METHOD FOR MEASURING THE CONCENTRATION OF A CHEMICAL COMPOUND CONTAINED IN A FLUID, BY MEANS OF AN OPTICAL MEASURING SYSTEM

Abstract

The present invention relates to a method for determining the concentration of a chemical compound in a fluid, by means of an optical measurement system for measuring a light intensity spectrum, such that: (i) by means of a training set comprising a plurality of measured training light intensity spectra for a plurality of training fluids of predetermined concentration, a model is built to determine the concentration of at least one chemical compound in a fluid by training a one-dimensional convolutional neural network on the training set, the network being such that the number of convolutional filters per layer decreases with the position of the layer in the network; (ii) by means of the optical measurement system, at least one light intensity spectrum of the fluid is measured; (iii) the constructed model is applied to the measured light intensity spectrum, and the concentration of the chemical compound in the fluid is determined.Figure 2 to be published.

Description

Title of the invention: METHOD FOR MEASURING DETERMINING THE CONCENTRATION OF A CHEMICAL COMPOUND CONTAINED IN A FLUID, BY MEANS OF AN OPTICAL MEASUREMENT SYSTEM technical field

[0001] The present invention relates to the field of measuring the concentration of at least one chemical compound contained in a fluid, by means of an optical measurement system for measuring light intensity spectra.

[0002] The present invention is advantageously applicable, but not limited to, the field of engine exhaust gas treatment or industrial flue gas treatment, or to the field of monitoring pollutant emissions from engine exhaust gases or industrial flue gases. The present invention may also, but not limited to, be applied to the field of air or bathing water quality monitoring, biogas treatment, gas leak monitoring, and more generally, industrial accidents.

[0003] Emission standards for pollutants from combustion (in internal combustion engines, turbines, industrial furnaces, incineration plants, thermal power plants, etc.) now require, in most cases, the reduction of the concentrations of these substances, for example, by means of post-combustion catalytic treatments. Pollutants in exhaust gases are typically unburned hydrocarbons, carbon monoxide (CO), nitrogen monoxide (NO), and nitrogen dioxide (NO2), commonly referred to by the acronym NOx, nitrous oxide (N2O), ammonia (NH3), sulfur compounds such as hydrogen sulfide (H2S) or sulfur oxides such as sulfur dioxide (SO2), ozone (O3), etc. Some of these substances are subject to strict regulations, according to which limit concentrations must be respected during atmospheric release.NOx, for example, has a detrimental impact directly on human health and indirectly through the secondary formation of tropospheric ozone. Regulations already mandate monitoring of the concentration of the NOx group formed by these nitrogen oxides, and may evolve to require monitoring of the concentrations of each nitrogen oxide individually, for example, NO2. Regarding NOx, it is known to implement a catalytic reduction system on the exhaust system. Selective combustion reduction (SCR) allows for the selective reduction of NOx to nitrogen through the action of a reducing agent, typically injected upstream of the SCR system. Examples include ammonia or a compound that generates ammonia through decomposition, such as urea, which can be advantageously used in aqueous solution. Ammonia is therefore a substance for which concentration monitoring can be useful, not only for controlling concentrations mandated by current or future standards, but also for diagnosing and / or monitoring the emissions control system. In this context, the ability to measure the concentration of pollutants is beneficial, particularly for monitoring and / or diagnosing post-combustion emissions control systems.For the purpose of controlling combustion, it may also be relevant to measure other combustion products such as carbon dioxide (CO2), oxygen (O2) and water (H2O).

[0004] In the field of air quality monitoring (for example, in confined or semi-confined spaces such as underground car parks, road or rail tunnels), ventilation and air extraction and / or supply systems are generally installed. However, these systems often prove insufficient to guarantee adequate air quality. For example, in metro-type rail networks, urban outside air is drawn in to renew the air supply. However, urban air can already present a significant level of pollution, particularly in terms of fine particulate matter but also in terms of gaseous pollutants (such as NOx, VOCs, ozone (O3), or sulfur compounds like sulfur oxides (SOx)), notably due to road traffic, district heating, or nearby industrial activities.As for confined spaces such as road tunnels or parking lots, where the air is laden with particles and polluting gases due to underground road traffic, this polluted air is extracted and released outside, thus polluting the outdoor atmosphere near these confined spaces. There is therefore a real need to estimate the concentration of pollutants entering and / or being extracted from these confined spaces in order to improve air quality within these spaces and / or in their vicinity.

[0005] Biogas is produced by the anaerobic decomposition of organic waste, such as sewage sludge, agricultural waste, and landfill waste. Biogas is primarily composed of methane (40 to 70%), CO2, and water vapor, but it also contains impurities such as sulfur compounds (H2S, SO2, etc.), siloxanes, halogens, and VOCs (Volatile Organic Compounds). Biogas is therefore not directly usable. To be able to utilize biogas, it must be treated (or purified), particularly to remove Carbon dioxide and hydrogen sulfide, as well as other impurities, are removed. This process yields biomethane, which can be injected into the natural gas distribution network or used as a biofuel. Furthermore, the biogas produced may be treated with an odorant molecule, most often a sulfur compound (mercaptans, for example, tetrahydrothiophene (THT) in France), which requires precise dosing and analytical verification. A particular application of purified biogas is in fuel cells, for which tolerance thresholds for impurities or contaminants are particularly stringent to avoid damaging the system. Developing methods for measuring each pollutant is therefore of major importance, both for controlling the biogas treatment process and for qualifying the purified biomethane for use. Previous technique

[0006] The following documents will be cited during the description:

[0007] Kawamura, Kensuke; Nishigaki, Tomohiro; Andriamananjara, Andry; Rakotonindrina, Hobhniarantsoa; Tsujimoto, Yasuhiro; Moritsuka, Naoki et al. (2021) Using a One-Dimensional Convolutional Neural Network on Visible and Near-Infrared Spectroscopy to Improve Soil Phosphorus Prediction in Madagascar. In: Remote Sensing, vol. 13, no. 8, p. 1519. DOI: 10.3390 / rsl3081519.

[0008] Malek, Salim; Melgani, Farid; Bazi, Yakoub (2018) One-dimensional convolutional neural networks for spectroscopic signal regression. In: Journal of Chemometrics, vol. 32, no. 5, e2977. DOI: 10.1002 / cem.2977.

[0009] The classical technique for quantifying chemical compounds present in a fluid, used in absorption spectrometry techniques, is based on a chemometric approach. More precisely, absorption spectrometry techniques allow the quantification of chemical compounds by measuring the intensity of the electromagnetic radiation they absorb at different wavelengths (absorption spectrum), under the assumption that absorption is proportional to the concentration of these compounds. The absorption spectrum (A(A)) is obtained by applying Beer-Lambert's law to a reference light intensity spectrum, or baseline (jQ(yQ)), and to the light spectrum (Jç(A)) obtained in the presence of the sample of interest, according to the following equation: [00101 AW =

[0011] In the field of pollution control and / or monitoring of engine exhaust gases or industrial fumes, patent application WO 2019 / 020326 Al is known in particular, which relates to a system and a method for measuring the concentration of different chemical species contained in exhaust gases. More specifically, by measuring the intensity as a function of wavelength of UV radiation that has passed through exhaust gases within a measurement zone, an absorbance is determined as a function of the wavelength of the exhaust gases contained within that zone. Then, the concentration of each compound present in the gases is determined from the gas absorbance thus calculated, as well as from predetermined absorbance characteristics and information regarding the temperature and pressure of each of the chemical species to be quantified. Various types of algorithms can be used for concentration calculations, such as least-squares fitting algorithms applied to the absorbance signals themselves, to the derivatives of the absorbance signals, or to absorbance signals in the frequency domain (typically derived from a Fourier transform).Similarly, a number of statistical methods can be used for this process, such as principal component analysis (PCA) or partial least squares (PLS) algorithms.

[0012] The chemometric approach, theoretically simple, suffers from three robustness problems: - Nonlinear response of the system: the response of the real system may deviate from the linear regime, either due to an excessively high concentration of the species (high optical density), or due to a nonlinear response of the sensor (spectrometer). While the first aspect can be calibrated and taken into account in chemometric theory, the second aspect requires fine calibration of the spectrometer and a correction for each physical system. - Interference between different species: While each species is characterized by a specific absorption spectrum, several species absorb over the same spectral range. Differentiating these spectral signatures and quantifying them is a major challenge in this area. The classical approach involves searching for and defining spectral ranges of interest for each species to reduce interference. This is often addressed empirically and must be adapted according to the type of gas mixture expected and the target species, but it is time-consuming and can significantly impact the quality of the results. - Adaptation to system evolution: the physical measurement system can change over time due to, for example, aging, fouling, or misalignment of optical and active components (spectrometer, lamp). Empirical mathematical solutions can be implemented, but the effectiveness of this solution is strictly linked to the type of change expected.

[0013] More recently, methods for detecting or classifying molecules in liquids and sometimes gases, based on 1D convolutional neural networks and applied to spectroscopy, have been developed. In particular, the paper (Malek et al., 2018) is known, which concerns a method for measuring concentrations in food liquids (wine, orange juice) from near-infrared (NIR) spectra to which 1D convolutional networks are applied. More precisely, the method described in this paper uses 1D convolutional neural networks having at most 3 layers, a variable number of features less than or equal to 5, and a free convolution size limited to 10% of the input data size. The optimal structure is determined by an exhaustive calculation of all structure combinations on the training data.This phase is therefore very long, and the optimal structure may present risks of overfitting. The final structure of the network is unknown a priori.

[0014] Patent application CN114707598 A is also known, which relates to a method for identifying and quantifying species in a mixed gas, from spectra and using a 1D convolutional neural network. However, this method is silent on the structure of the neural networks used. Furthermore, it would appear that it is necessary to add advanced parameters ("meaningful high-level feature parameters") for the method to function.

[0015] We are also aware of the document (Kawamura et al. 2021), which concerns a method for quantifying phosphorus in soils, and which is based on an "increasing" structure of a convolutional neural network, that is, on an increasing number of filters as one goes deeper into the network (more precisely 32 filters, then 64, 128, and finally 256), and on a data size halved between each layer. However, this method has the disadvantage of having a very large number of parameters (from 1 to several million), which makes it not very robust and prone to overfitting on small training and validation datasets.

[0016] The present invention aims to overcome these drawbacks. More specifically, the present invention relates to a method for determining the concentration of at least one chemical compound in a fluid, using at least one optical measurement system capable of measuring a light intensity spectrum and a particular one-dimensional convolutional neural network. More specifically, the one-dimensional convolutional neural network according to the invention is characterized in that the number of convolutional filters per convolutional layer decreases with the position of the convolutional layer in said convolutional neural network. Such a structure allows for faster implementation of a method for quantifying a chemical compound in a fluid based on one-dimensional convolutional neural networks, which, moreover, requires less This results in less training data, less storage memory, and greater network robustness, meaning it is less susceptible to overfitting. Furthermore, compared to chemometric approaches, the method according to the invention does not require a baseline spectrum to generate absorption spectra, since the inputs to the method according to the invention are directly light intensity spectra. Summary of the invention

[0017] The invention relates to a method for determining the concentration of at least one chemical compound in a fluid, by means of at least one optical measurement system capable of measuring a light intensity spectrum.

[0018] Said process comprises at least the following steps:

[0019] A) by means of a training base comprising a plurality of training light intensity spectra measured for a plurality of training fluids whose concentration in at least said chemical compound is predetermined, a model is constructed to determine the concentration of at least one chemical compound in a fluid by training a one-dimensional convolutional neural network on said training base, said one-dimensional convolutional neural network being such that the number of convolutional filters per convolutional layer of said one-dimensional convolutional neural network decreases with a position of said convolutional layer in said one-dimensional convolutional neural network;

[0020] B) using said optical measuring system, at least one light intensity spectrum of said fluid is measured;

[0021] C) said model constructed to determine the concentration of at least one chemical compound of a fluid in step A) is applied to said at least one light intensity spectrum measured in step B), and said concentration of at least said chemical compound of said fluid is determined.

[0022] According to one embodiment of the invention, said optical measurement system may include at least one light source emitting radiation and at least one sensor, preferably a spectrometer, for measuring the evolution of the light intensity as a function of the wavelength of radiation transmitted through said fluid, and from which, by means of said light source, radiation can be emitted through said fluid, and then, by means of said sensor, at least the evolution of the light intensity can be measured as a function of the wavelength of said radiation transmitted through said fluid.

[0023] According to one embodiment of the invention, said optical measurement system can be selected from means suitable for measuring a light intensity spectrum of said fluid by means of Ultra-Violet spectroscopy, visible spectroscopy, Raman spectroscopy, Infrared spectroscopy or near-Infrared spectroscopy.

[0024] According to one embodiment of the invention, said number of convolutional filters per convolutional layer of said one-dimensional convolutional neural network can decrease by a multiple of two from one convolutional layer to another.

[0025] According to one implementation of the invention, said one-dimensional convolutional neural network may be of the ResNet type.

[0026] According to one embodiment of the invention, said one-dimensional convolutional neural network may further comprise at least one intermediate maxpooling layer disposed between two convolutional layers of said one-dimensional convolutional neural network.

[0027] According to one embodiment of the invention, a data flattening step can be performed between a last convolutional layer of said one-dimensional convolutional neural network and a fully connected layer of said one-dimensional convolutional neural network.

[0028] According to one embodiment of the invention, in step B), a plurality of light intensity spectra of said fluid can be measured by means of said optical measuring system, and step C) can be applied to an average of said plurality of measured light intensity spectra.

[0029] According to one embodiment of the invention, said training base can be diversified by means of at least one additional light intensity spectrum Icorr(^) resulting from a correction applied to one of said training light intensity spectra Imes(A) according to a formula of the type: !<,<,„( A) = lmes(A) eA' <A)

[0031] where A) is an absorption spectrum measured for one or more distinct chemical compounds of said at least one chemical compound and present in predetermined concentrations, and X is the wavelength.

[0032] The invention further relates to a computer program product downloadable from a communication network and / or recorded on a computer-readable medium and / or executable by a processor, comprising program code instructions for the implementation of at least steps A) and / or C) of the process as described above, when said program is executed on a computer. List of figures [Fig 1]

[0033] Fig. 1 schematically presents an embodiment of the optical measurement system that can be implemented in the process according to the invention. [Fig 2]

[0034] Figure 2 shows an example of a light intensity spectrum measured for a fluid whose concentration of NO, SO2, NH3, O2 and H2O is to be determined. [Fig. 3A]

[0035] Figure 3A shows the comparison between the H2O and O2 concentration values ​​determined by the method according to the invention and by a reference method. [Figure 3B]

[0036] Figure 3B shows the comparison between the values ​​of the concentration of NH3, NO, and SO2 determined by the process according to the invention and by a reference method. Description of the implementation methods

[0037] The method according to the invention relates to a method for determining the concentration of at least one chemical compound of a fluid, by means of at least one optical measurement system capable of measuring a light intensity spectrum of the fluid.

[0038] The fluid according to the invention may be a gas such as engine exhaust, biogas, or industrial fumes (from, for example, an incineration plant, a biomass processing plant, etc.). The fluid according to the invention may also be a liquid, such as bathing water, drinking water, or industrial wastewater (from, for example, a dye works, a water treatment plant, oil production facilities, etc.).

[0039] By optical measuring system capable of measuring a light intensity spectrum, we mean an optical measuring system capable of producing, by measurement, a curve representing the evolution of light intensity as a function of wavelength. It is clear that the optical measuring system is at least capable of emitting radiation and measuring the intensity of radiation within the wavelength range characteristic of the chemical compound whose concentration is being sought. By wavelength range characteristic of the chemical compound, we mean a range of wavelengths within which the absorption spectrum of the chemical compound (obtained by applying Beer-Lambert's law to the light intensity spectra) has non-zero (or measurable) absorbance values.

[0040] According to one embodiment of the invention, the optical measurement system may include at least one source emitting radiation and a sensor for measuring the evolution of the light intensity as a function of the wavelength of the radiation. From such an optical measurement system, radiation can be emitted through the fluid of interest, for example within a measurement zone, by means of the light source, and then, by means of the sensor, at least the evolution of light intensity as a function of the wavelength of the radiation that has passed through the fluid in the measurement area.

[0041] According to one embodiment of the invention, the optical measurement system may be selected from means suitable for Ultraviolet (UV) spectroscopy, visible spectroscopy, Raman spectroscopy, Infrared (IR) spectroscopy, or near-Infrared spectroscopy. Thus, according to this embodiment of the invention, the optical measurement system may include a spectrometer, such as a UV spectrometer, a visible spectrometer, a Raman spectrometer, or an IR spectrometer. Spectrometric measurement means are advantageous because they are compact, easy to implement, and readily usable online. More generally, according to this embodiment of the invention, the optical measurement system may include: - a light source to emit radiation at least in a range of wavelengths characteristic of the chemical compound; - a spectrometer to measure light intensity as a function of wavelength in at least the characteristic wavelength range of the chemical compound.

[0042] Advantageously, in the case where the chemical compound corresponds to ammonia (NH3), sulfur dioxide (SO2), nitrogen oxide (NO), nitrogen dioxide (NO2), oxygen (O2), or water (H2O), the optical measurement system may include a light source for emitting at least one UV radiation, and a UV spectrometer. Indeed, the absorption spectrum of these chemical compounds exhibits non-zero absorbance values ​​in the UV.

[0043] Advantageously, in the case where the chemical compound corresponds to carbon monoxide (CO), nitrous oxide (N2O), or carbon dioxide (CO2), the optical measurement system may include a light source for emitting at least one radiation in the IR, and an IR spectrometer or an IR detector. Indeed, the absorption spectrum of these chemical compounds exhibits non-zero absorbance values ​​in the IR.

[0044] Fig. 1 schematically presents an embodiment of the optical measurement system SMO that can be implemented in the process according to the invention, comprising a light source SL for emitting radiation RE, in a measurement zone ZM comprising a fluid FL, and a sensor SP for measuring the intensity as a function of the wavelength of the radiation RT having passed through the fluid FL in the measurement zone ZM along an optical path of length d.

[0045] According to one embodiment in which the method is intended to determine the concentration of a plurality of chemical compounds in a fluid, the optical measurement system used may include a plurality of light sources and / or a plurality of sensors to measure the evolution of light intensity as a function of the wavelength of radiation, this set of sources and sensors being able to emit radiation and to measure the intensity of radiation in the wavelength range characteristic of each of the chemical compounds whose concentration is being sought.

[0046] According to one example of an implementation of the invention, the light source can be a halogen-deuterium source, such as the AvaLight-DH-S-BAL model from Avantes (Netherlands) and / or the spectrometer can be a high-sensitivity fiber spectrometer, such as the AvaSpec-HS2048XL-EVO model from Avantes (Netherlands).

[0047] According to one embodiment of the invention, the optical measurement system may include processing means for measuring the evolution of light intensity as a function of the wavelength of radiation, and may include a computer on which the Avasoft software from Avantes (Netherlands) is installed. Advantageously, the optical measurement system may further include means for transmitting (for example, by electrical cable, optical fiber, or wireless communication system) the measurements taken by the sensor of the optical measurement system to the processing means of the optical measurement system.

[0048] According to one embodiment of the invention, the method may include a preliminary step of calibrating the optical measurement system. Such a calibration step may include measuring the light intensity as a function of the wavelength transmitted through a reference fluid that is ideally transparent in the spectral range considered, such as nitrogen (N2) or dry air. According to one embodiment of this invention, the calibration step may include the following steps:

[0049] - the emission of radiation through the reference fluid within a zone of measure ;

[0050] - the measurement of a luminous intensity as a function of the wavelength of radiation having passed through the reference fluid.

[0051] The method according to the invention comprises at least the following steps:

[0052] 1) Construction of a model to determine the concentration of at least a chemical compound of a fluid from at least one light intensity spectrum

[0053] 2) Measurement of at least one light intensity spectrum of the fluid using the optical measurement system

[0054] 3) Application of the constructed model to determine the concentration of minus a chemical compound in a fluid from at least one measured light intensity spectrum

[0055] Step 1) can be performed only once, and beforehand. In other words, once the model for determining the concentration of at least one chemical compound in a fluid from at least one light intensity spectrum is constructed, this model can be applied one or more times to acquired light intensity spectra. For example, if it is desired to determine the concentration of a chemical compound in different samples of the fluid of interest, steps 2 and 3 can be repeated for these different samples, as the model may only need to be constructed once. For example, the evolution over time of the concentration of at least one chemical compound in a fluid, such as one circulating in an exhaust pipe, a chimney at an industrial site, etc., can thus be monitored.

[0056] At least steps 1) and / or 3) can be implemented by computer means, in particular a computer, a processor or a calculator.

[0057] The steps of the process according to the invention are detailed below.

[0058] 1) Construction of a model to determine the concentration of at least a chemical compound of a fluid from at least one light intensity spectrum

[0059] During this step, by means of a training set comprising a plurality of training light intensity spectra measured for a plurality of training fluids whose concentration in at least said chemical compound is predetermined, a model is built to determine the concentration of at least one chemical compound in a fluid by training a one-dimensional convolutional neural network (i.e. a 1D convolutional neural network) on said training set.

[0060] According to the invention, the one-dimensional convolutional neural network is such that the number of convolutional filters per convolutional layer decreases with the position of the convolutional layer in the one-dimensional convolutional neural network. Indeed, convolutional neural networks are made up of a plurality of convolutional layers, each layer containing convolutional filters.

[0061] In a classic manner in the field of convolutional neural networks in the broadest sense (i.e., including in fields other than determining the concentration of a chemical compound in a fluid), the number of filters in the convolutional layers of a convolutional neural network increases with the layer's position in the network. For example, the first convolutional layer might contain 4 convolutional filters, the second convolutional layer might contain 8 convolutional filters, the third convolutional layer might contain 16 convolutional filters, and so on, with one or two fully connected neural network layers at the network's end. This is justified by the fact that simple operations are performed at the beginning, and then we process an increasing number of "features" (properties determined by the network) according to the diversity of the input population.

[0062] Surprisingly, the Applicant observed that a decreasing number of convolutional filters per convolutional layer, with the position of the convolutional layer in the one-dimensional convolutional neural network, allows for faster implementation of a method for quantifying a chemical compound in a fluid. This method also requires less training data, less storage memory, and offers greater robustness to extrapolation, while still ensuring sufficient accuracy. This is due to the modular (pre-established stacking of layers) and parsimonious (far fewer parameters than conventionally used networks) structure of the convolutional neural network according to the invention.Hereafter, the structure of the one-dimensional convolutional neural network according to the invention is referred to as the "decreasing structure", and the structure classically used in the state of the art in the broad sense is referred to as the "increasing structure".

[0063] Indeed, in the case of spectra (here, light intensity spectra), the input population is fixed (a few molecules), and the task is to predict their quantity. Therefore, the number of convolutional layers required is not as large. By adopting a decreasing structure, the size of the intermediate data (between two layers of the network) decreases all the more rapidly, and the total number of parameters remains limited, particularly that of the final, fully connected layer.

[0064] According to one embodiment of the invention, the ResNet type network can be used as a one-dimensional convolutional neural network. This type of network is a robust evolution of convolutional neural networks, reducing vanishing gradient problems. The blocks of such networks have parallel paths at each layer that add the subsampled input information to the output of the neural layer. This type of network is described, for example, in the document (He et al., 2015).

[0065] Advantageously, the number of convolutional filters from one convolutional layer to another of the convolutional neural network decreases by a multiple of two from one convolutional layer to another.

[0066] According to one embodiment of the invention, the one-dimensional convolutional neural network may further comprise at least one intermediate layer, called a "maxpooling" layer, situated between two convolutional layers. Such a layer makes it possible to reduce the size of the data while preserving its main features. Generally, each "maxpooling" layer reduces the size of the intermediate data by half, and the number of features is also reduced with each layer, preferably also in a ratio of 2 (for example, 32 / 16 / 8 / 4 features respectively). This drastically reduces the number of parameters and the amount of computation in the neural network. This improves the network's efficiency and prevents overfitting.

[0067] Advantageously, a data flattening step can be performed between the last convolutional layer and the final layer of the convolutional neural network, called the fully-connected layer.

[0068] In step 1), the model of the process according to the invention is built by training the one-dimensional convolutional neural network as described above on a training set, the training set being formed by a plurality of training light intensity spectra measured for a plurality of training fluids whose concentration, in at least the chemical compound, is predetermined. The term "training set data" refers to all the pairs formed by a light intensity spectrum and a concentration relative to each of the training fluids in the training set.

[0069] Advantageously, the training set can be constructed from at least 100 training fluids, preferably between 100 and 300 training fluids. Indeed, such a training set proves sufficient to solve the problem posed.

[0070] According to one embodiment of the invention, the method according to the invention may include a preliminary step of constructing the learning base.

[0071] According to one embodiment of the invention, the training set can be constructed as follows: from a plurality of training fluid samples for which the concentration of at least the chemical compound of interest is predetermined, a light intensity spectrum is measured for each sample of the plurality of training fluids, preferably using the optical measurement system according to any one of its embodiments as described above. According to one embodiment of the invention, each sample of the plurality of training fluid samples may have a chemical composition identical to that of the fluid of interest, but with varying concentrations of the chemical compound of interest.The concentration of at least the chemical compound of interest can be predetermined by measurement, for example using reference analytical methods such as Tunable Diode Laser analyzers (TDLs) or Fourier Transform Infrared Spectroscopy (FTIR). Alternatively, the concentration of at least the chemical compound of interest can be predetermined during the preparation of the various training fluid samples. Advantageously, it is possible to construct... the training basis under pressure and / or temperature conditions close (for example, within 10%) to those of the implementation of the process according to the invention.

[0072] Advantageously, the training basis can be divided into 2, preferably 3, subsets:

[0073] - a first subset, comprising for example between 70% and 85%, of 80% of the training set data is used for training the network: this data will directly determine the network estimation error measure and how to adjust the parameter values ​​of the neural network; this is the training set.

[0074] - a second subset, comprising for example between 15% and 30%, of Preferably, 20% of the total training dataset is used to validate the learning process during its execution. This data is used to monitor the success of the learning operation but does not directly influence the adjustment of the network parameters; this is the validation dataset. If the results obtained are unsatisfactory, it may mean that the training dataset is not sufficiently representative and that it may be necessary to augment the training dataset with training data covering a wider range of concentrations of the chemical compound of interest.

[0075] - optionally but advantageously, a third subset, comprising a number of data points from the training set in a proportion equivalent to the validation set, intended to test the learning performance at the end of the learning process; this is the test set.

[0076] According to one embodiment of the invention, to constitute the subsets described above, a random draw can be made from the set of data in the training database.

[0077] According to one embodiment of the invention, the training dataset can be diversified to increase the robustness of the model determined by training. According to a first alternative, the training dataset can be diversified by means of light intensity spectra measurements taken at regular intervals and / or irregular intervals. For example, in the case of an online analysis of the composition of exhaust gas flowing through an exhaust pipe, light intensity spectra measurements can be taken over several non-consecutive days. This can take into account the evolution over time of the composition of the exhaust gas being analyzed and / or of the optical measurement system (for example, its fouling). Generally, a person skilled in the art knows how to determine the appropriate time intervals based on the fluid being analyzed and the expected variations in composition over time.According to a second alternative, which can however be combined with the . As described above, the first alternative involves diversifying the training dataset by means of light intensity spectra measurements taken using separate optical measurement systems. This allows for the consideration of inherent defects in each optical measurement system. A third alternative, which can also be combined with the first and / or second alternatives described above, involves diversifying the training dataset by means of additional light intensity spectra Icorr(A) resulting from a correction applied to the training light intensity spectra Inies(A) (i.e., the spectra forming part of the training dataset) according to a formula of the type: [00781 leorrU) = ImeSW eA'W

[0079] where (A) is an absorption spectrum measured for one or more chemical compounds different from the one or those of interest (for example, from the literature or previous studies), these compounds being present at known concentrations. In this way, the presence of unmeasured species whose absorption spectra are available from previous studies or in the literature can be mathematically simulated to increase the training base.

[0080] Thus, at the end of this step, we obtain a model to determine the concentration of at least one chemical compound of a fluid.

[0081] 2) Measurement of at least one light intensity spectrum of the fluid using the optical measurement system

[0082] In this step, at least one light intensity spectrum of the fluid of interest is measured using the optical measuring system according to the invention. Advantageously, a plurality of light intensity spectra of the fluid of interest can be measured using the optical measuring system according to the invention, and an average of the spectra thus measured can be obtained, to which step 3) described below can be applied.

[0083] For this step, any one of the embodiments of the optical system according to the invention described above can be implemented. As a reminder, the optical measurement system according to the invention is at least capable of emitting radiation and measuring the intensity of radiation in the characteristic wavelength range of at least one chemical compound whose concentration is to be determined.

[0084] According to an embodiment of the invention in which the concentration of a plurality of chemical compounds in the fluid of interest is to be determined, an optical measurement system comprising a plurality of sources and sensors capable of emitting radiation and measuring the intensity of radiation in the characteristic wavelength range of the set of chemical compounds whose concentration we are trying to determine.

[0085] 3) Application of the constructed model to determine the concentration of minus a chemical compound in a fluid from at least one measured light intensity spectrum

[0086] In this step, the model is applied to determine the concentration of at least one chemical compound in a fluid constructed in step 1) described above, based on at least one light intensity spectrum measured in step 2 described above. In other words, this step involves applying the one-dimensional convolutional neural network trained as described in step 1) to the spectra measured in step 2) above, optionally averaged. This determines the concentration of at least one chemical compound in the fluid under consideration.

[0087] According to an embodiment in which the process aims to determine the concentration of a plurality of chemical compounds, and if the model for determining the concentration of at least one chemical compound of a fluid has been constructed in step 1) so as to determine the concentration of the plurality of chemical compounds, this step can be carried out only once.

[0088] Alternatively, if the model for determining the concentration of at least one chemical compound in a fluid has been constructed in step 1) so as to determine the concentration of a single chemical compound, steps 1) to 3) as described above are repeated for each chemical compound in the plurality of chemical compounds whose concentration is to be determined.

[0089] Furthermore, the invention relates to a computer program product downloadable from a communication network and / or stored on a computer-readable medium (for example, an embedded computer) and / or executable by a processor. This program includes program code instructions for implementing the method as described above, in particular steps 1) and / or 3) described above, when the program is executed on a computer. Examples

[0090] The characteristics and advantages of the method according to the invention will become clearer upon reading the application example below.

[0091] For this application example, the fluid of interest is a mixture of gases with variable and random concentrations. The chemical compounds whose concentration we seek to determine are: NO, SO2, NH3, O2, and H2O. These gases have the characteristic of being absorbent in a UV range from 200 to 400 nm.

[0092] The present application example was implemented using a portable optical measurement system comprising a Xenon source and a compact spectrometer (weight < 50g) sensitive in the UV range 200 - 400 nm, and spectral resolution 0.3 nm. In other words, the optical measurement system used for this application example makes it possible to cover the characteristic wavelength range of each of the chemical compounds whose concentration we wish to determine.

[0093] A training dataset consisting of a set of 265 light intensity spectra and the concentrations of the chemical compounds of interest associated with each of these light intensity spectra was constructed. The light intensity spectra were measured using the optical measurement system used for this application example on 265 training fluids characterized by varying concentrations of the chemical compounds of interest. The concentrations were determined using a reference measuring device, in this case an FTIR (Fourier Transform Infrared Spectroscopy) analyzer.A first subset, comprising 80% of the total data in the training set, is intended for training the network (training dataset), and a second subset, comprising 20% ​​of the total data in the training set, is intended for validating the learning process (validation dataset).

[0094] The model for determining the concentration of the chemical compounds of interest was built by training a one-dimensional convolutional neural network according to the invention on the learning basis described above. More specifically, the one-dimensional convolutional neural network is a ResNet-type network, consisting of a stack of convolutional layers in which the number of convolutional filters per layer is halved from one convolutional layer to the next. Between each layer, a data size reduction of the "maxpooling" type is performed until the intermediate data size, before the final fully connected layer, is on the order of 10 to 30 times the number of output data points to be predicted (For example, to predict 5 concentrations, the order of magnitude of the number of intermediate data points before the last layer should be approximately 50 to 150).A data flattening step ("flatten step") is performed between the last ResNet layer and the final fully connected neural network layer.

[0095] The model for determining the concentration of the chemical compounds of interest constructed as described above was applied to an average of a plurality of light intensity spectra measured using the optical measurement system according to the invention.

[0096] Figure 2 shows an example of a light intensity spectrum, therefore representing the evolution of the light intensity I as a function of the wavelength L, measured for a fluid whose concentration in NO, SO2, NH3, O2 and H2O is to be determined.

[0097] Figure 3A shows the comparison between the concentration values ​​C of H₂O (denoted H₂O in the figure) and O₂ (denoted O₂ in the figure) determined by the process according to the invention (Histograms INV) and the concentration measured by a reference method (Histograms REF) for these same compounds. Figure 3B shows the comparison between the concentration values ​​C of NH₃ (denoted NH₃ in the figure), NO, and SO₂ (denoted SO₂ in the figure) determined by the process according to the invention (Histograms INV) and the concentration measured by a reference method (Histograms REF) for these same compounds. A perfect correspondence can be observed between the concentration value determined by the process according to the invention and the concentration measured by a reference method for the compounds NO, SO₂, NH₃, O₂, and H₂O.An absolute difference on the order of ppm and a relative difference of approximately 5% are observed between the concentration values ​​determined by the method according to the invention and the concentrations measured by a reference method for NO. These differences can be explained by the uncertainty in the measured data and the uncertainty of the model itself for the lowest absolute values. These results were obtained in less than ten minutes on an Intel Core i7 processor with 16 GB of memory.

[0098] Thus, the process according to the invention makes it possible to determine the concentration of chemical compounds in a fluid in a precise, rapid and efficient manner.

[0099] More specifically, the present invention makes it possible, compared to conventional methods based on chemometrics, to: - Simplify and automate the implementation of a gas or liquid concentration measurement system by spectroscopy by avoiding the use of chemometrics, and therefore reduce development times without sacrificing the accuracy of the response. - Improve the consideration of spectral interference between different species - Limit characterization tests, and potentially improve accuracy, for example through compensation for non-linearities. - Add the measurement of new species by numerically increasing the training set, potentially with a very limited number of calibration measurements.

[0100] Furthermore, the present invention makes it possible, compared to existing methods based on convolutional neural networks, to: - simplify the implementation of the convolutional neural network with a parameter-sparse structure (approximately a factor of 100 compared to the state of the art) limit the amount of training data, to limit the risk of overlearning, require less computation time and less memory capacity, which allows for embedded use of the algorithm.

Claims

Demands

1. A method for determining the concentration of at least one chemical compound in a fluid, by means of at least one optical measurement system (OMS) capable of measuring a light intensity spectrum of said fluid, said method being characterized in that it comprises at least the following steps: A) By means of a training set comprising a plurality of training light intensity spectra measured for a plurality of training fluids whose concentration in at least said chemical compound is predetermined, a model is constructed for determining the concentration of at least one chemical compound in a fluid by training a one-dimensional convolutional neural network on said training set,said one-dimensional convolutional neural network being such that the number of convolutional filters per convolutional layer of said one-dimensional convolutional neural network decreases with the position of said convolutional layer in said one-dimensional convolutional neural network, said training set being diversified by means of at least one additional light intensity spectrum Icorr(A) resulting from a correction applied to one of said measured training light intensity spectra Imes(A) according to a formula of the form: = eA'a where Ay(A) is an absorption spectrum measured for one or more distinct chemical compounds of said at least one chemical compound present in predetermined concentrations, and X is the wavelength. B) by means of said optical measurement system (OMS),(a) at least one light intensity spectrum of said fluid is measured; (b) said model constructed to determine the concentration of at least one chemical compound of a fluid in step A) is applied to said at least one light intensity spectrum measured in step B), and said concentration of at least said chemical compound of said fluid is determined.

2. A method according to claim 1, wherein said optical measuring system (OMS) comprises at least one light source (LS) emitting radiation (RE) and at least one sensor (SP), preferably a spectrometer (SP), to measure the evolution of the light intensity as a function of the wavelength of a transmitted radiation (FL) through said fluid (FL), and from which, by means of said light source (SMO), emitted radiation (RE) is emitted through said fluid (FL), then, by means of said sensor (SP), at least the evolution of the light intensity as a function of the wavelength of said radiation (RT) transmitted through said fluid (FL) is measured.

3. A method according to any one of the preceding claims, wherein said optical measurement system (OMS) is selected from means suitable for measuring a light intensity spectrum of said fluid (FL) by means of Ultra-Violet spectroscopy, visible spectroscopy, Raman spectroscopy, Infrared spectroscopy or near-Infrared spectroscopy.

4. A method according to any one of the preceding claims, wherein said number of convolutional filters per convolutional layer of said one-dimensional convolutional neural network decreases by a multiple of two from one convolutional layer to another.

5. A method according to any one of the preceding claims, wherein said one-dimensional convolutional neural network is of the ResNet type.

6. A method according to any one of the preceding claims, wherein said one-dimensional convolutional neural network further comprises at least one intermediate maxpooling layer disposed between two convolutional layers of said one-dimensional convolutional neural network.

7. A method according to any one of the preceding claims, wherein a data flattening step is performed between a final convolutional layer of said one-dimensional convolutional neural network and a fully connected layer of said one-dimensional convolutional neural network.

8. A method according to any one of the preceding claims, wherein, in step B), a plurality of light intensity spectra of said fluid are measured by means of said optical measuring system, and wherein step C) is applied to an average of said plurality of measured light intensity spectra.

9. Product: computer program downloadable from a communications network and / or stored on a computer-readable medium and / or executable by a processor, comprising instructions for program code for the implementation of at least steps A) and / or C) of the method according to one of the preceding claims, when said program is executed on a computer.

Images