阅读说明：本技术 一种气敏-气相色谱信息融合和电子鼻仪器在线分析方法 (Gas-sensitive-gas chromatography information fusion and electronic nose instrument on-line analysis method ) 是由 高大启 盛明健 王泽建 万培耀 贺德贵 邢利民 张小勤 于 2020-01-23 设计创作，主要内容包括：本发明提供一种气敏-气相色谱多感知信息选择、融合和发酵/恶臭过程电子鼻在线分析方法。气体进样单周期T<Sub>0</Sub>＝300-600s,气敏传感器阵列和毛细管气相色谱柱二模块被测气体进样时间不同步,累积进样量不相等,感知信息选择与分析时间同步；计算机控制与分析模块从单个气敏传感器时长60s响应曲线中选择稳态电压峰值、出峰时间、曲线下面积这3个感知信息,从时长T<Sub>0</Sub>-10s半分离色谱图中选择10个最大峰值、10个保留时间、1个曲线下面积共21个感知信息；电子鼻仪器用模块化深度卷积神经网络实现循环周期最大为T＝5T<Sub>0</Sub>的5个发酵罐或恶臭污染点长期在线检测与分析,包括气味类型识别、气味整体强度和主要成分浓度量化预测。(The invention provides an online analysis method for an electronic nose in gas-sensitive-gas chromatography multi-perception information selection, fusion and fermentation/malodor processes. Gas sample introduction monocycle T 0 The sample introduction time of the gas to be detected by the gas sensor array and the capillary gas chromatographic column two module is not synchronous, the accumulated sample introduction amount is not equal, and the sensing information selection and the analysis time are synchronous for 600 s; the computer control and analysis module selects 3 pieces of sensing information of steady state voltage peak value, peak-off time and area under the curve from a single gas sensor time length 60s response curve, and the time length T is selected 0 Selecting 21 pieces of perception information including 10 maximum peaks, 10 retention times and 1 area under a curve from a 10s semi-separation chromatogram; maximum cycle period T-5T realized by using modular deep convolution neural network for electronic nose instrument 0 5 pieces of hairAnd (3) long-term online detection and analysis of the fermenter or the foul smell pollution point, including odor type identification, odor overall intensity and main component concentration quantitative prediction.)

1. A gas-sensitive-gas chromatography multi-perception information selection, fusion and electronic nose instrument on-line analysis method is characterized in that an electronic nose instrument comprises a gas-sensitive sensor array module I, a capillary gas chromatography column module II, a measured gas automatic sample introduction module III, a computer control and analysis module IV and an auxiliary gas source V, and long-term circulation automatic on-line detection and intelligent analysis of a plurality of biological fermentation processes or a plurality of odor pollution points are realized;

the gas sensor array module I comprises: the gas sensor array I-1, the gas sensor array annular working cavity I-2, the resistance heating element I-3, the heat insulation layer I-4, the fan I-5 and the partition plate I-6 are positioned in the right middle part of the electronic nose instrument; capillary gas chromatography column module II comprises: the capillary gas chromatographic column II-1, the detector II-2, the amplifier II-3, the recorder II-4, the sample inlet II-5, the resistance heating wire II-6, the fan II-7 and the heat insulation layer II-8 are positioned at the upper right part of the electronic nose instrument;

the gas automatic sample injection module III comprises: the device comprises first to fifth two-position two-way electromagnetic valves III-1 to III-5, a first purifier III-6, a first micro vacuum pump III-7, a first flow meter III-8, a first two-position two-way electromagnetic valve III-9, a first throttle valve III-10, a two-position three-way electromagnetic valve III-11, a three-position four-way electromagnetic valve III-12, a second micro vacuum pump III-13, a seventh two-position two-way electromagnetic valve III-14, an eighth two-position two-way electromagnetic valve III-15 and a pressure stabilizing valve III-16; a first pressure reducing valve III-17, a second throttling valve III-18 and a first purifier III-19; the second pressure reducing valve III-20, the second purifier III-21, the third throttle valve III-22, the second flow meter III-23, the fourth throttle valve III-24 and the fifth throttle valve III-25 are positioned at the right lower part of the electronic nose instrument;

the computer control and analysis module IV comprises a computer mainboard IV-1, an A/D data acquisition card IV-2, a driving and control circuit board IV-3, a 4-path precise direct current stabilized power supply IV-4, a display IV-5 and a WIFI module IV-6, and is positioned on the left side of the electronic nose instrument;

a biological fermentation process/fermentation tank or a foul smell pollution monitoring point, hereinafter referred to as a 'detection point'; the single period of the electronic nose instrument for sampling the detected gas at one detection point is T₀300 + 600s, default T₀480 s; in a gas sampling monocycle T₀In the device, a gas to be detected is respectively pumped to a gas sensor array module I and a capillary gas chromatographic column module II by 2 micro vacuum pumps III-7 and III-13, the gas sensor array I-1 and the capillary gas chromatographic column II-1 generate sensitive responses, and an electronic nose instrument obtains 1 group of gas sensor array response curves and 1 gas chromatogram map, which is a gas-sensitive/gas chromatographic analog signal obtained by sensing a gas sample to be detected by the electronic nose instrument;

in a gas sampling monocycle T₀In the method, the computer control and analysis module IV selects a steady state peak value v from each voltage response curve of the gas sensor array I-1 with the duration of 60s_{gs_i}(τ), corresponding time to peak t_{gs_i}(τ), area under curve Ags_{_i}(tau) the 3 pieces of sensing information meet the triangle stability principle, and the qualitative and quantitative capacity of the gas sensor array is improved; if an array is formed by 16 gas sensors, i is 1,2, … and 16, and the computer control and analysis module IV samples gas in a single period T₀The response curve of the gas sensor array is totally divided into 16 × 3-48 sensing components;

in a gas sampling monocycle T₀In the interior, the electronic nose instrument does not pursue the complete separation of chromatogram peak/peak, and the computer control and analysis module IV selects the first 10 maximum chromatographic peak values v from the semi-separation chromatogram_gci(τ) and corresponding retention time t_gci(τ), area A under chromatogram Curve_gc(τ), i ═ 1,2, …,10, yielding a total of 21 perceptual components, to improve the on-line detectability of the gas chromatography column;

in a gas sampling monocycle T₀In the method, a computer control and analysis module IV fuses 48 sensing components extracted from 16 response curves of a gas sensor array I-1 and 21 sensing components extracted from a capillary chromatographic column II-1 semi-separation chromatogram to obtain a sensing vector x (tau) ∈ R with dimension m being 48+21 being 69⁶⁹The method is the basis for analyzing the biological fermentation process or the malodorous pollution monitoring point by an electronic nose instrument;

in a gas sampling monocycle T₀In the instrument, the electronic nose instrument senses the measured gas in a biological fermentation process or a foul smell pollution point to obtainTo an m-dimensional perceptual vector x (τ) ∈ R^mReferred to as a sample; the electronic nose instrument samples n (less than or equal to 5) biological fermentation processes or n (less than or equal to 5) stink monitoring points with the gas sampling cycle period of T-nT₀Sequentially obtaining n samples, sequentially storing the n samples in n corresponding data files of a computer hard disk, and sending the sample data to a cloud end and a specified fixed/mobile terminal through a WIFI routing module; if T₀When the sample is 480s, the cycle sample injection period of the measured gas is T-nT₀N 480s, corresponding to one fermenter or one foul odor contamination spot, was detected every n 480 s;

the electronic nose instrument forms a main body of odor big data X through long-term online detection of a plurality of biological fermentation processes and a plurality of foul odor pollution points in the year; the data set X also comprises off-line detection data of conventional analytical instruments such as gas chromatography, mass spectrometry, spectrophotometry and the like, odor concentration OU data obtained by olfactive identification in a laboratory of professionals, biological fermentation type data of penicillin, erythromycin, table vinegar, soy sauce, cooking wine, monosodium glutamate and the like recorded by operators and odor pollution monitoring area type data of chemical industrial parks, refuse landfills, sewage treatment plants, livestock and poultry farms and the like; a part of subsets of the data set X establishes corresponding relations between gas-sensitive/chromatographic responses and a plurality of biological fermentation processes/odor pollution types and main component concentrations;

in the learning stage, normalization preprocessing is carried out on all perception components of the odor big data X, and a machine learning model of the computer control and analysis module IV learns the odor big data X off line to determine the structure and the parameters of the odor big data X; in a decision stage, a machine learning model online learns gas-sensitive-chromatography recent response to finely adjust model parameters, a plurality of biological fermentation processes and odor pollution types are determined online according to a gas-sensitive/gas-phase recent sensing time sequence array, and the concentration of main components of fermentation liquor in the biological fermentation processes or ammonia NH specified by national standard GB14554 are quantitatively predicted₃Hydrogen sulfide H₂S, carbon disulfide CS₂Trimethylamine C₃H₉N, methyl mercaptan CH₄S, dimethyl sulfide C₂H₆S, dimethyldisulfide C₂H₆S₂Styrene C₈H₈These 8 malodorous compounds and odors were strongThe degree OU (odor unit) value is 8+1 odor pollutant concentration index values.

2. The gas-gas chromatography multi-perception information selection, fusion and electronic nose on-line analysis method as claimed in claim 1, characterized in that a single period T is introduced in gas sample₀Interior, [ T ]₀-10s,T₀]The time interval is an information selection and analysis time period with the duration of 10s, and the computer control and analysis module IV simultaneously performs sensing information selection and analysis processing operations on the gas sensor array module I and the capillary gas chromatographic column module II; the computer control and analysis module IV is arranged in [ T ] from the gas sensor array I-1₀-75s,T₀-15s]Selecting a selected steady state peak value v in each voltage response curve of a time period, i.e. duration 60s_{gs_i}(τ), corresponding time to peak t_{gs_i}(τ), area under curve Ags_{_i}(τ) these 3 perceptual information components, at [0, T ] from capillary gas chromatography column II-1₀-10s]Time period, i.e. duration T₀-10s of a semi-separated chromatogram, the first 10 maximum chromatographic peaks v being selected_{gc_i}(τ) and 10 corresponding retention times t_{gc_i}(τ), area A under 1 chromatogram plot_gc(τ) a total of 21 perceptual response components stored in a temporary file on the computer hard disk.

3. The gas-gas chromatography multi-perception information selection, fusion and electronic nose on-line analysis method as claimed in claim 1, characterized in that a single period T is introduced in gas sample₀If time length T₀-the number q of chromatographic peaks of the semi-separation chromatogram of 10s is less than 10, the computer control and analysis module IV selects the first q from the semi-separation chromatogram<10 maximum chromatographic peaks v_gci(τ), corresponding Retention time t_gci(τ) and area A under chromatogram Curve_gc(τ), insufficient chromatographic peaks and corresponding retention times are zero-filled, the resulting chromatographic perception information then being x_gc(τ)＝{(h_gc1(τ),h_gc2(τ),…,h_gc,q(τ),0,…,0)；(t_gc1(τ),t_gc2(τ),…,t_gc,q(τ)_,0,…,0)；A_gc(τ)}。

4. The gas-gas chromatography multi-perception information selection, fusion and electronic nose on-line analysis method as claimed in claim 1, wherein gas injection is performed for a single period T₀Last 10s namely [ T ]₀-10s,T₀]In the interval information processing and analyzing time period, a modularized machine learning model of a computer control and analysis module IV carries out odor type identification and intensity and main component quantitative prediction on a biological fermentation process or an odor pollution monitoring point according to a gas-sensitive/chromatographic recent sensing time sequence matrix X (tau-q), wherein the odor type identification and intensity and main component quantitative prediction comprise biological fermentation process type and odor pollution type identification, biological fermentation process cell concentration, substrate concentration and product concentration quantitative estimation and precursor substance concentration quantitative estimation such as n-propanol and phenethyl alcohol, and 8+1 odor pollutant concentration index value quantitative prediction specified by GB 14554; here, τ is the current time, q is the time that has passed recently, and τ -q is the recent time interval.

5. The gas-gas chromatography multi-perception information selection, fusion and electronic nose online analysis method as claimed in claim 1, wherein the smell big data X further comprises: the electronic nose instrument gas-sensitive/chromatographic sensing data of the headspace volatile gas of a plurality of single compounds with the concentration of 0.1-1,0000ppm, and the off-line detection data of conventional analytical instruments such as gas chromatography, mass spectrum and spectrophotometry; professional laboratory sniffing data; the single compounds include especially n-propanol and phenylacetic acid as precursors of the biological fermentation process, 8 malodorous compounds as specified in GB14554, and n-butanol as the OU standard reference substance for the odor concentration as specified in European Standard EN 13725.

6. The gas-gas chromatography multi-perception information selection, fusion and electronic nose on-line analysis method as claimed in claim 1, characterized in that a single period T is introduced in gas sample₀In the case where only one biofermentation process or one malodorous contamination point is detected, then the gas detection and analysis cycle is T ═ T₀(ii) a If k biological fermentation processes/foul smell pollution points are detected simultaneously, one of the processes is usedThe cycle detection and analysis period of the fermentation process/malodor contamination point is T ═ k T₀(ii) a If one of the biological fermentation process/odor pollution points is withdrawn in the long-term circulation monitoring process, the gas circulation detection and analysis period is changed into T ═ k-1 × (T)₀(ii) a Similarly, if a new biofermentation process/foul smell contamination point is added midway during long-term cycle monitoring, the cycle detection and analysis period becomes T ═ k +1 × T₀(ii) a And (3) recording the corresponding change of the period of the corresponding data file from the moment when a biological fermentation process/stink pollution point exits/joins.

7. The gas-gas chromatography multi-perception information selection, fusion and electronic nose online analysis method as claimed in claim 1, wherein the machine learning model is composed of a plurality of modular deep convolutional neural networks; the number of the single-output deep convolution neural network modules is equal to the predicted number of the main components of the fermentation liquor in the biological fermentation process, the number of the indexes of the main concentration of the malodorous pollutants and the number of the types of the measured objects, and the numbers correspond to one another; a single-output deep convolutional neural network is composed of an input layer, 3 convolutional layers, 2 down-sampling layers and 1 output unit, and the activation functions of all the hidden layers and the output layers are Sigmoid modified activation functionsIn a learning stage, each single-output deep convolutional neural network adopts an error back-transmission off-line layer-by-layer learning algorithm, the error back-transmission off-line layer-by-layer learning algorithm is mainly used for learning labeled data and odor big data with known components in the odor big data and necessary intelligence, the size of a convolutional layer scanning window is 5 × 5, the overlapping scanning step length is 1, the convolution kernel is a combination of a sine kernel, a cosine kernel, a polynomial kernel, a Gaussian kernel, a Sigmoid kernel, a wavelet kernel and a L aplace kernel, the size of a downsampling layer scanning window is 2 × 2, non-overlapping scanning, namely the step length is 2, maximum values, mean values and mean square error characteristics are extracted, in a decision stage, n single-output deep convolutional neural network models carry out odor type identification and estimate and predict current time tau and future tau +1 one by one according to a gas-sensitive/gas-chromatography current time and recently-occurring time sequence sensing matrix X (tau-q) in a decision stage,The intensity of the odor at the time of tau +2 and tau +3 and the concentration value of the main component.

8. The gas-gas chromatography multi-perception information selection, fusion and electronic nose online analysis method applying the gas-gas chromatography multi-perception information selection, fusion and electronic nose online analysis method as claimed in claims 1-6, wherein the electronic nose instrument carries out long-term circulation online analysis and quantitative prediction on a plurality of biological fermentation processes/malodor pollution points, and comprises the following steps:

(1) starting up: preheating the instrument for 30 min;

modification of screen menu' gas sample introduction monocycle T₀"set, Default value T₀8 min; the gas circulation sampling period of the 5 detection points is T-5T₀；

(2) Beginning a gas circulation sample introduction period: the electronic nose instrument can carry out circulating on-line detection on at most 5 detection points in sequence, and the computer control and analysis module IV automatically generates 5 text files so as to store the response data of the gas sensor array I-1 and the capillary gas chromatographic column module II to the gases at the 5 detection points;

(3) detecting the start of a gas sample introduction single period at a point k; by T₀As an example, 8 min:

(3.1) information perception and recording phase [0-470s ]:

(3.1a), sequentially carrying out six gas injection stages of preliminary recovery 360s of ①, accurate calibration of ② 40s, balance of ③ 5s, headspace injection of ④ 60s, transition of ⑤ 5s and cleaning of ⑥ 10s on the gas sensor array module I;

(3.1b), the capillary gas chromatographic column module II sequentially goes through ① 1s of headspace sampling, ② 469s of chromatographic separation and ③ 10s of emptying and cleaning to total 3 gas sampling stages;

(3.1c), the computer control and analysis module IV records the sensing data of the gas sensor array module I stage ④ time length 60s and the tubule gas chromatographic column module II stage ① + ② time length 470s in corresponding temporary files;

(3.2) information selection and processing stages [470- & ltwbr/& gts & 480s]: in sample introduction monocycle T₀470-Selecting a steady state peak value v in the response curve_{gs_i}(τ), time to peak t_{gs_i}(τ), area under curve Ags_{_i}(τ) the 3 pieces of perceptual information; array I-1 of 16 gas sensors in period T₀Obtaining 16 × 3 ═ 48 perception components; at the same time, the computer control and analysis module IV selects the first 10 maximum chromatographic peaks v from the semi-separation chromatogram of instant length 470s_{gc_i}(τ) and corresponding retention time t_{gc_i}(τ), area A under chromatogram Curve_gc(τ) resulting in 21 perceptual components in total; in each one-cycle T₀In the system, the computer control and analysis module IV obtains 1 perception vector x (tau) ∈ R with 69 dimensions from the perception information of the gas sensor array I module and the capillary chromatographic column module II⁶⁹；

The modularized deep convolution neural network model of the computer control and analysis module IV carries out odor type identification and intensity and main component quantitative prediction according to a recent multi-sensing time sequence matrix X (tau-q) of the gas-sensitive-gas chromatography, wherein the odor type identification and intensity and main component quantitative prediction comprise biological fermentation process and odor pollution type identification, biological fermentation process precursor substance and product concentration quantitative estimation, and odor concentration OU value and 8 odor compound concentration index value quantitative prediction; the monitor displays the monitoring and predicting results and transmits the results to the central control room and the plurality of fixed/mobile terminals through the Internet network;

(3.3) ending the detection point k and starting the next detection point;

(4) and (4) repeating the steps (2) to (3), and realizing gas circulation online detection, identification and multi-term concentration quantitative prediction of 5 detection points by the electronic nose instrument.

Technical Field

The invention discloses an online analysis method of an electronic nose instrument integrating a gas sensor array and a capillary gas chromatographic column, which aims at the automatic continuous online detection and analysis requirements of processes such as biological fermentation and environmental odor pollution which are characterized by long-term dynamic change of odor, relates to the technical fields of artificial intelligence, computers, environmental protection, bioengineering, analytical chemistry and the like, and mainly solves the problems of poor sensitivity of the gas sensor array, poor selectivity of the gas sensor array caused by a single perception information extraction mode, poor linearity of a chromatography caused by a peak-peak complete separation analysis mode, selection and integration of multiple perception information of gas sensitivity and gas chromatography and long-term circulation online analysis of the electronic nose instrument.

Background

The long-term smelling of the foul smell can cause serious damage to the body, and the artificial smelling of the tail gas to analyze the biological fermentation process is unrealistic, which is against the desire of people to pursue a beautiful life and the artificial intelligence. Moreover, the method for quantitatively determining indexes such as odor concentration, food and flavor odor intensity and the like by smelling is subject to scaling due to very complicated process, high cost, poor objective fairness and poor operability. The online detection of complex odor and the simultaneous online quantitative prediction of various components thereof are complex theoretical problems, and are more urgent technical and application problems to be solved.

The electronic nose technology has a main development trend that an array is formed by a plurality of gas sensitive devices with necessary sensitivity, and the qualitative and quantitative analysis capability of complex odor, including odor type identification and intensity and main component concentration quantitative prediction, is realized by mainly utilizing big data and artificial intelligence technology. The electronic nose instrument mainly adopts a working mode of long-term continuous on-line detection and analysis on application objects such as a biological fermentation process, environmental odor pollution monitoring and the like, and is characterized in that only one biological fermentation process (fermentation tank) or one odor pollution point is sensed and analyzed in one gas sampling single period; observation of multiple fermentors or malodorsThe gas sampling, sensing and analyzing operation is carried out circularly and continuously day and night, and the duration of the whole process is often days, weeks, months or even years. It is considered that the gas injection single cycle of the electronic nose instrument, i.e. the detection and analysis single cycle of the electronic nose instrument, should not exceed T₀And (3) sampling gas of a plurality of fermentation tanks or a plurality of odor monitoring points, and carrying out online detection and analysis for a cycle period T of not more than 1 hour, so that whether the detection and analysis method is online is judged reasonably.

A large number of experiments have shown that_nO₂Metal Oxide Semiconductor (MOS) gas sensors represented by materials have a fast response speed to some odors, for example, only 2 seconds are required for ethanol volatile gas; and very slow response to other odours, even up to 60s or more, for example gamma undecalactone C, a standard odours specified in GB/T14675₁₁H₂₀O₂This is true of the perception of volatile gases. This phenomenon tells us that although the steady state maximum of the response curve of the same gas sensor to both odors may be the same, the time to peak and the area under the curve may be different; or the area under the curve may be the same but the steady state maximum and the time to peak may be different, etc. In summary, the shape of the response curve of the gas sensor is related to the odor composition, and relates to many factors such as molecular weight, carbon number, polarity and functional groups.

The stability of the triangle means that three sides (straight lines) are connected end to form a stable structure and have the characteristic of no deformation under stress. The parallelogram is easy to deform under stress and is unstable; similarly, polygons with a number of edges greater than 3 are all unstable. The triangle stability principle gives us the inspiration that only two parameters (three conditions of 2 sides long, 2 included angles, 1 side long and 1 included angle) can not determine a triangle structure; needless to say, it is not feasible to know only one parameter (two cases of 1 side length and 1 included angle).

The single type gas sensor array has poor selectivity, limited overlapping sensing range and insufficient sensitivity, and does not meet the online detection requirements of objects such as biological fermentation, odor pollution and the like. Therefore, gas-chromatography (GC) methods have attracted a high degree of attention. The chromatographic method has the advantages of high sensitivity and good selectivity, and has the disadvantages of long separation time, i.e. detection period, complex instrument structure and harsh working conditions, and the existing method is not suitable for on-line detection of odor at all. It must be noted that the differences of "good GC column selectivity and poor MOS gas sensor selectivity" are only relative, and the "qualitative ability" of gas chromatography on unknown samples is still "weak". In the case of no internal/external standard sample spectrogram, the composition and composition of an unknown sample cannot be determined by only one measured spectrogram. The second drawback of gas chromatography is that the column "selectivity" is not universal. A particular column is sensitive to a particular sample only under particular conditions, i.e. a particular column can only detect a particular range of a particular sample. When one of the sample feeding condition, the testing condition and the self parameter of the chromatographic column is changed, the chromatographic sensing parameter of the specific sample is changed along with the change of the sample feeding condition, the testing condition and the self parameter of the chromatographic column.

A third disadvantage of gas chromatography is that it is difficult or impossible to achieve "complete separation" of the multicomponent chromatographic peaks. The more components, the more similar the polarity between the components, and the closer the retention time, the more difficult it is to completely separate the peaks. We believe that complete separation of chromatogram multicomponent peaks is relative, rare; in contrast, incomplete separation of multicomponent chromatographic peaks is absolute and common. From an operating parameter point of view, both increasing the degree of chromatographic separation and reducing the retention time are sometimes contradictory.

Experiments have shown that the chromatographic retention times for 8 malodorous compounds, specified in GB14554, are mostly less than 8min, with all odour components having molecular weights less than 300 Dalton. These properties of odor facilitate on-line detection and analysis by capillary gas chromatography. To increase the detection speed of gas chromatography, we can choose a capillary column with a larger inner diameter. Thus, within a given detection and analysis period, we obtain a finite duration T₀Semi-isolated multimodal plot ≦ 10 min. As the name implies, "semi-separated chromatogram" or "incompletely separated chromatogram" refers to a chromatogram in which complete separation between peak and peak is not achieved at a given time interval. This is achieved byThe semi-separation/incomplete separation phenomenon is a result of combined action of a plurality of factors including measured gas components, self characteristics of a chromatographic column, working parameter setting of a chromatograph, detector performance and recording time of a recorder. Incomplete or semi-complete separation between chromatographic peaks is a common phenomenon, and complete separation is only an ideal or limiting case. The more components of the object being tested, the more difficult it is to completely separate the peaks and peaks, and at the cost of long detection time.

For some time of the marathon race, although the guanaman did not occur, the win-loss trend was already divided and the guanaman was in the "team running ahead of the race team". The method is a biological basis for on-line detection and analysis by using a semi-separation chromatogram in a gas chromatography. The semi-separated chromatogram is a portion of the full separated chromatogram, corresponding to the marathon game "team running ahead of the team of the game". As long as the tested sample components and the chromatographic column testing conditions are kept unchanged, the semi-separation chromatogram obtained by testing the same sample at different times is kept unchanged, and the position relationship between the semi-separation chromatogram and the full-separation chromatogram is also kept unchanged. That is, we can use the half-separation chromatogram to estimate some major characteristics of the full-separation chromatogram, for example, to estimate the presence, absence and content of some components with long retention time that are not present on the half-separation chromatogram. For the process analysis of biological fermentation, odor pollution and the like, it is enough to obtain information reflecting main state parameters, the semi-separation chromatogram actually contains main information of the full-separation chromatogram, and the key is how to obtain required information from the chromatogram and explain the information.

It has been pointed out previously that the sensing range of both single column and single type gas sensor arrays is limited. In order to realize the on-line sensing and analysis of fermentation process and odor pollutant in a wider range, the problem to be solved is how to combine a gas sensor array and a gas chromatographic column, the advantages are complementary, and the gas sample introduction single period T is realized₀Long-term cyclic on-line detection for 5-10 min. In order to realize the gas-sensitive-gas chromatography multi-perception information selection and fusion and the online analysis method of an electronic nose instrument, the following odor perception theory and analysis technical problems need to be solved:

(A) gas sensor array and gas chromatographic column online perception information selection and fusion problem

Inspired by the principle of triangular stability, a plurality of characteristic information should be simultaneously extracted from a single gas sensor response curve, for example, the characteristic information such as the maximum value of 'steady-state' response, the peak time value and the area under the curve is simultaneously selected, which is equivalent to the improvement of the selectivity of an electronic nose instrument from the pretreatment angle. The response speed of the chromatographic column is at least one order of magnitude lower than that of the gas sensor, and the chromatographic peak/peak complete separation method is achieved, so that the chromatographic column does not meet the requirement of odor on-line detection. Inspired by marathon game life prototype, can be from a specified time interval (e.g., T)₀About 10 min) and the corresponding retention time, and adding the area under the spectrogram curve as the perception information characteristic of the capillary chromatographic column to the fermentation object or the malodorous pollutant, so as to improve the response speed of the gas chromatography, namely, the linearity.

How to select and fuse a plurality of characteristic information from the response curve of the gas sensor array and the semi-separation chromatogram map so as to improve the online qualitative and quantitative analysis capability of the electronic nose instrument is a main problem to be solved by the invention.

(B) Electronic nose instrument online analysis capability and intelligent problem based on big data and machine learning

There is no multisource perception data generated by on-line testing of a large amount of odors, no component detection data of conventional instruments such as olfactory identification data and color/mass spectra, no expert experience data and no professional field recording data, and it is unrealistic to attempt to simply rely on a single type gas sensor array, a single gas chromatographic column and a single machine learning model to estimate the complex odor intensity and the concentration of various components on line. Many electronic noses do so today, but the role of the resulting test data is quite limited and the results obtained are therefore not reliable.

Due to odor complexity and environmental variability, small data is not sufficient to train efficient machine learning models to identify multiple odor types and quantitatively predict complex odor components. The odor big data is established on the basis of gas-sensitive/chromatographic multi-source perception data, smell identification data, expert experience data, professional field recording data, color/mass spectrum and other conventional instrument detection data. With the big smell data, the machine learning method of the electronic nose instrument can identify the smell type and quantitatively predict the concentration of a plurality of components by data mining according to the current perception information. The big data and the on-line prediction of the odor components are two contradictory aspects, and the effective solution way is to deeply research and provide a simple and effective machine learning model and algorithm to realize the identification of various odor types and the real-time quantitative prediction of the odor intensity and the concentration of various main components.

Disclosure of Invention

The invention discloses a gas-sensitive-gas-chromatography multi-sensing information selection and fusion and electronic nose instrument online analysis method based on the existing invention patents of a multi-point centralized online monitoring and analysis system and method for malodorous gas (see application number: 2018104716131), a large data-driven multi-point centralized electronic nose instrument online analysis method for malodorous gas (see application number: 2018104717083) and a multi-channel integrated olfactory analog instrument and biological fermentation process online analysis method (see application number: 201310405315.X), so as to solve the problems of long-term online monitoring of a plurality of biological fermentation processes or a plurality of malodorous monitoring points, fermentation and malodorous pollution types, and online quantitative prediction of odor intensity qualitative indexes and a plurality of concentration control indexes.

In order to achieve the above purpose, the invention provides the following technical scheme:

the electronic nose instrument comprises a gas sensor array module I, a capillary gas chromatographic column module II, a measured gas automatic sample introduction module III, a computer control and analysis module IV and an auxiliary gas source V, and realizes long-term circulating automatic online detection and intelligent analysis of a plurality of biological fermentation processes or a plurality of stink pollution monitoring points.

The gas sensor array module I includes: the gas sensor array I-1, the gas sensor array annular working cavity I-2, the resistance heating element I-3, the heat insulation layer I-4, the fan I-5 and the partition plate I-6 are positioned in the right middle of the electronic nose instrument. Capillary gas chromatography column module II comprises: the capillary gas chromatographic column II-1, the detector II-2, the amplifier II-3, the recorder II-4, the sample inlet II-5, the resistance heating wire II-6, the fan II-7 and the heat insulation layer II-8 are positioned at the upper right part of the electronic nose instrument.

The gas autoinjection module III includes: the device comprises first to fifth two-position two-way electromagnetic valves III-1 to III-5, a first purifier III-6, a first micro vacuum pump III-7, a first flow meter III-8, a first two-position two-way electromagnetic valve III-9, a first throttle valve III-10, a two-position three-way electromagnetic valve III-11, a three-position four-way electromagnetic valve III-12, a second micro vacuum pump III-13, a seventh two-position two-way electromagnetic valve III-14, an eighth two-position two-way electromagnetic valve III-15 and a pressure stabilizing valve III-16; a first pressure reducing valve III-17, a second throttling valve III-18 and a first purifier III-19; the second pressure reducing valve III-20, the second purifier III-21, the third throttle valve III-22, the second flow meter III-23, the fourth throttle valve III-24 and the fifth throttle valve III-25 are positioned at the lower right of the electronic nose instrument.

The computer control and analysis module IV comprises a computer mainboard IV-1, an A/D data acquisition card IV-2, a drive and control circuit board IV-3, a 4-path precise direct current stabilized power supply IV-4, a display IV-5 and a WIFI module IV-6, and is positioned on the left side of the electronic nose instrument.

A biological fermentation process/fermentation tank or a foul smell pollution monitoring point, hereinafter referred to as a "detection point". The single period of the electronic nose instrument for sampling the detected gas at one detection point is T₀300 + 600s, default T₀480 s. In a gas sampling monocycle T₀In the device, a gas to be detected is respectively pumped to a gas sensor array module I and a capillary gas chromatographic column module II by 2 micro vacuum pumps III-7 and III-13, the gas sensor array I-1 and the capillary gas chromatographic column II-1 generate sensitive responses, and an electronic nose instrument obtains 1 group of gas sensor array response curves and 1 gas chromatogram map, which is a gas-sensitive/gas chromatographic analog signal obtained by sensing a gas sample to be detected by the electronic nose instrument.

In a gas sampling monocycle T₀In the method, the computer control and analysis module IV selects a steady state peak value v from each voltage response curve of the gas sensor array I-1 with the duration of 60s_gsi(τ)、Corresponding peak-off time t_gsi(τ), area under curve Ags_iAnd (tau) the 3 pieces of sensing information meet the triangular stability principle, and the qualitative and quantitative capacity of the gas sensor array is improved. If an array is formed by 16 gas sensors, i is 1,2, … and 16, and the computer control and analysis module IV samples gas in a single period T₀The total of 16 x 3 to 48 sensing components are obtained from the response curve of the gas sensor array.

In a gas sampling monocycle T₀In the interior, the electronic nose instrument does not pursue the complete separation of chromatogram peak/peak, and the computer control and analysis module IV selects the first 10 maximum chromatographic peak values v from the semi-separation chromatogram_gci(τ) and corresponding retention time t_gci(τ), area A under chromatogram Curve_gc(τ) to obtain 21 sensing components to improve the on-line detection capability of the gas chromatography column.

In a gas sampling monocycle T₀In the method, a computer control and analysis module IV fuses 48 sensing components of a plurality of response curves of a gas sensor array I-1 and 21 sensing components of a semi-separation chromatogram of a capillary chromatographic column II-1 to obtain a sensing vector x (tau) ∈ R with dimension m being 48+21 being 69⁶⁹The method is the basis for analyzing the biological fermentation process or the malodorous pollution monitoring point by the electronic nose instrument.

In a single period T₀In the method, an electronic nose instrument senses a measured gas in a biological fermentation process or a foul smell pollution point to obtain an m-dimensional sensing vector x (tau) ∈ R^mReferred to as samples. The electronic nose instrument samples n (less than or equal to 5) biological fermentation processes or n (less than or equal to 5) stink monitoring points with the gas sampling cycle period of T-nT₀And sequentially obtaining n samples, sequentially storing the n samples in n corresponding data files of the computer hard disk, and sending the sample data to the cloud and the appointed fixed/mobile terminal through the WIFI routing module. If T₀When the sample is 480s, the cycle sample injection period of the measured gas is T-nT₀N 480s, corresponding to a fermenter or a foul odor spot, was detected every n 480 s.

The electronic nose instrument forms a main body of big odor data X through long-term online detection of a plurality of biological fermentation processes and a plurality of foul odor pollution points over the years. The data set X also comprises off-line detection data of conventional analytical instruments such as gas chromatography, mass spectrometry, spectrophotometry and the like, odor concentration OU data obtained by olfactive identification in a laboratory of professionals, biological fermentation type data of penicillin, erythromycin, vinegar, soy sauce, cooking wine, monosodium glutamate and the like recorded by operators and odor pollution monitoring area type data of chemical industrial parks, refuse landfills, sewage treatment plants, livestock and poultry farms and the like. A subset of the data set X establishes a correspondence of gas/chromatographic response to multiple biofermentation processes/malodour contamination types and principal component concentrations.

In the learning stage, normalization preprocessing is carried out on all perception components of the odor big data X, and a machine learning model of the computer control and analysis module IV learns the odor big data X off line to determine the structure and the parameters of the odor big data X; in a decision stage, a machine learning model online learns gas-sensitive-chromatography recent response to finely adjust model parameters, a plurality of biological fermentation processes and odor pollution types are determined online according to a gas-sensitive/gas-phase recent sensing time sequence array, and the concentration of main components of fermentation liquor in the biological fermentation processes or ammonia NH specified by national standard GB14554 are quantitatively predicted₃Hydrogen sulfide H₂S, carbon disulfide CS₂Trimethylamine C₃H₉N, methyl mercaptan CH₄S, dimethyl sulfide C₂H₆S, dimethyldisulfide C₂H₆S₂Styrene C₈H₈These 8 malodorous compounds and odor concentration OU (odor unit) values total 8+1 odor pollutant concentration index values.

In gas sample introduction monocycle T₀Interior, [ T ]₀-10s,T₀]The time interval is the information selection and analysis time period with the duration of 10s, and the computer control and analysis module IV simultaneously performs sensing information selection and analysis processing operations on the gas sensor array module I and the capillary gas chromatographic column module II. The computer control and analysis module IV is arranged in [ T ] from the gas sensor array I-1₀-75s,T₀-15s]Selecting a selected steady state peak value v in each voltage response curve of a time period, i.e. duration 60s_gsi(τ), corresponding time to peak t_gsi(τ), area under curve Ags_i(τ) These 3 perceptual information components; from capillary gas chromatography column II-1 at [0, T₀-10s]Time period, i.e. duration T₀-10s of a semi-separated chromatogram, the first 10 maximum chromatographic peaks v being selected_gci(τ) and corresponding retention time t_gci(τ), area A under chromatogram Curve_gc(τ) a total of 21 perceptual response components stored in a temporary file on the computer hard disk.

In gas sample introduction monocycle T₀If time length T₀-the number q of chromatographic peaks of the semi-separation chromatogram of 10s is less than 10, the computer control and analysis module IV selects the first q from the semi-separation chromatogram<10 maximum chromatographic peaks v_gci(τ), corresponding Retention time t_gci(τ) and area A under chromatogram Curve_gc(τ), insufficient chromatographic peaks and corresponding retention times are zero-filled, the resulting chromatographic perception information then being x_gc(τ)＝{(h_gc1(τ),h_gc2(τ),…,h_gc,q(τ),0,0)；(t_gc1(τ),t_gc2(τ),…,t_gc,q(τ),0,0)；A_gc}。

Gas sample introduction monocycle T₀Last 10s namely [ T ]₀-10s,T₀]In the interval information processing and analyzing time period, a modularized machine learning model of a computer control and analysis module IV carries out odor type identification and intensity and main component quantitative prediction on a biological fermentation process or an odor pollution monitoring point according to a gas-sensitive/chromatographic recent sensing time sequence matrix X (tau-q), wherein the odor type identification and intensity and main component quantitative prediction comprise biological fermentation process type and odor pollution type identification, biological fermentation process cell concentration, substrate concentration and product concentration quantitative estimation and precursor substance concentration quantitative estimation such as normal propyl alcohol and phenethyl alcohol, and 8+1 odor pollutant concentration index value quantitative prediction specified by GB 14554.

The odor big data X also comprises the gas-sensitive/chromatographic sensing data of the electronic nose instrument of the headspace volatile gas of a plurality of single compounds with the concentration of 0.1-1,000 ppm; off-line detection data of conventional analytical instruments such as gas chromatography, mass spectrometry and spectrophotometry; professional laboratory sniff data. The single compounds include especially n-propanol and phenylacetic acid as precursors of the biological fermentation process, 8 malodorous compounds as specified in GB14554, and n-butanol as the OU standard reference substance for the odor concentration as specified in European Standard EN 13725.

In gas sample introduction monocycle T₀In the case where only one biofermentation process or one malodorous contamination point is detected, then the gas detection and analysis cycle is T ═ T₀. If k biological fermentation processes/malodorous contamination points are detected simultaneously, the cycle detection and analysis period of one of the biological fermentation processes/malodorous contamination points is T ═ k T₀. If one of the biological fermentation process/odor pollution points is withdrawn in the long-term circulation monitoring process, the gas circulation detection and analysis period is changed into T ═ k-1 × (T)₀. Similarly, if a new biofermentation process/foul smell contamination point is added midway during long-term cycle monitoring, the cycle detection and analysis period becomes T ═ k +1 × T₀. And (3) recording the corresponding change of the period of the corresponding data file from the moment when a biological fermentation process/stink pollution point exits/joins.

The machine learning model consists of a plurality of modular deep convolutional neural networks; the number of the single-output deep convolution neural network modules is equal to the predicted number of the main components of the fermentation liquor in the biological fermentation process, the number of the indexes of the main concentration of the malodorous pollutants and the number of the types of the measured objects, and the numbers correspond to one another; a single-output deep convolutional neural network is composed of an input layer, 3 convolutional layers, 2 down-sampling layers and 1 output unit, and the activation functions of the hidden layers and the output layers are modified Sigmoid activation functionsIn the learning stage, each single-output deep convolutional neural network adopts an error back-transmission off-line layer-by-layer learning algorithm, the error back-transmission off-line layer-by-layer learning algorithm is mainly used for learning labeled data and odor big data with known components in the odor big data and necessary intelligence, the size of a convolutional layer scanning window is 5 × 5, the overlapping scanning step length is 1, the convolution kernel is a combination of a sine kernel, a cosine kernel, a polynomial kernel, a Gaussian kernel, a Sigmoid kernel, a wavelet kernel and a L aplace kernel, the size of a downsampling layer scanning window is 2 × 2, non-overlapping scanning is 2, the step length is 2, the maximum value, the mean value and the mean variance characteristics are extracted, and in the decision stage, n single-output deep convolutional neural networks are used for extracting theAnd (3) carrying out odor type identification through a network model according to the current time tau of the gas-sensitive/gas chromatography and a recently-generated time series sensing matrix X (tau-q), and estimating and predicting the odor intensity and the concentration value of the main composition at the current time tau and the future time tau +1, tau +2 and tau +3 one by one.

The long-term circulation on-line analysis and quantitative prediction of the electronic nose instrument on a plurality of biological fermentation processes/odor pollution points comprises the following steps:

(1) starting up: the instrument was preheated for 30 min.

Modification of screen menu' gas sample introduction monocycle T₀"set, Default value T₀8 min; the gas circulation sampling period of the 5 detection points is T-5T₀。

(2) Beginning a gas circulation sample introduction period: the electronic nose instrument can carry out circulating on-line detection on at most 5 detection points in sequence, and the computer control and analysis module IV automatically generates 5 text files so as to store the response data of the gas sensor array I-1 and the capillary gas chromatographic column module II to the gases at the 5 detection points.

(3) And (5) detecting the start of a gas sample introduction single period at a point k. By T₀As an example, 8 min:

(3.1) information perception and recording phase [0-470s ]:

(3.1a), the gas sensor array module I sequentially goes through six gas injection stages of initial recovery of ① 360s, accurate calibration of ② 40s, balance of ③ 5s, headspace injection of ④ 60s, transition of ⑤ 5s and cleaning of ⑥ 10 s.

(3.1b), capillary gas chromatography column module II sequentially through ① 1s of headspace injection, ② 469s of chromatographic separation and ③ 10s of vent and cleaning total 3 gas injection stages.

(3.1c), the computer control and analysis module IV records the sensing data of the stage ④ time length 60s of the gas sensor array module I and the stage ① + ② time length 470s of the tubule gas chromatography column module II in corresponding temporary files.

(3.2) information selection and processing stages [470- & ltwbr/& gts & 480s]: in gas sample introduction monocycle T₀470-480s, the computer control and analysis module IV samples from the gas sensor array module I in the gas headspaceSelecting steady state peak value v in each voltage response curve with duration of 60s_gsi(τ), time to peak t_gsi(τ), area under curve Ags_i(τ) the 3 pieces of perceptual information; array I-1 of 16 gas sensors in period T₀A total of 16 × 3 — 48 perceptual components are obtained. At the same time, the computer control and analysis module IV selects the first 10 maximum chromatographic peaks v from the semi-separation chromatogram of instant length 470s_gci(τ) and corresponding retention time t_gci(τ), area A under chromatogram Curve_gc(τ), a total of 21 perceptual components are obtained. In each one-cycle T₀In the system, the computer control and analysis module IV obtains 1 perception vector x (tau) ∈ R with 69 dimensions from the perception information of the gas sensor array I module and the capillary chromatographic column module II⁶⁹。

And the modularized deep convolution neural network model of the computer control and analysis module IV carries out odor type identification and intensity and main component quantitative prediction according to a recent multi-sensing time sequence matrix of the gas-sensitive gas chromatography, wherein the odor type identification and intensity and main component quantitative prediction comprises biological fermentation process and odor pollution type identification, quantitative estimation of precursor substance and product concentration in the biological fermentation process, and quantitative prediction of an odor concentration OU value and 8 odor compound concentration index values. The monitor displays the monitoring and prediction results and transmits them to the central control room and a plurality of fixed/mobile terminals through the Internet network.

(3.3) end of detection point k and start of the next detection point.

(4) And (4) repeating the steps (2) to (3), and realizing the circulating online detection, identification and multi-term concentration quantitative prediction of the gas at the 5 detection points by the electronic nose instrument.

Drawings

FIG. 1 is a schematic diagram of the principle of the composition of an electronic nose instrument, which is a gas-sensitive-gas chromatography multi-perception information selection and fusion and an electronic nose instrument on-line analysis method.

FIG. 2 is a gas-sensitive gas chromatography multi-perception information selection, fusion and electronic nose instrument on-line analysis method-gas sample injection single period T₀The information sensing, selecting and processing conditions of the gas sensor array and the capillary gas chromatographic column are schematically shown in 480 s.

FIG. 3 is the gas-sensitive-gas chromatography multi-perception information selection, fusion and electronic nose instrument on-line analysis method-gas sample injection single period T₀Within 480s, the response curve of the gas sensor is a multi-feature selection schematic diagram.

FIG. 4 is the gas-sensitive-gas chromatography multi-perception information selection, fusion and electronic nose instrument on-line analysis method-gas sample injection single period T₀Within 480s, the semi-separation chromatogram is schematically selected in a multi-feature mode.

FIG. 5 is a gas-sensitive gas chromatography multi-perception information selection, fusion and electronic nose instrument on-line analysis method-gas sample injection single period T₀Within 480s, two semi-separation chromatograms are schematic for multi-feature selection.

FIG. 6 is a schematic diagram of the multi-parameter 'divide-and-conquer' quantitative prediction of a modular deep convolutional neural network model oriented to a 'continuous on-line' analysis mode, which is a gas-sensitive-gas chromatography multi-perception information selection, fusion and electronic nose instrument on-line analysis method of the invention.

Detailed Description

The present invention is described in further detail below with reference to the attached drawing figures.

FIG. 1 is a schematic diagram of the working principle of the gas-sensitive/gas-chromatographic integrated electronic nose instrument of the present invention. The electronic nose instrument mainly comprises: the gas sensor comprises a gas sensor array module I, a capillary gas chromatographic column module II, a gas automatic sample injection module III, a computer control and analysis module IV, and an auxiliary gas source, namely a hydrogen cylinder V-1 and a clean air cylinder V-2.

The gas sensor array module I mainly comprises the following components: the gas sensor array I-1, the gas sensor array annular working cavity I-2, the resistance heating element I-3, the heat insulation layer I-4, the fan I-5 and the partition plate I-6 are positioned in the right middle of the electronic nose instrument. The capillary gas chromatographic column module II mainly comprises the following components: the capillary gas chromatographic column II-1, the detector II-2, the amplifier II-3, the recorder II-4, the sample inlet II-5, the resistance heating wire II-6, the fan II-7 and the heat insulation layer II-8 are positioned at the upper right part of the electronic nose instrument.

The gas automatic sample introduction module III comprises the following components: first to fifth two-position two-way solenoid valves III-1 to III-5, a first purifier III-6, a first micro vacuum pump III-7, a first flow meter III-8, a sixth two-position two-way solenoid valve III-9, a first throttle valve III-10, a two-position three-way solenoid valve III-11, a three-position four-way solenoid valve III-12, a second micro vacuum pump III-13, a seventh two-position two-way solenoid valve III-14 and an eighth two-position two-way solenoid valve III-15, the device comprises a pressure maintaining valve III-16, a first pressure reducing valve III-17, a second throttling valve III-18, a first purifier III-19, a second pressure reducing valve III-20, a second purifier III-21, a third throttling valve III-22, a second flow meter III-23, a fourth throttling valve III-24 and a fifth throttling valve III-25. And the gas automatic sample introduction module III is positioned at the right lower part of the electronic nose instrument.

The computer control and analysis module IV mainly comprises the following components: the computer mainboard IV-1, the A/D data acquisition card IV-2, the driving and control circuit board IV-3, the 4-path precise direct current stabilized voltage supply IV-4, the display IV-5 and the WIFI module IV-6 are positioned on the left side of the electronic nose instrument. And the WIFI module IV-6 is used for transmitting the sensing information of the gas sensor array module I and the capillary gas chromatographic column module II to a specified fixed/mobile terminal in real time.

FIG. 2 shows the electronic nose instrument in the gas sample injection period T₀The information sensing, selecting and processing conditions of the gas sensor array module I and the capillary gas chromatography column module II are schematically shown in 480 s. Although the two modules have different shapes of sensing response curves and data recording time lengths, the single period T is₀The last 10s of information selection and analysis processing are performed simultaneously.

FIG. 3 is a gas injection monocycle T₀Within 480s, the response curve of the gas sensor is a multi-feature selection schematic diagram. The figure shows an example of the curves for 3 gas sensors, TGS822, TGS826 and TGS832, respectively, for a petroleum wax sample, 2,000ppm ethylene gas and 5,000ppm ethanol boil-off gas. Wherein the voltage response curves of FIG. 3(b) and FIG. 3(c) have the same steady state maximum value, i.e., v_a＝v_b. If only according to the conventional single steady-state maximum characteristic selection method of the voltage response curve, the electronic nose instrument cannot distinguish 2,000ppm of ethylene gas from 5,000ppm of ethanol volatile gas. Through careful observation, we send outNow, fig. 3(b) and 3(c) show case 1: although the maximum values of the voltage response steady states are equal, the peak values correspond to unequal peak-out times, and the areas under the curves are also unequal. Similarly, there is also case 2: the peak-out time is equal, but the peak value is not equal to the area under the curve. Case 3: the area under the curve is equal, but the peak time and the peak value are not equal to each other.

According to fig. 3, the invention proposes that the steady-state maximum v of the voltage response is simultaneously selected from the response curve of a gas sensor i (═ 1,2, …,16)_gsi(τ) the corresponding time t of the peak from the beginning of the sampling of the headspace of the gas under test_gsi(τ) plus the area under the curve A of the measured gas headspace sampling period 60s_gsi(τ). If the gas sensor array consists of 16 sensitive elements, the gas sampling period T is one₀In an information selection and processing area with the duration of 10s, the computer control and analysis module IV sequentially selects 3, 16 and 48 characteristic values from 16 response curves as primary sensing information of the gas sensor array module I on the gas to be detected, and the primary sensing information is marked as x_gs(τ)＝{(v_gs1(τ),v_gs2(τ),…,v_gs16(τ))；(t_gs1(τ),t_gs2(τ),…,t_gs16(τ))；(A_gc1(τ),A_gc2(τ),…,A_gc16(τ))}。

FIG. 4 shows a gas injection single period T₀And selecting information of a semi-separation chromatogram within 480 s. In gas sample introduction monocycle T₀In an information selection area with the duration of 5s, the computer control and analysis module IV sequentially selects 10 groups of { peak height h } from the semi-separation chromatogram_gci(τ), retention time t_gci(τ) } (i ═ 1,2, …,10) and area a under the graph curve for the specified duration 470s_gc(tau) 21 characteristic values are taken as the primary sensing information of the capillary chromatographic column module II on the measured gas and are recorded as x_gc(τ)＝{(h_gc1(τ),h_gc2(τ),…,h_gc10(τ))；(t_gc1(τ),t_gc2(τ),…,t_gc10(τ))；A_gc(τ)}。

FIG. 5 shows a gas injection single period T₀Within 480s, two semi-separated chromatogramsFigure schematic representation of feature selection. The semi-separation chromatogram of FIG. 5(a) has only 8 chromatographic peaks, and only 8 peak values h can be obtained_gci(τ) (i ═ 1,2, …,8) and corresponding retention time t_gci(τ) (i ═ 1,2, …,8), plus area a under the curve of the semi-resolved chromatogram_gc(τ). Our approach is to zero out an insufficient number of chromatographic peaks and corresponding retention times, so that the resulting chromatographic perception information is x according to fig. 5(a)_gc(τ)＝{(h_gc1(τ),h_gc2(τ),…,h_gc8(τ),0,0)；(t_gc1(τ),t_gc2(τ),…,t_gc8(τ),0,0)；A_gc(τ) }. The semi-separation chromatogram of fig. 5(b) has more than 10 chromatographic peaks, and we can select the top 10 maximum chromatographic peaks from the chromatogram.

The invention regards the semi-separation chromatogram as a part of the perception information, namely the mode, of the electronic nose instrument, combines the perception information of the gas sensor array, establishes the big smell data, and realizes the unknown smell identification, qualitative analysis and main component quantitative prediction by means of an artificial intelligence machine learning method. In a gas single sampling period T₀The information selection and processing area lasts for 10s, the computer control and analysis module IV fuses the gas sensor sensing information of the gas sensor array module I and the capillary chromatographic column module II in different time periods, and normalization preprocessing is carried out, so that a sensing information vector x (tau) x (x) of the electronic nose instrument to a gas sample to be detected is obtained_gs(τ)+x_gc(τ)＝{(v_gs1(τ),v_gs2(τ),…,v_gs16(τ))；(t_gs1(τ),t_gs2(τ),…,t_gs16(τ))；(A_gc1(τ),A_gc2(τ),…,A_gc16(τ))；(h_gc1(τ),h_gc2(τ),…,h_gc10(τ))；(t_gc1(τ),t_gc2(τ),…,t_gc10(τ))；A_gc}∈R⁶⁹The perception vector x (τ) ∈ R⁶⁹The method is a basis for online type recognition and main component quantitative prediction of odor of a biological fermentation process and a foul odor pollution point by an electronic nose instrument.

FIG. 6 is a multi-parameter divide-and-conquer method of deep convolutional neural network machine learning model facing to a continuous online analysis mode "And (5) quantizing the prediction schematic diagram. The method comprises the following specific steps: according to a time sequence matrix X (tau-q) recently sensed by the gas sensor array module I and the capillary gas chromatographic column module II, the fermentation and malodor pollution types, the odor intensity and the concentration value of the main components are predicted one by a plurality of single-output deep convolution neural networks. Here, τ is the current time, q is the time that has passed recently, and τ -q is the recent time interval. Thus, the time series matrix X (τ -q) has a dimension scale of R^69×(τ-q+1). The value of q is generally suitable for the recent time length of about 6 hours in the fermentation or stink pollution process.

In order to determine the modular convolutional neural network model structure and parameters, the primary task is to establish odor big data, including: the gas sensor array module I and the capillary gas chromatographic column module II sense data of a large number of biological fermentation processes and odorous polluted areas on line in the year-round period; off-line monitoring data of conventional instruments such as a chromatograph, a mass spectrometer and a spectrophotometer; odor label data of known type and composition; and sensory evaluation data.

What is needed next is the fusion of the sensory data of the gas sensor array and the sensory data of the capillary gas chromatographic column, including normalization and dimensionality reduction preprocessing. In order to reduce the difficulty of odor big data analysis, a 'divide and conquer' strategy is adopted to decompose a complex multi-odor qualitative and quantitative analysis problem, namely a complex multi-odor type identification problem and a complex multi-odor intensity and composition quantitative estimation problem into a plurality of odor types one-by-one identification and a plurality of simpler single odor intensities and important composition quantitative prediction problems one-by-one, namely, a problem of n curves/curved surface integral fitting is decomposed into a problem of n curves/curved surfaces one-by-one fitting, and the problem is solved by n single-output deep convolution neural network models in a one-to-one correspondence manner.

The invention adopts a plurality of modularized single-output deep convolutional neural networks to realize multi-parameter online quantitative prediction. A single-output deep convolutional neural network is composed of an input layer, 3 convolutional layers, 2 downsampling layers and 1 output unit, and mainly learns labeled data and data with known components in smell big data. Each hidden layer andthe output layer activation functions are all modified Sigmoid activation functionsThe method comprises the steps of adopting an error back-propagation off-line layer-by-layer learning algorithm, enabling the size of a convolutional layer scanning window to be 5 × 5, enabling the step length of overlapping scanning to be 1, enabling convolution kernels to be a combination of sine kernels, cosine kernels, polynomial kernels, Gaussian kernels, Sigmoid kernels, wavelet kernels and L aplace kernels, enabling the size of a down-sampling layer scanning window to be 2 × 2, enabling non-overlapping scanning to be 2, enabling the step length to be 2, extracting maximum values, mean values and mean square deviations from each scanning window, and enabling n single-output deep convolutional neural network models to predict multiple quantitative index values of the to-be-generated moments tau +1, tau +2, tau +3 and the like one by one according to a gas-sensitive/gas-chromatographic near-term sensing time sequence matrix X (tau-q) in a decision stage, wherein the multiple quantitative index values comprise odor types, intensity.

The specific elements of the gas sensitive/gas chromatography near-term perception time series matrix X (tau-q) are:

the invention takes m as 69 and q as 9.

When only 1 point is detected, namely the gas circulation sampling period T and the gas single sampling period T₀Setting "q" 9 "corresponds to predicting the possible changes of the odor intensity and the main composition at 8min, 16min and 24min in the future according to the gas-sensitive/gas chromatography time-series sensing response matrix from the current time τ to the past 1.2 hour period, and the change of the fermentation process or the odor environment at the past 1.2 hour period. When 5 points are detected, namely the gas circulation sample introduction period T is 5T₀Setting "q" 9 "is equivalent to predicting the possible changes in odor intensity and major components 40min, 80min, and 120min in the future from the gas sensitive/gas chromatograph time-series perceptual response matrix for the time period from the current time τ to the past 6 hours.