Grain grain mildew detector based on gas sensor array

Abstract

The invention discloses a cereal grain mildew detector based on a gas sensor array, the cereal grain mildew detector comprises a volatile matter accurate sensing unit and a volatile matter characteristic analysis unit, the volatile matter accurate sensing unit comprises a gas sensor array, a sensing chamber and a volatile matter sensing channel, and the volatile matter accurate sensing unit is used for accurately acquiring volatile matter characteristics; the volatile matter feature analysis unit comprises a response feature conversion module and a mildew detection module, the response feature conversion module converts response features of the gas sensor array into storable and processable digital signals, and whether the grains are mildewed or not is recognized in real time in combination with a mildew detection model. The method solves the problem of poor sensing data stability caused by the sensitivity consistency of a sensor array, can stably and accurately sense the characteristics of grain grain mildew volatile matters, and analyzes grain grain mildew information in real time in combination with a grain grain mildew detection model.

Description

Technical Field

[0001] The invention relates to a grain mildew detector based on a gas sensitive sensor array, belonging to the technical field of agricultural product detection. Background Art

[0002] Cereals are rich in carbohydrates, proteins, fibers, multiple vitamins and minerals, which can provide energy and balanced nutrition for humans. For a long time, people have mainly pursued high yields of cereals to meet their survival needs. However, with the improvement of living standards, people have higher and higher requirements for the quality of cereals. Color and smell are important manifestations of the safety quality of cereal grains. In the national standard, the classification of cereal quality grades is based on normal safety quality. Mildew is an important factor affecting the safety quality of cereal grains. During the production process, due to the influence of extreme climate such as high temperature and high humidity, cereal grains will be infected by molds such as Aspergillus, Penicillium, Fusarium and Black Mold, and mildew will occur, producing toxic and harmful substances. Among them, deoxynivalenol produced by Fusarium graminearum and aflatoxin produced by Aspergillus flavus are the most common toxic and harmful substances, which seriously affect food security. The growth and metabolic activities of molds in cereal grains will produce some volatile substances, which are important characteristics of cereal mildew detection.

[0003] At present, the detection of volatiles in grain mold is mainly carried out through gas chromatography-mass spectrometry (GC-MS), color sensors and gas sensors. GC-MS is a commonly used laboratory detection method, but it requires sample pretreatment, is time-consuming and labor-intensive, and is complex and professional. Color sensors are sensors based on material color changes. They are highly selective, simple and efficient, but are greatly affected by the environment and the materials are difficult to reuse. Gas sensors are a method of converting the volume fraction of volatiles into electrical signals for detection. They have the advantages of easy integration, high sensitivity and real-time non-destructive detection.

[0004] The composition of grain mold volatiles is complex and the concentration is low. The characteristic signals of volatiles obtained by a single gas sensor are weak and incomplete, making it difficult to accurately identify grain mold. The use of multiple gas sensors to form an array is a key technology to accurately identify complex volatiles. Gu et al. used the PEN2 sensor array (Airsense, Germany) to detect fungal-infected rice. Through ANOVA-partial least squares regression analysis, it was found that sensors S2 (nitrogen oxides), S6 (methane), S7 (sulfide), S8 (alcohols) and S9 (aromatic compounds, sulfide, chloride) had a strong positive correlation with volatiles. Zhao et al. used the Fox 3000 sensor array (Alpha MOS, France) to detect mold-infected wheat. Through the signal radar map, it was found that except for sensors LY2 / LG (oxidative volatiles), LY2 / gCT (hydrogen sulfide) and PA / 2 (ethanol, ammonia, organic amines), other sensors were more sensitive to their volatiles. Leggieri et al. used PEN3 (Airsense, Germany) to identify whether aflatoxin B1 (AFB1) and fumonisin (FB) in corn exceeded the legal limit. 2 The statistical test classified the correlation between sensor signals and mold into grades, and found that sensors W1C (aromatic compounds), W3C (ammonia, aromatic compounds), W5C (alkanes), W1S (methane), and W2S (alcohols, aldehydes, and ketone compounds) were all related to both mycotoxins. These commercial gas sensor arrays obtain volatile characteristics from multiple regions and angles, reducing the uncertainty of information and improving the accuracy and performance of classification and recognition. However, due to the generation of a large amount of redundant data, the reliability of classification and recognition results is reduced. Moreover, the cost is high and it is difficult to develop secondary. Domestic and foreign research institutions have conducted a lot of research on optimizing sensor arrays and reducing data redundancy. Lin et al. used 8 gas sensors to detect samples with different degrees of mold and different percentages of moldy wheat. They removed two redundant sensors through principal component analysis and constructed a BP neural network model. The prediction correlation coefficients for the total number of sample colonies and the percentage of moldy wheat were 0.91 and 0.94, respectively. Rutolo et al. used an array of 11 gas sensors to detect potatoes before and after soft rot, and identified three sensors through principal component analysis, achieving 100% detection accuracy before and after infection. Most of these studies optimized the sensor array from the perspective of reducing the redundancy of sensor data, without considering the consistency of sensitivity of each sensor in the array, resulting in poor stability of each sensor data.

[0005] The sensitivity consistency of the gas sensor array is closely related to the concentration and distribution of volatiles. The sensor chamber is the reaction site between the gas sensor array and volatiles. The chamber shape, sensor array layout, volatile pressure and flow rate will affect the concentration and distribution of volatiles. Marco et al. designed a centrally symmetrical annular array, in which each sensor board is evenly and symmetrically inserted on the main board at the bottom of the chamber. However, this array structure is large in size and has low chamber space utilization, making it difficult to determine the concentration of volatiles, which is not conducive to low-concentration detection. Wu et al. designed an array with four sensors evenly and symmetrically distributed in two installation methods. Through computational fluid dynamics (CFD) simulation, an array and chamber with more uniform pressure were selected, but the concentration of volatiles could not be determined. Wang et al. designed a chamber that imitated the nasal cavity of mammals and a sensor array arranged in a plane. Through CFD analysis, the pressure and flow rate in different chambers were compared to determine the chamber that enables the array to better obtain the characteristics of volatiles, but the uniformity of concentration distribution was not analyzed. Existing research has determined the sensor array and its reaction site (chamber) through theoretical design and CFD simulation, which has improved the sensitivity consistency of the array to a certain extent. However, without analyzing the sensor array distribution (array distribution in the chamber is also an important source of flow resistance) and the volatile flow resistance, it is difficult to further determine the volatile concentration and improve its distribution uniformity. Summary of the invention

[0006] The purpose of the present invention is to provide a grain grain mold detector based on a gas sensor array, which completes the following tasks: (1) Taking wheat as an example, the volatile components and relative contents of its grain mold are determined based on GC-MS, and the gas sensors sensitive to volatiles are selected through response characteristic analysis; (2) The sensor chamber is designed based on gas flow analysis characteristics, and the sensor array layout that can accurately obtain volatile characteristics is determined through CFD simulation; (3) A wheat grain mold detection model is constructed based on the linear discriminant analysis (LDA) method, and a detection device that can identify mold in real time is developed through model integration.

[0007] To achieve the above object, the technical solution adopted by the present invention is: a grain mildew detector based on a gas sensor array, which includes a volatile accurate sensing unit and a volatile feature analysis unit.

[0008] The volatile accurate sensing unit comprises a gas sensor array, a sensing chamber and a volatile sensing path, which is used to accurately obtain the characteristics of the volatiles;

[0009] The volatile feature analysis unit includes a response feature conversion module and a mildew detection module. The response feature conversion module converts the response features of the gas sensor array into a digital signal that can be stored and processed, and combines the mildew detection model to identify whether the grain is mildewed in real time.

[0010] Furthermore, the volatile sensing pathway includes a solenoid valve, an air pump and an air pipe for connection, the sensing chamber is arranged on the air pipe pathway, the gas sensitive sensor array is arranged in the sensing chamber, and the air pump is used to pump the volatiles in the sample into the sensing chamber through the solenoid valve to react with the gas sensitive sensor and then discharge them from the outlet.

[0011] Furthermore, the sensing chamber adopts a gradually expanding tube with a small inlet diameter.

[0012] Furthermore, the pipe diameter of the gradually expanding pipe is set to 8-10 mm, the pipe length is set to 18-62 mm, and the gradually expanding angle is set to 5° to 15°.

[0013] Furthermore, the gas sensor array is arranged in a surface mounting manner in the sensing chamber.

[0014] Furthermore, the response feature conversion module includes a main control board, a sensor board and a data storage card. The microcontroller on the main control board connects to the sensor board through its analog input and IIC interface to obtain the volatile response characteristics and convert them into digital signals. The digital signals are stored in the data storage card and converted into grain mold information in combination with the mold detection model.

[0015] Furthermore, for mold detection, the code conversion tool MATLAB Coder is used to convert the grain kernel mold detection model into C / C++ code and integrate it into the main controller, and then the response characteristics of volatiles are converted into grain kernel mold information in real time through the model.

[0016] Furthermore, the response time of the gas sensor array is the time from when the array detects changes in the concentration of volatile components to when it outputs a stable response value; when CV is less than 10%, it indicates that the data variation is small and the data stability is high, and the response value is considered stable.

[0017] Furthermore, the array response time calculation process is:

[0018] (1) Calculate the coefficient of variation from 1 to 60 seconds. If CV < 10%, the response time is 1 second. If CV > 10%, continue calculating.

[0019] (2) Calculate the coefficient of variation from n to 60 seconds. If CV < 10%, the response time is n seconds. If CV > 10%, continue calculating.

[0020] Furthermore, the mold detection module includes a grain kernel mold detection model constructed based on the linear discriminant analysis method, specifically: taking different mold quantities as response variables and volatile response characteristics as predictor variables, a grain kernel mold detection model is constructed by the linear discriminant analysis method, and 10-fold cross validation is used to evaluate the generalization ability of the classification model.

[0021] The beneficial effects of the present invention are as follows:

[0022] (1) The grain mildew detector of the present invention can stably and accurately sense the volatile characteristics of grain mildew, and analyze the grain mildew information in real time in combination with the grain mildew detection model.

[0023] (2) The present invention proposes a method for constructing a gas sensor array based on sensitivity consistency to solve the problem of poor stability of sensor data caused by sensitivity consistency of the sensor array. Based on the analysis and simulation of the flow characteristics of volatiles, a gradually expanding tube sensing chamber was designed, and a gas sensor array with sensitivity consistency (CV of 0.3083%) was determined.

[0024] (3) The present invention constructs a grain mold detection model based on the linear discriminant analysis method. Based on different mold quantities as response variables and volatile response characteristics as predictor variables, a grain mold detection model is constructed by the linear discriminant analysis method, and the generalization ability of the classification model is evaluated using 10-fold cross validation.

[0025] (4) The wheat grain mildew detector was tested. The recognition accuracy of different mildew types (N, I, II, III, IV, V) of different mildewed wheat (S1: Ningmai 13 inoculated with Fusarium graminearum, S2: Ningmai 13 inoculated with Aspergillus flavus, S3: Yangmai inoculated with Fusarium graminearum, S4: Yangmai inoculated with Aspergillus flavus) was 97.00%, 96.67%, 99.00%, and 99.67%, respectively. According to the grading requirements in the national standard, the classification accuracy of whether different mildewed wheat (S1, S2, S3, S4) is mildewed is 100%. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is the overall design block diagram of the real-time detector for grain mildew.

[0027] Figure 2 It is a schematic diagram of the overall structure of a real-time detector for grain mildew.

[0028] Marked in the figure: 1-control board, 2-debugging and charging interface, 3-SD card, 4-inlet, 5-solenoid valve, 6-button, 7-interaction board, 8-outlet, 9-sensing chamber, 10-sensor array, 11-air pump, 12-lithium battery, 13-sample chamber.

[0029] Figure 3 It is the total ion current chromatogram of moldy wheat (S1~S4) and normal wheat (S5~S6) and their volatiles.

[0030] Figure 4 It is the radar chart of the response characteristics of the gas sensor to the volatiles of moldy wheat.

[0031] Figure 5 It is the energy loss diagram of gas in the pipeline.

[0032] Figure 6 Figure 2 is a diagram of a circular tube sensing chamber (a) and the dimensions of the circular tube sensing chamber with the sensor installed (b).

[0033] Figure 7 This is the CFD simulation process; (a) chamber geometry model, (b) meshing, (c) CFD simulation, and (d) CFD-POST post-processing.

[0034] Figure 8 Simulation of pressure distribution in circular tubes of the same volume but different diameters.

[0035] Fig. 9 It is the local resistance diagram of gas flow.

[0036] Fig.10 It is the pressure and flow velocity distribution of the sudden expansion pipe and the pressure and flow velocity distribution of the gradual expansion pipe.

[0037] Fig.11 It is the geometric model of the sensing chamber for different installation methods of single-row, two-row and four-row gas sensors.

[0038] Fig.12 The pressure (a), flow rate (b) and flow rate distribution diagram (c) of the sensing chamber with a gas sensor.

[0039] Fig.13 are the pressure and flow rate of a gradually expanding pipe; (a) pressure of different pipe diameters and pipe lengths (b) flow rate of different pipe diameters and pipe lengths (c) pressure of different gradually expanding angles (d) flow rate of different gradually expanding angles.

[0040] Fig.14 It is a gas sensor array and a sensing chamber.

[0041] Fig.15 It is the signal processing diagram of gas sensor array.

[0042] Fig.16 is the prepared wheat grain sample.

[0043] Fig.17 The response time calculation process (left) and the response time of 60 sets of data (right).

[0044] Fig.18 It is the CV of the data of different gas sensor channels (response time 1 second).

[0045] Fig.19 is the response value of different gas sensor channels.

[0046] Fig. 20 It is the confusion matrix of the moldy wheat classification model based on the LDA recognition method. DETAILED DESCRIPTION

[0047] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0048] like Figure 1 As shown in the figure, the design of a grain seed mold detector based on a gas sensor array mainly consists of three parts: volatile sensitive analysis, volatile accurate perception and volatile feature analysis, which is used for the classification and identification of normal grains and moldy grains. The volatile sensitive analysis part is mainly used to perceive the volatile characteristics of moldy grains. The composition and relative content of volatiles are analyzed by the GC-MS method, and then the gas sensor sensitive to the volatiles of moldy grains is determined according to the response characteristics of the samples. The volatile accurate perception part is mainly used to accurately obtain the characteristics of volatiles. The volatile concentration distribution is analyzed by gas flow characteristics to construct a gas sensor array with consistent sensitivity. The volatile feature analysis part is mainly used to convert the response characteristics of the gas sensor array into a digital signal that can be stored and processed, and combined with the mold detection model to identify whether the grain is moldy in real time.

[0049] like Figure 2 As shown, a grain seed mildew detector based on a gas sensor array has an overall size of 150mm*135mm*80mm, which is convenient to carry. The volatile sensitive analysis part is mainly composed of gas sensors that are sensitive to volatiles. The volatile accurate perception part is composed of a volatile perception system including a gas sensor array (located on the sensor board), a sensor chamber 9, and a volatile perception path (solenoid valve 4, air pump 11, and an air pipe for connection). The volatile feature analysis part includes response feature conversion (sensor board 10, control board 1 and SD card 3, etc.) and real-time mildew detection (grain seed mildew detection model and interactive board 7). The volatile feature perception process is as follows: First, use a sample container 13 to hold 100g of grain seed sample, and connect the container 13 and the inlet 4 through a hose. Secondly, the air pump 11 pumps the volatiles in the sample into the sensor chamber 9 through the solenoid valve 5 to react with the gas sensor 10 and then discharge it from the outlet 8. The process of mildew detection is as follows: the microcontroller of the main control board 1 connects the sensor board 10 through its analog input and IIC interface to obtain the volatile response characteristics and convert them into digital signals. The digital signals can be stored in the SD card 3 and converted into grain mildew information in combination with the mildew detection model.

[0050] The following is a further description of the overall design of the grain mildew detector:

[0051] 1. Gas sensor selection

[0052] The volatiles of moldy grains include a variety of substances such as alkanes, alcohols, esters and amides. Using multiple gas sensors to form an array is an important method to achieve the perception of volatiles with complex components. Randomly selected naturally moldy wheat grains (sample1~sample4, S1~S4) and normal wheat grains (sample5~sample6, S5~S6) were analyzed by GC-MS for their volatile components and relative contents. The wheat grains and their total ion flow chromatograms are shown in the figure. Figure 3 As shown. By analyzing and comparing the volatile components and contents of the six groups of wheat samples, it can be found that compared with 5-amino-1-methyl-1H-pyrazole-4-carboxamide in the two groups of normal wheat, 3TMS in the four groups of moldy wheat has a decreasing trend, and decamethylcyclopentasiloxane in the four groups of moldy wheat has an increasing trend. N-[4-bromo-n-butyl]-2-piperidone was present in both groups of normal wheat, but not in the four groups of moldy wheat.

[0053] The gas sensors preliminarily screened according to the volatile components are shown in Table 1. These gas sensors are used to sense the volatile characteristics of wheat samples (S1 to S6) and convert them into digital signals. Among them, Mics4514 and SGP gas sensors have two channels, Mics6814 has three channels, and the remaining gas sensors have only one channel. The radar diagram of the response characteristics of the channels of the selected gas sensors in the volatiles of moldy wheat is shown in Figure 4 As shown in the figure, GM102B, GM402B, GM502B, GM512B, GM602B, GM802B, Mics6814, TGS8100 and SGP30 have obvious response characteristics. However, the response values ​​of GM202B, GM302B, GM702B, Mics2714, Mics4514, Mics5524 and Mics5914 in the six groups of samples are less than 1% of their range, and it is difficult for them to sense the characteristics of wheat grain volatiles.

[0054] Table 1 Gas sensors for detecting moldy wheat

[0055]

[0056] 2. Construction of gas sensor array

[0057] The accurate sensing part of volatiles includes a gas sensor array, a sensing chamber, and a volatile sensing path (composed of a solenoid valve, an air pump, and an air pipe for connection). The gas sensor array is the key to sensing volatiles, and the consistency of its sensing sensitivity is an important indicator for achieving accurate sensing. The increase in the concentration of volatiles in the sensing chamber can effectively improve the sensitivity of the detector, while the improvement in the uniformity of its volatile distribution can achieve consistency. The gas sensor array and the sensing chamber will hinder the flow of volatiles, affecting the concentration and distribution of volatiles, and affecting accurate sensing.

[0058] The gas sensor packaging types on the market can be divided into pin-type and patch-type. The patch-type sensor based on micro-electromechanical system (MEMS) is small in size and can design a miniaturized sensing chamber. Under the same flow conditions, the response concentration of the miniaturized sensing chamber is higher and the distribution of multiple gas sensor arrays is more compact, and the response time difference is smaller.

[0059] 2.1 Analysis of volatile flow characteristics

[0060] The flow state of gas can be divided into laminar flow, transitional flow and turbulent flow. The gas flow in the laminar state is orderly and stable, and the gas molecules flow in parallel laminar lines without obvious staggered mixing. The gas flow in the transitional flow state has the characteristics of both laminar flow and turbulent flow, and the gas molecules begin to staggered mix, but the overall flow is still relatively orderly. The gas flow in the turbulent state is chaotic and unstable, and the gas molecules mix and move in a complex and random manner. The Reynolds number Re is used to judge the flow state of the gas, and then the inlet flow velocity of the sensing chamber is calculated.

[0061]

[0062] Among them, ρ is the density of the gas, v is the flow velocity, d is the pipe diameter, and μ is the dynamic viscosity coefficient.

[0063] In addition, the relationship between gas volume and pressure (3) can be deduced from the ideal gas state equation (2). Combining the gas molar concentration formula (4) with (3), it can be deduced that the gas molar concentration is proportional to the pressure. The derivation process is as follows:

[0064] pV=nRT (2)

[0065] Where R is the molar gas constant (approximately 8.31 J / (mol·K)), p is the pressure (Pa), and V is the gas volume (m 3 ), T is temperature (K), and n is the amount of gas substance (mol).

[0066] When the temperature is T n (K), pressure is p n (Pa), then the gas volume is:

[0067]

[0068] The molar concentration of the gas is:

[0069]

[0070] Substituting (3) into (4) we can obtain:

[0071]

[0072] According to formula (4), when the amount of gas n is constant, the smaller the volume of the sensing chamber, the greater the molar concentration of the substance. The uneven distribution of gas pressure is caused by energy loss during gas flow and other reasons. The sensing chamber of a certain volume is divided into several regions of equal volume V0. According to formula (5), the greater the molar concentration of the substance in the V0 region with high pressure, the greater the amount of substance n. Under the same volume, the greater the amount of substance, the greater the concentration of the gas. That is, the pressure is proportional to the concentration of volatiles.

[0073] In summary, the turbulent state can ensure that the complex components of volatiles are fully mixed, which is beneficial to detection, and the flow rate of volatiles at the inlet of the sensing chamber is calculated based on the Reynolds number. The pressure magnitude and distribution in the sensing chamber can characterize the concentration and uniformity of volatiles. In addition, the flow rate can be used to calculate the array sensing time, and the flow characteristics can be used to analyze the flow dead corners. The coefficient of variation (CV) is the ratio of the standard deviation to the mean, which reflects the absolute value of the degree of data dispersion and can be used to analyze the uniformity of gas pressure distribution.

[0074] 2.2 Sensing chamber design

[0075] The uneven distribution of volatiles in the sensor chamber is mainly due to resistance, which leads to energy loss. The energy loss hw of gas flowing in the pipeline is equal to the sum of all along-the-way losses hf and local losses hj, such as Figure 5 As shown. Along-the-line loss refers to the energy loss caused by friction between the gas and the pipe boundary and the gas itself when the gas flows in a straight pipe. Local loss refers to the energy loss of the fluid when the flow changes sharply due to changes in the pipe cross-section. Under the conditions of the same flow rate, the same pipe length and the same pipe material, theoretically, the along-the-line loss of a circular pipe is the smallest. This is because the circular pipe has the smallest flow surface area compared to pipes of other shapes and is subject to less frictional resistance. From the perspective of reducing energy loss, a straight circular pipe is the best. The geometric model of the sensor chamber is designed based on a straight circular pipe, as shown in Figure 6 (a) is shown. The size of the circular tube sensing chamber is determined according to the size of the sensor, as shown in Figure 6(b) shows that the blue (5*5*1.55mm) and yellow rectangular (5*7*1.5mm) geometric models are used to represent gas sensors, the blue one represents GM102B, GM402B, GM502B, GM512B, GM602B, GM802B and the yellow one represents Mics6814 (TGS8100 and SGP30 are smaller in size, so in order to simplify the design, one yellow geometric model is used to represent these two sensors).

[0076] By performing CFD simulation on the sensor chamber model, the pressure (concentration) and flow rate of the gas (volatiles) can be simulated and analyzed. Figure 7 As shown, the CFD simulation process is as follows: ① Use SolidWorks to design the geometric model of the sensing chamber of the same volume with inlet diameters of 5, 6, 7, and 8 mm respectively. ② Import the model through ICEM CFD, divide the boundary areas such as the outlet, inlet and wall, set the max element to 0.3mm and generate the mesh, and export the mesh to the CFD solver format (.msh). ③ Import the mesh file through ansys fluent, set the boundary conditions (select air as the fluid material, the inlet condition is the flow rate calculated according to the Reynolds number, and the outlet condition is the pressure set to 0Pa), select the turbulence model (k-epsilon model), run the simulation and save the results. ④ Use CFD-POST to open the simulation result file, set the field variable (pressure) to be displayed and the surface to display the field variable, generate the pressure distribution cloud map and export the pressure data for calculation. The gas pressure distribution of the same flow rate of gas flowing through the circular tube sensing chamber of the same volume is shown as follows Figure 8 As shown. Under the same flow rate, the smaller the diameter of the sensing chamber, the greater the pressure uniformity of the gas. However, the pressure uniformity of the circular tube sensing chamber is poor (high CV value), and there is a significant pressure drop from the inlet to the outlet, which is not conducive to detection.

[0077] 2.3 Sensing chamber optimization

[0078] In order to ensure the subsequent installation of gas sensors, the sensing chamber is optimized based on the R8mmL62mm round tube. As mentioned above, the smaller the inlet diameter is at the same flow rate, the greater the flow rate and the greater the pressure. Considering that the minimum diameter of the air pump hose is 3mm, the inlet diameters are set to 3, 4, 5, and 6mm respectively. Due to the change in the cross-sectional area from the inlet to the pipeline, the local resistance encountered by the gas during flow cannot be taken seriously. The gas local resistance diagram of different types of pipelines is shown in the figure below: Fig. 9 The shape of the sensing chamber is changed from a round tube to a sudden expansion tube and a gradual expansion tube to reduce energy loss and improve the uniformity of gas pressure distribution.

[0079] The simulation process of the sudden expansion pipe and the gradual expansion pipe is the same as above. The pressure and flow rate are displayed through CFD-POST settings, and the pressure distribution cloud map and flow rate distribution line map are generated and the pressure and flow rate data are exported. The gas pressure distribution cloud map and flow rate distribution line map of the gradual expansion pipe and the sudden expansion pipe are shown in the figure. Fig.10 As shown. From the pressure distribution cloud map, the pressure of the sudden expansion tube with the same inlet diameter is greater than that of the gradual expansion tube, and the pressure CV of the gradual expansion tube is more evenly distributed than that of the sudden expansion tube. From the flow velocity distribution line diagram, the flow velocity of the gradual expansion tube with the same inlet diameter is greater than that of the sudden expansion tube, and the presence of air flow vortices of varying degrees at the inlet of the sudden expansion tube can easily cause flow dead corners. Therefore, choosing a gradual expansion tube with a small inlet diameter (3mm) as the optimized sensing chamber can obtain higher pressure distribution uniformity and ensure secondary detection (no flow dead corners).

[0080] 2.3 Construction of gas sensor array

[0081] The gas sensor array is arranged in the optimized sensing chamber in different mounting modes (surface mount (SM) and built-in mount (BM)), such as Fig.11 As shown in the figure. In the meshing process of the simulation, the sensor surface is meshed as part of the wall surface. In the analysis of the simulation results, the sensor surface is used as the surface of the field variables (pressure and flow rate) to be displayed, and the pressure distribution uniform velocity and flow rate distribution line diagrams and corresponding data files are exported. The simulation analysis results of the gas pressure and flow rate of the sensing chamber with gas sensor are shown in the figure. Fig.12 As shown. Fig.10 and Fig.12 (a), the gas pressure in the sensing chamber with sensor is about 74% lower than that in the sensing chamber without sensor, and the flow rate is also greatly reduced. From the installation point of view, the pressure magnitude and distribution uniformity of the sensing chamber of BM are slightly better than those of SM ( Fig.12 (a)), while the flow rate of the SM sensing chamber is much greater than that of the BM sensing chamber ( Fig.12 (b)). From the velocity distribution line diagram, the BM installation method will lead to different degrees of dead spots in the sensor chamber where the sensor is installed, while the gas flow is smoother under the SM installation method. Since the installation method has little effect on the pressure magnitude and uniformity, choosing the SM installation method to install the gas sensor in the sensor chamber can ensure secondary detection.

[0082] The pipe diameter, length and expansion angle of the gradually expanding tube are all important factors that affect the distribution of volatiles. According to the installation method and array layout of the gas sensor, the pipe diameter is set to 8mm, 9mm, 10mm (the minimum installation diameter is 8mm), the pipe length is set to 18mm, 32mm, 62mm (the minimum pipe lengths for array layouts of 4 symmetrical rows, 2 symmetrical rows, and 1 row, respectively), and the gradually expanding angle is set to 5°, 10°, 15°, 20° (5° to 15° has the smallest local resistance coefficient). The simulation analysis results of the pressure and flow rate of the sensing chamber with different pipe diameters, lengths and gradually expanding angles are shown in the figure. Fig.13 As shown. From the perspective of pipeline diameter, the gas pressure in the sensing chamber with a small pipeline diameter (8mm) is slightly lower and the uniformity is slightly worse, but the flow rate is much faster than that of the sensing chamber with a large diameter, which can ensure rapid detection. From the perspective of pipeline length, the gas pressure uniformity in the sensing chamber with a pipeline length of 32mm is the best. From the perspective of gradual expansion, the gas pressure uniformity in the sensing chamber with a gradual expansion angle of 10° is high and the gradual expansion section is not too long. Since the installation of the sensor in the sensing chamber will still lead to small dead corners, the gradual expansion tube sensing chamber with an inlet diameter of 3mm, a pipeline diameter of 8mm, a pipeline length of 27mm, and a gradual expansion angle of 10° is further optimized. The gas pressure distribution in the optimized sensing chamber is more uniform CV=0.3083%, which can meet the design requirements of sensitivity consistency. According to the flow velocity of the volatiles and the length of the sensing chamber, it can be calculated that the time for the volatiles to flow from the inflow to the outflow of the sensing chamber is approximately 0.03s. The gas sensor array and the sensing chamber are shown in Fig.14 shown.

[0083] 2.4 Gas sensor array signal processing

[0084] The signal processing of the gas sensor array is to convert the response characteristics into digital signals and then combine them with the mold detection model to identify the moldy grains in real time. It includes two parts: response characteristic conversion and real-time identification of moldy grains. Fig.15As shown. The response characteristics of volatiles in the sensor array are mainly manifested as the resistance of the sensitive resistor in the sensor that changes with the concentration and composition of the volatiles. The response characteristics of volatiles can be divided into digital response characteristics (SGP30, highly integrated digital gas sensors) and analog response characteristics (other gas sensors). The main controller collects the response characteristics through the analog input pins and IIC interface and converts them into digital signals (digital response characteristics are converted into volume concentrations, and analog response characteristics are converted into ADC values) for further storage and processing. The real-time mold recognition part mainly uses the code conversion tool MATLAB Coder to convert the grain mold detection model into C / C++ code and integrate it into the main controller, and then converts the response characteristics of volatiles into grain mold information in real time through the model. Arduinomega 2560 is used as the main controller of the detector. It provides 54 digital input / output pins and 16 analog input pins to receive analog signals. It has 256KB flash memory (for program storage) and 8KB static random access memory (for variables and data storage). It can handle large and complex projects and supports 4 serial ports, 1 spi interface and 1 I2C interface.

[0085] 3. Construction of wheat grain mildew detection model based on LDA method

[0086] LDA is a statistical learning method used to solve classification problems. It projects the samples onto a straight line or hyperplane by linearly projecting the classified samples, so that the projection points of samples of the same category are as close as possible and the projection points of samples of different categories are as far apart as possible. The LDA process is as follows:

[0087] For an R n There are m samples in the space, x1, x2, ..., x m , that is, each x is a matrix with n rows, where n i Represents the number of samples belonging to class i. Assuming there are C classes, then

[0088] n1+n2+…+n c =m. (6)

[0089] ① Calculate the mean u of class i samples i And the overall sample mean u:

[0090]

[0091]

[0092] ② Calculate the within-class scatter matrix:

[0093]

[0094] ③ Calculate the within-class scatter matrix:

[0095]

[0096] ④Calculate the generalized eigenvalue problem:

[0097]

[0098] Where w is the projection vector and λ is the eigenvalue.

[0099] ⑤Select the projection direction to project the sample.

[0100] According to the size of the eigenvalue λ, the eigenvectors corresponding to the first k largest eigenvalues ​​are selected as the projection direction. For the sample x, it is projected onto the projection direction to obtain the sample after dimensionality reduction:

[0101] y=w T x(12)

[0102] Through the above steps, the dimension reduction process of LDA can be implemented, and the classification and identification of data can be achieved while retaining the most important features.

[0103] 3.1 Array response characteristics of normal wheat and moldy wheat

[0104] (1) Preparation of wheat samples: A total of 6 groups of wheat samples (Sample 1 to Sample 6, S1 to S6) were prepared. ① Surface sterilization of wheat. After surface sterilization, Ningmai 13 and Yangmai 23 were sterilized to obtain (S5) sterilized Ningmai 13 and (S6) sterilized Yangmai 23. ② Inoculation of wheat mold. Under sterile conditions, the spore suspensions of Fusarium graminearum and Aspergillus flavus were respectively transferred to the surface-sterilized wheat to obtain (S1) Ningmai 13 inoculated with Fusarium graminearum, (S2) Ningmai 13 inoculated with Aspergillus flavus, (S3) Yangmai 23 inoculated with Fusarium graminearum and (S4) Yangmai 23 inoculated with Aspergillus flavus. ③ Classification of wheat mold quantity. The wheat samples inoculated with mold (S1 to S4) were cultured in a constant temperature and humidity chamber at 28°C and 90% RH to achieve mild mold (the total number of mold colonies was determined according to national standards, 4.4*10 4 ~1.2*10 5CFU / g). The mildew content of sterilized wheat was recorded as N, and the mildew content of mildewed wheat (S1 to S4) was recorded as I. ④ The mildly mildewed wheat samples (S1 to S4) were cultured in a constant temperature and humidity chamber at 28°C and 90% RH for n days, and 100g of samples were taken from each group every day to sense the volatile characteristics. The mildew content of these samples was recorded as II, III, IV, and V, respectively. The prepared wheat grain samples were as follows Fig.16 shown.

[0105] (2) Response characteristics of gas sensor array

[0106] Use a gas sensor array to collect volatiles from each sample for 1 minute to obtain 60 data (1Hz). The response time of the gas sensor array refers to the time it takes for the array to detect changes in the concentration of volatile components and output a stable response value. Response time is an important response characteristic that determines whether the instrument can achieve real-time detection. Generally, when CV is less than 10%, it means that the data variation is small and the data stability is high, and the response value is considered stable. The response time calculation process of the array is as follows: Fig.19 (Left), as follows: ① Calculate the coefficient of variation from 1 to 60 seconds. If CV < 10%, the response time is 1 second. If CV > 10%, continue to calculate. ② Calculate the coefficient of variation from n to 60 seconds. If CV < 10%, the response time is n seconds. If CV > 10%, continue to calculate. Through the calculation of the response time, the data of the GM802B and SGP-1 channels of the 12 channels of the selected gas sensor are difficult to stabilize within 1 to 60 seconds. The response time of the 10 channels of the selected gas sensor to detect different moldy wheat volatiles is as follows Fig.17 (Right) (curves of different colors represent different groups of data, a total of 60 groups of data).

[0107] like Fig.18 As shown in the figure, when the response time of different gas sensor channels is 1 second, the CV of the collected data is mostly less than 10%, and the response value is stable. The response time of the data of a few channels is greater than 1 second (the data is marked with a dotted circle, and its response time has been marked in the figure). Among them, TGS8100 (No.9) and SGP-2 (No.10) are small in size and installed in the opened sensor slot (7*5*1.5mm), leaving a large installation gap, which will produce airflow vortex and lead to a longer response time.

[0108] 3.2 Classification model of wheat grain mildew

[0109] The classification model of wheat grain mildew based on LDA was constructed using the Classification Learner APP of MATLAB, and the generalization ability of the classification model was evaluated by cross-validation. LDA uses 6 categories such as different mildew variables (N, I, II, III, IV, V) as response variables, and the response signals of 10 channels as predictor variables. Each channel contains 50 data (11-60s, 1hz), a total of 300 data as predictor variables. Due to the small data set, 10-fold cross-validation was used. The data set was divided into 10 subsets, one of which was used as a test set and the remaining 9 were used as training sets. This process was repeated 10 times, and each subset was used as a test set once. Finally, the results of 10 cross-validations were averaged to obtain the final model performance evaluation.

[0110] The recognition accuracy rates of different moldy wheat (S1, S2, S3, S4) with different moldy wheat (N, I, II, III, IV, V) were 97.00%, 96.67%, 99.00%, 99.67% respectively. The recognition accuracy rates of moldy wheat inoculated with the same mold were 87.83% for wheat inoculated with Fusarium graminearum (S1 & S3) and 89.50% for wheat inoculated with Aspergillus flavus (S2 & S4). According to the requirements of GB 1351-2008 wheat, wheat was divided into two categories: normal and moldy. The classification accuracy rates of whether different moldy wheat (S1, S2, S3, S4) was moldy were all 100%, the classification accuracy rates of whether wheat inoculated with Fusarium graminearum (S1 & S3) was 100%, and the classification accuracy rates of whether wheat inoculated with Aspergillus flavus (S2 & S4) was 99.16%.

[0111] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the above embodiments do not limit the protection scope of the present invention in any form, and all technical solutions obtained by equivalent replacement and the like fall within the protection scope of the present invention. The parts not involved in the present invention are the same as the prior art or can be implemented by the prior art.

Claims

1. A grain mildew detector based on a gas sensor array, characterized in that: Including volatile accurate sensing unit and volatile characteristic analysis unit, The volatile accurate sensing unit comprises a gas sensor array, a sensing chamber and a volatile sensing path, which is used to accurately obtain the characteristics of the volatiles; The volatile feature analysis unit includes a response feature conversion module and a mildew detection module. The response feature conversion module converts the response features of the gas sensor array into a digital signal that can be stored and processed, and combines the mildew detection model to identify whether the grain is mildewed in real time.

2. A grain mildew detector based on a gas sensor array according to claim 1, characterized in that: The volatile sensing pathway includes a solenoid valve, an air pump and an air pipe for connection. The sensing chamber is arranged on the air pipe pathway, the gas sensor array is arranged in the sensing chamber, and the air pump is used to pump the volatiles in the sample into the sensing chamber through the solenoid valve to react with the gas sensor and then discharge them from the outlet.

3. The grain mildew detector based on the gas sensor array according to claim 1, characterized in that: The sensing chamber adopts a gradually expanding tube with a small inlet diameter.

4. The grain mildew detector based on the gas sensor array according to claim 2, characterized in that: The pipe diameter of the gradually expanding pipe is set to 8-10 mm, the pipe length is set to 18-62 mm, and the gradually expanding angle is set to 5° to 15°.

5. A grain mildew detector based on a gas sensor array according to claim 1 or 2, characterized in that: The gas sensor array is arranged in a surface mounting manner in the sensing chamber.

6. The grain mildew detector based on the gas sensor array according to claim 1, characterized in that: The response feature conversion module includes a main control board, a sensor board and a data storage card. The microcontroller on the main control board connects to the sensor board through its analog input and IIC interface to obtain the volatile response characteristics and convert them into digital signals. The digital signals are stored in the data storage card and converted into grain mildew information in combination with the mildew detection model.

7. The grain mildew detector based on the gas sensor array according to claim 6, characterized in that: Mildew detection is to use the code conversion tool MATLAB Coder to convert the grain seed mildew detection model into C / C++ code and integrate it into the main controller, and then use the model to convert the response characteristics of volatiles into grain seed mildew information in real time.

8. The grain mildew detector based on the gas sensor array according to claim 1, characterized in that: The response time of the gas sensor array is the time from when the array detects changes in the concentration of volatile components to when it outputs a stable response value; when the CV is less than 10%, it means that the data variation is small and the data stability is high, and the response value is considered stable.

9. The grain mildew detector based on the gas sensor array according to claim 8, characterized in that: The array response time calculation process is: (1) Calculate the coefficient of variation from 1 to 60 seconds. If CV < 10%, the response time is 1 second. If CV > 10%, continue calculating. (2) Calculate the coefficient of variation from n to 60 seconds. If CV < 10%, the response time is n seconds. If CV > 10%, continue calculating.

10. The grain mildew detector based on gas sensor array according to claim 1, characterized in that: The mildew detection module includes a grain mildew detection model constructed based on the linear discriminant analysis method, specifically: taking different mildew amounts as response variables and volatile response characteristics as predictor variables, a grain mildew detection model is constructed by the linear discriminant analysis method, and a 10-fold cross validation is used to evaluate the generalization ability of the classification model.

Images