WO2019085369A1 - 一种电子鼻仪器和烟草与烟草制品感官质量评价方法 - Google Patents

一种电子鼻仪器和烟草与烟草制品感官质量评价方法 Download PDF

Info

Publication number
WO2019085369A1
WO2019085369A1 PCT/CN2018/079389 CN2018079389W WO2019085369A1 WO 2019085369 A1 WO2019085369 A1 WO 2019085369A1 CN 2018079389 W CN2018079389 W CN 2018079389W WO 2019085369 A1 WO2019085369 A1 WO 2019085369A1
Authority
WO
WIPO (PCT)
Prior art keywords
tobacco
seconds
solenoid valve
way solenoid
cigarette
Prior art date
Application number
PCT/CN2018/079389
Other languages
English (en)
French (fr)
Inventor
高大启
宋佳敏
张小勤
Original Assignee
高大启
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Priority claimed from CN201711054914.6A external-priority patent/CN107807199B/zh
Priority claimed from CN201711060781.3A external-priority patent/CN107843695B/zh
Application filed by 高大启 filed Critical 高大启
Priority to US16/651,722 priority Critical patent/US10895560B2/en
Publication of WO2019085369A1 publication Critical patent/WO2019085369A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N33/00—Investigating or analysing materials by specific methods not covered by groups G01N1/00 - G01N31/00
    • G01N33/0004—Gaseous mixtures, e.g. polluted air
    • G01N33/0006—Calibrating gas analysers
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N27/00—Investigating or analysing materials by the use of electric, electrochemical, or magnetic means
    • G01N27/02—Investigating or analysing materials by the use of electric, electrochemical, or magnetic means by investigating impedance
    • G01N27/04—Investigating or analysing materials by the use of electric, electrochemical, or magnetic means by investigating impedance by investigating resistance
    • G01N27/12—Investigating or analysing materials by the use of electric, electrochemical, or magnetic means by investigating impedance by investigating resistance of a solid body in dependence upon absorption of a fluid; of a solid body in dependence upon reaction with a fluid, for detecting components in the fluid
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N31/00—Investigating or analysing non-biological materials by the use of the chemical methods specified in the subgroup; Apparatus specially adapted for such methods
    • G01N31/12—Investigating or analysing non-biological materials by the use of the chemical methods specified in the subgroup; Apparatus specially adapted for such methods using combustion
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N33/00—Investigating or analysing materials by specific methods not covered by groups G01N1/00 - G01N31/00
    • G01N33/0001—Investigating or analysing materials by specific methods not covered by groups G01N1/00 - G01N31/00 by organoleptic means
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N33/00—Investigating or analysing materials by specific methods not covered by groups G01N1/00 - G01N31/00
    • G01N33/0004—Gaseous mixtures, e.g. polluted air
    • G01N33/0009—General constructional details of gas analysers, e.g. portable test equipment
    • G01N33/0027—General constructional details of gas analysers, e.g. portable test equipment concerning the detector
    • G01N33/0031—General constructional details of gas analysers, e.g. portable test equipment concerning the detector comprising two or more sensors, e.g. a sensor array
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N33/00—Investigating or analysing materials by specific methods not covered by groups G01N1/00 - G01N31/00
    • G01N33/0004—Gaseous mixtures, e.g. polluted air
    • G01N33/0009—General constructional details of gas analysers, e.g. portable test equipment
    • G01N33/0027—General constructional details of gas analysers, e.g. portable test equipment concerning the detector
    • G01N33/0036—General constructional details of gas analysers, e.g. portable test equipment concerning the detector specially adapted to detect a particular component
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N33/00—Investigating or analysing materials by specific methods not covered by groups G01N1/00 - G01N31/00
    • G01N33/0098—Plants or trees
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N35/00—Automatic analysis not limited to methods or materials provided for in any single one of groups G01N1/00 - G01N33/00; Handling materials therefor
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N1/00—Sampling; Preparing specimens for investigation
    • G01N1/02—Devices for withdrawing samples
    • G01N1/22—Devices for withdrawing samples in the gaseous state
    • G01N1/24—Suction devices

Definitions

  • Just evaluation refers to the sensory identification of the quality and function of tobacco and tobacco products and tobacco flavors in a specific environment using specific organs such as the mouth, nose, and throat of a person, that is, the intrinsic quality of the cigarette - the sensory quality index.
  • the quality of tobacco and tobacco products is determined by human sensory evaluation. Compared with flavoring substances such as wine, tea, and food, tobacco and tobacco products are unique in that quality evaluation depends entirely on human senses. So far, there is no practical physical and chemical indicator detection and analysis method.
  • the industry standard YC/T138 stipulates that sensory evaluators of tobacco and tobacco products are classified into three types: primary appraisers, preferred appraisers and experts. After special selection and training, the basic conditions that the appraisers should meet are: (a) possessing expertise in tobacco and tobacco products; (b) having sensory appreciating interests and experience; (c) being healthy and having no sensory aspects Defects; (d) personal sensitivity is normal, sensitivity is consistent with each other; (e) physiological state is normal during the evaluation; (f) no spicy food such as spicy food, no alcohol consumption; (g) no obvious body odor Do not use scented cosmetics. To prevent sensory fatigue and discomfort, the number of consecutive appraisers for a single appraisers should generally be controlled within 25/day. The evaluation should be carried out in a quiet, non-interfering, ventilated, odor-free chamber.
  • the tar contains a cancer inducing agent and a cancer promoting agent component such as benzo[ réelle]pyrene. Not only that, phenols, alcohols, acids, aldehydes, CO, HCN, NH 3 and other substances in cigarette smoke are toxic and carcinogenic.
  • the sensory evaluation of the intrinsic quality of cigarettes generally adopts the “integral circulation evaluation method”: the sucker inhales the smoke into the oral cavity, and then swallows it through the throat and then slowly discharges it from the nasal cavity, using all the sucking sensory organs. Conduct an evaluation. Due to the sensitivity of the sensory organs and the different levels of capture of various indicators, differences in physiological and psychological conditions, and differences in environmental conditions, the differences in the description and judgment of the evaluation results between the appraisers are sufficient, which fully indicates that the sensory appraisal method is very Big limitations.
  • YC/T415-2011 the main purpose of the industry standard "Tomato-in-the-sensory evaluation method" YC/T415-2011 is the tobacco production process. Process setting optimization, equipment process parameter optimization and product quality control.
  • Each indicator sets a total of 9 integer values from 1 to 9.
  • the degree of change in aroma style depends on changes in the three characteristics of aroma, smoke, and mouthfeel.
  • the main purpose of YC/T415 is to investigate the effects of changes in process and work sections or changes in process parameters on the sensory quality of tobacco products, as well as fluctuations in product quality.
  • the standard adopts the comparative evaluation method, and the sample before the process or the section change or the normal production is used as the control sample to evaluate the sample quality after the process or the section change, that is, the degree of the aroma style change and the processing status of the process section.
  • the results of industrial use show that the evaluation index of YC/T415 can comprehensively reflect the style and quality characteristics of tobacco leaves. At present, most cigarette companies in China use this standard for raw material quality evaluation, new product development, cigarette formula replacement, and tobacco product quality inspection.
  • YC/T530-2015 Industry standard "Tobacco flue tobacco quality style sensory evaluation method" YC/T530-2015, according to the quality requirements of Chinese cigarettes, focusing on the evaluation of the quality style characteristics of flue-cured tobacco leaves.
  • the standard follows the YC/T415 and the relevant evaluation indicators of the Chinese cigarette style sensory evaluation method YC/T497-2014, combined with the quality characteristics of flue-cured tobacco leaves, and carries out the refinement of style characteristics and quality characteristics.
  • the standard classifies the two indicators of cleanness and sweetness into the “aftertaste” index, and proposes 16 kinds of fragrance, 3 aroma states and 9 types of impurity, which increase the degree of softness and roundness.
  • the index retains the characteristics of aroma, aroma, permeability, strength, concentration, miscellaneous, irritating, dryness and softness, and tries to reflect the style characteristics of tobacco more comprehensively.
  • the main contents of the three current tobacco industry standards, YC/T497, YC/T415 and YC/T138, are the quality evaluation of tobacco, tobacco-in-process and tobacco manufactured goods, relying on human senses, emphasizing human feelings and experiences.
  • the internal quality indicators such as aroma, coordination, miscellaneous, irritating, aftertaste, energy, and irritability are directly or indirectly related to olfactory sensation.
  • the first smoke can not be used as a basis for judgment, because the first smoke is very susceptible to fire odor, smoke flavor, initial combustion state and other factors.
  • the “suction amount” of the flue gas is an important factor affecting the accuracy of the judging judgment.
  • the evaluators are not machines, they can only control and maintain the same amount of smoke and smoke between individuals or each other. According to the national standard "Standards and Standard Conditions for Smoking Machines for Conventional Analysis" GB/T16450-2004, the average volume of smoke taken by a person is 35 ml, the flow rate of flue gas is 17.5 ml/sec, and the suction time is 2 seconds. It is the main basis of the invention.
  • the core of the electronic nose instrument—the gas sensor array has high sensitivity to organic volatile gases such as hydrocarbons, alkenes, alcohols, esters, acids, aldehydes, and reducing inorganic volatile gases such as CO and NH 3 . It is particularly worth noting that the SnO 2 semiconductor gas sensor directly produces a volt-level voltage response output to cigarette smoke, which does not require secondary instrument amplification, which is attractive for the quality assessment of tobacco and tobacco products. It is in this context that electronic nose instrument evaluation methods for the quality of tobacco and tobacco products are highly valued.
  • electronic nose instruments can be used to identify tobacco and tobacco product types, brands, authenticity, etc., and objectively and fairly quantify the sensory quality indicators such as aroma, coordination, miscellaneous, irritating, and aftertaste, and accordingly
  • the process setting and process parameters of tobacco in-process production process are optimized, the quality level of tobacco and tobacco products is assessed, and the method is incorporated into relevant industry standards and national standards.
  • the sensory quality evaluation of tobacco and tobacco products is characterized in that the first flue gas is not used as the evaluation object; there is a cigarette self/smoldering process between the first flue gas and the second flue gas.
  • the characteristics of cigarette smoke are: (1) the composition is complex and complicated; (2) some components have strong adhesion; (3) the components with very small content have great influence on sensory perception. These are the basis for the selection of gas sensor and optimization of working conditions.
  • the response speed of the gas sensor should be fast with the necessary sensitivity; on the other hand, the contact time of the gas sensor with the cigarette smoke should be as short as possible, the ambient air flushing flow should be as large as possible, and the flushing time should be As long as possible, to flush out the adhered smoke molecules as much as possible, to avoid adhesion to the sensitive membrane surface of the gas sensor and the inner wall of the pipeline, so as to facilitate the recovery of the gas sensor as soon as possible.
  • the ideal goal is to use the gas sensor array module and the flue gas precision injection system module.
  • the exhaust gas discharge module, power supply module, working and control circuit module, computer, display unit and peripherals are integrated in a small test box, and the tobacco electronic nose instrument with small size, light weight and easy operation is designed to facilitate the sensory quality of cigarettes. On-site inspection and assessment of indicators.
  • a neural network using the standard Sigmoid activation function f(x) 1/(1 + exp(-x)) typically transforms the data set to the [0, 1] range, which is the actual default component mean of about 0.5. If we transform the input components into a certain range, we can appropriately enlarge the interval between the tobacco brands in the sample input space, which is beneficial to the neural network to speed up the learning, improve the learning accuracy and improve the promotion ability.
  • the learning and promotion performance of the overall multi-input-multi-output machine learning model is often not ideal.
  • a holistic multi-input-multiple-output neural network can easily fall into local minima during the learning process.
  • the simultaneous identification of large-scale tobacco and tobacco products and the quantitative prediction of multiple sensory indicators involve the classification and function approximation (non-linear regression, curve fitting) in the field of machine learning.
  • New machine learning models and algorithms including task decomposition methods, model structure optimization methods, fast learning algorithms and decision methods.
  • the invention is a prior invention patent "a machine olfactory device and its olfactory simulation test method” (see patent application number: 02111046.8), "a method for identifying a olfactory odor based on a modular combined neural network” (see patent application) No.: 03141537.7), "A small automatic machine olfactory instrument and odor analysis method” (see patent application number: 200710036260.4), "A olfactory simulation instrument and qualitative and quantitative analysis methods for various odors” (see patent application number: 201010115026.2) , “Methods for Selecting, Replacing, and Correcting Gas Sensors for Olfactory Simulators” (see Patent Application No.: 201310419648.8) and "A Field Analysis Method for Olfactory Simulator and Specific Material Gas (Smell) Flavor Levels" (see Patent Application No.) On the basis of :201310315482.5), an electronic nose instrument and analysis method was invented to solve the problem of on-site automatic detection, identification
  • the present invention provides an electronic nose instrument and a sensory quality evaluation method for tobacco and tobacco products, wherein the tobacco electronic nose device comprises a gas sensor array module, a flue gas automatic sample introduction system, a computer control and data analysis system, Automatic ignition device for online detection, identification and sensory quality index score prediction for tobacco and tobacco products.
  • the tobacco electronic nose device comprises a gas sensor array module, a flue gas automatic sample introduction system, a computer control and data analysis system, Automatic ignition device for online detection, identification and sensory quality index score prediction for tobacco and tobacco products.
  • the gas sensor array module consists of 16 SnO 2 semiconductor gas sensors, which are uniformly placed in a sealed cavity with a diameter of ⁇ 140mm and a section size of 20mm ⁇ 16mm to form an annular working chamber of the gas sensor array; the annular working chamber is at 55 ⁇
  • the constant temperature chamber at 0.1 °C is located at the upper right of the tobacco electronic nose instrument.
  • the flue gas automatic sample introduction system includes a cigarette holder, a micro vacuum pump, a first to a sixth total of two two-position two-way solenoid valves, a first and a second total of two throttle valves, a flow meter, a gas pipeline, and an overflow smoke.
  • the gas discharge device is located at the lower right of the tobacco electronic nose instrument.
  • the automatic ignition device comprises an ignition coil, a cigarette butt, a moving iron core, an electromagnetic coil, a magnet guide, a compression spring, a spring seat, a cable and a support, and is located at the right front lower side of the tobacco electronic nose instrument.
  • the computer control and data analysis system includes a computer motherboard, a data acquisition card, a precision linear and switching power supply module, a drive and control circuit module, a hard disk, a network card, a graphics card, and a display, which are located on the left side of the tobacco electronic nose instrument.
  • a sample of tobacco and tobacco products has a smoke sampling period of 5 minutes.
  • the gas sensor array undergoes initial recovery (210 seconds), clean air calibration (40 seconds), balance (2 seconds), second smoke extraction (2 seconds), ambient air Flush (46 seconds) for a total of 6 stages.
  • the ignition coil of the automatic ignition device When the smoke is sampled, under the control of the computer, the ignition coil of the automatic ignition device is horizontally moved 9 mm to the left, and the cigarette sample is ignited at a temperature of 380 °C.
  • the micro vacuum pump of the flue gas automatic sample introduction system draws the second flue gas at a flow rate of 17.5 ml/sec or 1050 ml/min, and flows through the annular working chamber of the gas sensor array to pass over the surface of the gas sensor sensitive membrane. For 2 seconds, the gas sensor array thus produces a sensitive response. From the beginning of the self-balancing state, the computer control and data analysis system began to record the response data; the balance (2 seconds), the second smoke suction (2 seconds), and the ambient air flush (the first 36 seconds) were recorded in sequence. The phase of the gas sensor array voltage response data, the total duration of 40 seconds. Data at other times during the flue gas sampling cycle are not recorded.
  • the computer control and data analysis system uses a modular neural network cascade model to perform 1 identification and 2 sensory quality index score prediction for tobacco and tobacco product samples.
  • 1 Modular neural network cascading model The first level consists of n(n-1)/2 single-output neural networks juxtaposed to form n voting identification groups for the identification of n tobacco and tobacco products, including brands and origins. And the identification of true and false.
  • 2 Modular neural network cascading model The second level consists of n ⁇ 5 single-output neural networks juxtaposed, each of which is used for the aroma, coordination, miscellaneous, irritating and aftertaste of n kinds of tobacco and tobacco products. 5 sensory quality indicators score prediction.
  • the angle between the axis of the cigarette holder and the horizontal plane is 0° to +5°.
  • the operator Prior to auto-ignition, the operator inserts the cigarette sample butt into the cigarette holder with an insertion depth of 9 ⁇ 0.5 mm and an insertion operation of 15 seconds. Within 15 seconds after the second smoke extraction, the operator removes the residual cigarette butt from the cigarette holder, extinguishes and discards.
  • the ignition coil of the automatic ignition device has an operating voltage of 24V and a current of 5A.
  • the cigarette core shaft line, the moving iron core shaft line, the electromagnetic coil shaft line, and the compression spring shaft line are on the same horizontal line as the cigarette holder axis line.
  • the ignition coil is energized from the 5th second before the first smoke is pumped, and the temperature is raised to 380 ° C in 5 seconds.
  • the electromagnetic coil is de-energized, and under the action of the compression spring, the ignition coil fixed on the cigarette butt is moved 9 mm from the reference position horizontally to ignite the cigarette sample.
  • the ignition coil is de-energized, and the electromagnetic coil is energized. Under the action of the electromagnetic force of the electromagnetic coil, the ignition coil is out of contact with the ignited cigarette sample and returns to the reference position.
  • the first flue gas is directly discharged to the outside through the second two-position two-way solenoid valve, the first throttle valve and the flow meter at a flow rate of 17.5 ml/sec or 1050 ml/min. It does not pass through the annular working chamber of the gas sensor array at all for 2 seconds, and smokes 35 ml of smoking gas.
  • the ignited cigarette sample was erected/self-ignited for 20 seconds.
  • the second flue gas flows through the first two-position two-way solenoid valve, the gas sensor array annular working chamber, and the fifth two-position two-way valve at a flow rate of 17.5 ml/sec or 1050 ml/min.
  • the solenoid valve, the first throttle valve, and the flow meter were finally discharged to the outside for 2 seconds, and 35 ml of flue gas was collected.
  • an ambient air flushing process with a flow rate of 6500 ml/min begins.
  • the ambient air flows through the third two-position two-way solenoid valve, the gas sensor array annular working chamber, the fifth two-position two-way solenoid valve, and the fourth two-position two-way solenoid valve, and is finally discharged to the outside for 15 seconds.
  • the operator removes the residual cigarette butts and discards them.
  • the ambient air flows through the first two-position two-way solenoid valve, the gas sensor array annular working chamber, the fifth two-position two-way solenoid valve, and the fourth two-position two-way solenoid valve at a flow rate of 6500 ml/min. Exhausted to the outside for 31 seconds. At this point, one test cycle of the cigarette sample ends. If the next sample is to be detected, the computer automatically starts a new detection cycle and automatically transfers to the initial recovery process; otherwise, the operator clicks the "End Detection" button on the screen drop-down menu to force the detection process to end.
  • the tobacco electronic nose instrument test was evaluated by the evaluation team and a standard sample of tobacco and tobacco products with sensory scores of quality indicators was given.
  • the tobacco electronic nose instrument obtains the gas sensor array voltage response standard sample set X' ⁇ R N ⁇ 16 , and establishes X′ with aroma d 1 ⁇ R N , coordinates d 2 ⁇ R N There is a one-to-one correspondence between the sensory scores of five quality indicators, such as heterogeneous d 3 ⁇ R N , irritating d 4 ⁇ R N , and aftertaste d 5 ⁇ R N .
  • max(X') and min(X') are the maximum and minimum values of X', respectively, and x pi is the steady-state maximum value of the voltage response of the gas sensitive sensor i after proportional conversion of the p-th standard sample.
  • Max(X') and min(X') are stored as basic data in the computer.
  • X' is transformed into a training set, which is called X, after a proportional preprocessing transformation.
  • Each single-output neural network of the modular neural network cascade model experiences the learning phase of the training set X and the identification of the sample to be determined x and the sensory quality index score prediction stage.
  • Pair of training samples x p single output neural network
  • the actual output is:
  • Neural network The weighted sum of the output nodes to the actual output of all hidden nodes, namely:
  • Each score prediction group consists of 5 single-output neural networks, which respectively fit the response of the proportionally transformed gas sensor array with the corresponding sensory aroma, coordination, miscellaneous, irritating, and aftertaste.
  • the target output of X j is the quality index of the brand ⁇ j .
  • the sensory score is scaled to [0.15, 2.85].
  • the actual output score prediction group ⁇ j r th hidden node single-output neural network is h
  • the actual output of the r-th single-output neural network is the score prediction group ⁇ j
  • Each (n-1) single output neural networks form a voting recognition group representing a tobacco and tobacco product brand with a maximum number of votes n -1.
  • Each single-output neural network must participate in only two of the voting recognition groups, and n(n-1)/2 single-output neural networks thus form n voting identification groups, respectively, and use most voting rules. decision making.
  • the decision rule for identifying the sample x is that x belongs to the brand represented by the voting identification group with the highest number of votes. If two or more voting recognition groups vote for the same number of votes and are the highest number of votes, then the decision: x does not belong to any existing brand.
  • Modular neural network cascading model for predicting the scores of sensory quality indicators of tobacco and tobacco products.
  • the n ⁇ 5 single output neural networks are divided into n score prediction groups, which correspond one-to-one with n vote recognition groups.
  • the score prediction group ⁇ j the actual output of the rth single output neural network is z (jr) , then x belongs to the brand ⁇ j .
  • the predicted value of the r sensory quality index is:
  • the modular neural network cascade model first level from the existing n(n-1)
  • the /2 single output neural networks are increased to n(n+1)/2.
  • the newly added and learned single output neural network module is
  • the modular neural network cascade model adds five single output neural networks and learns from the second level, increasing from the existing n ⁇ 5 to (n+1). ⁇ 5.
  • a fake brand or an existing same brand produced by another manufacturer is considered as a separate brand for identification and sensory quality indicator score prediction.
  • Tobacco electronic nose instrument for the detection, identification and sensory quality index score prediction of cigarette samples, including the following steps:
  • the four-two two-way solenoid valve is finally discharged to the outside.
  • the gas sensor array is thermostated to a constant room temperature of 55 ⁇ 0.1 °C from room temperature.
  • the sampling phase of the flue gas begins: the operator clicks the “Start Detection” button on the drop-down menu of the screen, and the instrument enters the smoke sampling period of 5 minutes, and the computer automatically generates a text file named “xxx” in the specified folder. To record the response data of the gas sensor array to the flue gas.
  • clean air flows through the second throttle valve, the sixth two-position two-way solenoid valve, the gas sensor array annular working chamber, and the flow rate of 17.5 ml/sec or 1050 ml/min.
  • the three-two two-way solenoid valve is finally discharged into the room for 40 seconds. Clean air allows the gas sensor array to be accurately restored to the baseline state.
  • the first mouth smoke suction in the 230.00-232.00 seconds of the flue gas sampling period, that is, the accurate calibration state of the clean air is 20.00-22.00 seconds.
  • the flue gas is 17.5 ml/
  • the flow rate of 1050 ml/min in seconds is directly discharged to the outside through the second two-position two-way solenoid valve, the first throttle valve, and the flow meter for 2 seconds.
  • Ambient air flushing During the flue gas sampling period of 254.00-300.00 seconds, the indoor air flows through the annular working chamber of the gas sensor array at a flow rate of 6500 ml/min, and adheres to the surface of the gas sensor sensitive membrane and the inner wall of the pipeline. The smoke odor molecules are initially washed away and the gas sensor array enters a preliminary recovery state. among them,
  • Residual cigarette butt removal operation During the smoke sampling period of 254.00-269.00 seconds, the operator takes out the residual cigarette butt and discards it within 15 seconds. During this period, the third two-position two-way solenoid valve, the fourth two-position two-way solenoid valve, and the fifth two-position two-way solenoid valve are turned on, and the other three two-position two-way solenoid valves are disconnected, and the indoor ambient air is 6500 ml / min flow, through the third two-position two-way solenoid valve, gas sensor array annular working chamber, fifth two-position two-way solenoid valve, fourth two-position two-way solenoid valve, and finally discharged to the outside, lasting 15 seconds.
  • (9) Feature extraction: In a smoke sampling period, the computer extracts the steady-state maximum value of the voltage response of each gas sensor as a characteristic component from the "xxx" data record file with a duration of 40 seconds, which is essentially the second port. In response to the smoke, a sample of the tested tobacco product is thus converted into a 16-dimensional measurement sample and stored in a tobacco and tobacco product sample data set file on a computer hard disk.
  • Recognition and sensory quality index score prediction in the smoke sampling period of 290.00-300.00 seconds, that is, within 10 seconds of the end of data recording, the modular neural network cascade model first level - n voting identification groups based on majority voting The rule determines the brand, origin and true and false of the sample x, and the second level of the modular neural network cascade model—the score prediction group corresponding to the winning vote recognition group predicts the aroma, coordination, noise, irritancy, aftertaste of x. The scores of the five sensory quality indicators are displayed on the display.
  • Figure 1 is a schematic diagram of the operation of a tobacco nose electronic instrument in an electronic nose instrument and a sensory quality evaluation method for tobacco and tobacco products (second smoke suction state).
  • Figure 2 is a diagram of the present invention - an electronic nose instrument and a sensory quality evaluation method for tobacco and tobacco products - a working schematic diagram of a tobacco electronic nose instrument (second smoke suction state) - a number of major parts.
  • Figure 3 is a schematic diagram of the present invention - an electronic nose instrument and a sensory quality evaluation method for tobacco and tobacco products - a solenoid valve and an air circuit on-off state (precision calibration of clean air, automatic ignition of a cigarette sample, and first smoke suction state) ).
  • Figure 6 is a schematic view of the present invention - an electronic nose instrument and a sensory quality evaluation method for tobacco and tobacco products - a solenoid valve and an air circuit on/off state (ambient air flushing, residual cigarette butt removal).
  • Figure 7 is a schematic view of the present invention - an electronic nose instrument and a sensory quality evaluation method for tobacco and tobacco products - a solenoid valve and an air circuit on-off state (primary recovery of a gas sensor).
  • Figure 8 is a schematic illustration of the present invention - an electronic nose instrument and a sensory quality evaluation method for tobacco and tobacco products - a single gas sensor response and a change in gas flow and duration in a circular working chamber of a gas sensor array during a sampling cycle.
  • Figure 9 is a schematic view of the present invention - an electronic nose instrument and a sensory quality evaluation method for tobacco and tobacco products - an automatic ignition device (electromagnetic coil power-off state).
  • Figure 10 is a schematic view of the present invention - an electronic nose instrument and a sensory quality evaluation method for tobacco and tobacco products - an automatic ignition device (electromagnetic coil energization state).
  • Figure 11 is a schematic view of the present invention - an electronic nose instrument and a sensory quality evaluation method for tobacco and tobacco products - a two-position two-way solenoid valve, an ignition coil, and an electromagnetic coil on and off in a sampling period.
  • Figure 12 is a perspective view showing the appearance of an electronic nose instrument and a sensory quality evaluation method for tobacco and tobacco products - a tobacco electronic nose instrument.
  • Figure 14 is a perspective view of the present invention - an electronic nose instrument and a sensory quality evaluation method for tobacco and tobacco products - a tobacco electronic nose instrument.
  • Figure 15 is a perspective view of the present invention - an electronic nose instrument and a sensory quality evaluation method for tobacco and tobacco products - Modified sigmoid activation function And standard Sigmoid activation functions First-order partial derivative curve and its ratio change curve.
  • First-order partial derivative Curve solid line
  • First-order partial derivative Curve dashed line
  • first-order partial derivative ratio curve first-order partial derivative ratio curve.
  • Figure 16 is a view of the present invention - an electronic nose instrument and a sensory quality evaluation method for tobacco and tobacco products - when determining the brand of a pending sample, n(n-1)/2 single output neural networks are divided into n voting recognition groups to vote Case.
  • FIG 17 is the present invention - an electronic nose instruments and tobacco and tobacco products Sensory Quality Evaluation Method - Sensory Quality Index of five single-output neural network modules prediction group Score ⁇ j.
  • Figure 18 is a schematic diagram of a modular neural network cascade model decision process for the simultaneous identification of a tobacco and tobacco product sample and the prediction of the sensory quality index of a tobacco and tobacco product sample by an electronic nose instrument and a sensory quality evaluation method for tobacco and tobacco products. schematic diagram.
  • FIG. 1 is a schematic diagram of the operation of the tobacco electronic nose instrument of the present invention, wherein the position of the gas path and the solenoid valve is the second port smoke suction working state.
  • the tobacco electronic nose instrument includes a gas sensor array module I, a flue gas automatic sample introduction system II, a computer control and data analysis system III, an automatic ignition device IV, and a cigarette sample V to be tested.
  • Figure 2 is a number diagram showing the main parts of the tobacco electronic nose instrument of the present invention.
  • the main components of the gas sensor array module I include: a gas sensor array I-1, a gas sensor array annular working chamber I-2, an insulation layer I-3, and a resistance heating wire and a fan, which are located at the upper right portion of the tobacco electronic nose instrument. .
  • the gas sensor array is mainly composed of 16 gas sensors of TGS800 and TGS2000 series. Available models include TGS800, TGS813, TGS816, TGS821, TGS822, TGS823, TGS826, TGS830, TGS832, TGS2600, TGS2602, TGS2603, TGS2610, TGS2611. TGS2612, TGS2620, TGS3830, TGS2201, and PID-A1 photoion detectors.
  • the function of the gas sensor array module I is to convert the cigarette smoke of complex components into an analog voltage signal of 0-10V.
  • the flue gas automatic sample introduction system II consists of: the first two-position two-way solenoid valve II-1, the second two-position two-way solenoid valve II-2, the third two-position two-way solenoid valve II-3, the fourth two Position two-way solenoid valve II-4, fifth two-position two-way solenoid valve II-5, sixth two-position two-way solenoid valve II-6, cigarette holder (phone holder) II-7, micro vacuum pump II-8, The first throttle valve II-9, the flow meter II-10, the second throttle valve II-11, the clean air II-12, the soot tray II-13, and the overflow flue gas discharge device II-14.
  • the flue gas automatic sample introduction system II is located at the lower right of the tobacco electronic nose instrument.
  • the main components of computer control and data analysis system III are computer motherboard III-1, 16-channel data acquisition card III-2, drive and control module III-3, multi-channel DC voltage module III-4, display III-5, and hard disk. , network card, video card, mouse, keyboard, etc., located in the left part of the tobacco electronic nose instrument.
  • the main functions of computer control and data analysis system III are: (1) acquisition, analysis and processing of gas sensor array response signals; (2) multiple two-position two-way solenoid valves and micro vacuum pumps for flue gas automatic sample system II , automatic ignition device IV, and computer control and data analysis system III itself drive and control.
  • Figure 3 is a schematic diagram of the on-off state of the solenoid valve and the pneumatic circuit when the clean air is accurately calibrated and the cigarette sample is automatically ignited and the first port is smoked.
  • the tobacco electronic nose instrument should complete the ignition of the tested cigarette sample V, the first smoke suction of the flue gas automatic sample introduction system II, and the precise calibration of the clean air of the gas sensor array I. action.
  • the second two-position two-way solenoid valve II-2 is turned on, the first two-position two-way solenoid valve II-1, The fourth two-position two-way solenoid valve II-4 and the fifth two-position two-way solenoid valve II-5 are disconnected.
  • the ignition coil IV-1 of the automatic ignition device IV is energized, the electromagnetic coil IV-4 is de-energized, and the spool IV-3 is horizontally moved 9 mm to the left under the action of the compression spring IV-6, so that the ignition coil IV -1 is in contact with the cigarette sample V to be tested for 5 seconds; then, in the first smoke suction state, the cigarette smoke does not pass through the gas sensor array module I at all, but in the micro vacuum pump II-8 Under the action of suction, it is directly discharged to the outside through the second two-position two-way solenoid valve II-2, the first throttle valve II-9, and the flow meter II-10 at a flow rate of 17.5 ml/sec or 1050 ml/min. For 2 seconds, equivalent to smoking 35 ml.
  • the gas sensor array I-1 While the cigarette sample V auto-ignition and flue gas automatic sample introduction system II sucks the first cigarette, the gas sensor array I-1 is in the state of accurate calibration of the clean air, the two-position two-way solenoid valves II-3 and II- 6 conduction, clean air flows through the second throttle valve II-11, the sixth two-position two-way solenoid valve II-6, the gas sensor array annular working chamber I-2, and the third at a flow rate of 1050 ml/min. The two-position two-way solenoid valve II-3 is finally discharged into the indoor atmosphere. This is also the second stage of accurate calibration of clean air for 7 seconds.
  • Fig. 4 is a schematic view showing the state of the solenoid valve and the air passage when the electronic nose instrument sucks the second mouth smoke.
  • the first two-position two-way solenoid valve II-1 and the fifth two-position two-way solenoid valve II-5 are turned on, and the other four two-position two-way solenoid valves are disconnected, and the cigarette smoke is 17.5 ml.
  • / sec is 1050 cc / min flow, through the first two-position two-way solenoid valve II-1, gas sensor array I-1 and its annular working chamber I-2, fifth two-position two-way solenoid valve II- 5.
  • the first throttle valve II-9 and the flow meter II-10 are finally discharged to the outside for 2 seconds, which is equivalent to collecting 35 ml of flue gas.
  • the ignition coil IV-1 of the automatic ignition device IV is de-energized, the electromagnetic coil IV-4 is energized and the spool IV-3 is horizontally moved 9 mm to the right, so that the ignition coil IV-1 is separated from the measured cigarette sample V. contact.
  • the gas sensor array I-1 produces a sensitive response during the second cigarette flow.
  • This 16-dimensional voltage response steady-state maximum vector is the basis for the identification and sensory quality indicators of tobacco and tobacco products by tobacco electronic nose instruments.
  • Fig. 5 is a schematic view showing the state in which the cigarette sample V is inserted into the cigarette holder II-7 and the cathode/self-ignition state, and the solenoid valve and the air passage are turned on and off.
  • the third two-position two-way solenoid valve II-3 and the sixth two-position two-way solenoid valve II-6 are turned on, and the other four two-position two-way solenoid valves are disconnected.
  • Cigarette insertion is manual operation, timing is 15 seconds; cigarette smoke/self-ignition duration is 18+2 seconds, adjustable.
  • “+2" seconds refers to the balance time.
  • the gas sensor array I-1 is in a state of accurate calibration of the clean air
  • the two-position two-way solenoid valves II-3 and II-6 are turned on, and the clean air flows through the second throttle in sequence at a flow rate of 1050 ml/min.
  • This process together with the cigarette sample V ignition and the flue gas auto-injection system II of Figure 3, draws the first cigarette process, so that the clean air is accurately calibrated for a total duration of 40 seconds.
  • Fig. 6 is a schematic view showing the state in which the solenoid valve and the air passage are turned on and off during the removal of the residual cigarette.
  • the third two-position two-way solenoid valve II-3, the fourth two-position two-way solenoid valve II-4 and the fifth two-position two-way solenoid valve II-5 are turned on, and the other three two-position two-way The solenoid valves are all disconnected, and the indoor ambient air is passed through the third two-position two-way solenoid valve II-3, the gas sensor array I-1 and its annular working chamber I-2, and the fifth two at a flow rate of 6500 ml/min.
  • the two-way solenoid valve II-5 and the fourth two-position two-way solenoid valve II-4 are finally discharged into the outdoor atmosphere.
  • the manual operation time for taking out the residual cigarette butts is 15 seconds (adjustable). This phase is the first phase of ambient air flushing and can be considered as part of the initial recovery state of the gas sensor array I-1.
  • Fig. 7 is a schematic view showing the state in which the solenoid valve and the air passage are turned on and off after the residual cigarette butt is taken out.
  • the first two-position two-way solenoid valve II-1, the fourth two-position two-way solenoid valve II-4 and the fifth two-position two-way solenoid valve II-5 are turned on, and the other three two-position two-way The solenoid valves are all disconnected.
  • the ambient air passes through the cigarette holder II-7, the first two-position two-way solenoid valve II-1, the gas sensor array I-1 and its annular working chamber I-2, fifth in sequence at a flow rate of 6500 ml/min.
  • the two-position two-way solenoid valve II-5 and the fourth two-position two-way solenoid valve II-4 are finally discharged into the outdoor atmosphere.
  • This phase is the second phase of ambient air flushing, which lasts for 31 seconds and can still be seen as another part of the initial recovery state of the gas sensor array I-1.
  • the new flue gas sampling cycle begins automatically, or the operator clicks the “End Detection” button on the screen drop-down menu and the detection process ends.
  • Figure 8 is a diagram showing the variation of the voltage response of a single gas sensor of a tobacco and electronic nose instrument during a smoke sampling period and the change of gas flow and duration in the annular working chamber I-2 where the gas sensor array I-1 is located. .
  • the gas flow experienced three changes of 6500 ml/min, 1050 ml/min and 0 (equilibrium); the gas type experienced three initial flushes of ambient air, precise calibration of clean air and sampling of cigarette smoke. Variety. In the equilibrium state, all six 2/2-way solenoid valves are disconnected, and no gas flows in the gas sensor array I-1 and its annular working chamber I-2.
  • the smoke sampling period of a cigarette sample V is 300 seconds. From the beginning of the equilibrium state, that is, the 250th second of the flue gas sampling period, the computer control and data analysis system III records the transient voltage response data of the gas sensor array, the recording time is 40 seconds, including the balance phase 2 seconds, the second In the sampling phase of the mouth smoke gas for 2 seconds and 36 seconds before the ambient air flushing phase, the voltage response of the gas sensor array I-1 to the cigarette smoke, that is, the sampled data, is stored in a text file.
  • the steady state maximum value of the voltage response of the gas sensor i to the flue gas of the cigarette sample p is extracted as the characteristic component x pi ', thereby obtaining the gas sensor array for the pth cigarette sample
  • the response of the second smoke is called the voltage response sample x p ' ⁇ R 16 .
  • the computer control and data analysis system III gives the brand recognition result of the sample p and the scores of five sensory indicators such as aroma, coordination, noise, irritation and aftertaste according to the sample x p '. forecast result.
  • FIG. 9 is a schematic diagram of the components of the automatic ignition device IV (when the electromagnetic coil IV-4 is powered off).
  • the automatic ignition device IV is located at the right front lower side of the tobacco electronic nose instrument, and the constituent units include: ignition coil IV-1, cigarette butt iron IV-2, moving iron core IV-3, electromagnetic coil IV-4, magnet guide IV-5, compression spring IV-6, spring seat IV-7, cable IV-8, support IV-9.
  • the working voltage of the ignition coil IV-1 and the electromagnetic coil IV-4 is 24V DC, and the maximum operating current of the ignition coil is 5A.
  • the ignition coil IV-1 is energized and heated to 380 ° C in 5 seconds, at which time the solenoid IV-4 is energized.
  • the electromagnetic coil IV-4 is de-energized, and under the action of the compression spring IV-6, the ignition coil IV-1 and the cigarette butt tip IV-2 fixed on the movable iron core IV-3 are horizontally moved to the left from the reference position. 9 mm, in contact with the cigarette butt of the tested cigarette sample V and ignited.
  • FIG 10 is a schematic diagram of the components of the automatic ignition device IV (when the electromagnetic coil IV-4 is energized).
  • the electromagnetic coil IV-4 is energized and the ignition coil IV-1 is de-energized.
  • the compression spring IV-6 is compressed by 9 mm, and the ignition coil IV-1 and the cigarette butt head IV-2 fixed on the movable iron core IV-3 are returned to the reference position, which is artificial.
  • the period of energization of the electromagnetic coil IV-4 is: manually inserting the cigarette sample V into the cigarette holder II-7 for 15 seconds; after sucking the first mouth of smoke for 1 second ( 2nd second); cigarette sample V/spontaneous combustion lasts 18 seconds; 2 seconds in equilibrium; 2 seconds of suction of the second cigarette; 15 seconds of removal of residual cigarette butt for 53 seconds.
  • the ignition coil IV-1 is energized, and the rest of the time is powered off.
  • Figure 11 is a schematic diagram showing the relationship between the six 2/2-way solenoid valves, the ignition coil IV-1 and the electromagnetic coil IV-4 during a smoke sampling period.
  • Figure 11 (a) shows that the two-position two-way solenoid valve II-1 is energized in the first two seconds of the first 210 seconds and the last 31 seconds, except that the second port must be turned on during the second second of the smoke suction. It is convenient for the ambient air to flush the smoke molecules of the cigarette holder II-7, the sensitive membrane surface of the gas sensor array and the inner wall of the pipeline for a long time.
  • Figure 11 (b) and (f) illustrate that the two-position two-way solenoid valve II-2 only controls the suction of the first port of smoke, and the two-position two-way solenoid valve II-6 only controls the opening and closing of the clean air.
  • the role of the person is relatively simple.
  • the two-position two-way solenoid valve II-3 is turned on in the two stages of 40 seconds of accurate calibration of clean air and 15 seconds after the end of second-stage smoke suction, but the flow direction of the gas And the flow is different.
  • Figure 11(d) shows that the two-position two-way solenoid valve II-4 is disconnected during the three stages of accurate calibration, balance, and second-stage smoke suction, thereby forcing the first smoke and the second.
  • the flue gas flows through the throttle valve II-9 at a flow rate of 1050 ml/min.
  • the two-position two-way solenoid valve II-4 mainly controls the gas flow rate between 6500 ml/min and 1050 ml/min. Comparing Fig. 11(e) with Fig. 11(d), the two-position two-way solenoid valve II-5 and the two-position two-way solenoid valve II-4 are inconsistent only in the state of the second cigarette smoking phase.
  • the two-position two-way solenoid valve II-5 is turned on, and the two-position two-way solenoid valve II-4 is disconnected, thereby forcing the second port of flue gas to flow through the throttle valve II-9 at a flow rate of 1050 ml/min.
  • Figure 12 is a perspective view of a tobacco electronic nose instrument.
  • the gas sensor array module I is located in the upper right part of the tobacco electronic nose;
  • the computer control and data analysis system III is located in the left part of the tobacco electronic nose;
  • the flue gas automatic sample introduction system II and the automatic ignition device IV are located in the tobacco electronic nose. The lower right.
  • Figure 13 is a front elevational view of a tobacco electronic nose instrument. According to this figure, what can be seen from the outside is the display III-5 of the computer control and data analysis system III; the cigarette holder II-7 of the flue gas automatic sample introduction system II, the exhaust fan of the flue gas overflow device II-14, Cigarette tip IV-2 of the automatic ignition device IV, and the sample V of the cigarette to be tested.
  • FIG 14 is a schematic rear view of a tobacco electronic nose instrument.
  • the electronic nose instrument is equipped with an external display interface, 2 USB interfaces, a mouse interface, a keyboard interface, an Internet interface; an ambient air and a clean air inlet; a flue gas outlet discharged from the flue gas automatic sample introduction system II; and a smoke gas
  • the flue gas directly discharged from the discharge device II-14. Users can easily plug in the required external devices, such as large-screen displays, keyboards, and mice for data transmission, switching, and Internet remote transmission.
  • Figure 15 is in the interval Modified sigmoid activation function And standard Sigmoid activation functions First-order partial derivative curve and its ratio change curve.
  • Figure 15(a) shows the first-order partial derivatives of the two activation functions in this interval. versus Curve, solid line is the modified activation function First-order partial derivative Curve, dashed line is the standard Sigmoid activation function First-order partial derivative curve.
  • Figure 15(b) shows the ratio of the modified first-order partial derivative to the standard sigmoid activation function. curve. Why use Is this modified sigmoid activation function?
  • the single hidden layer neural network adopts the error back-propagation algorithm, the square of the error between the actual output and the expected output of the neural network without the oscillation, the larger the partial derivative (one step) of the weight and threshold parameters. The faster the neural network learns.
  • the consideration of this approach is that the mean value of each component of the training data set X is around 3.0 regardless of the sample distribution state.
  • Standard Sigmoid activation function The neural network typically transforms the data set to the [0,1] range, which actually has a default component mean of about 0.5.
  • the advantage of the input component transforming to the [0,6] range is that the sample interval and the inter-class spacing are magnified 6 times compared with the original [0,1], which is beneficial to the neural network to accelerate without oscillation. Learning speed, improving learning accuracy and promotion ability.
  • the present invention employs a one-against-one (OAO) task decomposition method.
  • Tobacco training set X is broken down into A binary-class subproblem.
  • the n(n-1)/2 sub-problems are learned and solved by n(n-1)/2 single-output neural network modules, respectively, thereby forming the first stage of the modular neural network cascade model.
  • Table 3 gives a list of n(n-1)/2 single output neural networks participating in the study.
  • Table 3 list of n(n-1)/2 single output neural networks participating in learning to identify n brands of tobacco and tobacco products
  • the n(n-1)/2 single-output neural network of the first stage of the modular neural network cascade model uses most voting rule decisions.
  • Table 4 gives a list of n(n-1)/2 single output neural network groups participating in the voting in order to determine which brand of the tobacco and tobacco products to be determined.
  • Table 3 is the main diagonal symmetry, a single output neural network module must participate in 2 voting groups. For example, a single output neural network module Both the voting in the voting recognition group ⁇ j and the voting in the voting recognition group ⁇ k vote.
  • Figure 16 is a perspective view of the present invention - an electronic nose instrument and a sensory quality evaluation method for tobacco and tobacco products - determining the brand of the cigarette sample x to be determined, n(n-1)/2 single output neural networks divided into n voting recognition groups In the case of voting, each group has n-1 members.
  • each score prediction group consists of 5 single-output neural networks.
  • FIG 17 is the present invention - an electronic nose instruments and tobacco and tobacco products Sensory Quality Evaluation Method - Sensory Quality Index of five single-output neural network modules prediction group Score ⁇ j.
  • the five members predicted the scores of the five sensory quality indicators of the aroma, coordination, miscellaneous, irritating, and aftertaste of the cigarette sample x belonging to the brand ⁇ j .
  • the sensory score is scaled to [0.15, 2.85].
  • Figure 18 is a schematic diagram of a modular neural network cascade model decision process for simultaneous 1 identification and 2 sensory quality index score prediction for a tobacco and tobacco product sample.
  • the tobacco sample x is branded, originated, and authenticated
  • the first level of n(n-1)/2 single output neural networks participates, and is divided into n voting identification groups, each group of n-1 members.
  • a single output neural network participates in two of the voting recognition groups.
  • the n vote recognition groups use most voting rule decisions, and the group with the most votes wins.
  • the decision rule is that the sample x does not belong to any existing brand.

Landscapes

  • Chemical & Material Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Pathology (AREA)
  • Immunology (AREA)
  • Biochemistry (AREA)
  • Analytical Chemistry (AREA)
  • Medicinal Chemistry (AREA)
  • Food Science & Technology (AREA)
  • Combustion & Propulsion (AREA)
  • Theoretical Computer Science (AREA)
  • Molecular Biology (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Software Systems (AREA)
  • Mathematical Physics (AREA)
  • General Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Computational Linguistics (AREA)
  • Evolutionary Computation (AREA)
  • Computing Systems (AREA)
  • Artificial Intelligence (AREA)
  • Botany (AREA)
  • Wood Science & Technology (AREA)
  • Chemical Kinetics & Catalysis (AREA)
  • Electrochemistry (AREA)
  • Manufacturing Of Cigar And Cigarette Tobacco (AREA)

Abstract

一种电子鼻仪器和烟草与烟草制品感官质量评价方法，电子鼻仪器中气敏传感器阵列模块、烟气自动进样系统、计算机控制与数据分析系统、自动点火装置集成在一个测试箱内，实现关键部件模块化、整体集成化与小型化和操作自动化；通过模拟专业人员的评吸过程，实现第一口烟气、阴/自燃、第二口烟气、洁净空气与环境空气之间和6500毫升/分钟与1050毫升/分钟气体流量之间的精确自动转换；通过测试大量标准试样，建立气敏传感器阵列响应与品牌标记和感官质量指标评定值关系的烟草大数据，将烟草感官质量评价问题转化为①识别和②感官质量指标预测问题，用模块化神经网络级联模型予以解决；通过提出修正的S型活化函数和基于模块化神经网络的大多数投票决策与量化规则，实现大规模烟草与烟草制品的在线检测、识别和感官质量指标得分预测。

Description

一种电子鼻仪器和烟草与烟草制品感官质量评价方法 技术领域
本发明—一种电子鼻仪器和烟草与烟草制品质量检测方法，涉及计算机、精密测量、精密机械、自动控制、分析化学、烟草领域，面向烟草与烟草制品产品生产过程质量控制和市场监管需求，主要解决电子鼻仪器自动化、集成化、小型化和烟草与烟草制品在线检测、识别与感官质量指标预测问题。为简明起见，以下将“电子鼻仪器”称为“烟草电子鼻”。
背景技术
“评吸”指的是在特定环境下，用人的口腔、鼻腔、喉部等特定器官对烟草与烟草制品以及烟用香精的质量和作用进行感官鉴定，即卷烟内在品质—感官质量指标评定。现阶段，烟草与烟草制品质量优劣完全靠人的感官评吸确定。与酒、茶、食品等呈味物质相比，烟草与烟草制品的独特之处是，质量评价完全依赖于人的感官，迄今为止尚无切实可行的理化指标检测与分析方法。
现行行业标准《烟草及烟草制品 感官评价方法》YC/T138-1998由国家烟草专卖局1998.03.12日发布，1998.05.01日实施，广泛应用于烟草企业产品生产和行业与市场管理。该标准规定，评吸员采用整体循环评吸法对烟草及烟草制品试样进行评价，并填写表1所示的各项感官质量指标计分。其中，光泽和谐调指标记分单位为0.5，香气、杂气、刺激性和余味指标记分单位均为1.0。以百分制计算，香气和杂气两项指标满分达52分，占比达52％。
除色泽这一外在质量指标外，表1中的香气、协调、杂气、刺激性、余味这5项内在质量指标都与嗅觉感受有直接或间接的关系。为了使感官评价结果尽可能客观、公正、准确，评吸人员必须经过专业理论培训和技术锻炼，保持正确的心理状态和良好的身体条件；评吸环境条件应达到标准要求，评吸前以校准试样校准和统一口径。评吸小组成员应足够多，一般应为7人及以上，小组所有成员打分平均值即为某个烟草试样感官质量指标最终得分，并用二项分布和χ 2分布进行显著性检验。
行业标准YC/T138规定，烟草及烟草制品感官评价者分为初级评吸员、优选评吸员和专家三种。经过专门选择与培训后，评吸员应满足的基本条件是：(a)具有烟草及烟草制品专业知识；(b)具有感官评吸兴趣和经验；(c)身体健康，没有任何感觉方面的缺陷；(d)个人敏感性正常，相互之间敏感性一致；(e)评吸期间生理状态正常；(f)评吸前不得吃辛辣等刺激性食物，不得饮酒；(g)没有明显体味，不使用有气味的化妆品。为防止产生感官疲劳和不适，单个评吸员连续评吸试样的数目一般应控制在25支/天以内。评吸应在安静、无干扰、通风、无异味的评吸室内进行。
目前，“评吸”不仅是确定烟草与烟草制品质量唯一可行的方法，而且也是确定烟草与烟草制品产
表1,卷烟(烟草制成品)感官质量检验原始记录表(YC/T138-1998)*
Figure PCTCN2018079389-appb-000001
*：括号里是满分值。
品配方结构的基础，是开发新产品、保持既有产品风格和质量稳定的必要且决定性手段。对消费者—吸烟者而言，吸烟是一种生理刺激和美好享受，不存在劳动问题；而对评吸人员而言，评吸鉴定则是一项极为艰苦、细致的工作，在评吸过程中思想要高度集中，全神贯注，要在短短几分钟内做出判断。
吸烟有害健康是人类共识。经过数十年漫长而繁琐的成分分析过程，烟草学术界终于发现，卷烟烟气是由4,000～5,000种成分构成的复杂气溶胶体。吸烟过程中，大部分烟气吸入肺部，有害物质或停留在肺部，或进入消化道，或进入血液循环，流向全身。科学研究指出，尼古丁是吸烟上瘾的主要原因，对人体尤其是心血管系统有一定的毒害作用。焦油中含有癌症诱发剂和癌症促进剂成分，例如苯并[а]芘。不仅如此，卷烟烟气中的酚类、醇类、酸类、醛类、CO、HCN、NH 3等物质是有毒和有致癌作用的。总之，若长久吸烟，烟气中的致癌和促癌物质将损伤正常细胞，大大提高患肺癌等疾病的几率。这就是说，长期抽吸烟草与烟草制品无疑会对评吸人员的身体产生严重伤害。
依据现行行业标准YC/T138，卷烟内在品质的感官评定一般采用“整体循环评吸法”：评吸员将烟气吸入口腔，通过喉部吞咽后再从鼻腔徐徐排出，利用全部评吸感觉器官进行评价。由于感觉器官灵敏度及对各项指标捕捉程度不同、生理和心理状况不同、环境条件差异等因素影响，导致评吸员之间评吸结果描述和判断的差异，这充分说明感官评吸法有很大的局限性。
如果说行业标准YC/T138主要关注的是烟草制成品或卷烟商品的感官质量，那么，行业标准《烟草在制品感官评价方法》YC/T415-2011的主要目的则是烟草在制品生产过程的工序设置优化、设备工艺参数优化和产品质量控制。YC/T415将感官评价指标分为香气、烟气、口感3大特性，再将每个特性细分为4个单项共3×4＝12单项，如表2所示。每项指标设置1～9共9个整数分值。香气风格的变化程度依赖于香气、烟气、口感3大特性的变化。
YC/T415主要目的在于考察工序与工段的变化或工艺参数的变化对烟草在制品感官质量的影响，以及在制品质量波动情况。该标准采用对比评吸法，以工序或工段变化前或正常生产的样品作为对照样品，来评价工序或工段变化后的样品质量，即香气风格变化程度和工序工段处理状况。工业使用结果表明，YC/T415的评价指标能较全面地反映烟叶风格特征和质量特征。目前，国内多数卷烟企业将该项标准用于原料质量评价、新产品开发、卷烟配方替代、烟叶在制品质量检验等。
行业标准《烤烟 烟叶质量风格特色感官评价方法》YC/T530-2015，根据中式卷烟质量需求，侧重于烤烟烟叶质量风格特色的评价。该标准沿用YC/T415以及《卷烟 中式卷烟风格感官评价方法》YC/T497-2014相关评价指标，结合烤烟烟叶质量特点，进行了风格特征、品质特征的细化分类。与YC/T415相比，该标准将干净程度、回甜这2项指标归入“余味”指标，提出16种香韵、3种香气状态和9种杂气类型，增加柔和程度、圆润感评价指标，保留香气质、香气量、透发性、劲头、浓度、杂气、刺激性、干燥感、柔和程度指标，力图较为全面地反映烟叶风格特征。
YC/T497、YC/T415、YC/T138这3个现行烟草行业标准主要内容是烟叶、烟草在制品与烟草制成品的质量评价问题，依赖的是人的感官，强调人的感觉和经验，其中的香气、协调、杂气、刺激性、余味、劲头、透发性等内在质量指标都与嗅觉感受有直接或间接关系。
必须指出，上述3个行业标准并没有规范评吸员的具体评吸过程。评吸就是烟气的口腔吸入、喉
表2,烟草在制品感官评价指标表(YC/T415-2011)
Figure PCTCN2018079389-appb-000002
部下咽、鼻腔呼出即所谓“吸”、“吞”、“呼”的过程。具体评吸步骤一般为：
(1)卷烟点燃。采用统一火源；
(2)第一口烟。将烟气以适当抽吸量吸入口腔，然后立即吐出；
(3)阴然。卷烟在环境空气中自我燃烧；
(4)第二口烟。将烟气以适当抽吸量吸入口腔，在口腔内停留2-4秒，吞下烟气，然后紧闭嘴唇，迫使烟气从鼻腔徐徐呼出。
大量试验指出，第一口烟气不能作为判断依据，因为第一口烟气很易受火源异味、烟用香精、初始燃烧状态等因素的影响。烟气“抽吸量”是影响评吸判断准确性的一个重要因素。评吸人员不是机器，只能尽量控制和保持个人或相互之间烟气抽吸量的一致。依据国家标准《常规分析用吸烟机定义和标准条件》GB/T16450-2004，人抽一口烟气的平均容积为35毫升，烟气流量为17.5毫升/秒，抽吸时间为2秒，这些数据是本发明的主要依据。
嗅觉是气味分子刺激鼻腔嗅细胞而产生的复杂感觉。现行气味质量评定标准强调人的嗅觉感受，不仅描述术语贫乏，而且受生理、心理、环境等因素的影响难以做到客观公正。色、质谱等成分分析方法能测定几百乃至几千种化学组成，但分析过程复杂；成分越多，分析时间越长。例如，卷烟烟气含有4000～5000种化学成分这一结果是美国学者和技术人员经过数十年色质谱成分分析得到的。不仅如此，呈味物质化学组成与气味类型及强度之间的关系还很少被人们理解和掌握。
电子鼻仪器的核心—气敏传感器阵列对烃、烯、醇、酯、酸、醛等有机挥发气和CO、NH 3等还原性无机挥发气有高灵敏度。特别值得指出的是，SnO 2半导体型气敏传感器对卷烟烟气直接产生伏级电压响应输出，不需要二次仪表放大，这对烟草和烟草制品质量评定来说是很有吸引力的。正是在这种情况下，烟草和烟草制品质量的电子鼻仪器评价方法受到高度重视。人们期待能用电子鼻仪器对烟草和烟草制品类型、品牌、真假等进行识别，对香气、协调、杂气、刺激性、余味等感官质量指标进行客观、公正的量化预测，并据此对烟草在制品生产过程工序设置与工艺参数优化，评定烟草和烟草制品质量等级，并将该方法纳入到相关的行业标准和国家标准中。
烟草和烟草制品感官质量评价的特点是，第一口烟气不作为评价对象；第一口烟气与第二口烟气之间有一个卷烟自/阴燃过程。为了将电子鼻仪器和方法用于烟草和烟草制品质量评价中，我们需要解决以下问题：
(A)气敏传感器选型与工作条件优化
卷烟烟气特点是，(1)成分繁多且复杂；(2)一些成分附着力很强；(3)含量甚微的成分对感官感受影响很大。这些是气敏传感器选型和工作条件优化的依据。一方面，气敏传感器在具有必要灵敏度的前提下响应速度应可能快；另一方面，气敏传感器与卷烟烟气的接触时间应尽可能短，环境空气冲洗流量应尽可能大，冲洗时间应尽可能长，以尽可能冲洗掉黏附的烟气分子，避免其在气敏传感器敏感膜表面和管道内壁附着，便于气敏传感器的尽快恢复。
(B)卷烟烟气采样过程精密自动化
操作自动化和操作简便对烟草电子鼻仪器尤其重要。为了模拟评吸人员的评吸过程，我们需要设计无火焰点火器，发明卷烟烟气精密自动采样系统，设计烟气废气排出系统，满足只有第二口烟气流过气敏传感器阵列环形工作腔，且流量17.5毫升/秒，持续2秒钟，烟气容积35毫升这些要求。除卷烟试样插入烟嘴和残留烟蒂取出丢弃这两个环节外，烟气采样整个过程由烟草电子鼻仪器自动完成。
(C)烟草电子鼻仪器关键部件模块化与小型化
我们需要将数据采集卡驱动电路、气敏传感器阵列工作电路、直流电源电路、多种元件驱动电路集成于一体，形成驱动与控制电路模块；在多个电磁阀和节流阀集成的基础上，我们需要发明创造烟气精密进样系统模块，实现小型化。为便捷安装和更换，我们还需要致力于气敏传感器阵列和多路直流电源 等部件的模块化与小型化。
(D)烟草电子鼻仪器集成化
为使烟草电子鼻走出实验室，在关键部件模块化与小型化的基础上，我们需要解决仪器整体集成化与小型化问题，理想目标是将气敏传感器阵列模块、烟气精密进样系统模块、废气排出模块、电源模块、工作与控制电路模块、计算机、显示单元、外设集成在一个小型测试箱内，设计研制出尺寸小、重量轻、操作简便的烟草电子鼻仪器，便于卷烟感官质量指标的现场检测与评定。
(E)面向大规模烟草与烟草制品现场检测、识别和质量等级预测的机器学习方法
卷烟品牌众多，烟草来源众多，烟用香料香精众多。识别卷烟品牌是人们对烟草电子鼻仪器的基本要求。如果在品牌识别的基础上，进一步要求烟草电子鼻实现烟草与烟草制品香气、协调、杂气、刺激性、余味等感官质量指标的量化预测，并据此其评定质量等级，这就形成了大数据分析处理问题，对现有机器学习方法提出了挑战。为了将电子鼻仪器和方法用于烟草和烟草制品感官质量指标评价中，我们需要发明面向大规模烟草与烟草制品现场检测、识别和感官质量等级预测的机器学习方法。
采用标准Sigmoid活化函数f(x)＝1/(1+exp(-x))的神经网络一般将数据集变换到[0,1]范围，这一做法实际默认分量均值约为0.5。如果我们将输入分量变换到一定范围，就可适当放大烟草品牌之间在样本输入空间的间隔，有利于神经网络加快学习速度，提高学习精度和提高推广能力。
对大数据多类别问题，整体型的多输入—多输出机器学习模型的学习和推广性能往往不理想。例如，整体型的多输入—多输出神经网络在学习过程中极易陷入局部极小点。不仅如此，对大规模烟草与烟草制品同时进行品牌识别和多项感官指标量化预测这种问题涉及机器学习领域里的分类与函数逼近(非线性回归，曲线拟合)两个研究方向，需要发明新的机器学习模型与算法,包括任务分解方法、模型结构优化方法、快速学习算法和决策方法。
发明内容
本发明是在现有发明专利《一种机器嗅觉装置及其嗅觉模拟测试方法》(参见专利申请号：02111046.8)、《一种基于模块化组合神经网络的机器嗅觉气味识别方法》(参见专利申请号：03141537.7)、《一种小型自动化机器嗅觉仪器与气味分析方法》(参见专利申请号：200710036260.4)、《一种嗅觉模拟仪器与多种气味定性定量分析方法》(参见专利申请号：201010115026.2)、《面向嗅觉模拟仪器的气敏传感器选择、更换与校正方法》(参见专利申请号：201310419648.8)和《一种嗅觉模拟仪器和特定物质气(嗅)味等级现场分析方法》(参见专利申请号：201310315482.5)的基础上，发明一种电子鼻仪器和分析方法以解决烟草与烟草制品现场自动检测、识别与感官质量指标量化预测问题。
为了实现上述目的，本发明—一种电子鼻仪器和烟草与烟草制品感官质量评价方法，其中的烟草电子鼻仪器包括气敏传感器阵列模块、烟气自动进样系统、计算机控制与数据分析系统、自动点火装置，实现烟草及烟草制品的在线检测、识别和感官质量指标得分预测。
气敏传感器阵列模块由16个SnO 2半导体型气敏传感器组成，均布于中径φ140mm、断面尺寸20mm×16mm的密封腔内，形成气敏传感器阵列环形工作腔；该环形工作腔处于55±0.1℃的恒温室内，位于烟草电子鼻仪器右上方。
烟气自动进样系统包括卷烟夹持器、微型真空泵、第一至第六共6个二位二通电磁阀、第一和第二共2个节流阀、流量计、气体管道、溢出烟气排出装置，位于烟草电子鼻仪器右下方。
自动点火装置包括点火线圈、点烟头、动铁芯、电磁线圈、导磁铁架、压缩弹簧、弹簧座、电缆、支座，位于烟草电子鼻仪器右前下方。
计算机控制与数据分析系统包括计算机主板、数据采集卡、精密线性与开关电源模块、驱动与控制电路模块、硬盘、网卡、显卡、显示器，位于烟草电子鼻仪器左侧。
一个烟草与烟草制品试样的烟气采样周期为5分钟。依据流经的气体类型的不同，气敏传感器阵列历经初步恢复(210秒)、洁净空气精确标定(40秒)、平衡(2秒)、第二口烟气抽吸(2秒)、环境空气冲洗(46秒)共6个阶段。
烟气采样时，在计算机控制下，自动点火装置的点火线圈水平向左移动9毫米，以380℃的温度点燃卷烟试样。烟气自动进样系统的微型真空泵以17.5毫升/秒即1050毫升/分钟的流量抽吸第二口烟气，使之流经气敏传感器阵列环形工作腔，掠过气敏传感器敏感膜表面，持续2秒，气敏传感器阵列因此产生敏感响应。自平衡状态开始之刻起，计算机控制与数据分析系统开始记录响应数据；依次记录平衡(2秒)、第二口烟气抽吸(2秒)、环境空气冲洗(前36秒)这3个阶段的气敏传感器阵列电压响应数据，总时长40秒。烟气采样周期其它时间的数据不记录。
在40秒的数据记录时间内，单个气敏传感器对第二口烟气的电压响应曲线稳态最大值被提取为特征分量，16个气敏传感器组成的阵列因此产生一个16维电压响应向量。在数据记录结束后的10秒内，计算机控制与数据分析系统依据这一响应向量对烟草与烟草制品试样进行品牌、产地、真假识别和香气、协调、杂气、刺激性、余味共5项感官质量指标得分预测。
计算机控制与数据分析系统采用模块化神经网络级联模型对烟草与烟草制品试样进行①识别和②感官质量指标得分预测。①模块化神经网络级联模型第一级由n(n-1)/2个单输出神经网络并列组成，形成n个投票识别组，用于n种烟草与烟草制品的识别，包括品牌、产地和真假的识别。②模块化神经网络级联模型第二级由n×5个单输出神经网络并列组成，每5个一组，用于n种烟草与烟草制品的香气、协调、杂气、刺激性、余味这5项感官质量指标得分预测。
卷烟夹持器轴心线与水平面夹角为0°～+5°。在自动点火之前，操作人员将卷烟试样烟蒂插入卷烟夹持器，插入深度为9±0.5mm，插入操作用时15秒。第二口烟气抽吸结束后的15秒内，操作人员从卷烟夹持器上取出残留的卷烟烟蒂，熄灭并扔弃。
自动点火装置的点火线圈工作电压24V，电流5A。点烟头轴心线、动铁芯轴心线、电磁线圈轴心线、压缩弹簧轴心线与卷烟夹持器轴心线位于同一水平线上。在计算机控制下，在第一口烟气抽吸前第5秒钟起，点火线圈通电，并在5秒钟内升温到380℃。与此同时，电磁线圈断电，在压缩弹簧作用下，固定在点烟头上的点火线圈随动铁芯自基准位置水平左移9毫米，点燃卷烟试样。在抽吸第一口烟气1秒之后，点火线圈断电，电磁线圈通电，在电磁线圈电磁力的作用下，点火线圈与被点燃的卷烟试样脱离接触，并回复到基准位置。
在微型真空泵抽吸作用下，第一口烟气以17.5毫升/秒即1050毫升/分钟的流量直接经第二二位二通电磁阀、第一节流阀、流量计直接被排出到室外，完全不经过气敏传感器阵列环形工作腔，持续2秒，抽吸烟气35毫升。接下来，被点燃的卷烟试样阴/自燃20秒。
在微型真空泵的抽吸作用下，第二口烟气以17.5毫升/秒即1050毫升/分钟的流量经第一二位二通电磁阀、气敏传感器阵列环形工作腔、第五二位二通电磁阀、第一节流阀、流量计，最后被排出到室外，持续2秒，采集烟气35毫升。
洁净空气精确标定过程与卷烟插入、自动点火、第一口烟气抽吸、阴/自燃过程同时进行，均历时40秒。在洁净空气精确标定阶段，洁净空气以17.5毫升/秒的流量依次流经第二节流阀、第六二位二通电磁阀、气敏传感器阵列环形工作腔、第三二位二通电磁阀，最后排出到室外。以下操作与此依次同时进行：首先，操作人员在15秒钟内将被测卷烟试样插入卷烟夹持器；随后，自动点火装置的点火线圈通电，左移9毫米与卷烟试样接触，并在5秒钟内升温到380℃；接下来是时长为2秒的第一口烟气抽 吸；最后是18秒的卷烟试样阴/自燃过程。
流量为1050毫升/分钟的第二口烟气抽吸过程一结束，流量为6500毫升/分钟的环境空气冲洗过程就随即开始。环境空气依次流过第三二位二通电磁阀、气敏传感器阵列环形工作腔、第五二位二通电磁阀、第四二位二通电磁阀，最终排出到室外，持续15秒。在此期间，操作人员取出残留卷烟烟蒂并丢弃。随后，环境空气以6500毫升/分钟的流量依次流过第一二位二通电磁阀、气敏传感器阵列环形工作腔、第五二位二通电磁阀、第四二位二通电磁阀，最终排出到室外，持续31秒。至此，卷烟试样的一个检测周期结束。若欲检测下一个试样，计算机自动开始新的检测周期，自动转入初步恢复过程；否则，操作人员点击屏幕下拉菜单的“结束检测”按钮强制结束检测过程。
在数据采集阶段，烟草电子鼻仪器测试经过评吸小组品评并给出质量指标感官得分的烟草与烟草制品标准试样。单个标准试样测量周期为5分钟，提取各个气敏传感器对第二口烟气的电压响应稳态最大值作为特征分量，对第p个标准试样得到一个16维的电压响应向量x p'＝(x p1’,…,x pi’,…,x p16’) T∈R 16。通过对N个标准试样的测试，烟草电子鼻仪器得到气敏传感器阵列电压响应标准样本集X′∈R N×16，并建立X′与香气d 1∈R N、协调d 2∈R N、杂气d 3∈R N、刺激性d 4∈R N、余味d 5∈R N共5个质量指标感官得分之间的一一对应关系。每一品牌A、B、C三个等级各测量10个标准试样。若品牌数为n，则N＝30n。
气敏传感器阵列响应标准样本集X′的所有特征分量经正比例预处理变换到[0.0,6.0]范围。设气敏传感器i对第p个标准试样的电压响应稳态最大值为x pi’，比例变换后的值为：
Figure PCTCN2018079389-appb-000003
这里，max(X′)和min(X′)分别为X′的最大值与最小值，x pi是气敏传感器i对第p个标准试样经比例变换后的电压响应稳态最大值，电压响应向量x p'因此变为16维样本x p＝(x p1,,…,x pi,…,x p16) T∈R 16。max(X′)和min(X′)作为基本数据存入计算机。X′经正比例预处理变换后被称为训练集，记为X。
在待定试样x的识别与感官质量指标得分预测阶段，气敏传感器i电压响应稳态最大值x pi’仍采用标准样本集的max(X′)与min(X′)以公式(1)进行正比例变换。
模块化神经网络级联模型的每个单输出神经网络均经历对训练集X的学习阶段和对待定样本x的识别与感官质量指标得分预测阶段。
模块化神经网络级联模型第一级—n(n-1)/2个单输出神经网络学习阶段，首先对训练集X施以一对一(one-against-one,OAO)分解，X被分解成
Figure PCTCN2018079389-appb-000004
个二分类(binary-class)训练子集。然后，这n(n-1)/2个子集分别由n(n-1)/2个单输出神经网络采用误差反传算法一一进行学习。所有单输出神经网络结构均为单隐层的，输入节点数m＝16，隐节点数s 1＝8，输出节点数为1。目标输出采用{0.0,3.0}编码，所有隐节点和输出节点的活化函数均为
Figure PCTCN2018079389-appb-000005
例如，品牌ω j和ω k这个二分类训练子集为X jk＝{X j,X k}，由烟草电子鼻仪器测试这两个品牌的全部标准试样而得到，样本数为N jk＝N j+N k＝60，由单输出神经网络
Figure PCTCN2018079389-appb-000006
学习，学习步长为η jk＝10/N jk＝0.17。
对训练样本x p＝(x p1,,…,x pi,…,x p16) T∈R 16，单输出神经网络
Figure PCTCN2018079389-appb-000007
隐节点h的实际输出为：
Figure PCTCN2018079389-appb-000008
式中，
Figure PCTCN2018079389-appb-000009
为隐节点h对样本x p所有输入分量的加权和：
Figure PCTCN2018079389-appb-000010
其中，
Figure PCTCN2018079389-appb-000011
为阈值，常数项x p0＝+6.0。
对训练样本x p，单输出神经网络
Figure PCTCN2018079389-appb-000012
的实际输出为：
Figure PCTCN2018079389-appb-000013
式中，
Figure PCTCN2018079389-appb-000014
为神经网络
Figure PCTCN2018079389-appb-000015
输出节点对所有隐节点实际输出的加权和，即：
Figure PCTCN2018079389-appb-000016
其中，
Figure PCTCN2018079389-appb-000017
为阈值，常数项
Figure PCTCN2018079389-appb-000018
在模块化神经网络级联模型第二级—n×5个单输出神经网络学习阶段，训练集X被分解成n个训练子集。每个训练子集由一个品牌的全部样本所组成，即N j＝30。每个得分预测组由5个单输出神经网络组成，分别拟合经比例变换的气敏传感器阵列响应与对应品牌香气、协调、杂气、刺激性、余味这5个感官质量指标得分之间的非线性关系。每个单输出神经网络结构为单隐层的，输入节点数m＝16，隐节点数s 2＝8，输出节点数为1。所有隐节点和输出节点活化函数仍为
Figure PCTCN2018079389-appb-000019
学习步长为η j＝5/N j＝0.17，学习算法仍为误差反传算法。
例如，训练子集X j只由来自品牌ω j的全部N j＝30个样本组成，得分预测组Λ j的5个单输出神经网络分别拟合X j与品牌ω j的香气、协调、杂气、刺激性、余味这5项感官质量指标得分之间的非线性关系。X j的目标输出为品牌ω j的质量指标感官得分经比例变换到[0.15,2.85]之范围。
对来自品牌ω j的样本x p，设第r个质量指标感官得分为
Figure PCTCN2018079389-appb-000020
则第r个单输出神经网络经比例变换后的目标输出为：
Figure PCTCN2018079389-appb-000021
Figure PCTCN2018079389-appb-000022
为品牌ω j所有N j＝30个标准试样的第r个质量指标感官得分向量。
得分预测组Λ j第r个单输出神经网络的隐节点h的实际输出为
Figure PCTCN2018079389-appb-000023
式中，
Figure PCTCN2018079389-appb-000024
为隐节点h对样本x p所有输入分量的加权和，即：
Figure PCTCN2018079389-appb-000025
其中，
Figure PCTCN2018079389-appb-000026
为阈值，常数项x p0＝+6.0。
得分预测组Λ j第r个单输出神经网络的实际输出为
Figure PCTCN2018079389-appb-000027
式中，
Figure PCTCN2018079389-appb-000028
为第r个单输出神经网络输出节点对所有隐节点实际输出的加权和，即
Figure PCTCN2018079389-appb-000029
其中，
Figure PCTCN2018079389-appb-000030
为阈值，常数项
Figure PCTCN2018079389-appb-000031
对烟草与烟草制品进行①识别的模块化神经网络级联模型第一级—每(n-1)个单输出神经网络组成一个投票识别组，代表一个烟草与烟草制品品牌，最高得票数为n-1。每个单输出神经网络必须且仅参 加其中2个投票识别组，n(n-1)/2个单输出神经网络因此分别组成n个投票识别组，并采用大多数投票(majority vote)规则进行决策。
例如，单输出神经网络
Figure PCTCN2018079389-appb-000032
必须参加第j、k两个投票识别组Ω j和Ω k的投票。在第j组里，若
Figure PCTCN2018079389-appb-000033
的实际输出y (jk)>1.5，则预测待定试样x属于品牌ω j的可能性假设得1票。在第k组里，若y (jk)<1.5，则预测x属于品牌ω k的可能性假设得1票。
对样本x进行识别的决策规则是，x属于得票数最多的那个投票识别组所代表的品牌。若两个或以上投票识别组投出的得票数相等且均为最高票数，则决策：x不属于现有任一品牌。
对烟草与烟草制品②感官质量指标得分预测的模块化神经网络级联模型第二级—每5个单输出神经网络组成一个得分预测组，分别负责预测一个对应品牌的香气、协调、杂气、刺激性、余味这5项感官质量指标得分。n×5个单输出神经网络分成n个得分预测组，与n个投票识别组一一对应。
在预测样本x的5项感官质量指标得分阶段，在模块化神经网络级联模型第一级的投票识别组Ω j得票数最多的前提下，只需级联模型第二级—代表品牌ω j的得分预测组Λ j参加预测即可，其它得分预测组不需参加。
设得分预测组Λ j第r个单输出神经网络的实际输出为z (jr)，则x属于品牌ω j第r个感官质量指标得分预测值为：
Figure PCTCN2018079389-appb-000034
若在现有n种品牌的基础上，增加识别一种新的品牌，只须增加n个单输出神经网络并学习，模块化神经网络级联模型第一级从现有n(n-1)/2个单输出神经网络增加到n(n+1)/2个。例如，对增加的品牌ω n+1，新增加并学习的单输出神经网络模块为
Figure PCTCN2018079389-appb-000035
相应地，为了对新增加品牌进行感官质量指标得分预测，模块化神经网络级联模型第二级新增加5个单输出神经网络并学习，从现有n×5个增加到(n+1)×5个。假的品牌或另一厂家生产的现有同一品牌均被看成单独的一种品牌进行识别和感官质量指标得分预测。
烟草电子鼻仪器对卷烟试样的检测、识别和感官质量指标得分预测，包括以下步骤：
(1)开机：仪器预热30分钟，环境空气以6500毫升/分钟的流量依次流经第一二位二通电磁阀、气敏传感器阵列环形工作腔、第五二位二通电磁阀、第四二位二通电磁阀，最终排出到室外。气敏传感器阵列恒温室内温度从室温达到恒定的55±0.1℃。
(2)烟气采样周期开始：操作人员点击屏幕下拉菜单的“开始检测”按钮，仪器进入历时5分钟的烟气采样周期，计算机在指定文件夹自动生成一个名为“xxx”的文本文件，以记录气敏传感器阵列对烟气的响应数据。
(3)初步恢复：在烟气采样周期第0.00-210.00秒，环境空气以6500毫升/分钟的流量依次流经第一二位二通电磁阀、气敏传感器阵列环形工作腔、第五二位二通电磁阀、第四二位二通电磁阀，最终排出到室外。在流量为6500毫升/分钟的环境空气冲洗作用下，粘附在气敏传感器敏感膜表面和管道内壁的烟气气味分子被初步冲走，气敏传感器阵列初步恢复到基准状态，历时210秒。
(4)洁净空气精确标定：在烟气采样周期第210.00-250.00秒，(a)洁净空气精确标定和(b)卷烟插入、自动点火、第一口烟气抽吸、阴/自燃这两个环节同时进行，均历时40秒：
(4)a洁净空气精确标定：洁净空气以17.5毫升/秒即1050毫升/分钟的流量依次流过第二节流阀、第六二位二通电磁阀、气敏传感器阵列环形工作腔、第三二位二通电磁阀，最后排出到室内，历时40秒。洁净空气使气敏传感器阵列精确恢复到基准状态。
(4)b.1卷烟插入：在烟气采样周期第210.00-225.00秒即洁净空气精确标定状态的最初15秒内，屏幕显示“卷烟插入”字样，操作人员将卷烟试样滤嘴端插入卷烟夹持器，插入深度为9.0±0.5毫米。
(4)b.2自动点火：在烟气采样周期第225.00-231.00秒即洁净空气精确标定状态第15.00-21.00秒，点 火线圈通电。与此同时，点烟头向左移动9毫米使点火线圈与被测试样接触并点燃试样，持续6秒。在烟气采样周期第231.00-269.00秒，电磁线圈通电，点火线圈断电并回复到基准位置，历时38秒，包括第一口烟气抽吸的后1秒(第2秒)、阴/自燃的18秒、平衡的2秒，第二口烟气抽吸的2秒，残留烟蒂取出操作的15秒。
(4)b.3第一口烟气抽吸：在烟气采样周期第230.00-232.00秒即洁净空气精确标定状态第20.00-22.00秒，在微型真空泵抽吸作用下，烟气以17.5毫升/秒即1050毫升/分钟的流量，依次经由第二二位二通电磁阀、第一节流阀、流量计，直接被排出到室外，持续2秒。
(4)b.4阴/自燃：在烟气采样周期第232.00-250.00秒即洁净空气精确标定状态第22.00-40.00秒，第二二位二通电磁阀断开，卷烟进入阴/自燃状态，历时18秒。
(5)平衡：在烟气采样周期第250.00-252.00秒，所有电磁阀均处于断开状态，气敏传感器阵列环形工作腔内无气体流动，卷烟仍处于阴/自燃状态，历时2秒。
(6)第二口烟气抽吸：在烟气采样周期第252.00-254.00秒，第一二位二通电磁阀和第五二位二通电磁阀导通，其余四个二位二通电磁阀均断开，卷烟烟气以17.5毫升/秒即1050毫升/分钟的流量，依次通过第一二位二通电磁阀、气敏传感器阵列环形工作腔、第五二位二通电磁阀、第一节流阀、流量计，最后被排出到室外，持续2秒，采集烟气35毫升。
(7)环境空气冲洗：在烟气采样周期第254.00-300.00秒，室内空气以6500毫升/分钟的流量流过气敏传感器阵列环形工作腔，粘附在气敏传感器敏感膜表面和管道内壁的烟气气味分子被初步冲走，气敏传感器阵列进入初步恢复状态。其中，
(7.1)残留烟蒂取出操作：在烟气采样周期第254.00-269.00秒，操作人员在15秒钟内取出残留烟蒂并丢弃。在此期间，第三二位二通电磁阀、第四二位二通电磁阀、第五二位二通电磁阀导通，其余三个二位二通电磁阀均断开，室内环境空气以6500毫升/分钟的流量，依次通过第三二位二通电磁阀、气敏传感器阵列环形工作腔、第五二位二通电磁阀、第四二位二通电磁阀，最终排出到室外，历时15秒。
(7.2)残留烟蒂取出后：在烟气采样周期第269.00-300.00秒，第一二位二通电磁阀、第四二位二通电磁阀、第五二位二通电磁阀导通，其余三个二位二通电磁阀均断开，环境空气以6500毫升/分钟的流量依次流经第一二位二通电磁阀、气敏传感器阵列环形工作腔、第五二位二通电磁阀、第四二位二通电磁阀，最终排出到室外，历时31秒。这一阶段的电磁阀位置和环境空气流动状况与气敏传感器阵列初步恢复状态完全相同。
(8)数据记录：从烟气采样周期第250.00秒起，即从平衡状态开始之刻起，计算机通过16通道16位高精度数据采集卡将16个气敏传感器产生的电压响应存储在“xxx”文本文件里，直至烟气采样周期第290.00秒即环境空气冲洗阶段第36.00秒止，包括第二口烟气抽吸、残留烟蒂取出操作、残留烟蒂取出后这3个过程，数据记录时长为40秒。
(9)特征提取：在一个烟气采样周期内，计算机从时长40秒的“xxx”数据记录文件里，提取各个气敏传感器电压响应稳态最大值作为特征分量，本质上是对第二口烟气的响应，一个被测烟草制品试样因此转化为一个16维的测量样本，并存入计算机硬盘的烟草与烟草制品样本数据集文件中。
(10)识别和感官质量指标得分预测：在烟气采样周期第290.00-300.00秒即数据记录结束的10秒内，模块化神经网络级联模型第一级—n个投票识别组依据大多数投票规则确定试样x的品牌、产地与真假，模块化神经网络级联模型第二级—与获胜投票识别组对应的那个得分预测组预测x的香气、协调、杂气、刺激性、余味这5项感官质量指标得分值，并通过显示器显示出来。
重复步骤(2)～(10)，烟草电子鼻仪器实现对多个烟草与烟草制品试样烟气的测试、识别和感官质量指标得分预测。一个完整的烟草与烟草制品试样测试周期为300秒。
附图说明
图1是本发明—一种电子鼻仪器和烟草与烟草制品感官质量评价方法—烟草电子鼻仪器工作原理图(第二口烟气抽吸状态)。
图2是本发明—一种电子鼻仪器和烟草与烟草制品感官质量评价方法—烟草电子鼻仪器工作原理图(第二口烟气抽吸状态)—主要零部件编号图。
图3是本发明—一种电子鼻仪器和烟草与烟草制品感官质量评价方法—电磁阀与气路通断状态示意图(洁净空气精确标定、卷烟试样自动点火和第一口烟气抽吸状态)。
图4是本发明—一种电子鼻仪器和烟草与烟草制品感官质量评价方法—电磁阀与气路通断状态示意图(第二口烟气抽吸状态)。
图5是本发明—一种电子鼻仪器和烟草与烟草制品感官质量评价方法—电磁阀与气路通断状态示意图(洁净空气精确标定、试样插入和阴/自燃状态)。
图6是本发明—一种电子鼻仪器和烟草与烟草制品感官质量评价方法—电磁阀与气路通断状态示意图(环境空气冲洗，残留烟蒂取出)。
图7是本发明—一种电子鼻仪器和烟草与烟草制品感官质量评价方法—电磁阀与气路通断状态示意图(气敏传感器初步恢复)。
图8是本发明—一种电子鼻仪器和烟草与烟草制品感官质量评价方法—一个采样周期内，单个气敏传感器响应和气敏传感器阵列环形工作腔内气体流量与持续时间变化情况示意图。
图9是本发明—一种电子鼻仪器和烟草与烟草制品感官质量评价方法—自动点火装置示意图(电磁线圈断电状态)。
图10是本发明—一种电子鼻仪器和烟草与烟草制品感官质量评价方法—自动点火装置示意图(电磁线圈通电状态)。
图11是本发明—一种电子鼻仪器和烟草与烟草制品感官质量评价方法—一个采样周期内，6个二位二通电磁阀、点火线圈、电磁线圈通断与持续时间示意图。
图12是本发明—一种电子鼻仪器和烟草与烟草制品感官质量评价方法—烟草电子鼻仪器外观立体示意图。
图13是本发明—一种电子鼻仪器和烟草与烟草制品感官质量评价方法—烟草电子鼻仪器正面。
图14是本发明—一种电子鼻仪器和烟草与烟草制品感官质量评价方法—烟草电子鼻仪器背面。
图15是本发明—一种电子鼻仪器和烟草与烟草制品感官质量评价方法—在区间
Figure PCTCN2018079389-appb-000036
修正的sigmoid活化函数
Figure PCTCN2018079389-appb-000037
和标准的Sigmoid活化函数
Figure PCTCN2018079389-appb-000038
一阶偏导数曲线及其比值变化曲线。
Figure PCTCN2018079389-appb-000039
的一阶偏导数
Figure PCTCN2018079389-appb-000040
曲线(实线)与
Figure PCTCN2018079389-appb-000041
Figure PCTCN2018079389-appb-000042
的一阶偏导数
Figure PCTCN2018079389-appb-000043
曲线(虚线)；(b)一阶偏导数比值
Figure PCTCN2018079389-appb-000044
曲线。
图16是本发明—一种电子鼻仪器和烟草与烟草制品感官质量评价方法—确定一个待定试样的品牌时，n(n-1)/2个单输出神经网络分成n个投票识别组投票的情况。
图17是本发明—一种电子鼻仪器和烟草与烟草制品感官质量评价方法—由5个单输出神经网络模块组成的感官质量指标得分预测组Λ j。
图18是本发明—一种电子鼻仪器和烟草与烟草制品感官质量评价方法—对一个烟草与烟草制品试样同时进行①识别和②感官质量指标得分预测的模块化神经网络级联模型决策过程示意图。
具体实施方式
下面结合附图对本发明作进一步的详细描述。
图1是本发明的烟草电子鼻仪器工作原理图，此时的气路和电磁阀的位置为第二口烟气抽吸工作状态。烟草电子鼻仪器包括气敏传感器阵列模块I、烟气自动进样系统II、计算机控制与数据分析系统III、自动点火装置IV和被测卷烟试样V。图2是本发明的烟草电子鼻仪器主要零部件编号图。
气敏传感器阵列模块I主要组成单元包括：气敏传感器阵列I-1、气敏传感器阵列环形工作腔I-2、保温层I-3、以及电阻加热丝和风扇，位于烟草电子鼻仪器右上部。其中，气敏传感器阵列主要由TGS800和TGS2000系列16个气敏传感器组成，可用型号包括TGS800、TGS813、TGS816、TGS821、TGS822、TGS823、TGS826、TGS830、TGS832、TGS2600、TGS2602、TGS2603、TGS2610、TGS2611、TGS2612、TGS2620、TGS3830、TGS2201，以及PID-A1光离子检测器。气敏传感器阵列模块I的作用是将复杂成分的卷烟烟气转化为0～10V的模拟电压信号。
烟气自动进样系统II组成单元包括：第一二位二通电磁阀II-1，第二二位二通电磁阀II-2，第三二位二通电磁阀II-3，第四二位二通电磁阀II-4，第五二位二通电磁阀II-5，第六二位二通电磁阀II-6，卷烟夹持器(烟嘴)II-7，微型真空泵II-8，第一节流阀II-9，流量计II-10，第二节流阀II-11，洁净空气II-12，烟灰盘II-13，以及溢出烟气排出装置II-14。烟气自动进样系统II位于烟草电子鼻仪器右下部。
计算机控制与数据分析系统III主要组成单元为计算机主板III-1、16路数据采集卡III-2、驱动与控制模块III-3、多路直流电压模块III-4、显示器III-5，以及硬盘、网卡、显卡、鼠标、键盘等，位于烟草电子鼻仪器左部。计算机控制与数据分析系统III的主要作用是，(1)气敏传感器阵列响应信号的采集、分析与处理；(2)烟气自动进样系统II的多个二位二通电磁阀与微型真空泵、自动点火装置IV、以及计算机控制与数据分析系III自身的驱动与控制。
图3是洁净空气精确标定与卷烟试样自动点火和第一口烟气抽吸时，电磁阀与气路通断状态示意图。在这一状态下，烟草电子鼻仪器要完成被测卷烟试样V的点火、烟气自动进样系统II的第一口烟气抽吸、气敏传感器阵列I的洁净空气精确标定这几个动作。
在卷烟试样V自动点火和烟气自动进样系统II抽吸第一口烟气状态，第二二位二通电磁阀II-2导通，第一二位二通电磁阀II-1、第四二位二通电磁阀II-4和第五二位二通电磁阀II-5断开。在自动点火状态，自动点火装置IV的点火线圈IV-1通电，电磁线圈IV-4断电，阀芯IV-3在压缩弹簧IV-6的作用下水平向左移动9毫米，使得点火线圈IV-1与被测卷烟试样V接触，持续时间5秒；随后，在第一口烟气抽吸状态，卷烟烟气完全不通过气敏传感器阵列模块I，而是在微型真空泵II-8的抽吸作用下，经由第二二位二通电磁阀II-2、第一节流阀II-9、流量计II-10，以17.5毫升/秒即1050毫升/分钟的流量直接被排出到室外，持续2秒，相当于抽吸烟气35毫升。
在卷烟试样V自动点火和烟气自动进样系统II抽吸第一口烟的同时，气敏传感器阵列I-1处于洁净空气精确标定状态，二位二通电磁阀II-3和II-6导通，洁净空气以1050毫升/分钟的流量依次流过第二节流阀II-11、第六二位二通电磁阀II-6、气敏传感器阵列环形工作腔I-2、第三二位二通电磁阀II-3，最后排出到室内大气中。这也是洁净空气精确标定第2阶段，持续7秒。
图4是电子鼻仪器抽吸第二口烟气时，电磁阀与气路通断状态示意图。在这一状态，第一二位二通电磁阀II-1和第五二位二通电磁阀II-5导通，其余四个二位二通电磁阀均断开，卷烟烟气以17.5毫升/秒即1050毫升/分钟的流量，依次通过第一二位二通电磁阀II-1、气敏传感器阵列I-1及其环形工作腔I-2、第五二位二通电磁阀II-5、第一节流阀II-9、流量计II-10，最后被排出到室外，持续2秒，相当于采集烟气35毫升。这时，自动点火装置IV的点火线圈IV-1断电，电磁线圈IV-4通电并迫使阀芯IV-3水平向右移动9毫米，使得点火线圈IV-1与被测卷烟试样V脱离接触。
在第二口卷烟烟气流动过程中，气敏传感器阵列I-1产生敏感响应。对一个具体的烟草与烟草制品 试样p，单个气敏传感器i的电压响应曲线稳态最大值被提取作U pi(max)为特征分量x pi’，即x pi’＝U pi(max)，16个气敏传感器组成的阵列I-1因此产生一个16维的电压响应向量x p’＝(x p1’,x p2’,…,x pi’,…,x p16’) T∈R 16。这一16维的电压响应稳态最大值向量是烟草电子鼻仪器对烟草与烟草制品进行识别和感官质量指标预测的依据。
图5是卷烟试样V插入卷烟夹持器II-7和阴/自燃状态，电磁阀与气路通断状态示意图。这一阶段，第三二位二通电磁阀II-3和第六二位二通电磁阀II-6导通，其余四个二位二通电磁阀均断开。卷烟插入是人工操作，定时15秒；卷烟阴/自燃持续时间18+2秒，可调节。这里，“+2”秒指的是平衡时间。
在此期间，气敏传感器阵列I-1处于洁净空气精确标定状态，二位二通电磁阀II-3和II-6导通，洁净空气以1050毫升/分钟的流量依次流过第二节流阀II-11、第六二位二通电磁阀II-6、气敏传感器阵列I-1及其环形工作腔I-2、第三二位二通电磁阀II-3，最后排出到室内大气中。这一阶段被称为洁净空气精确标定第1阶段，持续15+18＝33秒。这一过程与图3的卷烟试样V点火和烟气自动进样系统II抽吸第一口烟过程一起，使得洁净空气精确标定状态总时长为40秒。
图6是残留烟蒂取出期间，电磁阀与气路通断状态示意图。在这一阶段，第三二位二通电磁阀II-3、第四二位二通电磁阀II-4和第五二位二通电磁阀II-5导通，其余三个二位二通电磁阀均断开，室内环境空气以6500毫升/分钟的流量，依次通过第三二位二通电磁阀II-3、气敏传感器阵列I-1及其环形工作腔I-2、第五二位二通电磁阀II-5、第四二位二通电磁阀II-4，最终排出到室外大气中。取出残留烟蒂的人工操作时间长度为15秒(可调)。这一阶段是环境空气冲洗第一阶段，可以被看成气敏传感器阵列I-1初步恢复状态的一部分。
图7是残留烟蒂取出后，电磁阀与气路通断状态示意图。在这一阶段，第一二位二通电磁阀II-1、第四二位二通电磁阀II-4和第五二位二通电磁阀II-5导通，其余三个二位二通电磁阀均断开。环境空气以6500毫升/分钟的流量，依次通过卷烟夹持器II-7、第一二位二通电磁阀II-1、气敏传感器阵列I-1及其环形工作腔I-2、第五二位二通电磁阀II-5、第四二位二通电磁阀II-4，最终排出到室外大气中。
这一阶段是环境空气冲洗的第二阶段，历时31秒，仍可以被看成气敏传感器阵列I-1初步恢复状态的另一部分。环境空气冲洗这一阶段一结束，新的烟气采样周期就自动开始，或者操作人员点击屏幕下拉菜单的“结束检测”按钮，检测过程人为结束。
在周期为300秒的烟草电子鼻仪器烟气采样与分析过程中，环境空气冲洗和气敏传感器阵列I-1初步恢复这两个阶段总时间长度为300-44＝256秒，其目的是，(1)环境空气以6500毫升/分钟的较大流量冲洗掉第二口烟抽吸期间黏附在气敏传感器敏感膜表面、气敏传感器环形工作腔I-2内壁以及管道内壁的烟气分子；(2)带走气敏传感器工作时积聚的热量，使气敏传感器初步恢复到初始状态。
图8是在一个烟气采样周期内，烟草与电子鼻仪器的单个气敏传感器电压响应变化情况和气敏传感器阵列I-1所处的环形工作腔I-2内气体流量与持续时间变化情况示意图。在一个烟气采样周期内，气体流量经历了6500毫升/分钟、1050毫升/分钟和0(平衡)三种变化；气体类型经历了环境空气初步冲洗、洁净空气精确标定和卷烟烟气采样三种变化。在平衡状态下，所有6个二位二通电磁阀均断开，气敏传感器阵列I-1及其环形工作腔I-2内无气体流动。
一个卷烟试样V的烟气采样周期为300秒。从平衡状态开始之刻起，即烟气采样周期第250秒开始，计算机控制与数据分析系统III记录气敏传感器阵列瞬态电压响应数据，记录时长为40秒，包括平衡阶段2秒、第二口烟气采样阶段2秒、环境空气冲洗阶段前36秒，气敏传感器阵列I-1对卷烟烟气的电压响应即采样数据保存在一个文本文件内。在数据记录40秒长度内，气敏传感器i对卷烟试样p的烟气的电压响应稳态最大值被提取作为特征分量x pi’，由此得到气敏传感器阵列对第p个卷烟试样第二口烟气的响应，称之为电压响应样本x p’∈R 16。在数据记录结束后的10秒钟内，计算机控制与数据分析系统III依据样本x p’给出试样p的品牌识别结果和香气、协调、杂气、刺激性、余味等5项感官指标得分预测结果。
为什么将卷烟烟气采样流量定为17.5毫升/秒即1050毫升/分钟和采样时长定为2秒？大量统计结果指出，人抽一口烟气容积为35毫升，平均时长2秒。国家标准《常规分析用吸烟机—定义和标准条件》GB/T 16450-2004规定，卷烟烟气标准抽吸容量为35毫升，流量为17.5毫升/秒，标准的单口抽吸持续时间为2.00±0.02秒。本发明的卷烟烟气采样流量、容积和持续时间与国标GB/T 16450一致。为统一起见，洁净空气流量仍确定为1050毫升/分钟。
图9是自动点火装置IV组成单元示意图(电磁线圈IV-4断电时)。自动点火装置IV位于烟草电子鼻仪器右前下方，组成单元包括：点火线圈IV-1、点烟头IV-2、动铁芯IV-3、电磁线圈IV-4、导磁铁架IV-5、压缩弹簧IV-6、弹簧座IV-7、电缆IV-8、支座IV-9。点火线圈IV-1和电磁线圈IV-4工作电压为直流24V，点火线圈最大工作电流5A。在抽吸第一口烟前第5秒钟开始，点火线圈IV-1通电，并在5秒钟内升温到380℃，此时电磁线圈IV-4通电。与此同时，电磁线圈IV-4断电，在压缩弹簧IV-6的作用下，固定在动铁芯IV-3上的点火线圈IV-1和点烟头IV-2自基准位置水平向左移动9毫米，与被测卷烟试样V的烟头接触并点燃。
图10是自动点火装置IV组成单元示意图(电磁线圈IV-4通电时)。在这一状态，电磁线圈IV-4通电，点火线圈IV-1断电。在电磁线圈IV-4电磁力的作用下，压缩弹簧IV-6被压缩9毫米，固定在动铁芯IV-3上的点火线圈IV-1和点烟头IV-2回复到基准位置，为人工插入被测卷烟试样和人工取出残留烟蒂做准备。
必须指出，只要卷烟试样V不被插入卷烟夹持器Ⅱ-7，点火线圈IV-1和电磁线圈IV-4就一直处于断电状态，在压缩弹簧IV-6的作用下，点火线圈IV-1和点烟头IV-2均处于基准位置水平左移9毫米的位置。在一个烟气采样周期内，电磁线圈IV-4通电的时间段是：人工将卷烟试样V插入卷烟夹持器II-7持续的15秒；抽吸第一口烟气的后1秒(第2秒)；卷烟试样V阴/自燃持续的18秒；平衡状态的2秒；抽吸第二口烟气的2秒；取出残留烟蒂的15秒，共53秒。在卷烟试样V点火和烟气自动进样系统Ⅱ抽吸第一口烟期间的前1秒共持续6秒的时间内，点火线圈IV-1通电，其余时间均断电。
图11是一个烟气采样周期内，6个二位二通电磁阀、点火线圈IV-1与电磁线圈IV-4通断关系示意图。图11(a)表明，二位二通电磁阀II-1除第二口烟气抽吸的2秒期间必须导通以外，在最初210秒和最后31秒这两个时间段也是通电的，便于环境空气以较长的时间冲洗卷烟夹持器II-7、气敏传感器阵列敏感膜表面和管道内壁的烟气分子。
图11(b)和(f)说明，二位二通电磁阀II-2只控制第一口烟气的抽吸，二位二通电磁阀II-6只控制洁净空气的通与断，二者的作用较为单一。根据图11(c)，二位二通电磁阀II-3在洁净空气精确标定的40秒和第二口烟气抽吸结束后的15秒这两个阶段是导通的，但气体流动方向和流量是不同的。在前一阶段的40秒期间，流量为1050毫升/分钟的洁净空气在流过气敏传感器阵列环形工作腔I-2后，经由二位二通电磁阀II-3，被排出到室内大气中。在后一阶段的15秒期间，流量为6500毫升/分钟的室内环境空气先从二位二通电磁阀II-3流入气敏传感器阵列环形工作腔I-2，然后经由二位二通电磁阀II-5和二位二通电磁阀II-4被排出到室外。
图11(d)表明，二位二通电磁阀II-4在洁净空气精确标定、平衡、第二口烟气抽吸这三个阶段是断开的，从而强迫第一口烟气、第二口烟气以1050毫升/分钟的流量流过节流阀Ⅱ-9。二位二通电磁阀II-4主要控制气体流量在6500毫升/分钟与1050毫升/分钟之间变换。比较图11(e)与图11(d)，二位二通电磁阀II-5与二位二通电磁阀II-4只是在第二口烟抽吸阶段的状态不一致。在此期间，二位二通电磁阀II-5导通，二位二通电磁阀II-4断开，从而强迫第二口烟气以1050毫升/分钟的流量流过节流阀Ⅱ-9。
图11(g)指出，点火线圈IV-1只是在点火前5秒和第一口烟气抽吸期间的第1秒共连续6秒期间通电，其余时间均断开。图11(h)表明，电磁线圈IV-4在卷烟插入期间的15秒、第一口烟气抽吸的第2秒、阴/自燃的18秒、平衡的2秒、第二口烟气抽吸的2秒、残留烟蒂取出期间的15秒，共53秒时间通电，其余时间均断开。
图12是烟草电子鼻仪器立体示意图。由此图可以看到，气敏传感器阵列模块I位于烟草电子鼻右上部；计算机控制与数据分析系统III位于烟草电子鼻左部；烟气自动进样系统II和自动点火装置IV位于烟草电子鼻右下部。
图13是烟草电子鼻仪器正面示意图。根据此图，外部可以看到的是，计算机控制与数据分析系统III的显示器III-5；烟气自动进样系统II的卷烟夹持器II-7、烟气溢出装置II-14的排风扇、自动点火装置IV的点烟头IV-2，和被测卷烟试样V。
图14是烟草电子鼻仪器背面示意图。电子鼻仪器配有外接显示器接口、2个USB接口、鼠标接口、键盘接口、Internet接口；配有环境空气和洁净空气入口；从烟气自动进样系统II排出的烟气出口；以及从烟气排出装置II-14直接排出的烟气。用户可便捷地接插所需的外部设备，例如大屏幕显示器、键盘、鼠标，进行数据传输、交换和Internet远程传输。
图15是在区间
Figure PCTCN2018079389-appb-000045
修正的sigmoid活化函数
Figure PCTCN2018079389-appb-000046
和标准的Sigmoid活化函数
Figure PCTCN2018079389-appb-000047
一阶偏导数曲线及其比值变化曲线。图15(a)为这两个活化函数在此区间的一阶偏导数
Figure PCTCN2018079389-appb-000048
与
Figure PCTCN2018079389-appb-000049
曲线，实线为修正的活化函数
Figure PCTCN2018079389-appb-000050
的一阶偏导数
Figure PCTCN2018079389-appb-000051
曲线，虚线为标准Sigmoid活化函数
Figure PCTCN2018079389-appb-000052
的一阶偏导数
Figure PCTCN2018079389-appb-000053
曲线。图15(b)为修正的与标准的sigmoid活化函数的一阶偏导数的比值
Figure PCTCN2018079389-appb-000054
曲线。为什么采用
Figure PCTCN2018079389-appb-000055
这一修正的sigmoid活化函数？因为单隐层神经网络采用误差反传算法，在不发生振荡的前提下，神经网络实际输出与期望输出之间的误差平方和函数关于权值与阈值参数的偏导数(一阶梯度)越大，神经网络学习速度越快。
图15表明，活化函数
Figure PCTCN2018079389-appb-000056
的一阶梯度远比标准Sigmoid活化函数
Figure PCTCN2018079389-appb-000057
的大。当
Figure PCTCN2018079389-appb-000058
时，二者的一阶梯度比值达ρ＝732.63。与之相匹配，输入分量变换到[0,6]范围。这种做法的考虑是，无论样本分布状态如何，训练数据集X各分量均值均在3.0附近。采用标准Sigmoid活化函数
Figure PCTCN2018079389-appb-000059
的神经网络一般将数据集变换到[0,1]范围，这一做法实际默认分量均值约为0.5。与之相比，输入分量变换到[0,6]范围的优点是，样本间隔和类间间隔较原来的[0,1]被放大6倍，有利于神经网络在不发生振荡的前提下加快学习速度，提高学习精度和推广能力。
识别2种卷烟品牌{ω j,ω k}的单输出神经网络模块
Figure PCTCN2018079389-appb-000060
有m＝16个输入节点，s 1＝8个隐节点，1个输出节点；所有输入分量正比例变换到[0,6]范围，目标输出采用{0.0,3.0}编码；所有隐节点和输出节点活化函数均为
Figure PCTCN2018079389-appb-000061
神经网络模块
Figure PCTCN2018079389-appb-000062
采用误差反传算法进行学习，学习因子为η jk＝10/N jk＝0.17。
本发明采用一对一(one-against-one,OAO)任务分解方法。烟草训练集X被分解成
Figure PCTCN2018079389-appb-000063
个二分类(binary-class)子问题。这n(n-1)/2个子问题分别由n(n-1)/2个单输出神经网络模块一一学习和解决，由此组成模块化神经网络级联模型第一级。表3给出了参加学习的n(n-1)/2个单输出神经网络列表
Figure PCTCN2018079389-appb-000064
表3,为识别n个烟草与烟草制品品牌，参加学习的n(n-1)/2个单输出神经网络列表
Figure PCTCN2018079389-appb-000065
表4,确定n种烟草与烟草制品品牌时，参加投票的n(n-1)/2个单输出神经网络分组列表
Figure PCTCN2018079389-appb-000066
为了待定试样x的品牌，模块化神经网络级联模型第一级的n(n-1)/2个单输出神经网络采用大多数投票规则决策。表4给出了为确定待定试样x属于n种烟草与烟草制品的何种品牌时，参加投票的n(n-1)/2个单输出神经网络分组列表。表3是主对角对称的，一个单输出神经网络模块须参加2个投票组。例如，单输出神经网络模块
Figure PCTCN2018079389-appb-000067
既参加投票识别组Ω j的投票，也参加投票识别组Ω k的投票。
图16是本发明—一种电子鼻仪器和烟草与烟草制品感官质量评价方法—确定待定卷烟试样x的品牌时，n(n-1)/2个单输出神经网络分成n个投票识别组参加投票的情况，每个组有n-1个成员。
得分预测组共有n个，与投票识别组一一对应，每个得分预测组由5个单输出神经网络组成。图17是本发明—一种电子鼻仪器和烟草与烟草制品感官质量评价方法—由5个单输出神经网络模块组成的感官质量指标得分预测组Λ j。5个成员分别预测属于品牌ω j的卷烟试样x的香气、协调、杂气、刺激性、余味这5项感官质量指标得分。每个单输出神经网络结构为，m＝16个输入节点，s 2＝8个隐节点，1个输出节点；所有输入分量正比例变换到[0,6]范围，目标输出采用{0.0,3.0}编码；所有隐节点和输出节点活化函数均为
Figure PCTCN2018079389-appb-000068
目标输出为品牌ω j的质量指标感官得分经比例变换到[0.15,2.85]之范围。学习算法为误差反传算法，学习步长为η j＝0.17。
图18是对一个烟草与烟草制品试样同时进行①识别和②感官质量指标得分预测的模块化神经网络级联模型决策过程示意图。在对烟草试样x进行品牌、产地、真假识别时，第一级的n(n-1)/2个单输出神经网络全部参加，分成n个投票识别组，每组n-1个成员，一个单输出神经网络参加其中的两个投票识别组。n个投票识别组采用大多数投票规则决策，得票最多的组获胜。当两个或以上投票识别组的票数相等且均为最多时，决策规则是：试样x不属于现有任何品牌。对烟草试样x进行质量指标得分预测时，在第二级的5n个单输出神经网络中，只有与获胜投票识别组对应的得分预测组的5个参加，分别预测属于获胜投票识别组所代表的品牌的香气、协调、杂气、刺激性和余味这5项质量指标得分。

Claims (14)

  1. 一种电子鼻仪器和烟草与烟草制品感官质量评价方法，其特征是，烟草电子鼻仪器包括气敏传感器阵列模块、烟气自动进样系统、计算机控制与数据分析系统、自动点火装置，实现烟草及烟草制品的在线检测、识别和感官质量指标得分预测；
    所述的气敏传感器阵列模块由16个SnO 2半导体型气敏传感器组成，均布于中径φ140mm、断面尺寸20mm×16mm的密封腔内，形成气敏传感器阵列环形工作腔；该环形工作腔处于55±0.1℃的恒温室内，位于烟草电子鼻仪器右上方；
    所述的烟气自动进样系统包括卷烟夹持器、微型真空泵、第一至第六共6个二位二通电磁阀、第一和第二共2个节流阀、流量计、气体管道、溢出烟气排出装置，位于烟草电子鼻仪器右下方；
    所述的自动点火装置包括点火线圈、点烟头、动铁芯、电磁线圈、导磁铁架、压缩弹簧、弹簧座、电缆、支座，位于烟草电子鼻仪器右前下方；
    所述的计算机控制与数据分析系统包括计算机主板、数据采集卡、精密线性与开关电源模块、驱动与控制电路模块、硬盘、网卡、显卡、显示器，位于烟草电子鼻仪器左侧；
    一个烟草与烟草制品试样的烟气采样周期为5分钟；依据流经的气体类型的不同，气敏传感器阵列历经初步恢复(210秒)、洁净空气精确标定(40秒)、平衡(2秒)、第二口烟气抽吸(2秒)、环境空气冲洗(46秒)共6个阶段；
    烟气采样时，在计算机控制下，自动点火装置的点火线圈水平向左移动9毫米，以380℃的温度点燃卷烟试样；烟气自动进样系统的微型真空泵以17.5毫升/秒即1050毫升/分钟的流量抽吸第二口烟气，使之流经气敏传感器阵列环形工作腔，掠过气敏传感器敏感膜表面，持续2秒，气敏传感器阵列因此产生敏感响应；自平衡状态开始之刻起，计算机控制与数据分析系统开始记录响应数据；依次记录平衡(2秒)、第二口烟气抽吸(2秒)、环境空气冲洗(前36秒)这3个阶段的气敏传感器阵列电压响应数据，总时长40秒；烟气采样周期其它时间的数据不记录；
    在40秒的数据记录时间内，单个气敏传感器对第二口烟气的电压响应曲线稳态最大值被提取为特征分量，16个气敏传感器组成的阵列因此产生一个16维电压响应向量；在数据记录结束后的10秒内，计算机控制与数据分析系统依据这一响应向量对烟草与烟草制品试样进行品牌、产地、真假识别和香气、协调、杂气、刺激性、余味共5项感官质量指标得分预测；
    所述的计算机控制与数据分析系统采用模块化神经网络级联模型对烟草与烟草制品试样进行①识别和②感官质量指标得分预测；①模块化神经网络级联模型第一级由n(n-1)/2个单输出神经网络并列组成，形成n个投票识别组，用于n种烟草与烟草制品的识别，包括品牌、产地和真假的识别；②模块化神经网络级联模型第二级由n×5个单输出神经网络并列组成，每5个一组，用于n种烟草与烟草制品的香气、协调、杂气、刺激性、余味这5项感官质量指标得分预测。
  2. 根据权利要求1所述的电子鼻仪器和烟草与烟草制品感官质量评价方法，其特征是，卷烟夹持器轴心线与水平面夹角为0°～+5°；在自动点火之前，操作人员将卷烟试样烟蒂插入卷烟夹持器，插入深度为9±0.5mm，插入操作用时15秒；第二口烟气抽吸结束后的15秒内，操作人员从卷烟夹持器上取出残留的卷烟烟蒂，熄灭并扔弃。
  3. 根据权利要求1所述的电子鼻仪器和烟草与烟草制品感官质量评价方法，其特征是，自动点火装置的点火线圈工作电压24V，电流5A；点烟头轴心线、动铁芯轴心线、电磁线圈轴心线、压缩弹簧轴心线与卷烟夹持器轴心线位于同一水平线上；在计算机控制下，在第一口烟气抽吸前第5秒钟起，点火线圈通电，并在5秒钟内升温到380℃；与此同时，电磁线圈断电，在压缩弹簧作用下，固定在点烟头上的点火线圈随动铁芯自基准位置水平左移9毫米，点燃卷烟试样；在抽吸第一口烟气1秒之后，点火线圈断电，电磁线圈通电，在电磁线圈电磁力的作用下，点火线圈与被点燃的卷烟试样脱离接触，并回复到基准位置。
  4. 根据权利要求1所述的电子鼻仪器和烟草与烟草制品感官质量评价方法，其特征是，在微型真空 泵抽吸作用下，第一口烟气以17.5毫升/秒即1050毫升/分钟的流量直接经第二二位二通电磁阀、第一节流阀、流量计直接被排出到室外，完全不经过气敏传感器阵列环形工作腔，持续2秒，抽吸烟气35毫升；接下来，被点燃的卷烟试样阴/自燃20秒。
  5. 根据权利要求1所述的电子鼻仪器和烟草与烟草制品感官质量评价方法，其特征是，在微型真空泵的抽吸作用下，第二口烟气以17.5毫升/秒即1050毫升/分钟的流量经第一二位二通电磁阀、气敏传感器阵列环形工作腔、第五二位二通电磁阀、第一节流阀、流量计，最后被排出到室外，持续2秒，采集烟气35毫升。
  6. 根据权利要求1所述的电子鼻仪器和烟草与烟草制品感官质量评价方法，其特征是，洁净空气精确标定过程与卷烟插入、自动点火、第一口烟气抽吸、阴/自燃过程同时进行，均历时40秒；在洁净空气精确标定阶段，洁净空气以17.5毫升/秒的流量依次流经第二节流阀、第六二位二通电磁阀、气敏传感器阵列环形工作腔、第三二位二通电磁阀，最后排出到室外；以下操作与此依次同时进行：首先，操作人员在15秒钟内将被测卷烟试样插入卷烟夹持器；随后，自动点火装置的点火线圈通电，左移9毫米与卷烟试样接触，并在5秒钟内升温到380℃；接下来是时长为2秒的第一口烟气抽吸；最后是18秒的卷烟试样阴/自燃过程。
  7. 根据权利要求1所述的电子鼻仪器和烟草与烟草制品感官质量评价方法，其特征是，流量为1050毫升/分钟的第二口烟气抽吸过程一结束，流量为6500毫升/分钟的环境空气冲洗过程就随即开始；环境空气依次流过第三二位二通电磁阀、气敏传感器阵列环形工作腔、第五二位二通电磁阀、第四二位二通电磁阀，最终排出到室外，持续15秒；在此期间，操作人员取出残留卷烟烟蒂并丢弃；随后，环境空气以6500毫升/分钟的流量依次流过第一二位二通电磁阀、气敏传感器阵列环形工作腔、第五二位二通电磁阀、第四二位二通电磁阀，最终排出到室外，持续31秒；至此，卷烟试样的一个检测周期结束；若欲检测下一个试样，计算机自动开始新的检测周期，自动转入初步恢复过程；否则，操作人员点击屏幕下拉菜单的“结束检测”按钮强制结束检测过程。
  8. 根据权利要求1所述的电子鼻仪器和烟草与烟草制品感官质量评价方法，其特征是，在数据采集阶段，烟草电子鼻仪器测试经过评吸小组品评并给出感官质量指标得分的烟草与烟草制品标准试样；单个标准试样测量周期为5分钟，提取各个气敏传感器对第二口烟气的电压响应稳态最大值作为特征分量，对第p个标准试样得到一个16维的电压响应向量x p′＝(x p1’，...，x pi’，...，x p16’) T∈R 16；通过对N个标准试样的测试，烟草电子鼻仪器得到气敏传感器阵列电压响应标准样本集X′∈R N×16，并建立X′与香气d 1∈R N、协调d 2∈R N、杂气d 3∈R N、刺激性d 4∈R N、余味d 5∈R N共5个质量指标感官得分之间的一一对应关系；每一品牌A、B、C三个等级各测量10个标准试样；若品牌数为n，则N＝30n。
  9. 根据权利要求1所述的电子鼻仪器和烟草与烟草制品感官质量评价方法，其特征是，气敏传感器阵列响应标准样本集X′的所有特征分量经正比例预处理变换到[0.0，6.0]范围；设气敏传感器i对第p个标准试样的电压响应稳态最大值为x pi’，比例变换后的值为：
    Figure PCTCN2018079389-appb-100001
    这里，max(X′)和min(X′)分别为X′的最大值与最小值，x pi是气敏传感器i对第p个标准试样经比例变换后的电压响应稳态最大值，电压响应向量x p′因此变为16维样本x p＝(x p1，，...，x pi，...，x p16) T∈R 16；max(X′)和min(X′)作为基本数据存入计算机；标准样本集X′经正比例预处理变换后被称为训练集，记为X；
    在待定试样x的识别与感官质量指标得分预测阶段，气敏传感器i电压响应稳态最大值x pi’仍采用标准样本集的max(X′)与min(X′)以公式(1)进行正比例变换；
    模块化神经网络级联模型的每个单输出神经网络均经历对训练集X的学习阶段和对待定样本x的识别与感官质量指标得分预测阶段。
  10. 根据权利要求1所述的电子鼻仪器和烟草与烟草制品感官质量评价方法，其特征是，模块化神经网络级联模型第一级-n(n-1)/2个单输出神经网络学习阶段，首先对训练集X施以一对一(one-against-one，OAO)分解，X被分解成
    Figure PCTCN2018079389-appb-100002
    个二分类(binary-class)训练子集；然后，这nn(n-1)/2个子集分别由n(n-1)/2个单输出神经网络采用误差反传算法一一进行学习；所有单输出神经网络结构均为单隐层的，输入节点数m＝16，隐节点数s 1＝8，输出节点数为1；目标输出采用{0.0，3.0}编码，所有隐节点和输出节点的活化函数均为
    Figure PCTCN2018079389-appb-100003
    例如，品牌ω j和ω k这个二分类训练子集为X jk＝{X j，X k}，由烟草电子鼻仪器测试这两个品牌的全部标准试样而得到，样本数为N jk＝N j+N k＝60，由单输出神经网络
    Figure PCTCN2018079389-appb-100004
    学习，学习步长为η jk＝10/N jk＝0.17；
    对训练样本x p＝(x p1，，...，x pi，...，x p16) T∈R 16，单输出神经网络
    Figure PCTCN2018079389-appb-100005
    隐节点h的实际输出为：
    Figure PCTCN2018079389-appb-100006
    式中，
    Figure PCTCN2018079389-appb-100007
    为隐节点h对样本x p所有输入分量的加权和：
    Figure PCTCN2018079389-appb-100008
    其中，
    Figure PCTCN2018079389-appb-100009
    为阈值，常数项x p0＝+6.0；
    对训练样本x p，单输出神经网络
    Figure PCTCN2018079389-appb-100010
    的实际输出为：
    Figure PCTCN2018079389-appb-100011
    式中，
    Figure PCTCN2018079389-appb-100012
    为神经网络
    Figure PCTCN2018079389-appb-100013
    输出节点对所有隐节点实际输出的加权和，即：
    Figure PCTCN2018079389-appb-100014
    其中，
    Figure PCTCN2018079389-appb-100015
    为阈值，常数项
    Figure PCTCN2018079389-appb-100016
  11. 根据权利要求1所述的电子鼻仪器和烟草与烟草制品感官质量评价方法，其特征是，在模块化神经网络级联模型第二级-n×5个单输出神经网络学习阶段，训练集X被分解成n个训练子集；每个训练子集由一个品牌的全部样本所组成，即N j＝30；每个得分预测组由5个单输出神经网络组成，分别拟合经比例变换的气敏传感器阵列响应与对应品牌香气、协调、杂气、刺激性、余味这5个感官质量指标得分之间的非线性关系；每个单输出神经网络结构为单隐层的，输入节点数m＝16，隐节点数s 2＝8，输出节点数为1；所有隐节点和输出节点活化函数仍为
    Figure PCTCN2018079389-appb-100017
    学习步长为η j＝5/N j＝0.17，学习算法仍为误差反传算法；
    例如，训练子集X j只由来自品牌ω j的全部N j＝30个样本组成，得分预测组Λ j的5个单输出神经网络分别拟合X j与品牌ω j的香气、协调、杂气、刺激性、余味这5项感官质量指标得分之间的非线性关系；X j的目标输出为品牌ω j的质量指标感官得分经比例变换到[0.15，2.85]之范围；
    对来自品牌ω j的样本x p，设第r个质量指标感官得分为
    Figure PCTCN2018079389-appb-100018
    则第r个单输出神经网络经比例变换后的目标输出为：
    Figure PCTCN2018079389-appb-100019
    Figure PCTCN2018079389-appb-100020
    为品牌ω j所有N j＝30个标准试样的第r个质量指标感官得分向量；
    得分预测组Λ j第r个单输出神经网络的隐节点h的实际输出为
    Figure PCTCN2018079389-appb-100021
    式中，
    Figure PCTCN2018079389-appb-100022
    为隐节点h对样本x p所有输入分量的加权和，即：
    Figure PCTCN2018079389-appb-100023
    其中，
    Figure PCTCN2018079389-appb-100024
    为阈值，常数项x p0＝+6.0；
    得分预测组Λ j第r个单输出神经网络的实际输出为
    Figure PCTCN2018079389-appb-100025
    式中，
    Figure PCTCN2018079389-appb-100026
    为第r个单输出神经网络输出节点对所有隐节点实际输出的加权和，即
    Figure PCTCN2018079389-appb-100027
    其中，
    Figure PCTCN2018079389-appb-100028
    为阈值，常数项
    Figure PCTCN2018079389-appb-100029
  12. 根据权利要求1所述的电子鼻仪器和烟草与烟草制品感官质量评价方法，其特征是，对烟草与烟草制品进行①识别的模块化神经网络级联模型第一级-每(n-1)个单输出神经网络组成一个投票识别组，代表一个烟草与烟草制品品牌，最高得票数为n-1；每个单输出神经网络必须且仅参加其中2个投票识别组，n(n-1)/2个单输出神经网络因此分别组成n个投票识别组，并采用大多数投票(majority vote)规则进行决策；
    例如，单输出神经网络
    Figure PCTCN2018079389-appb-100030
    必须参加第j、k两个投票识别组Ω j和Ω k的投票；在第j组里，若
    Figure PCTCN2018079389-appb-100031
    的实际输出y (jk)＞1.5，则预测待定试样x属于品牌ω j的可能性假设得1票；在第k组里，若y (jk)＜1.5，则预测x属于品牌ω k的可能性假设得1票；
    对样本x进行识别的决策规则是，x属于得票数最多的那个投票识别组所代表的品牌；若两个或以上投票识别组投出的得票数相等且均为最高票数，则决策：x不属于现有任一品牌；
    对烟草与烟草制品②感官质量指标得分预测的模块化神经网络级联模型第二级-每5个单输出神经网络组成一个得分预测组，分别负责预测一个对应品牌的香气、协调、杂气、刺激性、余味这5项感官质量指标得分；n×5个单输出神经网络分成n个得分预测组，与n个投票识别组一一对应；
    在预测样本x的5项感官质量指标得分阶段，在模块化神经网络级联模型第一级的投票识别组Ω j得票数最多的前提下，只需级联模型第二级-代表品牌ω j的得分预测组Λ j参加预测即可，其它得分预测组不需参加；
    设得分预测组Λ j第r个单输出神经网络的实际输出为z (jr)，则x属于品牌ω j第r个感官质量指标得分预测值为：
    Figure PCTCN2018079389-appb-100032
  13. 根据权利要求1所述的电子鼻仪器和烟草与烟草制品感官质量评价方法，其特征是，若在现有n 种品牌的基础上，增加识别一种新的品牌，只须增加n个单输出神经网络并学习，模块化神经网络级联模型第一级从现有n(n-1)/2个单输出神经网络增加到n(n+1)/2个；例如，对增加的品牌ω n+1，新增加并学习的单输出神经网络模块为
    Figure PCTCN2018079389-appb-100033
    相应地，为了对新增加品牌进行感官质量指标得分预测，模块化神经网络级联模型第二级新增加5个单输出神经网络并学习，从现有n×5个增加到(n+1)×5个；假的品牌或另一厂家生产的现有同一品牌均被看成单独的一种品牌进行识别和感官质量指标得分预测。
  14. 根据权利要求1所述的电子鼻仪器和烟草与烟草制品感官质量评价方法，其特征是，烟草电子鼻仪器对卷烟试样的检测、识别和感官质量指标得分预测，包括以下步骤：
    (1)开机：仪器预热30分钟，环境空气以6500毫升/分钟的流量依次流经第一二位二通电磁阀、气敏传感器阵列环形工作腔、第五二位二通电磁阀、第四二位二通电磁阀，最终排出到室外；气敏传感器阵列恒温室内温度从室温达到恒定的55±0.1℃；
    (2)烟气采样周期开始：操作人员点击屏幕下拉菜单的“开始检测”按钮，仪器进入历时5分钟的烟气采样周期，计算机在指定文件夹自动生成一个名为“xxx”的文本文件，以记录气敏传感器阵列对烟气的响应数据；
    (3)初步恢复：在烟气采样周期第0.00-210.00秒，环境空气以6500毫升/分钟的流量依次流经第一二位二通电磁阀、气敏传感器阵列环形工作腔、第五二位二通电磁阀、第四二位二通电磁阀，最终排出到室外；在流量为6500毫升/分钟的环境空气冲洗作用下，粘附在气敏传感器敏感膜表面和管道内壁的烟气气味分子被初步冲走，气敏传感器阵列初步恢复到基准状态，历时210秒；
    (4)洁净空气精确标定：在烟气采样周期第210.00-250.00秒，(a)洁净空气精确标定和(b)卷烟插入、自动点火、第一口烟气抽吸、阴/自燃这两个环节同时进行，均历时40秒：
    (4)a洁净空气精确标定：洁净空气以17.5毫升/秒即1050毫升/分钟的流量依次流过第二节流阀、第六二位二通电磁阀、气敏传感器阵列环形工作腔、第三二位二通电磁阀，最后排出到室内，历时40秒；洁净空气使气敏传感器阵列精确恢复到基准状态；
    (4)b.1卷烟插入：在烟气采样周期第210.00-225.00秒即洁净空气精确标定状态的最初15秒内，屏幕显示“卷烟插入”字样，操作人员将被测卷烟试样滤嘴端插入卷烟夹持器，插入深度为9.0±0.5毫米；
    (4)b.2自动点火：在烟气采样周期第225.00-231.00秒即洁净空气精确标定状态第15.00-21.00秒，点火线圈通电；与此同时，点烟头向左移动9毫米使点火线圈与被测试样接触并点燃试样，持续6秒；在烟气采样周期第231.00-269.00秒，电磁线圈通电，点火线圈断电并回复到基准位置，历时38秒，包括第一口烟气抽吸的后1秒(第2秒)、阴/自燃的18秒、平衡的2秒，第二口烟气抽吸的2秒，残留烟蒂取出操作的15秒；
    (4)b.3第一口烟气抽吸：在烟气采样周期第230.00-232.00秒即洁净空气精确标定状态第20.00-22.00秒，在微型真空泵抽吸作用下，烟气以17.5毫升/秒即1050毫升/分钟的流量，依次经由第二二位二通电磁阀、第一节流阀、流量计，直接被排出到室外，持续2秒；
    (4)b.4阴/自燃：在烟气采样周期第232.00-250.00秒即洁净空气精确标定状态第22.00-40.00秒，第二二位二通电磁阀断开，卷烟进入阴/自燃状态，历时18秒；
    (5)平衡：在烟气采样周期第250.00-252.00秒，所有电磁阀均处于断开状态，气敏传感器阵列环形工作腔内无气体流动，卷烟仍处于阴/自燃状态，历时2秒；
    (6)第二口烟气抽吸：在烟气采样周期第252.00-254.00秒，第一二位二通电磁阀和第五二位二通电磁阀导通，其余四个二位二通电磁阀均断开，卷烟烟气以17.5毫升/秒即1050毫升/分钟的流量，依次通过第一二位二通电磁阀、气敏传感器阵列环形工作腔、第五二位二通电磁阀、第一节流阀、流量计，最后被排出到室外，持续2秒，采集烟气35毫升；
    (7)环境空气冲洗：在烟气采样周期第254.00-300.00秒，室内空气以6500毫升/分钟的流量流过气敏传感器阵列环形工作腔，粘附在气敏传感器敏感膜表面和管道内壁的烟气气味分子被初步冲走，气敏传感器阵列进入初步恢复状态；其中，
    (7.1)残留烟蒂取出操作：在烟气采样周期第254.00-269.00秒，操作人员在15秒钟内取出残留烟蒂并丢弃；在此期间，第三二位二通电磁阀、第四二位二通电磁阀、第五二位二通电磁阀导通，其余三个二位二通电磁阀均断开，室内环境空气以6500毫升/分钟的流量，依次通过第三二位二通电磁阀、气敏传感器阵列环形工作腔、第五二位二通电磁阀、第四二位二通电磁阀，最终排出到室外，历时15秒；
    (7.2)残留烟蒂取出后：在烟气采样周期第269.00-300.00秒，第一二位二通电磁阀、第四二位二通电磁阀、第五二位二通电磁阀导通，其余三个二位二通电磁阀均断开，环境空气以6500毫升/分钟的流量依次流经第一二位二通电磁阀、气敏传感器阵列环形工作腔、第五二位二通电磁阀、第四二位二通电磁阀，最终排出到室外，历时31秒；这一阶段的电磁阀位置和环境空气流动状况与气敏传感器阵列初步恢复状态完全相同；
    (8)数据记录；从烟气采样周期第250.00秒起，即从平衡状态开始之刻起，计算机通过16通道16位高精度数据采集卡将16个气敏传感器产生的电压响应存储在“xxx”文本文件里，直至烟气采样周期第290.00秒即环境空气冲洗阶段第36.00秒止，包括第二口烟气抽吸、残留烟蒂取出操作、残留烟蒂取出后这3个过程，数据记录时长为40秒；
    (9)特征提取；在一个烟气采样周期内，计算机从时长40秒的“xxx”数据记录文件里，提取各个气敏传感器电压响应稳态最大值作为特征分量，本质上是对第二口烟气的响应，一个被测烟草制品试样因此转化为一个16维的测量样本，并存入计算机硬盘的烟草与烟草制品样本数据集文件中；
    (10)识别和感官质量指标得分预测；在烟气采样周期第290.00-300.00秒即数据记录结束的10秒内，
    模块化神经网络级联模型第一级-n个投票识别组依据大多数投票规则确定试样x的品牌、产地与真假，模块化神经网络级联模型第二级-与获胜投票识别组对应的那个得分预测组预测x的香气、协调、杂气、刺激性、余味这5项感官质量指标得分值，并通过显示器显示出来；
    重复步骤(2)～(10)，烟草电子鼻仪器实现对多个烟草与烟草制品试样烟气的测试、识别和感官质量指标得分预测；一个完整的烟草与烟草制品试样测试周期为300秒。
PCT/CN2018/079389 2017-10-31 2018-03-18 一种电子鼻仪器和烟草与烟草制品感官质量评价方法 WO2019085369A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US16/651,722 US10895560B2 (en) 2017-10-31 2018-03-18 Electronic nose instrument for sensory quality evaluation of tobacco and tobacco product

Applications Claiming Priority (4)

Application Number Priority Date Filing Date Title
CN201711054914.6 2017-10-31
CN201711054914.6A CN107807199B (zh) 2017-10-31 2017-10-31 一种烟草与烟草制品电子鼻仪器检测方法
CN201711060781.3A CN107843695B (zh) 2017-10-31 2017-10-31 烟草与烟草制品感官质量的电子鼻仪器评价方法
CN201711060781.3 2017-10-31

Publications (1)

Publication Number Publication Date
WO2019085369A1 true WO2019085369A1 (zh) 2019-05-09

Family

ID=66333436

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/079389 WO2019085369A1 (zh) 2017-10-31 2018-03-18 一种电子鼻仪器和烟草与烟草制品感官质量评价方法

Country Status (2)

Country Link
US (1) US10895560B2 (zh)
WO (1) WO2019085369A1 (zh)

Cited By (15)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111652516A (zh) * 2020-06-05 2020-09-11 江苏中烟工业有限责任公司 一种基于配方功效的烟叶基地适用性评价方法
CN112749906A (zh) * 2021-01-14 2021-05-04 云南中烟工业有限责任公司 卷烟主流烟气光谱数据的感官评价方法
CN113124784A (zh) * 2021-03-17 2021-07-16 红塔烟草(集团)有限责任公司 一种基于激光扫描共聚焦技术的初烤烟部位判定方法
CN113688514A (zh) * 2021-08-17 2021-11-23 云南中烟工业有限责任公司 一种基于卷烟机烟条温度确定最优电烙铁温度的方法
CN114689746A (zh) * 2022-03-25 2022-07-01 江苏中烟工业有限责任公司 一种筛选烟草浸膏特征的方法、装置、电子设备及介质
CN114860012A (zh) * 2022-07-07 2022-08-05 湖北省烟草科学研究院 一种基于物联网的雪茄烟晾制智能监测与调控系统
CN114924038A (zh) * 2022-05-17 2022-08-19 潍坊颖逸食品有限公司 一种基于食品检测的电子鼻设备
US11636870B2 (en) 2020-08-20 2023-04-25 Denso International America, Inc. Smoking cessation systems and methods
US11760170B2 (en) 2020-08-20 2023-09-19 Denso International America, Inc. Olfaction sensor preservation systems and methods
US11760169B2 (en) 2020-08-20 2023-09-19 Denso International America, Inc. Particulate control systems and methods for olfaction sensors
US11813926B2 (en) 2020-08-20 2023-11-14 Denso International America, Inc. Binding agent and olfaction sensor
US11828210B2 (en) 2020-08-20 2023-11-28 Denso International America, Inc. Diagnostic systems and methods of vehicles using olfaction
US11881093B2 (en) 2020-08-20 2024-01-23 Denso International America, Inc. Systems and methods for identifying smoking in vehicles
US11932080B2 (en) 2020-08-20 2024-03-19 Denso International America, Inc. Diagnostic and recirculation control systems and methods
US12017506B2 (en) 2020-08-20 2024-06-25 Denso International America, Inc. Passenger cabin air control systems and methods

Families Citing this family (17)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110398438B (zh) * 2019-09-05 2024-07-09 中国烟草总公司郑州烟草研究院 一种评价造纸法再造烟叶烟粉黏附性的方法及检测装置
CN112082923B (zh) * 2020-09-21 2024-05-28 云南中烟工业有限责任公司 一种准确检测并量化表征卷烟抽吸轻松感的方法
CN112949704B (zh) * 2021-02-24 2021-11-02 北京优创新港科技股份有限公司 一种基于图像分析的烟叶成熟度状态识别方法及装置
CN113237998A (zh) * 2021-04-06 2021-08-10 北京盈盛恒泰科技有限责任公司 一种分析香烟口味的分析系统及方法
CN113238004A (zh) * 2021-05-10 2021-08-10 云南中烟工业有限责任公司 一种利用mlp神经网络模型预测酸味、甜味的方法
CN113379278B (zh) * 2021-06-23 2022-05-10 红云红河烟草(集团)有限责任公司 制丝批次全过程质量评测方法
CN113962554B (zh) * 2021-10-22 2024-07-23 浙江中烟工业有限责任公司 一种基于慢特征聚类的双重粒度烟支质量在线评价方法
CN114062589B (zh) * 2021-11-23 2024-01-26 云南中烟工业有限责任公司 一种卷烟灰烬特征的评价方法
CN114119267B (zh) * 2021-12-07 2024-05-31 浙江中烟工业有限责任公司 一种适用于加热卷烟烟叶原料的内在质量评价方法
CN114324760B (zh) * 2021-12-29 2023-07-21 山东和富工程检测有限公司 一种建材产品气味释放测试装置
CN114397297B (zh) * 2022-01-19 2024-01-23 河南中烟工业有限责任公司 一种烤烟淀粉含量的快速无损检测方法
CN115081890A (zh) * 2022-06-27 2022-09-20 湖北中烟工业有限责任公司 一种卷烟配方模块质量监测系统、方法及相关组件
CN115308366A (zh) * 2022-07-14 2022-11-08 红云红河烟草(集团)有限责任公司 一种基于多个感官参数的卷烟质量评分方法
CN115336782B (zh) * 2022-09-22 2023-08-11 四川金叶生物防治有限公司 一种微环境条件下的复烤片烟醇化方法及醇化控制系统
CN115437333B (zh) * 2022-11-07 2023-02-28 杭州安脉盛智能技术有限公司 一种基于感官质量的调节方法、装置、设备及存储介质
CN116189803B (zh) * 2023-03-31 2024-02-02 江苏中烟工业有限责任公司 一种卷烟叶组配方的稳质增效方法及其应用
CN117634993A (zh) * 2023-11-27 2024-03-01 红云红河烟草(集团)有限责任公司 烟丝质量评价方法、系统、电子设备及存储介质

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101101299A (zh) * 2007-06-25 2008-01-09 华东理工大学 一种并-串联模式识别方法及其在机器嗅觉中的应用
US7531137B2 (en) * 2004-04-30 2009-05-12 Honeywell International Inc. Substance detection system
CN103336090A (zh) * 2013-05-14 2013-10-02 湖南中烟工业有限责任公司 一种预测卷烟感官质量指标的支持向量机svm方法
CN105699437A (zh) * 2016-02-01 2016-06-22 浙江大学 一种基于电子鼻的不同新鲜度山核桃的快速无损区分方法
CN107807199A (zh) * 2017-10-31 2018-03-16 华东理工大学 一种电子鼻仪器和烟草与烟草制品感官质量检测方法

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP3703862B2 (ja) * 1994-04-25 2005-10-05 富士電機機器制御株式会社 ハイブリッドスイッチ
WO2001007907A1 (en) * 1999-07-23 2001-02-01 Cyrano Sciences, Inc. Food applications of artificial olfactometry
CN1115112C (zh) * 2000-12-21 2003-07-23 青岛海洋大学 建立卷烟感官质量评价的模糊—神经网络专家系统的方法
GB201611596D0 (en) * 2016-07-04 2016-08-17 British American Tobacco Investments Ltd Apparatus and method for classifying a tobacco sample into one of a predefined set of taste categories

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7531137B2 (en) * 2004-04-30 2009-05-12 Honeywell International Inc. Substance detection system
CN101101299A (zh) * 2007-06-25 2008-01-09 华东理工大学 一种并-串联模式识别方法及其在机器嗅觉中的应用
CN103336090A (zh) * 2013-05-14 2013-10-02 湖南中烟工业有限责任公司 一种预测卷烟感官质量指标的支持向量机svm方法
CN105699437A (zh) * 2016-02-01 2016-06-22 浙江大学 一种基于电子鼻的不同新鲜度山核桃的快速无损区分方法
CN107807199A (zh) * 2017-10-31 2018-03-16 华东理工大学 一种电子鼻仪器和烟草与烟草制品感官质量检测方法

Cited By (18)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111652516A (zh) * 2020-06-05 2020-09-11 江苏中烟工业有限责任公司 一种基于配方功效的烟叶基地适用性评价方法
US11813926B2 (en) 2020-08-20 2023-11-14 Denso International America, Inc. Binding agent and olfaction sensor
US11636870B2 (en) 2020-08-20 2023-04-25 Denso International America, Inc. Smoking cessation systems and methods
US11760170B2 (en) 2020-08-20 2023-09-19 Denso International America, Inc. Olfaction sensor preservation systems and methods
US12017506B2 (en) 2020-08-20 2024-06-25 Denso International America, Inc. Passenger cabin air control systems and methods
US11760169B2 (en) 2020-08-20 2023-09-19 Denso International America, Inc. Particulate control systems and methods for olfaction sensors
US11932080B2 (en) 2020-08-20 2024-03-19 Denso International America, Inc. Diagnostic and recirculation control systems and methods
US11881093B2 (en) 2020-08-20 2024-01-23 Denso International America, Inc. Systems and methods for identifying smoking in vehicles
US11828210B2 (en) 2020-08-20 2023-11-28 Denso International America, Inc. Diagnostic systems and methods of vehicles using olfaction
CN112749906A (zh) * 2021-01-14 2021-05-04 云南中烟工业有限责任公司 卷烟主流烟气光谱数据的感官评价方法
CN113124784B (zh) * 2021-03-17 2022-08-05 红塔烟草(集团)有限责任公司 一种基于激光扫描共聚焦技术的初烤烟部位判定方法
CN113124784A (zh) * 2021-03-17 2021-07-16 红塔烟草(集团)有限责任公司 一种基于激光扫描共聚焦技术的初烤烟部位判定方法
CN113688514A (zh) * 2021-08-17 2021-11-23 云南中烟工业有限责任公司 一种基于卷烟机烟条温度确定最优电烙铁温度的方法
CN113688514B (zh) * 2021-08-17 2024-03-12 云南中烟工业有限责任公司 一种基于卷烟机烟条温度确定最优电烙铁温度的方法
CN114689746B (zh) * 2022-03-25 2023-11-10 江苏中烟工业有限责任公司 一种筛选烟草浸膏特征的方法、装置、电子设备及介质
CN114689746A (zh) * 2022-03-25 2022-07-01 江苏中烟工业有限责任公司 一种筛选烟草浸膏特征的方法、装置、电子设备及介质
CN114924038A (zh) * 2022-05-17 2022-08-19 潍坊颖逸食品有限公司 一种基于食品检测的电子鼻设备
CN114860012A (zh) * 2022-07-07 2022-08-05 湖北省烟草科学研究院 一种基于物联网的雪茄烟晾制智能监测与调控系统

Also Published As

Publication number Publication date
US10895560B2 (en) 2021-01-19
US20200256837A1 (en) 2020-08-13

Similar Documents

Publication Publication Date Title
WO2019085369A1 (zh) 一种电子鼻仪器和烟草与烟草制品感官质量评价方法
CN107807199B (zh) 一种烟草与烟草制品电子鼻仪器检测方法
CN108133313B (zh) 一种人工智能感官评价食品风味系统及其构建方法
WO2021147275A1 (zh) 电子鼻仪器气敏-色谱多信息融合和风味物质现场检测与分析方法
CN107843695B (zh) 烟草与烟草制品感官质量的电子鼻仪器评价方法
Katurji et al. Direct measurement of toxicants inhaled by water pipe users in the natural environment using a real-time in situ sampling technique
CN107796782B (zh) 基于烟叶特征光谱一致性度量的复烤质量稳定性评价方法
CN101419209A (zh) 卷烟感官评吸和烟气指标的免疫神经网络预测方法
Olmedo et al. Metal exposure and biomarker levels among e-cigarette users in Spain
CN110308101A (zh) 一种片烟水提物吸光度值检测及判断醇化质量的方法
CN104680314A (zh) 一种烟叶风格特征感官评价的方法
Goodall et al. Evaluation of behavioural, chemical, toxicological and clinical studies of a tobacco heated product glo™ and the potential for bridging from a foundational dataset to new product iterations
Brinkman et al. Exposure to and deposition of fine and ultrafine particles in smokers of menthol and nonmenthol cigarettes
Lalonde et al. Characterization of a rapid condensate collection apparatus for in vitro assays of electronic nicotine delivery systems
CN108645746B (zh) 基于炉挥发物的电加热不燃烧卷烟感官质量分析方法
Eaton et al. E-cigarette devices, uses, and exposures
Malt et al. The product science of electrically heated tobacco products: A narrative review of the scientific literature
Debbaneh et al. Electronic cigarettes: a narrative review and cohort study of electronic cigarette users in the otolaryngology clinic
CN101701885B (zh) 联接吸烟机与电子鼻的集气瓶装置及其使用方法
Bernert et al. Increases in tobacco exposure biomarkers measured in non-smokers exposed to sidestream cigarette smoke under controlled conditions
JP7203124B2 (ja) 喫煙または吸引による慢性閉塞性肺疾患のリスクのインビトロ評価方法
CN111887473B (zh) 基于全基因组选择烟气有害成分释放量预测方法及应用
Proctor et al. Integrating chemical, toxicological and clinical research to assess the potential of reducing health risks associated with cigarette smoking through toxicant regulation
Valencia et al. Mobile ascertainment of smoking status through breath: A machine learning approach
CN114965815B (zh) 基于化学计量学-感官组学的赋香卷烟纸分类识别方法

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 18872865

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 18872865

Country of ref document: EP

Kind code of ref document: A1