Method for determining a modified operating strategy for a plurality of similar gas sensors and method for operating a gas sensor

Abstract

The invention relates to a method for determining a modified operating strategy for a plurality of similar gas sensors, wherein each of the plurality of similar gas sensors (100) is configured to acquire measured values for a basic number of features according to a regular operating strategy (302), comprising: providing (306) measured values (304) that have been acquired by the plurality of similar gas sensors with a measured number of features; performing (310) a correlation analysis (312) between the measured values of at least two of the measured number of features; checking (314), based on a result of the correlation analysis, whether a first of the measured number of features is correlated with at least a second of the measured number of features, at least within predetermined tolerances; depending on a result of the checking, determining (316) a modified operating strategy (318) for the plurality of similar gas sensors;and providing (320) the modified operating strategy (318) for use with the plurality of similar gas sensors;

Description

[0001] The present invention relates to a method for determining a modified operating strategy for a plurality of similar gas sensors, a method for operating a gas sensor, a computing unit and a computer program for carrying out the same, and a gas sensor. Background of the invention

[0002] Gas sensors can be used to determine information about gases, such as their composition or concentration, especially for trace gases and / or odors in air as a carrier gas. Gas sensors that can process different gases (or gas mixtures) are also called multi-gas sensors. These can also be referred to as so-called electronic noses. Disclosure of the invention

[0003] According to the invention, a method for determining a modified operating strategy for a plurality of similar gas sensors, a method for operating a gas sensor, a gas sensor, a computing unit, and a computer program for implementing the method are proposed, each having the features of the independent patent claims. Advantageous embodiments are the subject of the dependent claims and the following description.

[0004] The invention relates to gas sensors, in particular multi-gas sensors. Using a gas sensor, components of a gas atmosphere can be determined qualitatively or quantitatively. A gas atmosphere comprises a mixture of gases and / or vapors. It can also predominantly be air, which can contain other components in addition to natural components. These individual components can be described, for example, in terms of their chemical substances (elements, compounds) and their concentrations, for example (molar) concentrations of ethanol, carbon monoxide, nicotine or certain refrigerants (e.g. as leaks from cooling systems / vehicle air conditioning systems). Existing air humidity can also fall under this term of gaseous components as a chemical compound. The purpose of the gas sensor can therefore be, in particular, the determination of a subset of these chemical substances and their (molar) concentrations.If a typical carrier gas or carrier gas mixture is used in an application (e.g. air), the relevant information can be seen in particular in the additional substances; the basic composition of the carrier gas (air) can be assumed to be known per se or is known and therefore no longer relevant.

[0005] In general, this can be described as the determination or detection of a substance or target substance (which itself may comprise several different components) in a target environment. Target substance and target environment refer to a substance that is to be determined during subsequent application of the gas sensor in the desired environment.

[0006] The components of the gas atmosphere can also be viewed in a comprehensive manner, especially with odors. For example, in cigarette smoke, perfumes, or foodstuffs such as beer or coffee, a variety of chemical compounds can be present as odor carriers, which do not necessarily have a fixed molar ratio that is reproducible from sample to sample. Here, the purpose of the gas sensor can be, in particular, to identify the odor type from a set of target odors (e.g., categorically yes / no; or in a semi-quantitative gradation, e.g., on a standardized scale from 0 to 1 depending on the degree of dilution of the emitted odor from the odor carrier with air). Even with this alternative, comprehensive meaning (without explicitly naming the individual chemical substances), the purpose of the gas sensor can be the determination or detection of a subset of existing odors as a target substance in the target environment (e.g.,the gas atmosphere). Therefore, instead of a gas sensor, one can also speak of an odor sensor.

[0007] Gas sensors can therefore be used to determine information about one or more gaseous components – hereinafter also referred to as target substances – of a gas atmosphere or target environment in general. Depending on the context, this can be particularly relevant for chemical substances and molar concentrations or, as is particularly the case with odors in air as a carrier medium, for detecting the typical odor source. In the case of odors, the information can also be evaluated quantitatively, e.g., as a degree of dilution.

[0008] For this purpose, the gas sensor can capture measured values ​​in a target environment, i.e., a gas atmosphere, from which information about the target substance can then be determined. The gas sensor does not necessarily have to capture all components that become or could be part of the gas atmosphere in a given application. The information obtained can be limited to the target substance. The other components can also be referred to as interfering substances or interfering components, regardless of whether they are specifically identifiable chemical substances or a combination of potentially many chemical substances from an odor carrier.

[0009] A machine learning algorithm or model can be used to determine information about target substances or to identify them from measured values, for example, to classify and / or quantify odors. A machine learning model can, for example, include or be an artificial neural network. The machine learning model then receives the measured values ​​as input and determines the information about the gas or target substance, which the machine learning model then outputs.

[0010] In order to determine the most accurate information possible from the measured values, a machine learning model must be trained. This requires so-called training data, which typically consists of a large number of measured values ​​of a substance (training substance) in a training environment—that is, a substance that is later to be detected as a target substance in a target environment. It is understood that different training substances with corresponding measured values ​​can be used for different target substances to be detected. The terms "training substance" and "training environment" refer to a substance used in a suitable environment when collecting training data for the gas sensor.

[0011] In so-called supervised learning or training, in addition to the measured values ​​(so-called characteristics or features) themselves, the relevant information is also required so that the machine learning algorithm or its weights and thus the machine learning model can be adapted accordingly so that the correct information about a target substance is output for certain measured values.

[0012] This information is also referred to as annotations or labels and usually includes at least a description of the training substance. The measured values ​​and the annotations can be linked using time information or timestamps, meaning that a measured value recorded at a specific time is assigned information indicating whether or not the training substance was present in the training environment at that time.

[0013] Such a data set (the training data) is then accessible to a machine learning process. Multiple odors / substances can also be detected in parallel. Depending on the structure of the labels, models can be trained using artificial intelligence methods. As a result, the odors are classified or quantitatively represented based on sensor characteristics.

[0014] Gas or odor sensors can also be used in networked devices, e.g. for monitoring interior spaces in vehicles or buildings with regard to cleanliness, pollution, cigarette smoke, and the like.

[0015] Typical gas sensors record measured values ​​for a specific number of characteristics. Gas sensors based on a thermally modulated sensor, for example, are configured to run through a predefined temperature profile during operation and record measured values ​​for different characteristics at different temperatures. Alternatively or additionally, measured values ​​can be recorded at different times since a temperature change, i.e., different lengths of time are waited since the last temperature change before the measured value is recorded.

[0016] A gas sensor can, for example, have ten different characteristics for which measured values ​​are recorded and which are used to analyze or identify a substance. Additional characteristics for pressure, temperature, or humidity can also be present. During regular operation, i.e. according to a regular operating strategy, a gas sensor should record measured values ​​for a basic number of characteristics. In other words, a regular operating strategy means that the gas sensor is operated in such a way that measured values ​​are recorded for a specific number of characteristics (the basic number). The gas sensor can, for example, be manufactured in such a way that, as long as the gas sensor is not modified, it will operate according to the regular operating strategy.

[0017] This basic set of features then represents, for example, input values ​​for a previously trained machine learning model, which provides the odor information categorically or through concentration values ​​as an inference. In principle, this basic set of features can be generated from individual sensors using an electrical mode of operation (e.g., the aforementioned thermal modulation) or from multiple individual sensors integrated on the same chip or in completely different electronic components.

[0018] When operating gas or odor sensors in IoT applications, for example, it may be necessary to centrally evaluate raw data from gas sensors (e.g., from vehicles or buildings). Gas sensors usually involve a variety of signal characteristics, the aforementioned basic number of characteristics. However, the regular evaluation of a large number of sensor characteristics per device is associated with costs and effort, e.g., the energy consumption in the data-collecting sensor unit (relevant for battery-operated sensor units), the volume of data to be transmitted, e.g., via a mobile network, and any required computing power.

[0019] Instead of a regular operating strategy, however, a modified operating strategy for gas sensors may be indicated if, for example, it turns out that the application involves more or less variety of target odors or disturbances than for which the gas sensor was developed, for example on a laboratory basis.

[0020] Against this background, the invention proposes a possibility for determining a modified operating strategy for a plurality of similar gas sensors, wherein each of the plurality of similar gas sensors is configured to record measured values ​​for a basic number of features according to a regular operating strategy, as already mentioned.

[0021] For this purpose, measured values ​​that have been recorded by the plurality of similar gas sensors with a measured number of features are provided. In one embodiment, the measured number can correspond to the basic number, as will be explained in more detail later. A correlation analysis is then carried out between the measured values ​​of at least two of the measured number of features. In a simple case, this can, for example, comprise a correlation analysis between the measured values ​​of two of the measured number of features. However, this simple case can, for example, also be just a first step, and in further steps, a multiple correlation analysis can, for example, be carried out, specifically between the measured values ​​of, for example, three of the measured number of features. This can then be carried out successively in further steps with the measured values ​​of, for example, four of the measured number of features, etc. In general, the correlation analysis can, however, for example,can also only be carried out between the measured values ​​of three or four of the number of measured features.

[0022] In one embodiment, the measured values ​​are first checked to determine whether one of the basic features of the plurality of similar gas sensors is not required or whether at least one additional feature might be necessary. This can be done within the framework of a so-called cluster analysis. This can mean, in particular, that the relationships found indicate that a feature is redundant, i.e., does not provide any additional information, and is therefore not required in operation; however, it can also be the case, in particular, that the relationships found indicate, i.e., provide evidence, that even more information could be obtained with an additional feature. If all expected odors occur in such a cluster analysis, then no in-depth correlation analysis is required. In this case, the correlation analysis is only performed if the check is positive, i.e.,that one of the features is not needed or that an additional feature might be required.

[0023] As part of the correlation analysis, it is then checked whether a first of the measured features is correlated with at least a second of the measured features (I), at least within specified tolerances. It should be noted here that in the case of the aforementioned multiple correlation analysis, a correlation of a first feature with more than one second feature can occur. Depending on the result of the check, a modified operating strategy is then determined for the plurality of similar gas sensors, which is then made available for use with the plurality of similar gas sensors.

[0024] Such a changed operating strategy can, for example, save costs or detect new odors.

[0025] In one embodiment, the number of measurements corresponds to the base number, as already mentioned. If, in this case, the result of the test is positive, the modified operating strategy for the plurality of similar gas sensors involves changing the operation for a selected one of the first and at least one second of the base number of features. This allows, in particular, an originally developed required signal feature to be identified as unnecessary and omitted. Nevertheless, a machine learning model trained on the base number of features can continue to be used. In this case, the omitted feature can be replaced by one that correlates with the omitted feature.

[0026] For example, the changed operating strategy for the selected feature can include at least one of the following procedures. The gas sensor is no longer controlled to acquire measured values ​​for the selected feature, i.e., the feature in question is omitted. Or, no more measured values ​​are acquired for the selected feature. Or, no more measured values ​​are processed for the selected feature or transmitted for further processing. Or, the control of the gas sensor for acquiring measured values ​​for the selected feature is changed, e.g., to a different measurement time.

[0027] In general, the measured values ​​can be evaluated using cluster analysis (possibly preceded by principal component analysis). It may turn out that previously suspected smells or types of smell do not occur. In this case, as mentioned, correlation analysis is carried out, in which the basic number of sensor features are examined at least in pairs (generally in tuple form) and quantified, for example, using correlation coefficients. If, for example, a first feature is correlated with a second feature, both features contain equivalent information. The first feature can also only exhibit a correlation when the second feature is combined with a third (or fourth, etc.) feature. The triple (or four-tuple, etc.) of features then contains only two (or three, etc.) pieces of independent information, and the third (or fourth, etc.) pieces of information are derived from the other two (or three, etc.) pieces of information.

[0028] In both cluster analysis and correlation analysis, the number of rare exceptions can be appropriately assessed and considered in the context of expectations. For example, a missing cluster may be present due to a few exceptional measurement points, or the correlation curve may show a few outliers from a systematic relationship.

[0029] For example, if it is expected that 0.1% of all vehicle interior use cases will result in unacceptable contamination, then measured values ​​corresponding to 0.1% are not outliers that justify omitting the feature. If, on the other hand, contamination is expected in 50% of all cases, then the 0.1% would be a candidate for an event that actually occurs too rarely. Whether a feature can be omitted is therefore also an economic decision for rare events. The corresponding acceptance thresholds can be specified, for example, by the user (e.g., the operator of a vehicle fleet).

[0030] The analysis of the measured values ​​can take place in the central unit, e.g. in the cloud or in another computing system or unit. The analysis can be carried out, for example, for all of the numerous gas sensors or on a random sample. It can also be broken down according to systematic parameters, e.g. due to a limitation of the variability of the gas sensors, variability of the odor-forming substance or variability of interfering substances. This can include climatic characteristics broken down by region and time of year (or ambient temperatures and humidities that could influence the raw sensor signals). This can include inhomogeneous user behavior, e.g. for the target variable cigarette smoke, individual cigarette brands could smell different, and the individual brands could in turn be more or less popular in certain target markets or age groups.This may include regional variations in outdoor air quality, e.g. inner cities, industrial areas or agricultural areas.

[0031] This breakdown can result in an evolutionary split of operating modes into a specific number of groups. For example, a first group can continue to operate with the full set of features. In a second group, one of the features can be omitted. In a third group, a different feature can be used. In general, the modified operating strategy can therefore only be used for a selected portion of the large number of similar gas sensors.

[0032] By repeatedly applying the proposed procedure, two or more of the basic number of characteristics can be omitted for individual groups or in general.

[0033] As mentioned, the changed operating strategy may include not recording measured values ​​for a specific characteristic (omitting the characteristic). Likewise, the changed operating strategy may include no longer processing or transmitting measured values ​​for the specific or selected characteristic. For example, while a full measurement with the basic number of characteristics is performed, transmission of one less characteristic is reduced.

[0034] If a characteristic is omitted, deactivating a raw value can mean that a measured value in the gas sensor neither needs to be recorded, buffered, nor transmitted. However, with thermally modulated metal oxide sensors, this does not necessarily mean that a specific (intermediate) temperature can be omitted. If the quality of the remaining characteristics is influenced by a previous (intermediate) temperature, the temperature profile must be maintained unchanged over time.

[0035] If there are completely separate sensors / receptor layers of a physical sensor array and a single sensor or receptor layer exclusively provides the feature to be deleted, its heating element can again be switched off.

[0036] As mentioned, the control of the gas sensor for recording measured values ​​for the selected feature can also be modified. If, based on the considerations described above, a previous readout point (feature) of the gas sensor is a candidate for omission of this feature, a variation can be made instead of omission. For an existing temperature profile, for example, temperature gradations are approached in stages or with slow gradients. For each point (feature), it is conceivable to read it earlier, later, or at a different temperature in order to increase the information content.

[0037] In one embodiment, the measurement count is at least one greater than the base count. If the test result is negative, i.e., if there is no correlation, measurement values ​​for the measurement count of features are recorded in the modified operating strategy.

[0038] If, for example, an over-the-air configuration option is available on the gas sensor hardware, the procedure for checking whether a feature can be omitted can also be operated in reverse. For example, an additional feature is read out or generated. When generating, the above comments apply equally to thermally modulated gas sensors, i.e., a changed temperature profile to generate a new measured value or measuring point must not adversely affect the other readout points. The following options are therefore preferable: The temperature profile of the gas sensor remains unchanged. However, an additional readout or measuring point on the receptor layer is placed between two previously readout points. If individual heating pulses have previously been separated in time (e.g.If a pulse profile with a duration of 10 seconds that is repeated periodically once per minute), the additional readout point with a new temperature, if applicable, can be placed directly at the end of the previous profile. The new profile with the new readout or operating point and the previous profile with the basic number of operating points (for the features) can be distributed across the large number of gas sensors in two test groups. If no statistically significant difference can be identified or proven across both groups in the basic number of features or their measured values, it can be assumed that the new feature is neutral with regard to the previous features, and the new feature can be used without limiting the accuracy of the previous machine learning model with regard to the previous features.

[0039] However, it may not be sufficient to simply prove the neutrality of the existing features (base number) when introducing a new feature. It should also be proven or shown, for example, that the new feature provides additional information, i.e., it must provide additional points in the existing correlation of pairs of sensor features or through clustering that correspond to outlier points. Which odor is present and whether it is a uniform odor class or multiple odors can then be determined, for example, by selecting the timestamps and identifying the affected gas sensors (e.g., in vehicles), e.g., via a user survey. Accordingly, a new machine learning model can be trained with the new feature. The additional training data used can be the data for which an annotation of the odor is available (e.g., from the example user survey mentioned above).

[0040] In summary, for example, all characteristics (base number) of the large number of gas sensors, e.g. several thousand, can be examined together in an observation phase. If it turns out that at least one of the base number of characteristics is no longer required, this can be "switched off" in the future, e.g. by an over-the-air update, i.e. the changed operating strategy is used. This is particularly the case if a correlation analysis can show that one of the base number of characteristics (considered across the large number of gas sensors) can also be predicted by one fewer than the base number of characteristics and is therefore not an independent measured value. However, a characteristic can also be omitted, for example, if it can be proven that it lies above or below a measuring range limit of the sensor in the target application (so-called clipping).

[0041] Reasons for a lower diversity compared to the expected and developed characteristics can be, for example, that not all target odors occur in reality (e.g., a cigarette smell due to a smoking ban), or that (target) odors occur in lower concentrations than is relevant for monitoring, or that disturbances considered and to be compensated for during the development phase have a lower diversity in the application than was feared during the development phase. For example, compensation for temperature and humidity fluctuations may be intended, but the conditions at the site of use turn out to be less fluctuating.

[0042] To avoid having to train a new machine learning model for the now reduced number of features, the omitted feature is emulated based on the correlations, for example, and the previous model, which requires the basic number of input variables, is continued to be used.

[0043] An optional monitoring phase can be used to test a sample of the numerous gas sensors (e.g., 1%) with the full set of basic characteristics without deactivation to determine whether the deactivation remains appropriate. In other words, after the modified operating strategy has been determined, it is checked whether the modified operating strategy is still appropriate. For example, typical user behavior could change again, or fluctuations in temperature and humidity could occur again due to seasonal changes or expanded deployment locations.

[0044] Conversely, the method can also be used to add an additional feature (as a candidate, so to speak) to identify newly measurable odors. Alternatively, instead of removing or adding a feature, the existing feature could be varied. For example, by varying a temperature level or a temperature holding time, the feature can be read or recorded.

[0045] Repeating the process by switching off, on, and varying raw signal characteristics can result in an evolutionary optimization of the operating profile of, for example, a thermally modulated gas sensor. Depending on systematic variations in operating conditions, a "splitting" of operating modes can also occur depending on the application.

[0046] In one embodiment, a gas sensor is operated with a modified operating strategy determined according to one of the preceding variants, and using a machine learning model trained for the regular operating strategy. This applies in particular if the operation of the gas sensor is modified for one feature. In the case of an additional feature, however, a machine learning model trained for the modified operating strategy is expediently used. This operation preferably includes providing measured values ​​acquired by the gas sensor in accordance with the modified operating strategy in a test environment. Based on the machine learning model and the measured values, a target substance in the target environment is then determined, and information about the target substance is provided.

[0047] A computing unit according to the invention, e.g., a control unit of a gas sensor or a computer, is configured, in particular by programming, to carry out a method according to the invention. A gas sensor according to the invention comprises such a computing unit.

[0048] The implementation of a method according to the invention in the form of a computer program or computer program product with program code for carrying out all method steps is also advantageous, since this entails particularly low costs, in particular if an executing control unit is also used for other tasks and is therefore already present. Finally, a machine-readable storage medium is provided with a computer program stored thereon, as described above. Suitable storage media or data carriers for providing the computer program are, in particular, magnetic, optical and electrical memories, such as hard disks, flash memories, EEPROMs, DVDs, and others. Downloading a program via computer networks (Internet, intranet, etc.) is also possible. Such a download can be wired or cable-based or wireless (e.g. via a WLAN network, a 3G, 4G, 5G or 6G connection, etc.).

[0049] Further advantages and embodiments of the invention will become apparent from the description and the accompanying drawings.

[0050] The invention is illustrated schematically in the drawing using exemplary embodiments and is described below with reference to the drawing. Short description of the drawings Fig. 1 shows schematically a gas sensor to explain the invention. Fig. 2 shows schematically an arrangement for explaining the invention. Fig. 3a, Fig. 3b schematically show sequences of the method according to the invention in various preferred embodiments. Fig. 4, Fig. 5, Fig. 6 show schematic diagrams for explaining the invention. Embodiment(s) of the invention

[0051] In Fig. Figure 1 schematically shows a gas sensor 100 to explain the invention. By way of example, the gas sensor 100 has two sensor elements 102, 104, thus representing a multi-gas sensor. Furthermore, the gas sensor 100 has a computing or control unit 106, on which, for example, operating software can run to record and, if necessary, process measured values. By way of example, the control unit 106 is also configured for wireless communication, for example to send data to a higher-level server 120 (e.g., in a so-called cloud) and / or to receive data from there.

[0052] Furthermore, a target substance 110 (e.g., an odor source) in a target environment 112 is indicated by way of example, which target substance 110 can be detected by means of the gas sensor 100 or about which information can be determined using the gas sensor 100.

[0053] The Fig. The gas sensor 100 shown in Figure 1 can be used for an application to detect whether a target substance or odor source is present in the target environment, as well as for collecting or generating training data. A comparable or different gas sensor can also be used to generate the training data.

[0054] In Fig. 2, various vehicle types 200, 201, 202 are shown as examples in Figures (a), (b), and (c), as well as a higher-level server 220 (e.g., in a so-called cloud). This is intended to represent an example of a vehicle fleet that in practice may comprise a large number, e.g., thousands, of vehicles and includes variants that in turn correspond to different typical user behaviors. In each vehicle, a gas sensor, such as the gas sensor 110 according to Fig. 1 may be arranged or installed to detect odors in the respective vehicle.

[0055] The measurement data acquired by the gas sensors can be transmitted to the higher-level server 220 via an indicated wireless communication connection, e.g. mobile radio, as shown in Figure (a), as well as in Fig. 1. Various odors 210, representing target substances, are also indicated as symbols by way of example. The gas sensors can be configured to detect all of these different odors, according to a regular operating strategy.

[0056] Based on this analysis, a modified operating strategy can now be determined, which is then transmitted to the gas sensors in the vehicles, as indicated in Figure (b). For example, only certain odors will be detected, as indicated in Figure (c).

[0057] In Fig. Figure 3a schematically illustrates a method flow in one embodiment for determining a modified operating strategy for a plurality of similar gas sensors. Each of the plurality of similar gas sensors is configured to acquire measured values ​​for a basic number of features according to a regular operating strategy. These can be gas sensors as in Fig. 1 shown, which are used in vehicles such as Fig. 2 or in other devices.

[0058] For this purpose, in step 300, measured values ​​304 are recorded by each of the plurality of similar gas sensors with the respective regular operating strategy 302. These measured values ​​are transmitted, in step 306, to a higher-level server or other computing unit, where they are received and provided, as is also the case, for example, in Fig. 2 is shown in Figure (a). Each gas sensor records measured values ​​for, for example, ten characteristics or sensor characteristics. In one embodiment, this number "ten" represents the base number and the measurement number.

[0059] In an optional step 308, the measured values ​​are then checked, for example, to determine whether one of the ten features is not required. This is shown in diagrams (a), (b), (c) in Fig. 4 shows a cluster analysis. Such a cluster analysis can be performed, for example, using so-called principal component analysis. In the diagrams, a first principal component 401 is plotted to the right and a second principal component 402 is plotted upward.

[0060] Diagram (a) shows that three clusters are formed, each of which can be assigned to an odor 411, 412, 413. Diagram (b) shows that the cluster for odor 412 is not present; this means that the odor 412 in question was not present in the underlying measurement data. Diagram (c) shows that the cluster for odor 412 is only present with a few measured values; this may mean that the odor 412 in question was present significantly less frequently in the underlying measurement data than expected. In case (b), and possibly also in case (c), it can then be decided that one of the ten features is not required or at least may be unnecessary or superfluous.

[0061] If it is the case that one of the ten features is not required, a correlation analysis 312 is carried out in step 310. As mentioned, the cluster analysis is optional. If the cluster analysis, i.e., step 308, is not carried out, the correlation analysis, step 312, can be carried out directly after step 306. The correlation analysis is carried out between the measured values ​​of, for example, two of the ten features. Based on a result of the correlation analysis, a check is then carried out in step 314 to determine whether a first of the ten features is correlated with a second of the ten features, at least within predetermined tolerances. In general, however, a linear relationship is not required for a correlation, and a correlation can only be detected by including more than two features. For illustration purposes, the simple case of a pairwise correlation in Fig. 5 are shown.

[0062] For this purpose, in diagrams (a), (b), (c), (d) in Fig. Figure 5 shows such a correlation analysis and its results. In the diagrams, a first feature or its raw signal 501 is plotted to the right, and a second feature or its raw signal 502 is plotted upwards. The higher the value of the raw signal, the higher the respective odor intensity, for example.

[0063] Diagram (a) shows that the measured values ​​511, 512, and 513 for three different odors or odor types all lie on the same line. For all measured values, the measured values ​​of attributes 501 and 502 correlate. Therefore, one of the two attributes can be omitted, for example.

[0064] Diagram (b) shows that the measured values ​​512 for one of the three odors or odor types are not aligned with the others. For odors or odor types 511 and 513, the two features correlate and could be omitted. For odor or odor type 512, however, this correlation does not exist. Therefore, if odor or odor type 512 is to be recognized, neither of the two features should be omitted.

[0065] However, if odor or odor type 512 does not appear in the measured values, as can be seen in diagram (c), one of the two characteristics could be omitted.

[0066] If odor or odor type 512 only appears sporadically in the measured values, as shown in diagram (d), a decision could be made, depending on the situation, whether one of the two characteristics should be omitted.

[0067] In step 316, depending on the result of the check, a modified operating strategy 318 for the gas sensors is determined. In particular, this includes whether an operation for a selected one of the first and second of the ten features is or will be modified.

[0068] For this purpose, diagrams (a), (b), (c), (d) in Fig. Six operating strategies are shown. In each diagram, a temperature 602 is plotted against a time 601. The curves shown in the diagrams correspond to a temperature profile, ie, the temperature level is changed several times, for example, while remaining constant for a certain period of time. The points indicate the time at which a measured value is recorded—and thus a characteristic. Characteristics 611, 612, and 613 are described in more detail in the diagrams as examples, and these will be discussed below.

[0069] Diagram (a) shows six exemplary features, including features 611 and 612. This could, for example, represent a regular operating mode. Diagram (b) shows only five exemplary features, with feature 611 still present, but feature 612 not present—it has thus been omitted. This could therefore represent a modified operating strategy.

[0070] Diagram (c) shows seven exemplary features, with features 611 and 612 still present, as well as an additional feature 613. At the end of the temperature profile of the regular operating strategy, a further temperature level is added, at which measured value 613 is recorded (as a feature). This could therefore also represent a modified operating strategy.

[0071] Diagram (d) shows six features as examples, with features 611 and 612 still present, but the acquisition time of feature 612 has been shifted to a later date. This could also be a change in operating strategy.

[0072] The modified operating strategy is then provided, step 320, for use with the gas sensors, as also described in Fig. 2 is shown in Figure (b). In a step 322, it may then be checked whether the modified operating strategy is still appropriate.

[0073] In Fig. 3b schematically shows a sequence of a method in one embodiment, namely for operating a gas sensor with a modified operating strategy, which, for example, as in Fig. 3a, using a machine learning model trained for the regular operating strategy.

[0074] In step 330, measured values ​​332 are provided that were acquired by the gas sensor in a target environment according to the modified operating strategy. In step 334, a target substance in the target environment is determined based on a machine learning model 336 and the measured values; then, in step 338, information 340 about the target substance is provided.

[0075] For the machine learning model, which is trained on ten features and therefore requires measurements for ten features, only measurements for nine features are recorded or provided, for example. The measurement for the tenth feature can, as mentioned, be emulated, for example.

Claims

[1] A method for determining a modified operating strategy for a plurality of similar gas sensors (100), wherein each of the plurality of similar gas sensors (100) is configured to acquire measured values ​​for a basic number of features according to a regular operating strategy (302), comprising: Providing (306) measured values ​​(304) which have been detected by the plurality of similar gas sensors with a measured number of features; Carrying out (310) a correlation analysis (312) between the measured values ​​of at least two of the measured number of features; Checking (314), based on a result of the correlation analysis, whether a first of the measured number of features is correlated with at least a second of the measured number of features, at least within predetermined tolerances; depending on a result of the testing, determining (316) a modified operating strategy (318) for the plurality of similar gas sensors; and Providing (320) the modified operating strategy (318) for use with the plurality of similar gas sensors. [2] Method according to claim 1, wherein the gas sensors (100) are based on a thermally modulated sensor, wherein the gas sensor is configured to run through a predetermined temperature profile in the regular operating strategy, and for various features - to record measured values ​​at different temperatures, and / or - To record measured values ​​at different time periods since a change in temperature. [3] The method of claim 1 or 2, further comprising: Checking (308) the measured values ​​to determine whether one of the basic number of features of the plurality of similar gas sensors is redundant and thus not required, or whether there are indications that at least one additional feature could provide additional information; The correlation analysis is only performed if the verification is positive. [4] The method according to claim 1, wherein the measurement number corresponds to the basic number, and wherein, if the result of the checking is positive, in the changed operating strategy for the plurality of similar gas sensors, an operation for a selected one of the first and the at least one second of the measurement number of the features is changed. [5] The method of claim 4, wherein the modified operating strategy for the selected feature comprises at least one of the following: - The gas sensor is no longer controlled to record measured values ​​for the selected feature, - No more measured values ​​are recorded for the selected feature, - No more measured values ​​for the selected feature are processed or transmitted for further processing, - A control of the gas sensor for recording measured values ​​for the selected feature is changed. [6] Method according to one of claims 1 to 3, wherein the number of measurements is at least one greater than the basic number, and wherein, if the result of the test is negative, measured values ​​for the number of measurements of features are recorded in the changed operating strategy. [7] The method of claim 6, wherein the modified operating strategy comprises using a machine learning model (336) trained for the modified operating strategy. [8] The method of any one of claims 1 to 5, wherein the modified operating strategy comprises using a machine learning model (336) trained for the regular operating strategy. [9] Method according to one of the preceding claims, wherein the modified operating strategy is used only for a selected part of the plurality of similar gas sensors. [10] Method according to one of the preceding claims, wherein, after the changed operating strategy has been determined, it is checked (322) whether the changed operating strategy is still required. [11] A method for operating a gas sensor (100) with a modified operating strategy determined according to any one of the preceding claims with reference to claim 4, using a machine learning model (336) trained for the regular operating strategy, or with a modified operating strategy determined according to any one of the preceding claims with reference to claim 5, using a machine learning model (336) trained for the modified operating strategy. [12] A method according to claim 11, comprising: Providing (330) measured values ​​(332) that have been detected by the gas sensor in a target environment according to the changed operating strategy; Determining (334), based on the machine learning model (336) and the measured values, a target substance in the target environment; and Providing (338) information (340) about the target substance. [13] Computing unit (106, 120) configured to carry out a method according to any one of the preceding claims. [14] Gas sensor (100) with a computing unit (106) according to claim 13, wherein claim 13 is dependent on claim 11 or 12. [15] A computer program which causes a computing unit to carry out a method according to any one of claims 1 to 12 when executed on the computing unit. [16] A machine-readable storage medium having stored thereon a computer program according to claim 15.

Images