DE102022213273A1 - Method for detecting a fluid - Google Patents

Abstract

The invention relates to a method for providing a functional relationship between an operating variable, in particular a measured gas resistance, of an aged fluid sensor, in particular a gas sensor, a reference operating variable, in particular a reference gas resistance, and an age of the aged fluid sensor, in particular a gas sensor, comprising the steps:
- Measuring an operating variable, in particular a reference gas resistance, a fluid sensor,
- Artificially aging the fluid sensor to various artificial age stages, in particular by operating the fluid sensor in a high temperature environment, preferably at a high temperature between 50 °C and 400 °C,
- Measuring the operating variable, in particular the gas resistance, of the aged fluid sensor at the various artificial age stages,
- Determining the functional relationship based on the operating variables, in particular the measured gas resistances, of the aged fluid sensor, the measured reference operating variable, in particular the measured reference gas resistance, the artificial age stages of the aged fluid sensor and the actual age of the fluid sensor.
- Providing the functional context, especially for a user of fluid sensors.

Description

The invention relates to a method for providing a functional relationship between a measured operating variable, in particular a gas resistance, of an aged fluid sensor, a reference operating variable, in particular a reference gas resistance, and an age of the aged fluid sensor.

The invention further relates to a method for training a machine learning algorithm for detecting fluids by means of a fluid sensor.

Furthermore, the invention relates to a method for detecting a fluid.

Furthermore, the invention relates to a fluid sensor device for detecting fluids.

Although generally applicable to any operating size of fluid sensors, the following invention is explained using gas resistances of gas sensors such as metal oxide sensors.

State of the art

Gas sensors can be used to detect and quantify certain gases in the air. For example, volatile organic compounds, carbon monoxide or hydrocarbon compounds can be detected. To do this, the gas resistance is measured using a gas sensor.

When metal oxide sensors are operated for long periods of time, they can become inaccurate due to environmental influences and wear. This is particularly evident in sensor drift, where the measured gas resistances of an aged gas sensor differ by a factor from the measured gas resistances of a new sensor.

A machine learning algorithm can be used to classify the various gases. This must first be trained using suitable data. If only data from new sensors is available, the machine learning algorithm cannot make reliable statements about the behavior of aged sensors because the machine learning algorithm has not been trained to compensate for sensor drift. On the other hand, providing data from aged sensors is expensive and time-consuming.

Disclosure of the invention

• - Messen einer Betriebsgröße, insbesondere eines Referenzgaswiderstands, eines Fluidsensors,
• - Künstliches Altern des Fluidsensors auf verschiedene künstliche Altersstufen, insbesondere durch Betreiben des Fluidsensors in einer Hochtemperaturumgebung, vorzugsweise bei einer Hochtemperatur zwischen 50 °C und 400 °C,
• - Messen der Betriebsgröße, insbesondere des Gaswiderstands, des gealterten Fluidsensors zu den verschiedenen künstlichen Altersstufen,
• - Ermitteln des funktionalen Zusammenhangs anhand den Betriebsgrößen, insbesondere den gemessenen Gaswiderständen, des gealterten Fluidsensors, der gemessenen Referenzbetriebsgröße, insbesondere dem gemessenen Referenzgaswiderstand, den künstlichen Altersstufen des gealterten Fluidsensors und dem tatsächlichen Alter des Fluidsensors.
• - Bereitstellen des funktionalen Zusammenhangs, insbesondere für einen Nutzer von Fluidsensoren.

In one embodiment, the present invention provides a method for providing a functional relationship between an operating variable, in particular a measured gas resistance, of an aged fluid sensor, in particular a gas sensor, a reference operating variable, in particular a reference gas resistance, and an age of the aged fluid sensor, in particular a gas sensor, comprising the steps:

• - Measuring an operating variable, in particular a reference gas resistance, a fluid sensor,
• - Artificially aging the fluid sensor to various artificial age stages, in particular by operating the fluid sensor in a high temperature environment, preferably at a high temperature between 50 °C and 400 °C,
• - Measuring the operating variable, in particular the gas resistance, of the aged fluid sensor at the various artificial age stages,
• - Determining the functional relationship based on the operating variables, in particular the measured gas resistances, of the aged fluid sensor, the measured reference operating variable, in particular the measured reference gas resistance, the artificial age stages of the aged fluid sensor and the actual age of the fluid sensor.
• - Providing the functional context, especially for a user of fluid sensors.

• - Messen zumindest einer Betriebsgröße, insbesondere eines Referenzgaswiderstands, mittels des Fluidsensors,
• - Erstellen von künstlichen Messpunkten eines gealterten Fluidsensors durch Transformation der zumindest einen gemessenen Betriebsgröße, insbesondere des Referenzgaswiderstands, anhand des ermittelten funktionalen Zusammenhangs, wobei die künstlichen Messpunkte zu Betriebsgrößenwerten, insbesondere Gasmesswerten, von unterschiedlich gealterten Fluidsensoren korrespondieren,
• - Trainieren des Maschinenlernalgorithmus anhand der erstellten künstlichen Messpunkte.

In one embodiment, the present invention provides a method for training a machine learning algorithm for detecting fluids, in particular gases, by means of a fluid sensor, using a functional relationship provided in particular according to one of claims 1-4, comprising the steps:

• - measuring at least one operating variable, in particular a reference gas resistance, by means of the fluid sensor,
• - creating artificial measuring points of an aged fluid sensor by transforming the at least one measured operating variable, in particular the reference gas resistance, based on the determined functional relationship, wherein the artificial measuring points correspond to operating variable values, in particular gas measurement values, of differently aged fluid sensors,
• - Training the machine learning algorithm using the created artificial measurement points.

• - Messen einer Betriebsgröße, insbesondere eines Gaswiderstands, mittels eines Fluidsensors,
• - Detektieren des Fluids, Bestimmen der Konzentration und/oder der Art des Fluids anhand eines nach einem der Ansprüche 5-8 trainierten Maschinenlernmodells, dem Alter des Fluidsensors und der gemessenen Betriebsgröße, insbesondere dem gemessenen Gaswiderstand.

Furthermore, the present invention provides a method for detecting a fluid, in particular a gas, comprising the steps:

• - Measuring an operating variable, in particular a gas resistance, by means of a fluid sensor,
• - Detecting the fluid, determining the concentration and/or the type of the fluid based on a machine learning model trained according to one of claims 5-8, the age of the fluid sensor and the measured operating variable, in particular the measured gas resistance.

Furthermore, in one embodiment, the present invention provides a fluid sensor device for detecting fluids, comprising a computing unit, a storage unit and at least one fluid sensor, wherein an operating variable, in particular a gas resistance, of a fluid can be measured by means of the fluid sensor, wherein a functional relationship, determined in particular according to one of claims 1-4, between an operating variable, in particular a gas resistance, of an aged fluid sensor, a reference operating variable, in particular a reference gas resistance, and an age of the aged fluid sensor is stored on the storage unit, wherein the computing unit is designed to train a machine learning algorithm according to one of claims 5 to 8, wherein a fluid and/or a concentration of a fluid can be detected by means of the machine learning algorithm and the operating variable, in particular according to a method according to claim 9.

One of the advantages achieved is that it can be used to detect various fluids, especially gases, in a simple and cost-effective manner. Another advantage is that training data can be easily provided for training a machine learning model. Another advantage is that a functional context can be provided for creating the training data, so that, for example, a gas sensor can be trained directly in a desired environment.

The term "artificial age" is to be understood in the broadest sense and refers, in particular in the claims and preferably in the description, to an age of a fluid sensor, in particular a gas sensor, which has been artificially aged by an ageing process. An artificial age therefore does not correspond to the actual age of a sensor, but to an age of a reference sensor which, when aged, has the same sensor drift as the artificially aged sensor. For example, a gas sensor can be aged to an artificial age of one year by operating it for a hundred hours in a high-temperature environment. In particular, the temperature in the high-temperature environment can be changed and the sensor can be operated at different temperatures for certain periods of time.

The term "operating variable" is to be understood in the broadest sense and refers, in particular in the claims, preferably in the description, to a variable that can be determined by means of the fluid sensor. In particular, the gas resistance of a gas is a possible operating variable.

The term “fluid” is to be understood in the broadest sense and refers, in particular in the claims, preferably in the description, to liquids and/or gases and/or a combination of one or more liquids and one or more gases.

To create a functional relationship, an operating variable such as a reference gas resistance of a sensor is first determined. The sensor is then artificially aged and the operating variable such as the gas resistance is measured again after certain age stages. Based on this, a functional relationship can be determined that relates the aged operating variable(s), in particular gas resistances, to the reference operating variable, in particular reference gas resistance, and the age of the fluid sensor.

The functional relationship can then be used to artificially determine data from an aged sensor. Instead of a manually aged gas sensor, for example, the measured gas resistance values of a new gas sensor are simply transformed into new gas resistance values that could come from an aged gas sensor. These artificially aged gas resistance values are used to train a machine learning algorithm. This machine learning algorithm is then able to reliably detect different gases even with really aged gas sensors that exhibit sensor drift.

Further features, advantages and further embodiments of the invention are described below or will become apparent thereby.

According to an advantageous development of the invention, an ambient temperature, a relative humidity and/or the high temperature are used to determine the functional relationship. These additional parameters can be used to reduce environmental influences on the gas sensor. One advantage of this is that the accuracy of the functional relationship is improved.

According to an advantageous development of the invention, a linear regression model is used to determine the functional relationship. One advantage of this is that the functional relationship can be determined in a simple manner from the measured data.

According to an advantageous development of the invention, the measurement of the reference operating variable, in particular the reference gas resistance, and the operating variable, in particular the gas resistance of the aged fluid sensor, is carried out for different types of fluid and/or for different fluid concentrations. One advantage of this is that the accuracy of the functional relationship is improved because more data is available.

According to an advantageous development of the invention, a room temperature, a room humidity and/or a sensor temperature are taken into account to create the artificial measuring points. One advantage of this is that the artificial measuring points can be adapted to the location where a fluid sensor is used. For example, the sensor could be used in a humid environment, so that the machine learning model is trained with data that corresponds to a correspondingly humid environment.

According to an advantageous development of the invention, reference operating variables, in particular reference gas resistances, are measured from several fluids and/or in different fluid concentrations and/or by several fluid sensors. One advantage of this is that the machine learning algorithm can more reliably recognize and classify different fluids and fluid concentrations.

According to an advantageous development of the invention, the machine learning algorithm is trained using the measured reference operating variable, in particular using the measured reference gas resistance. One advantage of this is that the machine learning algorithm can reliably detect and classify fluids even with new sensors that do not yet exhibit sensor drift.

Further important features and advantages of the invention emerge from the subclaims, from the drawings and from the associated description of the figures.

It is understood that the features mentioned above and those to be explained below can be used not only in the combination specified in each case, but also in other combinations or on their own, without departing from the scope of the present invention.

Preferred embodiments and embodiments of the present invention are shown in the drawings and are explained in more detail in the following description.

It shows in schematic form

• 1 Steps of a method for providing a functional relationship according to an embodiment of the present invention;
• 2 Steps of a method for training a machine learning algorithm according to an embodiment of the present invention,
• 3 Steps of a method for detecting a gas according to an embodiment of the present invention,
• 4 a gas sensor according to an embodiment of the present invention,
• 5 an overview of the methods according to embodiments of the present invention, and
• 6a , 6b , 6c a comparison of the quality of different sensors according to an embodiment of the present invention.

1 shows in schematic form steps of a method for providing a functional relationship according to an embodiment of the present invention.

In a step S1, an operating variable, in particular a reference gas resistance, of a gas sensor is measured. The reference gas resistance of the gas sensor is in particular the measured gas resistance of a new gas sensor.

In a further step S2, the gas sensor is artificially aged by operating it in a high temperature environment. In particular, an ISO HTOL (High Temperature Operating Life) test can be used to age the sensor. Operating the gas sensor at a high temperature of around 400 °C for around 100 hours artificially ages the gas sensor by around one year. In this way, the gas sensor can be aged to different ages without the gas sensor actually being in use for years. It is also possible for the sensor to be operated at different temperatures and/or different air humidity for certain periods of time, for example with a fluctuating temperature between 100 °C and 320 °C.

In a further step S3, the gas resistance of the aged gas sensor is measured at different ages. For example, the gas resistance is measured after 1.68, 5 and 10 artificial Additional measurements can also be carried out with different gases, gas concentrations, relative ambient humidities, sensor temperatures and ambient temperatures to generate more relevant data. It is also possible to age several gas sensors and measure data in parallel to obtain greater variability in the measurement data. Mutual influences can also be determined.

Then, in a further step S4, a functional relationship is determined based on the measured gas resistances of the aged sensor, the measured reference gas resistance, the artificial age levels of the aged gas sensor and the actual age of the gas sensor. The functional relationship can be determined, for example, by linear regression. Using the functional relationship, an aged gas resistance can be calculated based on a reference gas resistance and an age of the sensor. In addition, the ambient temperature, the relative humidity or the sensor temperature can also be used for the calculation.

This functional relationship is provided in a further step S5. Since the functional relationship is independent of gases and gas concentrations, the functional relationship can be determined once and then used for several gas sensors.

2 shows in schematic form steps of a method for training a machine learning algorithm according to an embodiment of the present invention.

A gas sensor, for example a fluid sensor, can be trained to detect gases and/or gas concentrations using a machine learning algorithm. Training data is required for this. To ensure that the machine learning algorithm can reliably sense gases even in aged sensors, the training data can also include data from aged sensors.

For this purpose, a reference gas resistance is measured in a step S6 using a gas sensor.

In a further step S7, artificial measuring points of an aged gas sensor are determined by transforming the at least one measured reference gas resistance based on a determined functional relationship, wherein the artificial measuring points correspond to gas measurement values of differently aged gas sensors. The functional relationship can be determined in particular by a method according to steps S1 to S5 according to 1 provided. For example, the functional relationship can be stored in the gas sensor. The functional relationship can be used to determine artificial training data that approximately corresponds to the data that a regularly aged gas sensor would provide. This training data corresponds to data from differently aged gas sensors, for different gases, gas concentrations, relative humidities, sensor temperatures and/or ambient temperatures. This makes it possible to create a wide range of different training data for different applications.

In a further step S8, the machine learning algorithm is trained using the measurement points created in step S7. In this way, the machine learning algorithm can be trained with data from aged gas sensors without the gas sensors having to be aged in a complex manner.

3 shows in schematic form steps of a method for detecting a gas according to an embodiment of the present invention.

In a step S9, a gas resistance is measured using a gas sensor.

In a further step S10, a gas is detected using a machine learning algorithm, the concentration of the gas is determined and/or the type of gas is determined. The machine learning algorithm is designed to classify and detect the gases present based on a measured gas resistance and taking into account the age of the gas sensor. The machine learning algorithm can be trained in particular according to one of claims 5-8. In addition to the measured gas resistance, a relative humidity of the environment, a temperature of the environment or a sensor temperature can be input variables for the machine learning algorithm. In addition, the age of the gas sensor can be used as an input variable.

4 shows in schematic form a gas sensor according to an embodiment of the present invention.

The gas sensor device 1 comprises a gas sensor 2, a computing unit 3 and a storage unit 4. A functional relationship is stored on the storage unit 4, by means of which measuring points of an aged gas sensor can be calculated. The functional relationship can be determined by means of a method according to steps S1 to S5 of the 1 have been provided.

The computing unit 3 is designed to determine measurement data of an aged gas sensor by means of the functional relationship, in particular according to the method according to steps S6 to S8 according to 2 . For this purpose, operating variables such as gas resistance are measured using the gas sensor 2 and measurement data of an aged gas sensor are calculated from this using the functional relationship. The gas sensor device 1 can have additional sensors 2b with which the relative humidity, the ambient temperature or a sensor temperature can be measured. These additional parameters can be used when creating the measuring points so that the measuring points are aligned with the current environment of the gas sensor device 1.

The machine algorithm can then be trained using the measurement data from an aged gas sensor using the computing unit. This means that the machine algorithm can recognise and detect various gases, even if the gas sensor is aged, because the machine algorithm has also been trained using measurement data from an aged gas sensor.

By means of the machine learning algorithm, gases in the environment can then be detected and/or their concentration and/or type can be determined, in particular according to the method according to steps S9 to S10 according to 3 .

In particular, the functional relationship can be determined outside the gas sensor unit and the machine learning algorithm can be trained locally in a computing unit of the gas sensor unit.

5 shows an overview of the methods according to embodiments of the present invention.

First, input variables 6a such as a reference gas resistance, age stages of an artificially aged gas sensor, an actual age of the artificially aged gas sensor, an ambient humidity, ambient temperature and a gas sensor temperature are used in a linear regression model 5. The target variables are gas resistances 7 of a gas sensor artificially aged in a high-temperature process. This creates a functional relationship 8 for creating artificial sensor data.

Using the functional relationship 8 and input variables 6b, measurement data 10 of an artificially aged gas sensor are calculated in a calculation step 9. The input variables 6b can be the same input variables as the input variables 6a.

A trained machine learning algorithm 13 is then obtained in a training process 11 using the measurement data 10 and input variables 6c. The training data used are the calculated measurement data 10 and input variables 6c such as gas resistances measured by a gas sensor, a relative humidity and an ambient temperature. The target data 12 correspond to information about the type and concentration of the gases that correspond to the calculated measurement data 10.

Finally, in a classification process 14, real measurement data 6d measured by a gas sensor, such as a measured gas resistance, a relative humidity and an ambient temperature, can be analyzed by means of the trained machine learning algorithm 13 so that information 15 about the composition of the air measured by the gas sensor is obtained.

The 6a , 6b , 6c show in schematic form a comparison of the quality of different sensors according to an embodiment of the present invention.

The 6a , 6b and 6c show a comparison of the quality of different sensors. The X-axis 601 a, 601 b and 601 c show the sensor age in years. The Y-axis 602a of 6a shows the accuracy in %, i.e. the proportion of all correctly detected gases compared to all measured gases, the Y-axis 601 b of 6b shows the macro-average recall in %, i.e. the weighted proportion of all correctly identified gases compared to all existing gases and the Y-axis 601 c of 6c shows the macro-average F1 value in %, i.e. the harmonic mean of accuracy and macro-average recall.

Shown are a curve 603a, 603b, 603c for a sensor that was not trained with data from an aged sensor, a curve 604a, 604b, 604c for a sensor that was trained with data from an actually aged sensor, and a curve 605a, 605b, 605c for a sensor that was trained with artificially created data from an aged sensor according to an embodiment of the present invention.

It can be seen that the sensor that was trained with data from an actually aged sensor has the highest accuracy 604a, the highest recall 604b and the highest F1 value 604c. The sensor that was not trained with data from an aged sensor is significantly less accurate in comparison and shows a significant decrease in accuracy 603a, recall 603b and F1 value 603c over time. For example, the accuracy 603a of the sensor without aged measurement data has decreased by almost 15% after 10 years. lower than the accuracy 604a of a sensor trained with real data and the recall 603b is up to 30% lower than the recall 604b of a sensor trained with real data from an aged sensor.

The sensor trained with artificially created data from an aged sensor is almost as accurate as the sensor trained with real data. For example, the sensor trained with artificial data has an accuracy 605a that is only less than 3% lower after 10 years than the sensor trained with real data, 604a and the sensor trained with artificial data has a recall 605b that is less than 2.5% lower after 10 years than the recall 604b of the sensor trained with real data.

This means that by training a gas sensor with artificial measurement data according to an embodiment of the present invention, the sensor drift over time can be greatly reduced and thus the accuracy is almost as high as a sensor that was laboriously trained with data from an actually aged sensor.

Although the present invention has been described using preferred embodiments, it is not limited thereto but can be modified in many ways.

Claims (10)

Method for providing a functional relationship (8) between an operating variable, in particular a measured gas resistance, of an aged fluid sensor, in particular a gas sensor, a reference operating variable, in particular a reference gas resistance, and an age of the aged fluid sensor, in particular a gas sensor, comprising the steps: - measuring (S1) an operating variable, in particular a reference gas resistance, of a fluid sensor, - artificially aging (S2) the fluid sensor to various artificial age stages, in particular by operating the fluid sensor in a high-temperature environment, preferably at a high temperature between 50 °C and 400 °C, - measuring (S3) the operating variable, in particular the gas resistance (7), of the aged fluid sensor at the various artificial age stages, - determining (S4) the functional relationship (8) based on the measured operating variables, in particular the gas resistances (7), of the aged fluid sensor, the measured reference operating variable, in particular the measured reference gas resistance, the artificial age stages of the aged fluid sensor and the actual age of the fluid sensor. - Providing (S5) the functional relationship (8), in particular for a user of fluid sensors. Procedure according to Claim 1 , wherein an ambient temperature, a relative humidity and/or the high temperature is used to determine (S4) the functional relationship (8). Procedure according to one of the Claims 1 - 2 , whereby a linear regression model is used to determine (S4) the functional relationship (8). Procedure according to one of the Claims 1 - 3 , wherein the measurement (S3) of the reference operating variable, in particular the reference gas resistance, and the operating variable, in particular the gas resistance (7), of the aged fluid sensor is carried out for different types of fluid and/or for different fluid concentrations. Method for training a machine learning algorithm (13) for detecting fluids, in particular gases, by means of a fluid sensor (2), using a method, in particular according to one of the Claims 1 - 4 provided, functional relationship (8), comprising the steps: - measuring (S6) at least one operating variable, in particular a reference gas resistance (6b), by means of the fluid sensor (2), - creating (S7) artificial measuring points (10) of an aged fluid sensor by transforming the at least one measured operating variable, in particular the reference gas resistance (6b), based on the determined functional relationship (8), wherein the artificial measuring points (10) correspond to operating variable values, in particular gas measurement values, of differently aged fluid sensors, - training (S8) the machine learning algorithm (13) based on the created artificial measuring points (10). Procedure according to Claim 5 , wherein a room temperature, a room humidity and/or a sensor temperature are taken into account for creating (S7) the artificial measuring points (10). Procedure according to one of the Claims 5 - 6 , wherein reference operating variables, in particular reference gas resistances, are measured from several fluids and/or in different fluid concentrations and/or from several fluid sensors. Procedure according to one of the Claims 5 - 7 , wherein the machine learning algorithm (13) uses the measured reference operating variable, in particular which is trained using the measured reference gas resistance (6c). Method for detecting a fluid, in particular a gas, comprising the steps of: - measuring (S9) an operating variable, in particular a gas resistance (6d), by means of a fluid sensor (2), - detecting (S10) the fluid, determining the concentration and/or the type of the fluid using a sensor according to one of the Claims 5 - 8th trained machine learning model (13) and the measured operating variable, in particular the measured gas resistance (6d). Fluid sensor device (1) for detecting fluids, in particular gases, comprising a computing unit (3), a storage unit (4) and at least one fluid sensor (2), wherein an operating variable, in particular a gas resistance (6d), of a fluid can be measured by means of the fluid sensor (2), wherein a sensor, in particular according to one of the Claims 1 - 4 determined, functional relationship (8) between an operating variable, in particular a gas resistance (7), of an aged fluid sensor, a reference operating variable, in particular a reference gas resistance, and an age of the aged fluid sensor is stored, wherein the computing unit (3) is designed to implement a machine learning algorithm (13) according to one of the Claims 5 until 8th to train, wherein a fluid and/or a concentration of a fluid can be detected by means of the machine learning algorithm (13) and the operating variable (6d), in particular according to a method according to Claim 9 .

Images