Multi-component online real-time stink analysis method for fixed and alive excrement electronic nose instrument

Abstract

The invention provides a stink multi-concentration index online real-time analysis method based on a fixed-excrement and live-excrement electronic nose instrument. The instrument mainframe box is composed of a gas sensor array module, a headspace sampling module, a pressure multiplication module and a computer control and analysis module. The software module comprises upper and lower computer communication, a computer operation platform and a machine learning model and algorithm. And within the gas sampling period T = 180-600s, the electronic nose instrument detects at one time to obtain a 50-dimensional gas-sensitive sensing sample. An electronic nose instrument establishes an odor big data set X through a large number of tests, and converts a complex odor recognition and concentration index analysis task into a plurality of single-type recognition and a plurality of single-concentration index estimation and prediction sub-tasks based on a division and treatment strategy and a convention and inspiration rule; learning groups with single neurons, a shallow neural network and a deep neural network as members are used for completing one by one according to simple and effective laws, and online real-time estimation and prediction of multiple concentration indexes of the odor pollutants are flexibly realized.

Description

Technical Field

[0001] The present invention - an online real-time analysis method for multiple components of odor for fixed-term or live electronic nose instruments, is aimed at the needs of analyzing the odor pollution status in chemical parks, landfills, sewage treatment plants, livestock and poultry farms and other areas, and involves scientific and technological fields such as artificial intelligence, computers, analytical chemistry, and ecological environment. It mainly solves the problems of existing electronic nose instruments with unsatisfactory sensitivity and selectivity, increased uncertainty in perceived information, and poor online real-time capabilities, so as to flexibly realize continuous online or on-site real-time analysis and prediction of multiple components of odor pollutants. Background Art

[0002] Odor is an environmental hazard that seriously disturbs people and endangers health. The current National Environmental Protection Law lists it as one of the seven major pollutants that need to be prevented and controlled.

[0003] GB14554 "Emission Standards for Odor Pollutants" (Draft for Comments) published by the Ministry of Ecology and Environment on December 3, 2018 stipulates that the emission control indicators of odor pollutants are composed of 8 single component concentrations such as hydrogen sulfide H2S, ammonia NH3, methyl mercaptan CH4S, dimethyl disulfide C2H6S2, methyl sulfide C2H6S, carbon disulfide CS2, trimethylamine C3H9N, styrene C8H8 and 1 overall odor concentration, that is, 8+1 quantitative indicators. Both the current GB14554 standard and the new standard (Draft for Comments) stipulate that the three-point comparison odor bag method is used to determine the odor concentration, the gas chromatography method is used to determine the concentrations of H2S, CH4S, C2H6S, C2H6S2, C3H9N and C8H8, and the spectrophotometry method is used to determine the concentrations of NH3 and CS2. According to the industry standard HJ1262-2022 "Ambient air and exhaust gas - Determination of odor - Three-point comparison odor bag method", "odor concentration" is Odor Room Use odorless clean air On-site collection The dilution factor when the odor sample is continuously diluted to the olfactory threshold of the odor detector; the EU standard EN17325-2003 uses OU (odor unit) value for measurement.

[0004] Whether it is the conventional instrumental method for measuring the concentration of 8 single components such as malodorous H2S or the three-point comparison odor bag method for measuring odor concentration, the difficulties faced are: (1) the sensory measurement process is complicated, time-consuming and labor-intensive; (2) the consistency of the olfactory threshold is poor and not objective. One odor often contains dozens or hundreds of malodorous components, and dozens of them are more harmful to health. (3) Conventional instrumental methods such as chromatography have complex operating processes and harsh testing conditions; the common defects of the two are high cost, low efficiency, and poor real-time performance. Therefore, the machine olfaction method based on big data and artificial intelligence technology is particularly attractive.

[0005] One of the development trends of machine olfaction is to develop highly sensitive and highly selective gas sensors. It is encouraging that the gas sensitivity of SnO2 metal oxide semiconductor (MOS) material has reached 10 -8 The V / V (ppb) level does not require secondary amplification and directly produces a V-level response, which is very attractive for odor detection and analysis. The second trend is to discover new working principles and new structures of electronic nose instruments based on typical applications. An array is composed of multiple gas sensors with overlapping perception ranges, and the structure of multiple functional components such as the gas sensor array is optimized to develop a miniaturized, integrated, and automated electronic nose instrument. Its multiple working parameters are optimized to significantly improve the overall selectivity and sensitivity of the instrument. The third trend is to establish a large data set of complex odors and invent simple and efficient information enhancement and machine learning methods, including gas sensor response drift compensation, to achieve online real-time identification of complex odors and estimation and prediction of multiple concentration indicators.

[0006] The application prospects of machine olfactory methods are broad. The relevant search results are as follows: (1) Papers. There were only more than 60 papers before 1990, more than 500 papers before 2000, and now there are more than 13,000 papers in total, indicating that this research has been widely carried out since 2000 in this century. (2) Invention patents. More than 2,500 international patents and more than 700 domestic patents were mostly applied for or authorized in the past 10 years, indicating that the protection of machine olfactory intellectual property rights has been taken seriously. (3) Technical standards. In 2020, the IEEE Standards Association issued technical standards related to artificial olfactory equipment and systems. (4) Domestic related work is basically based on laboratory research on foreign electronic nose products. The above results show that the theoretical and applied research of machine olfactory needs to be deepened.

[0007] “ On-site real-time analysis "It means - a one-time or limited number of detection and analysis of odor or volatile gas of source substances on site, which can be tested as soon as the person arrives and as soon as the sample is prepared, and the detection and analysis cycle is 5-10 minutes;" Continuous online analysis "It refers to the continuous and multiple detection and analysis of malodorous gases on site, and the detection and analysis cycle is still 5-10 minutes;" Identification / Classification "refers to determining the aroma and its source, fragrance, origin or authenticity;" estimate "refers to quantifying the current intensity, level and concentration of odor components;" predict " refers to quantifying the intensity, level and component concentration of odor in the near future. The present invention is based on odor pollutants Continuous online Detection, estimation and prediction are the main application backgrounds, supplemented by Live real time Identification and estimation, thus organically combining the two working modes of the electronic nose instrument: continuous online and on-site real-time.

[0008] According to the Web of Science core database, there are very few papers on the application of machine olfactory methods in odor detection and analysis, only more than 200 papers, less than 2% of the total number of papers in this field, most of which are offline detection and data processing of indoor air, water, and soil samples. The existing odor monitoring system uses a "one point one nose" camera method. Long-term exposure to wind, sun and rain will inevitably cause fatigue and "poisoning" of gas-sensitive components, resulting in a series of problems such as short life, high cost, low sensitivity, and few concentration indicators. The Shanghai Laogang Landfill, which installed the French RQ-BOX electronic nose, and the Tianjin Dagang Industrial Zone, which spent 14.835 million yuan to install 102 German Olfosense electronic noses, have been listed in the top few environmental protection reports many times, which is a clear proof. These primary applications did not consider the fundamental problems such as "MOS components have poor selectivity, and the sensitivity of EC and PID components is 1-2 orders of magnitude lower than MOS", and their effects are very limited.

[0009] One of the key scientific issues and technical bottlenecks of electronic noses for odor detection and analysis applications is that big data and machine learning have not been studied from the perspective of complex odor analysis needs, resulting in low online real-time analysis capabilities of olfactory simulation technology, which has not been able to leave the laboratory. In order to apply olfactory simulation methods and technologies to online real-time detection and analysis of odor pollutants, we must solve the following scientific and technological problems:

[0010] (A) Gas sensing mechanism and information selection issues oriented towards typical applications

[0011] There is a tendency for the theory and practice to be out of touch in the research of gas sensing mechanism. On the one hand, material researchers are striving to find more sensitive gas sensing materials and develop more advanced manufacturing processes; on the other hand, there are so many gas compounds that it is too arbitrary to draw conclusions about whether a material is sensitive or not based on laboratory tests of a limited number of compounds. Therefore, it is particularly important to study the gas sensing mechanism with an application-oriented approach.

[0012] One of the scientific and technological problems to be solved is to establish odor big data to reveal the variation patterns of micro- and macro-characteristics of MOS gas sensors and typical compounds, such as molecular weight, structure, functional group, and strength: (i) 8 single components and odor concentrations specified by GB14554; (ii) 5 standard odorous liquids and n-butanol concentrations specified by HJ1262; (iii) ethanol concentrations. Table 1 is a detailed list of micro-indicators of typical compounds. Similar to the human body BMI index, we conceived that useful patterns can be discovered through the normalized combination of indicators such as molecular weight, structure, and functional group.

[0013] Will the steady-state response peaks of a gas sensor for 1ppm gas A and 10ppm gas B be exactly the same? It is entirely possible. This example tells us a truth: it is not feasible to select a single information from a single response curve because it does not satisfy the triangular stability principle. Therefore, we should study how to select multiple information components from a single response curve to achieve the goal of satisfying the triangular stability principle and helping to improve selectivity and sensitivity.

[0014] (B) Simple, efficient, flexible and changeable machine learning models and algorithms to solve the problem of online real-time analysis of multiple concentration indicators of many malodors

[0015] Table 1, Detailed list of molecular weight, structure, and functional group microscopic characteristics of typical malodorous compounds

[0016]

[0017] For industrial applications, the use of electronic nose instruments is only for smell. Identification , its role will be very limited, more importantly, it can perform online real-time analysis of odor intensity, quality level and concentration of multiple main components estimate and predict The machine learning method in the olfactory simulation system has its own particularity compared with other fields such as image and speech. The machine learning model and algorithm must not only complete the task of identifying the odor and its source substances, but also complete the current state of the complex odor. estimate and future state predict , it is necessary to flexibly complete multiple tasks such as recognition, recognition + estimation, estimation, estimation + prediction, etc. This invention focuses on the latter two.

[0018] Given the diversity of odor components and concentrations and the variability of the atmospheric environment, it is difficult to establish a reliable analysis model based on a small data set. It must be pointed out that the relevant papers are basically using conventional single models and algorithms in the laboratory to perform offline identification of small samples and a few categories of odors, which is far from people's requirements and expectations for the industrial application of electronic nose instruments, and indirectly explains why it has not been able to leave the laboratory.

[0019] Big data has profoundly changed people's lives and ways of thinking. In order to flexibly perform multiple types of analysis tasks such as identification, identification + estimation, estimation + prediction, we must propose simple and efficient machine learning models and algorithms, and combine offline learning of odor big data with online learning of recent small data sets to achieve the purpose of online real-time analysis and prediction of multiple parameters. Therefore, the scientific problems that need to be solved include: ① How to decompose the overall analysis problem of multiple complex odors into multiple simpler binary classification, single concentration and multiple single time series sub-datasets based on the principle of "divide and conquer" and the law of "building on the past and ushering in the future"; ② How to use multiple different types of machine learning sub-models and algorithms that are as simple and effective as possible to learn the aforementioned multiple sub-datasets one by one based on Occam's razor law - the "simple and effective" law; ③ Discover simple and effective decision rules to form a flexible machine learning model to complete multi-type odor analysis tasks.

[0020] Another scientific and technological problem that needs to be solved is how to use the new mechanism of gas sensor perception and simple and efficient machine learning methods to establish odor big data based on a large number of multi-scale on-site gas sensor perceptions, and use detailed and reliable analysis results as support to incorporate machine olfaction methods into the industry standards for odor pollution status and online real-time analysis of various odor pollutants, solve the problem of the vagueness of various odor pollutant indicators in the existing national / industry standards in the odor field, and realize real industrial applications.

[0021] In summary, the present invention is based on the general requirement of "demand-driven, breakthrough of bottlenecks", breaks through the bottleneck of olfactory simulation methods and instrument industrial application technology, reveals the changing laws of complex odor processes, proposes new machine learning models and algorithms, and realizes multi-scale online real-time detection of odor and intelligent analysis of multi-state parameters to meet the urgent needs of the environmental protection industry. Summary of the invention

[0022] The present invention is based on the existing invention patents "A system and method for multi-point centralized online monitoring and analysis of malodorous gases" (see patent application number: 2018104716131) and "Big data driven multi-point centralized online analysis method of malodorous gases by electronic nose instruments" (see patent application number: 2018104717083), and invents a multi-component online real-time analysis method for malodors for fixed-point and live electronic nose instruments to solve the problem of on-site or long-term continuous online monitoring or short-term real-time detection of malodor pollution, so as to realize online real-time quantitative estimation and prediction of odor concentration and concentration of multiple malodorous compounds.

[0023] In order to achieve the above object, the present invention provides the following technical solutions:

[0024] The hardware of the electronic nose instrument includes gas sensor array module I, gas automatic headspace sampling module II, booster cylinder III, computer control and analysis module IV, backup power supply V, and clean air bottle VI. Among them, the four components of gas sensor array module I, gas automatic headspace sampling module II, booster cylinder III, and computer control and analysis module IV constitute the main box; the software includes a computer operating platform and machine learning models and algorithms. The electronic nose instrument flexibly realizes fixed or mobile detection of odor pollution on site, identification of odor types, and quantitative estimation and prediction of the concentration of multiple odor compounds.

[0025] The main components of the gas sensor array module I are 16 gas sensors with partially overlapping sensing ranges, forming a gas sensor array. The gas sampling cycle of the fixed-duty electronic nose instrument for a malodor observation point is T = 180-600s, and the default value is T = 300s. In order to meet the triangular stability principle, in the kth cycle corresponding to the time kT, the computer control and analysis module IV selects the steady-state voltage peak value v from each response curve of the gas sensor array with a duration of 60s. i,0 (k), corresponding to the peak time t i,0 (k), area under the curve A i,0 (k) These three original sensing components, i = 1, 2, .., 16; therefore, the computer control and analysis module IV extracts 3*16=48 response components from the 16 response curves of the gas sensor array in period kT, and adds the local ambient temperature value T at that time. em,0 (k) and ambient humidity value H um,0 (k), the electronic nose instrument obtains a 50-dimensional initial perception vector x0(k) = (v 1,0 (k),…,v 16,0 (k),t 1,0 (k),…,t 16,0 (k),A 1,0 (k),…,A 16,0 (k),T em,0 (k),H um,0 (k)) T ∈R 50 , called the initial sample, is stored in the hard disk data file of the computer control and analysis module IV, and x0(k) is sent to multiple designated fixed / mobile terminals and the cloud through the WIFI board.

[0026] The Dinghuo two-hand electronic nose instrument forms odor big data X through on-site or long-term continuous online or short-term multiple detections of various odor pollution sites. Part of the data is used to establish the correspondence between the gas sensor array response samples and the odor types and the 8 single components specified in GB14554-1993 "Emission Standard of Odor Pollutants", including hydrogen sulfide H2S, ammonia NH3, methyl mercaptan CH4S, dimethyl disulfide C2H6S2, methyl sulfide C2H6S, carbon disulfide CS2, trimethylamine C3H9N, styrene C8H8 and 1 overall odor concentration, that is, 8+1 odor pollutant concentration control indicators.

[0027] In the learning stage, the machine learning cascade model software of the computer control and analysis module IV learns the odor big data X offline to determine its own structure and parameters, and grasps the correspondence between the gas sensor array response samples and the pollution type and 8+1 concentration control indicators. In the decision-making stage, based on the current perception vector x0(k) of the gas sensor array, the machine learning cascade model online and real-time quantitatively estimates the 8+1 odor pollutant concentration control index values ​​specified in GB14554, and determines the odor pollution type as needed. Furthermore, the machine learning cascade model predicts the perception response of the gas sensor array I-1 within the next 1 hour by learning the long-term series perception vectors generated by the gas sensor array within the last 72 cycles of about 6 hours online, and predicts the 8+1 odor pollutant concentration control index values ​​within the next 1 hour online based on this.

[0028] The online real-time detection and analysis of multiple components of odor by the Dinghuo Liangbian electronic nose instrument includes the following steps:

[0029] (1) Preparation: (i) For on-site real-time work scenarios with AC power supply, carry the main chassis and 2L clean air bottle VI-2; (ii) For scenarios with AC power supply and continuous online work, install the main chassis and 40L clean air bottle VI-1. (iii) For field work scenarios without AC power supply, carry the main chassis, backup power supply - 25.2V lithium battery pack V and 2L clean air bottle VI-2.

[0030] (2) Power on: Preheat the instrument for 30 minutes; set the “gas injection cycle T”, the default value is T = 5 minutes.

[0031] (3) The online real-time detection and analysis cycle of odorous gases begins:

[0032] (3.1) Malodorous gas detection:

[0033] In a gas sampling cycle T, depending on the gas type and flow rate, the gas sensor array module I undergoes: ① precise calibration, ②

[0034] There are 5 stages: 1. balance, 2. headspace injection, 3. pressure doubling, and 4. exhaust gas emission;

[0035] From the beginning of the 6th second of the ⑤ exhaust gas emission phase to the end of the 35th second of the cycle T, the electronic nose instrument enters the information processing and analysis phase, which lasts for 30 seconds. From the beginning of the 6th second of the ② equilibrium phase, including the entire ③ headspace injection and ④ pressure doubling phases, until the end of the 5th second of the ⑤ exhaust gas emission phase, the computer control and analysis module IV records and stores the sensing information of the gas sensor array I-1, which lasts for 70 seconds, and then selects the steady-state peak value v from each gas sensor voltage response curve with a duration of 70 seconds. i,0 (k), peak time t i,0 (k), area under the curve A i,0 (k) These three sensing information components, the 16*3=48 sensing information components of the gas sensor array module plus the local ambient temperature T em,0 (k) and ambient humidity H um,0 (k) together form the initial sample x0(k)∈R 50 .

[0036] (3.2) Malodorous gas analysis:

[0037] (a) The k=1th gas sampling cycle: The electronic nose instrument performs the first detection on a certain odor pollution point and obtains the initial sample x0(1), which is normalized to x(1). The machine learning cascade model of the computer control and analysis module IV simultaneously identifies the odor pollution type and quantifies and predicts the concentration indicators of 8+1 typical odor compounds.

[0038] (b) k=2 to 71 consecutive gas sampling cycles: The electronic nose instrument conducts online real-time detection of a certain odor pollution point for k<72 cycles, and obtains normalized finite time series samples {x(1), x(2),…, x(k<72)} including x(1). When the odor pollution type is still k=1, the first level of the machine learning cascade model - voting to identify the winning group represented, the second level of the cascade model - the quantitative prediction group corresponding to the winning group only needs to perform quantitative estimation of 8+1 typical compound concentration indicators.

[0039] (c) k ≥ 72 consecutive cycles of gas sampling: The electronic nose instrument performs online real-time detection on a certain odor pollution point for k ≥ 72 cycles, and obtains a normalized time series response sample set {x(1),…,x(72),x(73),…,x(k)}; the latest 72 consecutive samples are intercepted from them to form a moving long time series sample set X(k-τ) = {x(k-τ)}∈R with Δk = 72 72 ×50 ,τ=0,1,…,71. For example, a recent Δk=72 moving long time series is

[0040] {x(5),…,x(72),…,x(k=76)}. Next, to predict the response values ​​of the future k+2, k+6, and k+12 cycles, 102 supervised ▽k=12 short time series convolution training subsets {X i (k+2)∈R 59×12 ,X i (k+6)∈R 55×12 ,X i (k+12)∈R 49×12},i=1,…,16,

[0041] 33,…,50. Step 3: 102 12-5-1 single-output neural networks learn the 102 ▽k=12 short time series convolution training subsets online in real time. Once the learning is completed, the 102 12-5-1 single-output neural networks immediately learn the ▽k=12 short time series {x i (k-11),x i (k-10),…,x i (k-1),x i (k)} Predict the response component values ​​and temperature and humidity values ​​of the gas sensor array in the future k+2, k+6, and k+12 cycles, and combine them with the peak time components that are not predicted to generate sample prediction values ​​for the future k+2, k+6, and k+12 cycles Finally, the second level of the machine learning cascade model - the quantitative prediction group corresponding to the winning group predicts the 8+1 concentration control index values ​​at the corresponding moment based on the three sample prediction values.

[0042] (4) Once the gas sampling cycle kT-⑤ exhaust gas emission phase ends at the same odor pollution point, or enters the next cycle (k+1)T,

[0043] Repeat process (3), or move to the next odor contamination point to start new detection and analysis.

[0044] The present invention uses 8 kinds of odor pollutants specified in GB14554 - hydrogen sulfide H2S, ammonia NH3, methyl mercaptan CH4S, dimethyl disulfide C2H6S2, methyl sulfide C2H6S, carbon disulfide CS2, trimethylamine C3H9N, styrene C8H8, and 5 kinds of standard odor liquids specified in HJ1262-2022 for the qualification examination of odor detectors - methylcyclopentanone C6H 10 O, β-phenylethanol C8H 10 O, γ-undecanoic acid lactone C 11 H 20 O2, β-methylindole C9H9N, isovaleric acid C5H 10O2 and n-butanol C4H used to test the olfactory ability of odorants 10 A total of 14 typical malodorous compounds were selected as analysis objects. Single or mixed samples with concentrations of 0.1 to 100 ppm were prepared in turn. The volatile gases of the samples were detected one by one using an electronic nose instrument. The relationship between the multi-dimensional perception information of the gas sensor array and the microscopic characteristics of the molecular weight, structure, functional groups and the macroscopic characteristics of the odor concentration of the 14 malodorous compounds was established.

[0045] The odor big data set X includes: (i) offline detection data of the above 14 typical odor compound samples by electronic nose instruments; (ii) a large amount of online real-time detection data of electronic nose instruments at multiple odor pollution sites; (iii) laboratory olfactory data of odor concentration of a large number of odor samples collected on site by a team of odor detectors; (iv) laboratory component analysis data of a large number of odor samples collected on site by conventional instruments such as chromatography and mass spectrometry; (v) atmospheric temperature / humidity data at the odor pollution detection site; (vi) complaint data from nearby residents; and (vii) data on the correspondence between multi-dimensional perception samples of gas sensor arrays and odor types and 8+1 concentration control indicators.

[0046] In the kth gas sampling cycle, the computer control and analysis module IV simultaneously selects the steady-state peak value v from the voltage response curve of the i-th gas sensor i,0 (k), peak time t i,0 (k), total area under the curve A i,0 (k) The normalized value calculation formulas of these three initial perception information components are: (a) Steady-state peak value v i (k):

[0047]

[0048] In formula (1), v max =10.0V is the working circuit voltage value of the gas sensor.

[0049] (b), peak time t i (k):

[0050]

[0051] In formula (2), t max =30+30=60s is the sum of the duration of the two stages ③ head space injection and ④ pressure doubling after the computer control and analysis module IV issues the "head space injection" command.

[0052] (c), total area under the curve A i (k):

[0053]

[0054] In formula (3), it is assumed that the maximum area is t max *v max =10*60=600.

[0055] In addition, assuming the maximum ambient temperature is 45°C, the normalized ambient temperature is T em (k) = T em,0 (k) / 45; the normalized ambient humidity is directly the relative humidity value, that is, H um (k) = H um,0 (k) / 100.

[0056] The normalized sample is x(k)=(v1(k),…,v 16 (k),t1(k),…,t 16 (k),A1(k),..,A 16 (k),T em (k),H um (k)) T =(x1(k),…,x 16 (k),x 17 (k),…,x 32 (k),x 33 (k),..,x 48 (k),x 49 (k),x 50 (k)) T ∈R 50 , after normalization, all input components are dimensionless.

[0057] The present invention flexibly uses the "divide and conquer" strategy to transform the pollution type identification implied by the odor large data set X and the estimation and prediction of 8+1 pollutant concentration control indicators into multiple single target identification and multiple single concentration indicator estimation and prediction problems, that is, decomposing the n odor pollution type identification problem into C n 2 =n(n-1) / 2 binary classification sub-problems, decomposing the estimation and prediction problem of concentration indicators of n types r=8+1 odor pollutants into n×r single indicator one-to-one estimation and prediction problems. According to Occam's "simple and effective" law, n(n-1) / 2+nr "three-person" teams are formed by single neurons with priority 1, single-output single hidden layer neural networks with priority 2, and single-output deep neural networks with priority 3, corresponding to n(n-1) / 2 binary classification subsets + nr single concentration indicator subsets; team members learn n(n-1) / 2+nr subsets formed by "divide and conquer" decomposition of the odor large data set X in order of priority.

[0058] The basic parameter settings for the neural network learning process are as follows: Based on the conventional Sigmoid, ReLU, and Softmax activation functions, a new Sigmoid activation function f(σ) = 3(1 + exp(-σ / 3)) is discovered, where σ is the total input of a certain hidden node or output node. A dynamic learning mechanism and algorithm that allows the most necessary samples to participate are established. The input components are proportionally amplified to [0, 6], the target output components for classification are amplified from {0, 1} to {0, 3}, and the target output components for concentration estimation and prediction are amplified from (0, 1) to (0, 3). The learning step factor η = 0.05 / N, where N is the number of samples in the training subset. The learning termination condition is that the root mean square error δ ≤ 0.075, and the maximum number of learning steps is ζ -1 = 10,000; The number of hidden nodes s in a single-output single-hidden layer neural network is 8; The deep neural network adopts a layer-by-layer learning algorithm based on peer single-hidden layer networks. max According to Occam's "simple and effective" law, in any "three-person" group, once the learning accuracy of a single neuron with a priority of 1 - the root mean square error δ meets the requirements, it automatically becomes the winner, and the other two members do not need to learn anymore, and so on. Thus, each group selects a winner with a simple structure and can effectively solve problems to participate in the decision-making; n(n - 1) / 2 classification winners form n voting decision-making groups, each group consists of n - 1 winners, and each classification winner participates and must participate in 2 of the voting decision-making groups; nr concentration estimation and prediction winners form n quantization prediction groups, each group consists of r winners.

[0059] The fixed-term and current-account electronic nose instrument conducts real-time or continuous online detection of the malodorous pollution site. The computer control and analysis module IV performs the online real-time analysis task of the concentration index values of 8 + 1 malodorous pollutants specified in GB14554-1993. According to the size of the continuous detection cycle number k, it is divided into the following 3 cases: (a) k = 1: Corresponding to the sample x(1), the winning group Ω of the first level of the machine learning cascade model - n voting decision-making groups

[0060] determines the type of malodorous pollution. The second level of the cascade model - the quantization prediction group Λ j corresponding to Ω j estimates the 8 + 1 concentration control index values one by one for the 8 + 1 winners. j

[0061] (b) 1 < k ≤ 72: Corresponding to the samples x(2) to x(72), the first level of the cascade model no longer votes, and the type of malodorous pollution is defaulted to Ω j , and the second level - the quantization prediction group Λ j of the 8 + 1 winners continues to estimate the 8 + 1 concentration control index values one by one.

[0062] (c) k>72: This is equivalent to continuous detection and sample acquisition time kT>6h when T=300s. The first level of the cascade model still does not vote, and the odor pollution type defaults to Ω j Following the principle of "building on the past and ushering in the future", the long time series response generated by the gas sensor array I-1 in the last 72 cycles (about 6 hours) is used to predict the multiple concentration index values ​​of odor pollutants in the next hour, with a focus on predicting the 8+1 concentration control index values ​​in the future k=2 (10 minutes), k=6 (30 minutes) and k=12 cycles (60 minutes).

[0063] When the number of consecutive cycles k>72 of the fixed-life electronic nose instrument sampling gas from a malodorous pollution point, the gas sensor array generates a long time series normalized data array {X(k-71),…,X(k)} with a response length of Δk=72 in the latest 72 cycles. Each component of the gas sensor array automatically generates the target output from a part of its own time series to form three short time series training sub-arrays with a time length of ▽k=12. i (k-71),…,x i (k-τ),…,x i (k)}∈R 72 For example, as shown in Table 2, X i (k+2), X i (k+6), X i The three (k+12)▽k=12 short time series convolution training subsets all have a dimension of 12 and the number of samples are 59, 55 and 49 respectively, which are used for online learning of three single-output shallow neural networks with a structure of 12-5-1, so as to take kT as the starting point and online predict the value of the i-th perception component at the future (k+2)T, (k+6)T and (k+12)T moments.

[0064]

[0065] Table 2c, time series training set X for predicting perception component i at time k+12 i (k+12)

[0066]

[0067] To speed up learning, assume that the normalized peak time t iThe time series of the gas sensor perception component corresponding to (k) remains unchanged at time (k+2)T, (k+6)T, and (k+12)T, and no time series convolution training subset is generated. Therefore, there are 2*16*3+2*3=102 short time series convolution training subset arrays, which are completed online within 30s by 102 single-output shallow neural networks with a structure of 12-5-1. Based on this, the normalized time series response {x(k+2), x(k+6), x(k+12)} of the gas sensor array in the future k+2 cycle, i.e., the 10th minute, the k+6 cycle, i.e., the 30th minute, and the k+12 cycle, i.e., the 60th minute is predicted online; and Ω j The corresponding quantitative prediction group Λ j Based on this, 8+1 concentration control index values ​​at the future (k+2)T, (k+6)T, and (k+12)T moments are predicted in turn.

[0068] In particular, it is pointed out that an important capability of the electronic nose instrument used in the online real-time detection and analysis of odor pollution in industry should be that once a perception vector x(k) is obtained through on-site real-time detection, the instrument can estimate the 8+1 concentration control index values ​​of odor pollutants in real time within 1 minute. As for the type of odor pollution, it is relatively secondary. Here, this can be achieved only by the second level of the machine learning cascade model - the quantitative prediction group. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 The present invention is a method for online real-time analysis of malodor multi-components of electronic nose instruments for fixed-life feces and feces, and the development and application of electronic nose instruments for fixed-life feces and feces.

[0070] Schematic diagram of the technical route for online real-time detection and analysis of multiple odor components.

[0071] Figure 2 The present invention is a method for online real-time analysis of multiple components of odor for a fixed-life electronic nose instrument and a working principle of a multifunctional odor electronic nose instrument.

[0072] Schematic diagram 1: The headspace injection position of malodorous gases, with continuous online working state with alternating current.

[0073] Figure 3 The present invention is a method for online real-time analysis of multiple components of malodor by an electronic nose instrument for fixed or active defecation. Within a period of T = 360s, the gas sensor array

[0074] Schematic diagram of the changes in gas injection parameters and gas sensor response in the working chamber I-2.

[0075] Figure 4 The present invention - a method for online real-time analysis of malodor multi-components of electronic nose instruments for fixed and live feces - a basic component of the machine learning cascade model

[0076] Meta-schematic diagrams, which are: single neuron, single-output single hidden layer neural network and single-output deep neural network.

[0077] Figure 5 The present invention is a method for online real-time analysis of odor multi-components for electronic nose instruments for fixed and live feces and stools - a new Sigmoid activation function and input

[0078] The combination of moderate amplification of the output components can significantly accelerate the learning speed of the neural network.

[0079] Figure 6 The present invention is a method for online real-time analysis of multiple components of malodor by an electronic nose instrument for fixed-life and dual-feces use, and a machine learning method for malodor type identification.

[0080] Schematic diagram of the first level of the cascade model - voting recognition group component units.

[0081] Figure 7 The present invention is a method for online real-time analysis of multiple components of odor for electronic nose instruments for both fixed and live feces, which is used for odor type identification and 8+1

[0082] Schematic diagram of the machine learning cascade model for online real-time quantitative estimation of odor pollutant concentration control indicators.

[0083] Figure 8 The present invention is a method for online real-time analysis of odor multi-components for fixed-life electronic nose instruments, which is used for controlling the concentration of 8+1 odor pollutants.

[0084] Schematic diagram of the machine learning cascade model for online real-time quantitative prediction of control indicators. DETAILED DESCRIPTION

[0085] The present invention is further described in detail below in conjunction with the accompanying drawings.

[0086] Figure 1 It is a schematic block diagram of the technical route for developing a fixed-function and active-function electronic nose instrument and the technical route for online real-time detection and analysis of multiple components of malodor of the present invention.

[0087] The specific invention contents include: (1) New methods for revealing the mechanism of gas-sensing perception and selecting and fusing multiple information. The law between gas-sensing perception and the micro- and macroscopic characteristics of odor is revealed, and a new method for enhancing the concentration of trace components and selecting and fusing multiple information of gas-sensing perception is proposed, focusing on realizing the high selectivity and high sensitivity perception of complex odors from the perspective of instrument design and application objects. (2) New models and algorithms for flexible machine learning. A large data set of odor is established, and the "divide and conquer" strategy, Occam's simple and effective law and the law of inheritance are comprehensively applied. Offline and online learning are combined, and recent and current perception information are combined. A flexible machine learning cascade model and algorithm integrating single neurons, shallow neural networks and deep neural networks are proposed to flexibly realize the multi-type recognition of odor and the estimation and prediction of multiple concentration indicators of odor pollutants. (3) Multi-parameter optimization combination and industrial application of electronic nose instruments for fixed and active use. A multi-working parameter optimization combination method for instruments characterized by large sample flow, small working chamber volume and rapid adsorption and desorption is invented to significantly improve the sensitivity, selectivity and stability of electronic nose instruments, and realize the online and real-time detection, recognition and estimation and prediction of multiple concentration indicators of fixed and active use odor.

[0088] Figure 2 This is one of the working principle diagrams of the multifunctional fixed-action and dual-action malodor electronic nose instrument of the present invention. The instrument is in the malodor gas headspace sampling position and is in a continuous online working state with alternating current. The main hardware components of the electronic nose instrument include: gas sensor array module I, gas automatic headspace sampling module II, booster cylinder III, computer control and analysis module IV, backup power supply - 25.2V lithium battery pack V, 40L clean air bottle VI-1, 2L clean air bottle VI-2.

[0089] Figure 3 The diagram is a schematic diagram of the gas sampling type parameters of the gas sensor array working chamber I-2 and the sensing response changes of one of the gas sensors in a gas sampling cycle T=360s of the fixed-action and active-action electronic nose instrument of the present invention. Figure 3(a) A gas sampling cycle T = 300s includes five stages: ① Precise calibration: Under the action of 1,000ml / min clean air purge for 60s, the gas sensor array I-1 accurately recovers to the standard state; ② Balance: There is no gas flow in the annular working chamber I-2 of the gas sensor array, which lasts for 10s; ③ Headspace sampling: Under the suction action of the micro vacuum pump II-1 for 30s, the gas sensor array I-1 responds sensitively to the 1,000ml / min malodorous gas under test; ④ Pressure doubling: After 30s of pressure doubling, the gas sensor array I-1 responds sensitively to the 1,000ml / min malodorous gas under test; s of 0.2Mpa clean air thrust, the booster cylinder III presses the 78ml of stored malodorous gas back to the annular working chamber I-2 of the gas sensor array, increasing the pressure in the chamber by 1 times, and the response of the gas sensor array I-1 is further increased; ⑤ Exhaust gas emission: Under the suction action of the micro vacuum pump II-1 at a flow rate of 6,000ml / min for 170s, the exhaust gas from the annular working chamber I-2 of the gas sensor array, the booster cylinder III and related pipelines is extracted, and the gas sensor array I-1 gradually returns to its initial state.

[0090] in accordance with Figure 3 (b) In a gas sampling cycle T, only 70s of gas sensor response information is actually recorded by the computer control and analysis module IV, including the last 5s of the equilibrium phase, the total of 60s of the complete headspace sampling phase and the pressure doubling phase, and the first 5s of the exhaust gas emission phase. After the data recording is completed, the information selection and analysis interval lasting 30s, the computer control and analysis module IV selects the steady-state voltage peak value v from the 70s response curve of the gas sensor i. i,0 (k), peak time t i,0 (k), area under the curve A i,0 (k) these three original perception components, i = 1, 2, .., 16. Therefore, the computer control and analysis module IV extracts 3*16=48 response components from the 16 response curves of the gas sensor array I-1 in one period T, and obtains a 48-dimensional initial perception vector or sample x0(k) = (…, v i,0 (k),…,t i,0 (k),…,A i,0 (k),…) T ∈R 48 , stored in the data file of the computer motherboard IV-3 hard disk, and the sample data x(k) is sent to the designated fixed / mobile terminal and the cloud through the WIFI board IV-5, which serves as the main basis for the machine learning cascade model to identify the type of odor pollution and quantify and predict the 8+1 concentration control indicators. The computer control and analysis module IV gives the results of odor type identification and quantification and prediction of 8+1 concentration control indicators within the 30s information selection and analysis interval.

[0091] The present invention uses a flexible machine learning cascade model to flexibly complete four major tasks of online real-time analysis of odor pollution: ① Type Identification ; ② Type Identification And 8+1 concentration control indicators estimate ;③8+1 concentration control indicators estimate ;④8+1 concentration control indicators estimate and predict . The present invention comprehensively uses the "divide and conquer" strategy, Occam's simple and effective theorem, and the law of continuity, combines offline and online learning, and combines current and recent perception information, creatively using a flexible machine learning cascade model to flexibly complete the above four major categories of odor pollutant analysis tasks, reflecting "flexibility". In the specific implementation process, the present invention decomposes the n odor pollution type identification problems into n(n-1) / 2 binary classification sub-problems, decomposes the n-type r (=8+1) odor pollutant concentration index estimation and prediction problems into n×r single index estimation and prediction sub-problems, and uses single neurons, single-output single hidden layer neural networks, and single-output deep neural networks as basic members of the machine learning cascade model to solve these n(n-1) / 2+n×r sub-problems one by one according to priority.

[0092] Figure 4 The structure of the machine learning cascade model is given: a single neuron, a single-output shallow neural network, and a single-output deep neural network. A single shallow neural network has a single output and a single hidden layer, and a single deep neural network has a single output and three hidden layers. The activation functions of all hidden and output layers of neurons, shallow neural networks, and deep neural networks are f(σ)=3(1+exp(-σ / 3)) -1 . Figure 5 This is a schematic diagram of the combination of the new Sigmoid activation function and the moderate amplification of the input and output component values. The new Sigmoid activation function f(σ)=3(1+exp(-σ / 3)) -1 Replace the original f0(σ)=(1+exp(-σ)) -1 After that, the significant change range of the first-order partial derivative becomes much wider, assisted by the increase of the input component from [0,1] to [0,6], and the output component from {0,1} to {0,3}, which increases the learning speed of the neural network by more than an order of magnitude and greatly improves the promotion ability!

[0093] Based on Occam's law of simplicity and effectiveness, the complexity of the neural network varies with the difficulty of the task. The present invention forms a learning group in the order of priority of (i) neurons, (ii) single-output shallow neural network, and (iii) single-output deep neural network to learn a subtask, using a dynamic learning and error back propagation algorithm with the most necessary samples, and the learning accuracy is uniformly set to the root mean square error ρ = 0.05, and the maximum learning factor ηmax =5 / N, where N is the number of samples in a subset. Once a member with a higher priority reaches the set accuracy, the learning process of the group is terminated, and the winner is generated, which again reflects flexibility. If a subtask can be completed by a single neuron, a more complex single-output shallow neural network will no longer be used, and so on. Therefore, a flexible machine learning cascade model is composed of a mixture of winners from multiple groups, which include neurons, single-output shallow neural networks, and single-output deep neural networks.

[0094] Figure 6 This is a schematic diagram of the constituent units of the first level of the machine learning cascade model for odor type identification of the present invention, the voting identification group. Assuming there are n types of odor, it is necessary to form n voting identification groups, each with n-1 members. In the learning stage, the n(n-1) / 2 single-output neural network model ω of the first level jk (j, k = 1, 2, ..., n and j ≠ k) n(n-1) / 2 normalized binary classification subsets are learned in priority order, and the winners participate in the subsequent voting decision. In the decision stage, a winner only needs the votes of two recognition groups, for example, the winner ω jk Participate in voting identification group Ω j and Ω k According to the majority voting rule, the group with the most votes among the n voting identification groups determines the type of sample x.

[0095] Figure 7 The machine learning cascade model of the present invention is used for both odor type identification and online real-time quantitative estimation of 8+1 odor pollutant concentration control indicators - voting winning group Ω j and the corresponding quantitative prediction group Λ j Schematic diagram. Quantitative prediction group Λ j It is composed of winners with simpler structures and single-concentration training subsets that can pass learning normalization and reach the set learning accuracy, and estimates the concentrations of 8+1 pollutants of odor type j one by one.

[0096] Figure 8 This is a schematic diagram of the machine learning cascade model for online real-time quantitative prediction of 8+1 odor pollutant concentration control indicators of the present invention. This model is still cascaded in form, but it no longer performs type recognition. The default type is known, and only continuous online quantitative prediction of 8+1 odor pollutant concentration control indicators is performed. The core of this invention is that the first level of the machine learning cascade model - the convolutional neural network layer online real-time learning Δk = 72 long time series response data array X(k-τ)∈R generated by the gas sensor array in the recent 6 hours 72×50, based on which the short-time series response of the gas sensor array within the next hour is predicted; on this basis, the second level of the cascade model - the quantitative prediction group with known odor types predicts 8+1 concentration control index values ​​of odor pollutants within the next hour based on the prior knowledge obtained from offline learning of the odor big data set X.

[0097] The specific steps to achieve the above invention objectives are as follows:

[0098] (1) Supervised short time series convolution training subset generation. The long time series response data array X(k-τ) = {x(k-τ)}∈R generated by the gas sensor array plus the local temperature and humidity sensor at that time is Δk = 72. 72×50 , τ=0,1,…,71 (adjustable), automatically generate several ▽k=12 short time series convolution training subsets with target output. In order to shorten the online real-time learning time, only 8+1 concentration control index values ​​at the future (k+2)T, (k+6)T, and (k+12)T are selectively predicted. Correspondingly, each long time series response component x i (k-τ)∈R 72 Automatically generate 3 ▽k=12 short time series convolution training subsets X containing target output i (k+2), X i (k+6) and X i (k+12). To predict the perception component value x of the future (k+2)T period i (k+2) as an example, from x i The first training sample pair generated by (k-τ) is {x i (k-13),x i (k-12),…,x i (k-3),x i (k-2)}-->x i (k); The second training sample pair is {x i (k-14),x i (k-13),…,x i (k-4),x i (k-3)}-->x i (k-1), until the 59th training sample pair {x i (k-71),x i (k-70),…,x i (k-61),x i (k-60)}-->x i (k-58). Therefore, the short time series training subset X i (k+2)∈R 59×12 There are 59 samples and the dimension is 12. Similarly, X i(k+6)∈R 55 ×12 , X i (k+12)∈R 49×12 For the specific generated data subset, please refer to “Summary of the Invention” - Table 2.

[0099] To speed up the learning process, we only analyze the 16 normalized steady-state voltage peaks v i (k), 16 areas under the curve A i (k), ambient temperature value T em (k) and ambient humidity value H um (k) is used to learn and predict the perceptual response time series, and the peak time t i The corresponding value of (k) remains unchanged because the latter does not change much in the next 1 hour. Therefore, a total of 2*16*3+2*3=102 short time series convolution training subsets are automatically generated.

[0100] (2) Online real-time learning of convolutional neural networks and prediction of the sensor response and temperature and humidity values ​​within the next 12 cycles of 1 hour

[0101] The present invention uses 102 single-output neural networks with a structure of 12-5-1 to form a Figure 8 The first level of the machine learning cascade model shown in Figure 1 is the convolutional neural network layer, which learns the above 102 short time series convolution training subsets one by one in real time. The convolutional neural network layer here is similar to Figure 6 and Figure 7 The single-output single-hidden-layer neural networks of the voting recognition group and the quantitative prediction group differ in functions, number of input and hidden nodes, training subsets, and winner selection. The similarities are single hidden layer, new Sigmoid activation function, learning accuracy, and step size factor.

[0102] After learning, these 102 12-5-1 single-output neural networks are based on ▽k=12 short time series {x i (k-11),x i (k-10),…,x i (k-1),x i (k)}

[0103] Predict the responses and temperature and humidity values ​​of the 102 gas sensor arrays one by one, that is, {x i (k+2), x i (k+6), x i (k+12)}, i=1,…,16,33,…,50. The perception component values ​​obtained by these predictions are taken as the main body, and the component values ​​of the peak time in the near future are added to equal the current time value, that is, {x i (k+2)=

[0104] xi (k), x i (k+6)=x i (k), x i (k+12)=x i (k)}, i = 17, ..., 32, perception vector prediction values ​​for the future periods (k+2)T, (k+6)T, (k+12)T

[0105] To be produced.

[0106] (3) Prediction of 8+1 concentration control index values ​​of odor pollutants within 1 hour in the next 12 cycles

[0107] Figure 8 The second stage of the machine learning cascade network shown is the quantitative prediction group Λ j Based on the prediction To predict the 8+1 concentration control index values ​​of odor pollutants.

[0108] In particular, an important capability that an electronic nose instrument used in the online real-time detection and analysis of odor pollution should have is that once a perception vector x(k) is obtained through on-site real-time detection, the instrument can estimate the 8+1 concentration control index values ​​of odor pollutants in real time within 1 minute. The type of odor pollution is relatively secondary. Figure 8 This is achieved by the second stage of the machine learning cascade model shown - the quantitative prediction group.

Claims

1. A method for online real-time analysis of multiple components of malodor for electronic nose instruments, characterized in that: The electronic nose instrument hardware includes a gas sensor array module I, a gas automatic headspace sampling module II, a pressurized cylinder III, a computer control and analysis module IV, a backup power supply V, and a clean air bottle VI; among them, the gas sensor array module I, the gas automatic headspace sampling module II, the pressurized cylinder III, and the computer control and analysis module IV constitute the main chassis; the software includes a computer operating platform and a machine learning model and algorithm; the electronic nose instrument flexibly realizes fixed or mobile detection of odor pollution on site, identification of odor types, and quantitative estimation and prediction of the concentration of multiple odorous compounds; The gas sensor array module I is mainly composed of 16 gas sensors with partially overlapping sensing ranges, forming a gas sensor array; the gas sampling cycle of the fixed-active electronic nose instrument for a malodor observation point is T = 180-600s, and the default value is T = 300s; in order to meet the triangular stability principle, in the kth cycle corresponding to the time kT, the computer control and analysis module IV selects the steady-state voltage peak value v from each response curve of the gas sensor array with a duration of 60s. i,0 (k), corresponding to the peak time t i,0 (k), area under the curve A i,0 (k) These three original sensing components, i = 1, 2, .., 16; therefore, the computer control and analysis module IV extracts 3*16=48 response components from the 16 response curves of the gas sensor array in period kT, and adds the local ambient temperature value T at that time. em,0 (k) and ambient humidity value H um,0 (k), the electronic nose instrument obtains a 50-dimensional initial perception vector x0(k) = (v 1,0 (k),…,v 16,0 (k),t 1,0 (k),…,t 16,0 (k),A 1,0 (k),…,A 16,0 (k),T em,0 (k),H um,0 (k)) T ∈R 50 , called the initial sample, is stored in the hard disk data file of the computer control and analysis module IV, and x0(k) is sent to multiple designated fixed / mobile terminals and the cloud through the WIFI board; The fixed-life electronic nose instrument forms a large odor data set X by conducting on-site or long-term continuous online or short-term multiple detections of various odor pollution. Part of the data is used to establish the corresponding relationship between the gas sensor array response samples and the odor types and the 8 single components and 1 overall odor concentration, namely 8+1 odor pollutant concentration control indicators specified in GB14554-1993 "Emission Standard of Odor Pollutants", including hydrogen sulfide H2S, ammonia NH3, methyl mercaptan CH4S, dimethyl disulfide C2H6S2, methyl sulfide C2H6S, carbon disulfide CS2, trimethylamine C3H9N, and styrene C8H8; In the learning stage, the machine learning cascade model software of the computer control and analysis module IV learns the odor large data set X offline to determine its own structure and parameters, and grasps the correspondence between the gas sensor array response samples and the pollution type and 8+1 concentration control indicators; in the decision-making stage, based on the current perception vector x0(k) of the gas sensor array, the machine learning cascade model online and real-time quantitatively estimates the 8+1 odor pollutant concentration control index values ​​specified in GB14554, and determines the odor pollution type as needed; further, the machine learning cascade model predicts the perception response of the gas sensor array I-1 within the next 1 hour by online real-time learning the long-time series perception vectors generated by the gas sensor array within the last 72 cycles of about 6 hours, and predicts the 8+1 odor pollutant concentration control index values ​​within the next 1 hour online and in real time based on this. The online real-time detection and analysis of multiple components of odor by the Dinghuo Liangbian electronic nose instrument includes the following steps: (1) Preparation: (i) For on-site real-time work scenarios with AC power supply, carry the main chassis and 2-6L clean air bottle VI-2; (ii) For scenarios with AC power supply and continuous online work, fix the main chassis and 40L large clean air bottle VI-1; (iii) For field work scenarios without AC power supply, carry the main chassis, backup power supply - 25.2V lithium battery pack V and 2-6L small clean air bottle VI-2; (2) Power on: Preheat the instrument for 30 minutes; set the "gas injection cycle T", the default value T = 5 minutes; (3) The online real-time detection and analysis cycle of odorous gases begins: (3.1) Malodorous gas detection: In a gas sampling cycle T, depending on the gas type and flow rate, the gas sensor array module I goes through five stages in sequence: ① precise calibration, ② balance, ③ headspace sampling, ④ pressure multiplication, and ⑤ exhaust gas emission; From the beginning of the 6th second of the ⑤ exhaust gas emission phase to the end of the 35th second of the cycle T, the electronic nose instrument enters the information processing and analysis phase, which lasts for 30 seconds. From the beginning of the 6th second of the ② equilibrium phase, including the entire ③ headspace injection and ④ pressure doubling phases, until the end of the 5th second of the ⑤ exhaust gas emission phase, the computer control and analysis module IV records and stores the sensing information of the gas sensor array I-1, which lasts for 70 seconds, and then selects the steady-state peak value v from each gas sensor voltage response curve with a duration of 70 seconds. i,0 (k), peak time t i,0 (k), area under the curve A i,0 (k) These three sensing information components, the 16*3=48 sensing information components of the gas sensor array module plus the local ambient temperature T em,0 (k) and ambient humidity H um,0 (k) together form the initial sample x0(k)∈R 50 ; (3.2) Malodorous gas analysis: (a) The k=1th gas sampling cycle: the electronic nose instrument performs the first detection on a certain odor pollution point to obtain the initial sample x0(1), which is normalized to x(1). The machine learning cascade model of the computer control and analysis module IV simultaneously identifies the odor pollution type and quantifies and predicts the concentration indicators of 8+1 typical odor compounds; (b) k = 2 to 71 consecutive gas sampling cycles: the electronic nose instrument performs online real-time detection on a certain odor pollution point for k < 72 cycles, and obtains a normalized finite time series sample {x(1), x(2), ..., x(k < 72)} including x(1). When the odor pollution type is still k = 1, the first level of the machine learning cascade model - voting to identify the winning group, the second level of the cascade model - the quantitative prediction group corresponding to the winning group only needs to perform quantitative estimation of 8 + 1 typical compound concentration indicators; (c) k ≥ 72 gas sampling cycles: The electronic nose instrument performs k ≥ 72 cycles of continuous online real-time detection on a certain odor pollution point. Get the normalized time series response sample set {x(1),…,x(72),x(73),…,x(k)}; extract the latest 72 consecutive samples from it to form a moving long time series sample set X(k-τ) = {x(k-τ)}∈R with Δk = 72 72×50 , τ=0,1,…,71; for example, a recent Δk=72 moving long time series is {x(5),…,x(72),…,x(k=76)}; Next, to predict the response values ​​of the future k+2, k+6, and k+12 periods, 102 supervised ▽k=12 short time series convolution training subsets {X i (k+2)∈R 59×12 ,X i (k+6)∈R 55×12 ,X i (k+12)∈R 49×12 }, i = 1, ..., 16, 33, ..., 50; Step 3, 102 12-5-1 single-output neural networks learn these 102 Short time series convolution training subset; once the learning is completed, these 102 12-5-1 single output neural networks are immediately based on Short time series {x i (k-11),x i (k-10),…,x i (k-1),x i (k)} Predict the response component values ​​and temperature and humidity values ​​of the gas sensor array in the future k+2, k+6, and k+12 cycles, and combine them with the peak time components that are not predicted to generate sample prediction values ​​for the future k+2, k+6, and k+12 cycles Finally, the second level of the machine learning cascade model - the quantitative prediction group corresponding to the winning group predicts the 8+1 concentration control index values ​​at the corresponding moment based on the three sample prediction values; (4) Once the gas sampling cycle kT-⑤ exhaust gas emission stage at the same odor pollution point is completed, either enter the next cycle (k+1)T and repeat the process (3), or move to the next odor pollution point to start new detection and analysis.

2. The method for online real-time analysis of multiple components of malodor for electronic nose instruments for both fixed and live feces according to claim 1 is characterized in that: The eight odor pollutants specified in GB14554 are hydrogen sulfide H2S, ammonia NH3, methyl mercaptan CH4S, dimethyl disulfide C2H6S2, methyl sulfide C2H6S, carbon disulfide CS2, trimethylamine C3H9N, and styrene C8H8, and the five standard odor liquids specified in HJ1262-2022 for the olfactory qualification examination are methylcyclopentanone C6H 10 O, β-phenylethanol C8H 10 O, γ-undecanoic acid lactone C 11 H 20 O2, β-methylindole C9H9N, isovaleric acid C5H 10 O2 and n-butanol C4H used to test the olfactory ability of odorants 10 A total of 14 typical malodorous compounds were analyzed, and single or mixed samples of 0.1 to 100 ppm were prepared in turn. The volatile gases of the samples were detected one by one using an electronic nose instrument to establish the relationship between the multi-dimensional perception information of the gas sensor array and the molecular weight, structure, functional group microscopic characteristics of the 14 malodorous compounds and the macroscopic characteristics of odor concentration; The odor big data set X includes: (i) offline detection data of the above 14 typical odor compound samples by electronic nose instrument; (ii) A large amount of online real-time detection data from electronic nose instruments at multiple odor pollution sites; (iii) Laboratory olfactory data on odor concentration of a large number of odor samples collected on site by a team of odor identifiers; (iv) Laboratory component analysis data of a large number of odor samples collected on site by conventional instruments such as chromatography and mass spectrometry; (v) Atmospheric temperature / humidity data at the odor pollution detection site; (vi) Complaint data from nearby residents; (vii) Correspondence data between multi-dimensional perception samples of gas sensor arrays and odor types and 8+1 concentration control indicators.

3. The method for online real-time analysis of malodor multiple components for electronic nose apparatus for both fixed and active defecation according to claim 1 is characterized in that In the kth gas sampling cycle, the computer control and analysis module IV simultaneously selects the steady-state peak value v from the voltage response curve of the i-th gas sensor i,0 (k), peak time t i,0 (k), total area under the curve A i,0 (k) The normalized value calculation formulas for these three initial perception information components are: (a), Steady-state peak value v i (k): In formula (1), v max =10.0V is the working circuit voltage value of the gas sensor; (b), peak time t i (k): In formula (2), t max =30+30=60s is the sum of the duration of the two stages of ③ headspace injection and ④ pressure doubling after the computer control and analysis module IV issues the "headspace injection" command; (c), total area under the curve A i (k): In formula (3), it is assumed that the maximum area is t max *v max =10*60=600; In addition, assuming the maximum ambient temperature is 45°C, the normalized ambient temperature is T em (k) = T em,0 (k) / 45; the normalized ambient humidity is directly the relative humidity value, that is, H um (k) = H um,0 (k) / 100; The normalized sample is x(k)=(v1(k),…,v 16 (k),t1(k),…,t 16 (k),A1(k),..,A 16 (k),T em (k),H um (k)) T =(x1(k),…,x 16 (k),x 17 (k),…,x 32 (k),x 33 (k),..,x 48 (k),x 49 (k),x 50 (k)) T ∈R 50 , after normalization, all input components are dimensionless.

4. The method for online real-time analysis of malodor multiple components for electronic nose apparatus for both fixed and active defecation according to claim 1 is characterized in that: Using the "divide and conquer" strategy, the pollution type identification and 8+1 pollutant concentration control index estimation and prediction problems implied in the odor big data set X are transformed into multiple single target identification and multiple single concentration index estimation and prediction problems, that is, the n odor pollution type identification problem is decomposed into C n 2 =n(n-1) / 2 binary classification sub-problems, decomposing the estimation and prediction problem of n types r=8+1 odor pollutant concentration indicators into n×r single indicator one-to-one estimation and prediction problems; according to Occam's "simple and effective" law, n(n-1) / 2+nr "three-person" teams are formed by single neurons with priority 1, single-output single hidden layer neural networks with priority 2, and single-output deep neural networks with priority 3, corresponding to n(n-1) / 2 binary classification subsets + nr single concentration indicator subsets; team members learn the n(n-1) / 2+nr subsets formed by the "divide and conquer" decomposition of the odor large data set X in order of priority; The basic parameters of the neural network learning process are set as follows: Based on the conventional Sigmoid, ReLU, and Softmax activation functions, a new Sigmoid activation function f(σ)=3(1+exp(-σ / 3)) is found -1 , σ is the total input of a hidden node or output node; establish a dynamic learning mechanism and algorithm that allows the most necessary samples to participate; the input component is proportionally enlarged to [0,6], the target output component for classification is enlarged from {0,1} to {0,3}, and the target output component for concentration estimation and prediction is enlarged from (0,1) to (0,3), the learning step factor η=0.05 / N, where N is the number of training subset samples, the learning termination condition—root mean square error δ≤0.075, and the maximum number of learning steps is ζ max =10,000; single-output single hidden layer neural network hidden nodes s = 8; deep neural network adopts layer-by-layer learning algorithm based on peer-to-peer single hidden layer network; According to Occam's "simple and effective" theorem, in any "three-person" group, once the single neuron learning accuracy - root mean square error δ with a priority of 1 meets the requirements, it automatically becomes the winner, and the other two members do not need to learn anymore, and so on; thus, each group selects a winner with a simple structure and the ability to effectively solve problems to participate in decision-making; n(n-1) / 2 classification winners form n voting decision groups, each group consists of n-1 winners, and each classification winner participates and must participate in 2 of the voting decision groups; nr concentration estimation and prediction winners form n quantitative prediction groups, each group consists of r winners.

5. The method for online real-time analysis of malodor multiple components for electronic nose instruments for both fixed and live feces according to claim 1 is characterized in that: The electronic nose instrument can detect odor pollution on site in real time or continuously online. The computer control and analysis module IV performs the online real-time analysis task of the concentration index values ​​of 8+1 odor pollutants specified in GB14554-1993. It is divided into the following three situations according to the number of continuous detection cycles k: (a) k = 1: corresponding to sample x(1), the winning group Ω of the first level of the machine learning cascade model—n voting decision groups j Determine the type of odor pollution, the second stage of the cascade model - with Ω j The corresponding quantitative prediction group Λ j The 8+1 winners estimate the 8+1 concentration control index values ​​one by one; (b)1 < k ≤ 72: Corresponding to samples x(2) to x(72), the first stage of the cascade model no longer votes, and the malodor pollution type is defaulted to Ω j , and the second stage - the quantization prediction group Λ j The 8 + 1 winners continue to estimate the 8 + 1 concentration control index values one by one; (c) k>72: This is equivalent to continuous detection and sample acquisition time kT>6h when T=300s. The first level of the cascade model still does not vote, and the odor pollution type defaults to Ω j Following the principle of "building on the past and ushering in the future", the long time series response generated by the gas sensor array I-1 in the last 72 cycles (about 6 hours) is used to predict the multiple concentration index values ​​of odor pollutants in the next hour, with a focus on predicting the 8+1 concentration control index values ​​in the future k=2 (10 minutes), k=6 (30 minutes) and k=12 cycles (60 minutes).

6. The method for online real-time analysis of malodor multiple components for electronic nose instruments for both fixed and live feces according to claim 1 is characterized in that: When the number of continuous cycles k>72 of the fixed-activity electronic nose instrument sampling gas from a malodorous pollution point, the gas sensor array generates a long time series normalized data array {X(k-71),…,X(k)} with a response length of Δk=72 in the latest 72 cycles. Each component is automatically generated by a part of its own time series to generate the target output and form three time series with the same length. The short time series training subarray of ; the i-th long time series component {x i (k-71),…,x i (k-τ),…,x i (k)}∈R 72 For example, as shown in Table 1a-c, X is generated i (k+2), X i (k+6), X i (k+12) these 3 The short time series convolution training subsets have a dimension of 12 and the number of samples is 59, 55, and 49, respectively, for online learning of three single-output shallow neural networks with a structure of 12-5-1, so as to online predict the value of the i-th perception component at the future time (k+2)T, (k+6)T, and (k+12)T starting from kT; To speed up learning, assume that the normalized peak time t i The time series of the gas sensor perception component corresponding to (k) remains unchanged at time (k+2)T, (k+6)T, and (k+12)T, and no time series convolution training subset is generated; therefore, there are 2*16*3+2*3=102 short time series convolution training subset arrays, which are completed online within 30s by 102 single-output shallow neural networks with a structure of 12-5-1. Based on this, the normalized time series response {x(k+2), x(k+6), x(k+12)} of the gas sensor array in the future k+2 cycle, i.e., the 10th minute, the k+6 cycle, i.e., the 30th minute, and the k+12 cycle, i.e., the 60th minute is predicted online; and Ω j The corresponding quantitative prediction group Λ j Based on this, 8+1 concentration control index values ​​at the future (k+2)T, (k+6)T, and (k+12)T moments are predicted in turn.

Images