AI-accelerated analysis method, evaluation unit, gas chromatograph, analysis system and computer program product
Patent Information
- Application Number
- DE502022002092
- Authority / Receiving Office
- DE · DE
- Patent Type
- Patents
- Current Assignee / Owner
- Priority Date
- 2021-08-26
- Filing Date
- 2022-07-06
- Publication Date
- 2024-11-28
- Estimated Expiration
- 2042-07-06
AI Technical Summary
Gas chromatographic methods used in the chemical industry for monitoring exhaust gases and product purity are too slow to determine the composition of material samples with sufficient precision, necessitating an accelerated and accurate analysis method.
A method utilizing a gas chromatograph in conjunction with artificial intelligence, where the AI determines the concentrations of multiple components in a sample by processing partial measurement values obtained during a single measurement cycle, allowing for increased accuracy and speed without additional hardware, using techniques such as neural networks or fuzzy logic, and incorporating historical data for improved estimation.
This approach enables rapid and precise determination of sample composition, allowing for timely detection of improper compositions and reducing inactive phases, while being adaptable to interference-sensitive applications and capable of identifying anomalies, thus enhancing the reliability of automated manufacturing processes.
Description
[0001] The invention relates to an analysis method for determining the composition of a substance sample, which is accelerated by artificial intelligence. The invention also relates to an evaluation unit for carrying out such a method and a gas chromatograph having such an evaluation unit. Likewise, the invention relates to an analysis system based on a correspondingly trained artificial intelligence. Furthermore, the invention relates to a computer program product embodied as a corresponding artificial intelligence.
[0002] Patent application CN 111443160 A discloses a device that can be used to monitor fermentation processes or exhaust gas compositions. The device comprises a gas sensor array and a gas chromatograph. Measurement values from the gas sensor array are linked to measurement values from the gas chromatograph via data analysis. This is done using a machine learning algorithm. Patent application US-A-2017074840 describes a device for determining the composition of a substance sample with multiple components using the concentration of a first component of the substance sample.
[0003] Gas chromatography methods are used to determine the composition of a substance sample in various manufacturing processes in the chemical industry, for monitoring exhaust gases, or for product purity. Gas chromatography measurements are too slow for many processes to monitor the composition of the corresponding substance samples with sufficient accuracy. The invention is based on the objective of providing a method for determining the composition of a substance sample more quickly and with high measurement accuracy.
[0004] The problem is solved by a method according to the invention for determining the composition of a substance sample. The substance sample is in the form of a mixture of at least three components, the respective concentrations of which are to be determined. The method is carried out using a gas chromatograph in which the substance sample is fractionated into the at least three components by a separation device in such a way that these reach a detector at different times. The substance sample can be in the form of a mixture of gases and / or liquids. The method comprises a first step in which the concentration of a first component of the substance sample is quantitatively determined. The first component is the component of the substance sample that reaches the detector first after passing through the separation device. The quantitative determination is to be understood as a measurement with the detector.The result of the quantitative recording of the concentration of the first component represents a first partial measurement value, which is passed on to an artificial intelligence for further use, i.e. output.
[0005] A second step follows, in which the artificial intelligence determines the concentration of each of the second and third components. The artificial intelligence determines the concentration of the second and third components based on the first partial measurement. To determine the concentrations of the second and third components, the first partial measurement serves as the input value for the artificial intelligence, which then calculates the concentrations of the second and third components. This calculation can be implemented as an experience-based estimate based on historical data on the concentrations of the first, second, and third components.
[0006] The method also includes a third step in which the concentration of the second component of the sample is quantitatively determined. The second component is the component of the sample that reaches the detector after the first component and is quantitatively detected by it. The result of this quantitative detection is to be understood as a second partial measurement value, which is forwarded to the artificial intelligence for further use, i.e., output. The artificial intelligence then has access to the first and second partial measurement values, which are then used together.
[0007] In a fourth step of the method, the concentration of the third component of the substance sample is determined using artificial intelligence. With the first and second partial measured values, the artificial intelligence has a broader basis of measurement data at its disposal, so that fewer concentrations need to be determined by the artificial intelligence. The fourth step is carried out using the first and second partial measured values. As a result, the concentration of the third component can be determined with increased accuracy. According to the invention, the first to fourth steps of the method take place within a first measurement cycle of the gas chromatograph in which the substance sample is examined. The first measurement cycle begins with the introduction of the substance sample into the separation device and ends when a new substance sample is injected into the separation device.
[0008] The method according to the invention is therefore suitable for specifying a concentration for the second and third component several times during the first measuring cycle, which concentration is determined by the artificial intelligence. During the measuring cycle, the concentrations determined by the artificial intelligence become increasingly more accurate. The invention is based, among other things, on the surprising discovery that, using artificial intelligence, missing partial measured values during the first measuring cycle can be temporarily replaced with sufficient accuracy by calculated concentrations. Each time a partial measured value is recorded or during each calculation run by the artificial intelligence, a composition of the substance sample can thus be specified. This leads, among other things, to the surprising effect that the more components the substance sample has to be detected, the more compositions are determined per measuring cycle.Overall, the complexity of the analysis task to be solved accelerates the applied artificial intelligence. This allows, for example, the accelerated detection of an improper composition of the substance sample. Likewise, the claimed method can be used to implement automated manufacturing processes in which the composition of a process medium is critical and at the same time sensitive to interference. Furthermore, the accelerated determination of the composition of the substance sample is achieved without additional hardware expenditure and can also be easily implemented in existing gas chromatographs during a retrofit, for example, by replacing an evaluation unit.
[0009] In one embodiment of the claimed method, the artificial intelligence is embodied as a so-called neural network, a statistics module, or a combination thereof. A neural network is understood, in particular, to be an algorithm capable of determining at least one output value from incomplete and / or ambiguous input values, which output value is to be regarded as the most probable correct result based on the input values. Neural networks can be largely independently configured for their intended task using a so-called training data set. Likewise, neural networks can be further trained independently during operation using collected data. For example, the neural network in the claimed method can already be configured for the method using a training data set and can continuously develop further during this time.This makes it easy to increase the informative value of artificial intelligence for specific applications. The same neural network can be used for each step in which concentrations of components in the sample are to be determined, with nodes of the neural network being automatically adjusted with each quantitatively determined partial measurement value. This achieves a greater degree of homogeneity in the behavior of the artificial intelligence throughout an entire measurement cycle. Alternatively, a separate neural network can be used as artificial intelligence for each of the outlined steps. This achieves increased step-specific adaptation of the artificial intelligence for each corresponding step. Alternatively or additionally, the artificial intelligence in the claimed method can also be implemented as fuzzy logic.The statistics module is, for example, a computing unit for carrying out statistical evaluations, for example using Bayesian statistics, least squares regression or similar numerical methods.
[0010] Furthermore, the artificial intelligence can be trained using a training data set. The training data set can contain data from historical determinations of the compositions of substance samples. For example, the data from the historical determinations can include values for at least one of the partial measured values, in particular the first, second, and / or third partial measured value. In particular, the data from the historical determinations of composition can be determined on gas chromatographs of the same or similar type. A gas chromatograph of the same type has the same technical properties as the gas chromatograph on which the claimed method is carried out. A gas chromatograph of a similar type is technically similar to the gas chromatograph on which the claimed method is carried out insofar as the respective partial measured values are transferable.
[0011] Furthermore, the claimed method can also comprise a fifth step in which the concentration of the third component of the substance sample is quantitatively determined. For this purpose, the third component is detected by the detector in the same way as the first and second components. Furthermore, in the fifth step, the quantitatively determined concentration of the third component is output as a third partial measurement value. In particular, the third partial measurement value can be output to the artificial intelligence. The concentration of the third component determined by the artificial intelligence in the fourth step can thus be verified using the third partial measurement value. For example, a corresponding marking on a graphical user interface can emphasize to a user of the gas chromatograph that a composition based entirely on partial measurement values is now displayed.Furthermore, after the fifth step, the end of the first measurement cycle can be detected and a subsequent second measurement cycle can be initiated. This minimizes inactive phases in the claimed method.
[0012] Furthermore, in the claimed method, partial measurement values and corresponding concentrations of corresponding components determined by the artificial intelligence can be made available to the artificial intelligence for machine learning. The partial measurement values and the corresponding concentrations can be stored in a data set related to one another, which can serve as a training data set for the artificial intelligence or supplement such a data set. For each pass of the method, i.e., each measurement cycle, a relatively compact data set can be provided, which is particularly suitable for machine learning and thus for the independent further development of the artificial intelligence. Consequently, the claimed method is designed for a particularly targeted further development.
[0013] Furthermore, the claimed method can carry out a sixth step in which the concentration of a residue, i.e., a residual component, of the substance sample is determined. The residue can comprise one or more components and is not among the components of the substance sample detected by the detector or the artificial intelligence. The concentration of the residue results from the difference between the sum of the concentrations determined by the detector or the artificial intelligence and the total concentration of the substance sample. Furthermore, the concentration of the residue, i.e., the residual component, thus determined can be compared with an adjustable threshold value. If the concentration of the residue exceeds the adjustable threshold value, in particular if it exceeds the absolute value, a warning can be issued. The threshold value can be specified, for example, by the user, a table, or an algorithm.The claimed method thus exhibits a tolerance for the unspecified residue that can be adapted to the specific application. The claimed method is therefore also applicable in applications with disruptive operating conditions in which unexpected contamination of the material samples can occur. In particular, the third component of the material sample can be the residual component.
[0014] In a further embodiment of the claimed method, the concentration of the second and / or third component of the substance sample determined by artificial intelligence can be output with an error margin and / or confidence interval. Concentrations determined only by artificial intelligence but not yet supported by partial measured values are inherently subject to uncertainty. By outputting the associated error margin or confidence interval, a user and / or a control system can decide whether further decisions should be made based on the composition thus determined or whether further partial measured values should be awaited. The claimed method can therefore also be used with sufficient reliability in interference-sensitive applications. The time until a determined composition of the substance sample is sufficiently reliable or precise is shortened overall by the claimed method.
[0015] Likewise, in the claimed method, a difference can be determined between a concentration determined by the artificial intelligence and the corresponding partial measured value. If the difference exceeds an adjustable anomaly threshold, a warning can be issued. The warning can be issued to a user or to a control program. This makes it possible to detect, for example, an abrupt increase or decrease in the concentration of a component. This can be caused by a closed feed valve, a stoppage of an upstream process from which the substance sample is taken, or by a defect in the detector and / or the evaluation unit. The claimed method can also be designed to detect the cause of such an anomaly using the artificial intelligence or a separate method coupled with the artificial intelligence.This can be done, for example, using an experience database. Furthermore, the user can minimize the issuance of unnecessary warnings by appropriately setting the anomaly threshold. The claimed method is therefore also suitable for ensuring the reliable operation of an automation system in which the method is used and for identifying defects early on.
[0016] Furthermore, in the claimed method, the second and / or fourth step can be carried out taking into account a process parameter. The process parameter can be, for example, a temperature, a pressure, a pH value, an electrical quantity such as impedance, conductivity, and / or an optical quantity that is present in the process medium from which the substance sample is taken or in a process by which the substance sample is produced. For example, in the case of an exhaust gas that serves as a substance sample, the associated combustion temperature can be taken into account by the artificial intelligence as a process parameter, i.e., used as an input variable when determining component concentrations. For a variety of fuels, the combustion temperature determines the concentrations of carbon dioxide or nitrogen oxides produced. The claimed method is therefore capable of dynamically adapting to a given operating situation.Accordingly, the claimed method is also suitable for regulating or controlling sensitive processes where rapid reactions are required.
[0017] Furthermore, in the claimed method, an intermediate measurement value for at least one component of the substance sample can be provided by an auxiliary analysis device during the first measurement cycle. An auxiliary analysis device can be, for example, a continuous gas analyzer, a Raman spectrometer, a laser spectrometer, a non-dispersive infrared sensor (NDIR sensor for short), or a mass spectrometer. Such auxiliary analysis devices offer reduced measurement accuracy compared to a gas chromatograph. The auxiliary analysis device can provide a measured value, which is to be understood as an auxiliary measurement value, between the determination of a concentration of a component via the artificial intelligence and the quantitative detection, for example, by means of the detector. This allows the plausibility of a concentration of a component determined by the artificial intelligence to be verified during the method.Furthermore, auxiliary measurement values can also be made available for machine learning by the artificial intelligence. In particular, the concentrations of all components of the substance sample that are determined by the artificial intelligence or measured via the detector in the claimed method can be determined by the auxiliary analysis device. This further supports the claimed method and allows any defects in the gas chromatograph to be identified.
[0018] In one embodiment of the claimed method, the auxiliary analysis device can have a measurement cycle duration that is shorter than the first measurement cycle. Using the auxiliary analysis device, multiple measurements can thus be performed during one measurement cycle, thus providing two or more auxiliary measurement values for each component. The artificial intelligence not only determines the concentrations of previously unmeasured components, but also determines them when at least one new auxiliary measurement value is available. This allows the artificial intelligence to perform more calculation runs per measurement cycle, making the determination of the unmeasured concentrations more precise.
[0019] The problem outlined above is also solved by an evaluation unit according to the invention. The evaluation unit is designed to interact with an analysis device that serves to determine the composition of a substance sample. For this purpose, the evaluation unit, in the assembled state, is connected to at least one detector of the analysis device and is suitable for receiving measurement signals from the detector. According to the invention, the evaluation unit is designed to carry out at least one embodiment of the method outlined above. For this purpose, the evaluation unit can, for example, be equipped with a computer program product that implements the corresponding method. The evaluation unit can also be designed as a combination of several hardware platforms that are communicatively connected to one another and interact to carry out the corresponding method, for example in a computer cloud.The evaluation unit according to the invention can also be retrofitted as part of an upgrade to an existing system. This allows the method according to the invention to be easily transferred to existing systems.
[0020] Likewise, the underlying object is achieved by a gas chromatograph according to the invention, which is designed to determine the composition of a substance sample. For this purpose, the gas chromatograph comprises a separation device, for example a so-called separation column, and a detector. The detector can be designed, for example, as a flame ionization detector, a thermal conductivity detector, a photoionization detector, a mass spectrometer, or an ion mobility spectrometer. The detector is connected to an evaluation unit, which is designed to receive and evaluate measurement signals from the detector. According to the invention, the evaluation unit is designed according to one of the embodiments presented above. By means of the evaluation unit according to the invention, the method described above can be implemented on the gas chromatograph according to the invention.This increases the measurement speed, i.e., the frequency with which the compositions of the substance sample to be measured are output. At the same time, the increased measurement accuracy of a gas chromatograph is maintained. The gas chromatograph according to the invention requires no additional hardware components, so the technical advantages outlined can be achieved economically.
[0021] The problem described above is also solved by an analysis system according to the invention, which is designed to determine the composition of a substance sample. The analysis system comprises an auxiliary analysis device designed as a Raman spectrometer, a continuous gas analyzer, a laser spectrometer, a non-dispersive infrared sensor (NDIR sensor for short), or a mass spectrometer. The auxiliary analysis device is connected to an evaluation unit suitable for receiving and evaluating measurement signals from the auxiliary analysis device. An artificial intelligence is executably stored in the evaluation unit, by means of which the composition of the substance sample is to be determined. According to the invention, the artificial intelligence is trained by a method according to one of the embodiments described above.This means that the recognition behavior of the artificial intelligence is predetermined by a training data set obtained by the method described above. In particular, the artificial intelligence is suitable for using measurement signals received from the auxiliary analysis device as partial measurement values within the meaning of the method according to the invention. The invention is based, among other things, on the surprising finding that such artificial intelligence can be used to realize a virtual gas chromatograph that has the measurement accuracy of a real gas chromatograph. In such cases, the gas chromatograph according to the claimed method is therefore a virtual gas chromatograph. However, the analysis system according to the invention only requires one or more auxiliary analysis devices, which as such have a lower measurement accuracy than a gas chromatograph.The analysis system can thus be designed without a gas chromatograph, i.e., a real gas chromatograph. The technical advantages of a gas chromatograph can be achieved cost-effectively with the analysis system according to the invention, which can be considered a virtual gas chromatograph, even with simpler hardware.
[0022] Furthermore, the object described above is achieved by a computer program product according to the invention, which is designed to determine the composition of a substance sample. The computer program product comprises artificial intelligence, which can be designed, for example, as a neural network, and a training data set. The training data set is suitable for specifying the behavior of the artificial intelligence when determining the composition of the substance sample through so-called machine learning. The training data set can be expanded during operation of the computer program product with additional values, in particular partial measured values and corresponding determined concentrations of components of the substance sample. According to the invention, the computer program product is designed to implement at least one embodiment of the method set out above.In particular, the computer program product is suitable for serving as artificial intelligence within the meaning of the method. The computer program product can be monolithic, i.e., executable on a hardware platform. Alternatively, the computer program product can comprise a plurality of subprograms that interact communicatively to implement the desired functionality, for example, in a computer cloud. The computer program product can be implemented entirely or partially as software or hard-wired, for example, in a chip, microcontroller, or FPGA. The computer program product according to the invention can be easily provided for an existing evaluation unit, so that the method according to the invention can also be implemented on existing evaluation units.
[0023] The invention is explained in more detail below using individual embodiments in the figures. The figures are to be read as complementary to one another, in that identical reference numerals in different figures have the same technical meaning. The features of the different embodiments can be combined with each other and with the features outlined above. They show in detail: FIG 1 shows the structure of an embodiment of the claimed gas chromatograph; FIG 2 shows a sequence of a first embodiment of the claimed method; FIG 3 shows a sequence of a second embodiment of the claimed method; FIG 4 shows an embodiment of the claimed analysis system.
[0024] The structure of an embodiment of the claimed gas chromatograph 10 is shown in FIG 1 shown schematically. The gas chromatograph 10 comprises an injection device 12 into which a carrier gas 11 can be introduced. The carrier gas 11 is in contact with a substance sample 15 in order to introduce this into a separation device 14, which is designed as a separation column. The separation device 14 divides the substance sample 15 into individual components 31, 32, 33, 34. The components 31, 32, 33, 34 reach a detector 16 one after the other, which is designed to quantitatively detect the individual components 31, 32, 33, 34. For this purpose, measurement signals 25 are output from the detector 16 to an evaluation unit 18, which has an artificial intelligence 40. The artificial intelligence 40 serves to evaluate the measurement signals 25 from the detector 16 in order to determine the composition 17 of the substance sample 15. For this purpose, the artificial intelligence 40 is designed to implement at least one embodiment of the claimed method 100.
[0025] The gas chromatograph 10 is equipped with an auxiliary analysis device 20, which can be configured as a continuous gas analyzer or Raman spectrometer. The same substance sample 15 as the injection device 12 is fed to the auxiliary analysis device 20. The auxiliary analysis device 20 is suitable for determining the composition 17 of the substance sample 15, but with less precision than the detector 16. The auxiliary analysis device 20, on the other hand, is suitable for performing such determinations of the composition 17 of the substance sample 15 more quickly than the detector 16. Corresponding measurement signals 25 can be output from the auxiliary analysis device 20 to the evaluation unit 18 to support the artificial intelligence 40. Additionally, the evaluation unit 40 is coupled to a sensor 26, by which a process parameter 45 can be detected. The process parameter 45 influences the composition 17 of the substance sample 15.Measurement signals 25, which represent the process parameter 45, are output by the sensor 26 to the evaluation unit 18 in order to support the artificial intelligence 40.
[0026] The artificial intelligence 40 can be coupled to a training data set 22, which enables the artificial intelligence 40 to determine the composition 17 of the substance sample 15 from measurement signals 25 from the detector 16, the auxiliary analysis device 20, and the sensor 26. For this purpose, the training data set 22 contains data on historical determinations of compositions 17 of substance samples 15 on gas chromatographs 10 of the same or similar type. The training data set 22 can also be supplemented with measurement signals 25 and results of self-performed determinations of compositions 17 of substance samples 15. The determined compositions 17 are output on an output unit 19, which is designed as a screen. The artificial intelligence 40 and the training data set 22 together form a computer program product 60, which enables the implementation of the claimed method 100.
[0027] In FIG 2 A sequence of a first embodiment of the claimed method 100 is shown in a diagram 50, which has a horizontal time axis 51. In the lower part, the diagram 50 has a vertical magnitude axis 52, which represents a curve of concentrations 35 and a curve of a process parameter 45. The diagram 50 shows a first measuring cycle 23, in which the composition 17 of a substance sample 15 is determined, which has a first component 31, a second component 32, a third component 33 and a residual component 34. The beginning of the first measuring cycle 23 is by injecting the substance sample 15 into a separation device 14, as in FIG 1 shown, defined. After a retention time 21, which is characteristic of the first component 31 in combination with the separation device 14, it reaches the detector 16 and is quantitatively detected in a first step 110. The result of this detection represents a first partial measured value 41, which represents the concentration 35 of the first component 31 at the beginning of the first measurement cycle 23. In the first step 110, the first partial measured value 41 is also output to an artificial intelligence 40, which is designed as a neural network 47 and uses the first partial measured value 41 as input.
[0028] This is followed by a second step 120 in which the artificial intelligence 40 determines the composition 17 of the material sample 15 based on the first partial measurement value 41. As in FIG 2 symbolized by the empty circles, the artificial intelligence 40 has no partial measured values for the second component 32, the third component 33 or the residual component 34 at the time of the first step 110. The concentrations 35 of the second component 32, the third component 33 and the residual component 34 are determined in the second step 120 by means of the artificial intelligence 40, which is trained by a training data set 22. During the determination in the second step 120, the process parameter 45, which as in FIG 1 shown, is determined by a sensor 26. After completion of the second step 120, a result for the composition 17 of the substance sample 15 is available, which, with the first partial measured value 41, is based on a real measurement with the detector 16 and three concentrations 35 resulting from the artificial intelligence 40. This composition 17 determined in the second step 120 can be output via an output unit 19.
[0029] After a further period of time, which is characterized by a retention time 21 of the second component 32 in the separation device 14, the second component 32 is quantitatively detected by the detector 16. The quantitative detection of the second component 32 takes place in a third step 130. The quantitative detection in the third step 130 produces a second partial measured value 42 which represents the concentration 35 of the second component 32 at the beginning of the first measuring cycle 23. The second partial measured value 42 is output to the artificial intelligence 40, which uses it as further input. This is followed by a fourth step 140, in which the artificial intelligence 40 has the first and second partial measured values 41, 42 available, but no information about the concentrations 35 of the third component 33 or the residual component 34, as symbolized by the empty circles.In the fourth step 14, the concentration 35 of the third component 33 and the residual component 34 is determined via the artificial intelligence 40 based on the first and second partial measured values 41, 42. In doing so, the artificial intelligence 40 relies on its training data set 22 and the process parameter 45. In the fourth step 140, a composition 17 of the substance sample 15 at the beginning of the first measurement cycle 23 is determined, which composition is based on the first and second partial measured values 41, 42 and is supplemented by concentrations 35 for the third component 33 and the residual component 34, which are determined by the artificial intelligence 35. Because two partial measured values 41, 42 are already available in the fourth step 140, the concentrations 35 of the third component 33 and the residual component 34 can be determined with increased accuracy, particularly compared to the second step 120.The composition 17 determined in the fourth step 140 can also be output via an output unit 19 of the gas chromatograph 10.
[0030] Analogous to the first and third steps 110, 130, after the fourth step 140, a fifth step 150 is also carried out, in which a concentration 35 of the third component 33 is determined by means of the detector 16. The result of the fifth step 150 is forwarded as a third partial measurement value 43 to the artificial intelligence 40, which uses this as input. Taking into account the first, second, and third partial measurement values 41, 42, 43, the artificial intelligence 40, including the process parameter 45, also determines a concentration 35 for the residual component 44 in the fifth step 150. This results in a composition 17 of the substance sample 15, which comprises three partial measurement values 41, 42, 43 and a concentration 35 for the residual component 34 determined by the artificial intelligence 40. The composition 17 thus obtained can also be output via the output unit 19 of the gas chromatograph 10.
[0031] When determining the concentration of the residual component 34, an adjustable threshold value 44 is taken into account during at least one of steps 110, 120, 130, 140, 150. If the concentration 35 of the residual component 34 exceeds the adjustable threshold value 44, a warning is issued, for example, to a user via the output unit 19. The fifth step 150 is followed by an inactive phase 27, in which no further measurements are taken by the detector 16 or determinations are made by the artificial intelligence 40. The onset of the inactive phase 27 can be recognized by the fact that a result is available for the composition 17 of the substance sample 15 that can no longer be further specified by additional measurements with the detector 16. During the inactive phase 27, for example, the separation device 14 can be rinsed. Alternatively, the inactive phase 27 can be shortened by injecting a new substance sample 15.By injecting a new substance sample 15, a second measuring cycle 24 is initiated, in which the described steps 110, 120, 130, 140, 150 from the first measuring cycle 23 can be carried out again. Compared to known solutions, three results are output for the composition 17 of the substance sample 15 in the first measuring cycle 23 instead of just one result. This allows, for example, an accelerated reaction to an improper operating state in an automation system. By including the process parameter 45, the artificial intelligence 40 can make a precise prediction for the first, second and third components 31, 32, 33 as to which concentrations 35 can be expected at the beginning of the second measuring cycle 24. This represents an automatic further development of the training data set 22. The data set shown in . FIG 2 The method 100 shown is implemented by means of a computer program product 60 which comprises the artificial intelligence 40 and the training data set 22.
[0032] A second embodiment of the claimed method 100 is FIG 3 shown schematically. The sequence of the second embodiment of the claimed method 100 is shown in a diagram 50, which has a horizontal time axis 51. In the lower part, the diagram 50 has a vertical magnitude axis 52, which represents a curve of concentrations 35. The diagram 50 shows a first measuring cycle 23, in which the composition 17 of a substance sample 15 is determined, which has a first component 31, a second component 32, a third component 33 and a residual component 34. The beginning of the first measuring cycle 23 is by injecting the substance sample 15 into a separation device 14, as in FIG 1 shown, defined. After a retention time 21, which is characteristic of the first component 31 in combination with the separation device 14, it reaches the detector 16 and is quantitatively detected in a first step 110. The result of this detection represents a first partial measured value 41, which represents the concentration 35 of the first component 31 at the beginning of the first measurement cycle 23. In the first step 110, the first partial measured value 41 is also output to an artificial intelligence 40, which is designed as a neural network 47 and uses the first partial measured value 41 as input.
[0033] This is followed by a second step 120 in which the artificial intelligence 40 determines the composition 17 of the material sample 15 based on the first partial measurement value 41. As in FIG 2 symbolized by the empty circles, at the time of the first step 110, the artificial intelligence 40 has no partial measured values for the second component 32, the third component 33, or the residual component 34. The concentrations 35 of the second component 32, the third component 33, and the residual component 34 are determined in the second step 120 using the artificial intelligence 40, which is trained using a training data set 22. After completion of the second step 120, a result is available for the composition 17 of the substance sample 15, which, with the first partial measured value 41, is based on a real measurement with the detector 16, and three concentrations 35 that result from the artificial intelligence 40. This composition 17 determined in the second step 120 can be output via an output unit 19.
[0034] Between the second step 120 and a third step 130 of the method 100, the composition 17 of the substance sample 15 is determined by means of an auxiliary analysis device 20, as in FIG 1 shown. The auxiliary analysis device 20 generates an auxiliary measured value 37 for each of the second component 32, the third component 33, and the residual component 34. The auxiliary measured values 37 represent the concentrations 35 of the corresponding components 32, 33, 34, but have a reduced measurement accuracy compared to partial measured values 41, 42, 43. Based on the first partial measured value 41 and the auxiliary measured values 37 for the second component 32, the third component 33, and the residual component 34, the artificial intelligence 40 determines a composition 17 of the substance sample 15. This result, analogous to the result from the second step 120, can also be output via the output unit 19.
[0035] After a further period of time, which is characterized by a retention time 21 of the second component 32 in the separation device 14, the second component 32 is quantitatively detected by the detector 16. The quantitative detection of the second component 32 takes place in a third step 130. The quantitative detection in the third step 130 produces a second partial measured value 42 which represents the concentration 35 of the second component 32 at the beginning of the first measuring cycle 23. The second partial measured value 42 is output to the artificial intelligence 40, which uses it as further input. This is followed by a fourth step 140, in which the artificial intelligence 40 has the first and second partial measured values 41, 42 available, but no information about the concentrations 35 of the third component 33 or the residual component 34, as symbolized by the empty circles.In the fourth step 14, the concentration 35 of the third component 33 and the residual component 34 is determined via the artificial intelligence 40 based on the first and second partial measured values 41, 42. In this process, the artificial intelligence 40 relies on its training data set 22. In the fourth step 140, a composition 17 of the substance sample 15 at the beginning of the first measurement cycle 23 is determined, which composition is based on the first and second partial measured values 41, 42 and is supplemented by concentrations 35 for the third component 33 and the residual component 34, which are determined by the artificial intelligence 35. Because two partial measured values 41, 42 are already available in the fourth step 140, the concentrations 35 of the third component 33 and the residual component 34 can be determined with increased accuracy, particularly compared to the second step 120.The composition 17 determined in the fourth step 140 can also be output via an output unit 19 of the gas chromatograph 10. The results for the composition 17, which arise in the second step 120, in the fourth step 140, and in between based on the auxiliary measured values 37, are comparable to one another and can thus be mutually checked for plausibility. Such a plausibility check can be carried out, for example, by forming differences between the determined concentrations 35 for corresponding components 32, 33, 34 of the substance sample 15 and evaluating them in terms of their magnitude. From this, a characteristic value (not shown in detail) can be determined, which is at least comparable in terms of magnitude to an anomaly threshold value.
[0036] In an analogous manner, between the third step 130 and the fifth step 150, a composition 17 of the substance sample 15 is again determined using the auxiliary analysis device 20 and made available to the artificial intelligence 40 in the form of auxiliary measured values 37. Between the third and fifth steps 130, 150, only auxiliary measured values 37 for the concentrations 35 of the third component 33 and the residual component 34 are determined or forwarded to the artificial intelligence 40. Based on the first and second partial measured values 41, 42 and the auxiliary measured values 3 for the third component 33 and the residual component 34, concentrations 35 are determined for these by the artificial intelligence 40. These concentrations 35, together with the first and second partial measured values 41, 42, form an indication of a composition 17 of the substance sample 15. The composition 17 thus determined can be output via the output unit 19.
[0037] Analogous to the first and third steps 110, 130, after the fourth step 140, a fifth step 150 is also carried out, in which a concentration 35 of the third component 33 is determined by means of the detector 16. The result of the fifth step 150 is forwarded as a third partial measurement value 43 to the artificial intelligence 40, which uses this as input. Taking into account the first, second, and third partial measurement values 41, 42, 43, the artificial intelligence 40, including the process parameter 45, also determines a concentration 35 for the residual component 44 in the fifth step 150. This results in a composition 17 of the substance sample 15, which comprises three partial measurement values 41, 42, 43 and a concentration 35 for the residual component 34 determined by the artificial intelligence 40. The composition 17 thus obtained can also be output via the output unit 19 of the gas chromatograph 10.
[0038] The auxiliary analysis device 20 has, as in FIG 3 shown, suitable for a measurement cycle duration 39 which is shorter than the first measurement cycle 23. The auxiliary analysis device 20 is suitable for outputting auxiliary measurement values 37 to the artificial intelligence 40 several times per measurement cycle 23, 24. The auxiliary measurement values 37 are determined between the acquisition of two partial measurement values 41, 42, 43. As a result, in addition to the compositions 17 determined in the second, fourth and fifth steps 120, 140, 150, additional compositions 17 are determined. After FIG 3 Thus, five compositions 17 are determined per measurement cycle 23, 24. The auxiliary measured values 37 and the concentrations 35 determined from them can be stored and added to the training data set 22 of the artificial intelligence 40. Such enrichment of the training data set 22 leads to accelerated machine learning of the artificial intelligence.
[0039] The fifth step 150 is followed by an inactive phase 27, in which no further measurements are taken. In the inactive phase 27, the separating device 14 can be FIG 1 shown, rinsed. Alternatively, the onset of the inactive phase 27 can be detected by the artificial intelligence 40 and shortened accordingly. The inactive phase 27, and thus also the first measurement cycle 23, ends with the start of the second measurement cycle 24. The second measurement cycle 24 begins when a new substance sample 15 is injected. The steps 110, 120, 130, 140, 150 shown with regard to the first measurement cycle 23 and the use of the auxiliary analysis device 20 can be transferred analogously to the second measurement cycle 24.
[0040] FIG 4 shows an embodiment of the claimed analysis system 70, which has an auxiliary analysis device 20 designed as a Raman spectrometer or as a continuous gas analyzer. The auxiliary analysis device 20 is connected to an evaluation unit 18. Furthermore, the auxiliary analysis device 20 is designed to examine a substance sample 15, which is designed as a substance flow, for its composition 17. Corresponding measurement data are transmitted as auxiliary measurement values 37 to the evaluation unit 18. The evaluation unit 18 is equipped with an artificial intelligence 40 for determining the composition 17. The artificial intelligence 40 is trained using a training data set 22, i.e., for determining the composition 17 of the substance sample 15. The training data set 22 is at least partially generated from a method 100 according to one of the embodiments described above.Thus, the artificial intelligence 40 is at least partially trained by such a method 100. Consequently, the artificial intelligence 40 is able to determine the composition 17 of the substance sample 15 with increased precision without the need for a detector 16, such as in a gas chromatograph 10 according to FGI 1. The analysis system 70 in . FIG 4 is to be understood as a virtual gas chromatograph, which, free of detectors 16, i.e. free of real gas chromatographs 10, offers a similar or equivalent measurement accuracy. The analysis system 70 is particularly cost-effective and suitable for simulating the functionality of a real gas chromatograph 10 as in FIG 1 with reduced hardware effort.
Claims
1. Method (100) for ascertaining a composition of a substance sample (15) with at least three components (31, 32, 33, 34) by means of a gas chromatograph (10), comprising the steps: a) quantitatively ascertaining a concentration (35) of a first component (31) of the substance sample (15) and outputting the concentration (35) as a first partial measured value (41) to an artificial intelligence (40); characterised in that the method (100) further comprises the steps: b) ascertaining a concentration (35) of the second and third component (32, 33) of the substance sample (15) by means of the artificial intelligence (40) using the first partial measured value (41) and outputting the ascertained concentrations (35); c) quantitatively ascertaining a concentration (35) of the second component (32) of the substance sample (15) and outputting the concentration (35) as a second partial measured value (42) to the artificial intelligence (40); d) ascertaining the concentration (35) of the third component (33) of the substance sample (15) by means of the artificial intelligence (40) using the first and second partial measured value (41, 42) and outputting the ascertained concentration (35); wherein steps a) to d) are performed within a first measurement cycle (23) of the gas chromatograph (10) in which the substance sample (15) is analysed.
2. Method (100) for ascertaining a composition of a substance sample (15) according to claim 1, characterised in that the artificial intelligence (40) is embodied as a neural network (47), as fuzzy logic and / or as a statistics module.
3. Method (100) for ascertaining a composition of a substance sample (15) according to claim 1 or 2, characterised in that the artificial intelligence (40) is trained by a training dataset (22).
4. Method (100) for ascertaining a composition of a substance sample (15) according to claim 3, characterised in that the training dataset (22) comprises data on historical ascertainments of compositions (17) of substance samples (15) on gas chromatographs (10) of the same or similar construction.
5. Method (100) for ascertaining a composition of a substance sample (15) according to one of claims 1 to 4, characterised in that, in a step e), the concentration (35) of the third component (33) of the substance sample (15) is ascertained quantitatively and the concentration (35) of the third component (33) is output as a third partial measured value (43).
6. Method (100) for ascertaining a composition of a substance sample (15) according to one of claims 1 to 5, characterised in that partial measured values (41, 42, 43) and corresponding concentrations (35) of components (31, 32, 33, 34) of the substance sample (15) ascertained by the artificial intelligence (40) are provided for machine learning of the artificial intelligence (40).
7. Method (100) for ascertaining a composition of a substance sample (15) according to one of claims 1 to 6, characterised in that, in a step f), a concentration (35) of a residual component (34) of the substance sample (15) is detected and a warning is issued if the concentration (35) of the residual component (34) exceeds an adjustable threshold value (44).
8. Method (100) for ascertaining a composition of a substance sample (15) according to one of claims 1 to 7, characterised in that the concentrations (35) of the second and / or third component (32, 33) of the substance sample (15) ascertained by the artificial intelligence (40) are output with an error margin and / or a confidence interval.
9. Method (100) for ascertaining a composition of a substance sample (15) according to one of claims 1 to 8, characterised in that a difference is ascertained between a concentration (35) of a component (31, 32, 33) ascertained by the artificial intelligence (40) and the corresponding partial measured value (41, 42, 43) and a warning is issued if the difference exceeds an adjustable anomaly threshold value in terms of amount.
10. Method (100) for ascertaining a composition of a substance sample (15) according to one of claims 1 to 9, characterised in that step b) and / or d) is performed taking account of a process parameter (45).
11. Method (100) for ascertaining a composition of a substance sample (15) according to claim 10, characterised in that the process parameter (45) is a temperature, a pressure, a pH value and / or an electrical quantity.
12. Method (100) for ascertaining a composition of a substance sample (15) according to one of claims 1 to 11, characterised in that an auxiliary measured value (37) of an auxiliary analysis apparatus (20) is provided during the first measurement cycle (23) for the concentration (35) of at least one component (31, 32, 33, 34) of the substance sample (15).
13. Method (100) for ascertaining a composition of a substance sample (15) according to claim 10, characterised in that the auxiliary analysis apparatus (20) has a measurement cycle duration (39) that is shorter than the first measurement cycle (23).
14. Computer program product (60) for ascertaining a composition of a substance sample (15) comprising an artificial intelligence (40) and a variable training dataset (22), characterised in that the computer program product (60) comprises commands, which prompts the evaluation unit from claim 15 to perform at least one method (100) according to one of claims 1 to 13.
15. Evaluation unit (18) for an analysis apparatus (10), which is embodied to ascertain a composition (17) of a substance sample (15) and can be connected to at least one detector (16), characterised in that the evaluation unit (18) is embodied to perform at least one method (100) according to one of claims 1 to 13.
16. Gas chromatograph (10) for ascertaining a composition (17) of a substance sample (15), which has at least one separating apparatus (14) and one detector (16), wherein the detector (16) is connected to an evaluation unit (18), characterised in that the evaluation unit (18) is embodied according to claim 15.
17. Analysis system (70) for ascertaining a composition (17) of a substance sample (15), comprising an auxiliary analysis apparatus (20), which is embodied as a Raman spectrometer or continuous gas analyser and connected to an evaluation unit (18), in which an executable artificial intelligence (40) is stored, characterised in that the evaluation unit (18) is embodied according to claim 15.
