, the structure of the spatial attention branch is as follows Figure 7 As shown, the grouping features are combined with spatial attention to enhance the feature representation. Spatial attention can automatically focus on more important sensor positions and ignore noise and irrelevant information, thereby improving the quality of feature representation.

[0082] Furthermore, for a given feature map F∈R C×H×W , F is divided into G groups in the channel dimension, that is, F = [F1, F2, ..., F G], where specific semantic responses are gradually captured by each sub-feature F during the training process. Then, each sub-feature is pooled in parallel using global average pooling in the channel dimension, and the spatial relationship of the features is used to generate corresponding spatial attention for each sub-feature, and they are connected to construct an effective spatial descriptor F s Next, for F s Apply 3×3 convolution to prevent coefficients from being different between samples.

[0083] Amplitude bias, each group introduces a pair of parameters to move the weight values ​​of different groups. Finally, F s Use the sigmoid activation function to generate the spatial attention map M S (F)∈R 1×H×W . Its specific expression is:

[0084] M S (F) = σ(Conv2d(GAP C (F),k) (7);

[0085] In the formula, σ is the sigmoid function; Conv2d is the ordinary two-dimensional convolution; GAP C is the global average pooling; k is the convolution kernel size.

[0086] In one embodiment, for the derived channel attention map and spatial attention map, each attention map first acts on the original feature to obtain two refined feature maps. Then, the two features are fused using element-wise sum. Finally, a new attention module GCHAM is implemented, and its final expression is:

[0087] F'=(F×M C (F)+F×M S (F)) / 2 (8).

[0088] On the other hand, this embodiment also discloses a tea quality identification system based on gas, which applies the above method and includes:

[0089] A data acquisition module, used to obtain sensor array data;

[0090] The classification module is used to input the sensor array data into a pre-trained classification model to obtain the classification results; the classification model is constructed by multi-layer convolution and introduces a grouped convolution hybrid attention mechanism.

[0091] In one embodiment, the data acquisition module includes a plurality of tin dioxide series sensors.

[0092] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0093] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for identifying tea quality based on gas, characterized in that: The specific steps are: Acquire sensor array data; The sensor array data is input into a pre-trained classification model to obtain a classification result; the classification model is constructed by multi-layer convolution and a grouped convolution hybrid attention mechanism is introduced.

2. A method for identifying tea quality based on gas according to claim 1, characterized in that: The steps of acquiring the sensor array data are as follows: Collect gas data of samples to be identified based on multiple sensors to obtain raw data; The raw data is intercepted according to a preset threshold to obtain the sensor array data.

3. The method for identifying tea quality based on gas according to claim 1, characterized in that: The classification model includes a layer of two-dimensional convolution and multiple layers of depth-wise separable convolution.

4. The method for identifying tea quality based on gas according to claim 1, characterized in that: The grouped convolution hybrid attention mechanism includes a channel attention branch, a spatial attention branch and a feature fusion module, and the channel attention branch generates a channel attention map based on multiple convolution modules; The spatial attention branch generates a spatial attention map based on the grouping module and the spatial attention module; The feature fusion module performs feature fusion on the channel attention map and the spatial attention map to obtain a fused feature.

5. The method for identifying tea quality based on gas according to claim 4, characterized in that: The expression of the grouped convolutional mixed attention mechanism is: F'=(F×M C (F)+F×M S (F)) / 2; Where F is the given feature map; M C (F) is the channel attention map; M S (F) is the spatial attention map.

6. The method for identifying tea quality based on gas according to claim 4, characterized in that: The expression of the channel attention map is: M C (F)=σ(Conv1d(F,k)); Where M C (F) is the channel attention map; σ is the sigmoid function; Conv1d is a one-dimensional convolution; F is the given feature map; k is the convolution kernel size.

7. The method for identifying tea quality based on gas according to claim 4, characterized in that: The expression of the spatial attention map is: M S (F)=σ(Conv2d(GAP C (F),k); Where M S (F) is the spatial attention map; σ is the sigmoid function; Conv2d is the two-dimensional convolution; GAP C is the global average pooling; F is the given feature map; k is the convolution kernel size.

8. A gas-based tea quality identification system, characterized in that: include: A data acquisition module, used to obtain sensor array data; A classification module, used for inputting the sensor array data into a pre-trained classification model to obtain a classification result; The classification model is constructed by multi-layer convolution, and a grouped convolution hybrid attention mechanism is introduced.

9. A gas-based tea quality identification system according to claim 8, characterized in that: The data acquisition module includes a plurality of tin dioxide series sensors.

10. A gas-based tea quality identification device, characterized in that: include: A collection end and a processing end, wherein the collection end is integrated with multiple sensors for collecting gas data of the sample to be identified; The processing end stores a classification model for performing grade classification according to the gas data using the classification model.

Images