Patent Application
20230110328
This application claims the benefit of European Patent Application No. 21202164, filed on Oct. 12, 2021, which application is hereby incorporated herein by reference.
Embodiments relate to a gas sensing device for sensing one or more gases in a mixture of gases. Further embodiments relate to a method for operating such gas sensing device. More particular, the disclosure deals with the estimation of gas concentrations through the use of chemo-resistive gas sensors.
Literature on chemo-resistive gas sensors is generally limited to a simple model for proof of sensor functionality or costly data acquisition methodologies using geographically distributed sensor systems with impractical implementations. In order to distinguish between different gases, the use of selective physical gas filters or additional non-chemo-resistive gas sensors has been proposed. However, such use has a significant impact on the product sizes and cost.
A gas sensing device for sensing one or more gases in a mixture of gases is provided. The gas sensing device comprises:
one or more chemo-resistive gas sensors, wherein each of the gas sensors is configured for generating signal samples corresponding to concentrations of the one or more gases in the mixture of gases;
one or more heat sources, wherein the one or more heat sources are controlled in such way that the gas sensors are each heated according to one or more temperature profiles;
a preprocessing processor configured for receiving the signal samples from each of the gas sensors and for preprocessing the received signal samples in order to generate preprocessed signal samples for each of the gas sensors;
a feature extraction processor configured for receiving the preprocessed signal samples and for extracting a set of feature values from each of the received preprocessed signal samples of the gas sensors based on characteristics of the received preprocessed signal samples of the gas sensors;
a discriminative embedding network processor configured for receiving the sets of feature values and for creating for each of the sets of feature values a set of embedded feature values, wherein the discriminative embedding network processor comprises a first trained model based algorithm processor and a first trained model for the first trained model based algorithm processor, wherein the first trained model is configured for applying a loss function using discriminate weights to the sets of feature values in order to create the sets of embedded features values;
a classification processor configured for receiving the sets of embedded feature values and for creating a classification value for each set of the embedded feature values, wherein the classification value indicates a class of the mixture of gases, wherein the classification processor comprises a second trained model based algorithm processor and a second trained model for the second trained model based algorithm processor, wherein the sets of embedded feature values are fed to an input of the second trained model based algorithm processor, wherein the classification values are provided at an output of the second trained model based algorithm processor; and
a quantification processor configured for receiving the sets of embedded feature values and the classification values, wherein the quantification processor is configured for creating for each of the gases a sensing result for each of the sets of embedded feature values, wherein the quantification processor comprises a third trained model based algorithm processor and a plurality of third trained models for the third trained model based algorithm processor, wherein the sets of embedded feature values are fed to an input of the third trained model based algorithm processor, wherein the sensing result are provided at an output of the third trained model based algorithm processor, wherein one third trained model of the plurality of third trained models is selected for creating the sensing results based on the classification values.
The one or more chemo-resistive gas sensors may be graphene gas sensors or reduced graphene gas sensors, where the base material is functionalized with specific chemicals, e.g. with platinum (Pt), or manganese dioxide (MnO2), so that each of the gas sensors is sensitive for gases, e.g. for nitrogen dioxide (NO2), ozone (O3) or carbon monoxide (CO). In doing so, the interaction between graphene sheets and absorbed gas analytes influences the electronic structure of the material depending on the mixture of gases, resulting in altered charge carrier concentration and changed electrical conductance.
In case of multi-gas sensing, a multi-gas sensor array comprising a plurality of chemo-resistive gas sensors having dissimilar selectivity may be used. Due to the different sensitivity towards various gas molecules, resistances of the gas sensors change in disparate patterns, making it possible to analyze complicated gas mixtures with one single sensor array.
A signal sample is a sequence consisting of time-discrete signal values, wherein the signal values are output by one of the gas sensors.
Each of the gas sensors may be heated by one or more heat sources. The heat sources may be electrically powered resistive heating elements or radiators emitting light, in particular with ultra violet light. Each of the one or more heat sources is controlled according to one or more temperature profiles during operational phases. Each of the temperature profiles modulates a temperature of one or more of the gas sensors between a maximum temperature and a minimum temperature.
For example, according to one of the temperature profiles the temperature of the one or more heating elements may be pulsed between a maximum temperature and a minimum temperature. The maximum temperature may be, for example, set to a value between 150° C. and 300° C., whereas the minimum temperature may be, for example, set to a value between 50° C. and 200° C.
In other embodiments, other temperature profiles such as ramps may be used.
The temperature modulation could be the same for all sensors or different for at least some of the sensors.
The temperature modulation improves repeatability and stability of the sensing results.
The term processor refers to an electronic device configured for specific task. A processor may comprise hardware or a combination of hardware and software. Different processors may share hardware components and/or software components.
The preprocessing processor is configured for suppressing and/or compensating of artifacts in the signal samples and/or noise in the signal samples and/or invalid signal samples due to malfunctioning gas sensors and/or errors in the signal samples due to drifts of the gas sensors in order to produce more reliable filtered signal samples.
The feature extraction processor is configured for receiving the preprocessed signal samples and for extracting one or more feature values from the received preprocessed signal samples of each of the gas sensors based on characteristics of the received preprocessed signal samples of the respective gas sensor. The features may be based on dynamic characteristics of the signal samples. To this end, the modulated nature of the responses of the gas sensors is leveraged and characteristics are extracted which rely on the dynamic evolution of the gas sensors.
According to embodiments, a gas sensing device has an improved detection mechanism, where a discriminative embedding network processor is used followed by a classification processor and a quantification processor so as to reduce cross-sensitivity and to improve estimation accuracy for each single target gas in a mixture of gases.
A trained model based algorithm processor is a processor, which is capable of machine learning. The machine learning is done in a preoperational training phase in which trained models are developed by comparing actual output values of the trained model based algorithm stage with desired output values of the trained model based algorithm stage for defined inputs of the trained model based algorithm stage. The trained models have a predefined structure, wherein a parametrization of the predefined structure is done during the training phase. The trained models comprise the learned content after the training phase is finished. In an operational phase for producing processing results one or more of the trained models from the training phase are used to process their input data.
In the training phase, the plurality of trained models can be established and afterwards stored at the gas sensing device. The trained models may differ in the structures and/or the parameters. During the operation phase the most appropriate trained model may be selected depending on the specific use-case.
The discriminative embedding network processor is configured for embedding the sets of feature values into a new space so that the sets of embedded feature values have better separability. In other words, the discriminative embedding network processor applies a discriminative mapping to the sets of feature values in order to generate sets of embedded feature values so that embedded feature values belonging to the same gas or gas mixtures are closer to each other and those belonging to different gases or gas mixtures are further apart.
The classification processor is configured for creating a classification value for each set of the embedded feature values, wherein the classification value indicates a class of the mixture of gases. In other words, the classification processor predicts for each of the sets of embedded feature values a class or a use-case and recommends a regression mode to be used by the quantification processor for the quantification of the gas concentration(s).
The quantification processor is configured for producing sensing results depending on the sets of embedded feature values and depending on the classification value. In other words, the quantification processor is a regressor which applies the recommended regression mode to the sets of embedded values in order to obtain one or multiple gas concentration estimates.
The sensing results may be alphanumeric terms, for example alphanumeric terms on a scale from “high” to “low”. In particular, the terms of an air quality index system, for example terms of the European air quality index, may be used for outputting the sensing results. In other embodiments, the sensing results may be physical quantities such as “4% by volume”.
The combination of the discriminative embedding network processor, the classification processor and the quantification processor improves the accuracy of the sensing results. As the quantification is performed depending on the class of the mixture of gases, this is valid for different mixtures of gases. Furthermore, the gas sensor may be adapted for the detection of additional gases without changing the hardware, as it is sufficient to modify the third trained models of the quantification processor. In addition, it is possible to adapt the gas sensing device to additional mixtures of gases, as it is sufficient to modify the second trained model of the classification processor.
The proposed gas sensing device provides an end to end solution for multi-gas adsorption sensors which is versatile, widely-applicable to multiple applications and uses cases (indoor or outdoor air quality monitoring, health checks, such as breath analysis or disease diagnostics, etc.) and can be embedded in a smart portable device.
According to embodiments of the disclosure the one or more gas sensors are alternately operated in recovery phases and in sense phases;
wherein the one or more heat sources are controlled in such way that the gas sensors are each heated according to one or more first temperature profiles of the one or more temperature profiles during the recovery phases and according to one or more second temperature profiles of the one or more temperature profiles during the sense phases, wherein for each of the gas sensors a maximum temperature of the respective first temperature profile is higher than a maximum temperature of the respective second temperature profile.
Each of the one or more heat sources may be controlled according to one or more temperature profiles during the recovery phases and according to a second temperature profile during the sense phases, wherein a maximum temperature of the first temperature profile is higher than a maximum temperature of the second temperature profile.
For example, the temperature of the one or more heating elements may be pulsed between a first temperature during the recovery phases of the gas sensors and a second temperature during the sense phases of the gas sensors, wherein the first temperature is higher than the second temperature. The first temperature may be, for example, set to a value between 150° C. and 300° C., whereas the second temperature may be, for example, set to a value between 50° C. and 200° C.
The temperature modulation could be the same for all sensors or different for at least some of the sensors.
In order to improve repeatability and stability of the sensing results, at least some of the signal samples of each of the gas sensors may represent at least one of the recovery phases and at least one of the sense phases.
According to embodiments of the disclosure a number of the chemo-resistive gas sensors is greater than one, wherein at least some of the chemo-resistive gas sensors have different sensitivities towards one or more of the gases. Such features further improve the accuracy of the sensing results.
According to embodiments of the disclosure the preprocessing processor is configured for executing a baseline calibration algorithm for the signal samples received from the gas sensors. Baseline manipulation is the transformation of a signal sample of one of the gas sensors into a relative resistance change with respect to sensor response to a reference analyte, wherein such sensor response is called a baseline. Synthetic air is a very common baseline as it is easily applicable and realistic in a real world scenario. The purpose of a baseline is to potentially create a more stable and reproducible sensing result by removing some of the drift caused by long term gas exposure and ageing of the sensor. As shown in Equation (1), subtracting the sensor response by its baseline R0 removes additive drift while division removes multiplicative drift. Using both operations combined results in the relative resistance change ΔR/R0:
ΔR/R0=(R−R0/R0 (1).
According to embodiments of the disclosure the preprocessing processor is configured for executing a filtering algorithm for the signal samples received from the gas sensors. The filtering algorithm may, for example, be implemented as a high pass filter or a noise filter. Such features further improve the accuracy of the sensing results.
According to embodiments of the disclosure the feature extraction processor is configured for extracting from the received preprocessed signal samples a normalized sensor sensitivity ΔR/R0 as one of the feature values for each of the gas sensors. The normalized sensor sensitivity ΔR/R0 may be calculated according to Equation (1).
Using the normalized sensor sensitivities ΔR/R0, as feature values improves the accuracy of the sensing results.
According to embodiments of the disclosure the feature extraction processor is configured for extracting from the received preprocessed signal samples a slope R′(t) of one of the preprocessed signal samples as one of the feature values for each of the gas sensors. The slope R′(t) or derivative may be calculated according to Equation (2):
R′(t)=ΔR(t)/Δt (2).
Using the slopes R′(t) as feature values improves the accuracy of the sensing result.
According to embodiments of the disclosure the feature extraction processor is configured for extracting from the received preprocessed signal samples for each of the gas sensors a time correlation between a first of the preprocessed signal samples of the respective gas sensor and a second preprocessed signal sample of the respective gas sensor as one of the feature values for the respective gas sensor.
According to embodiments of the disclosure the feature extraction processor is configured for extracting from the received preprocessed signal samples for each of the gas sensors a spatial correlation between one of the preprocessed signal samples of the respective gas sensor and one of the preprocessed signal sample of another of the gas sensors as one of the feature values for the respective gas sensor.
Given the dynamic behavior of the gas sensors, the availability of several transient in the sensor responses and the characteristic array structure with different functionalizations, it makes sense to introduce metrics which exploits such time and spatial properties. This can be achieved introducing a time autocorrelation function of the normalized sensor responses of the type (and its derivative)
R
τ=Σk=1nxkyk (3)
where x and y indicate the normalized response at different moments in time (or, alternatively, their derivatives) and n is the window size being used to calculate the autocorrelation. Particularly:
Similarly, the correlation among the different gas sensors should also be exploited with a spatial correlation matrix of the type:
According to embodiments of the disclosure the first trained model based algorithm processor is implemented as a first artificial neural network.
According to embodiments of the disclosure the second trained model based algorithm processor is implemented as a second artificial neural network, in particular as a fully connected artificial neural network.
According to embodiments of the disclosure the third trained model based algorithm processor is implemented as a third artificial neural network, in particular as a second gated recurrent unit followed by a fully connected artificial neural network.
An artificial neural network is a parameterized statistic model, in which a number of logistic regressions are combined non-linearly. Such systems “learn” to perform tasks by considering examples, generally without being programmed with any task-specific rules. A neural network is based on a collection of connected nodes called artificial neurons. Each connection can transmit a signal from one artificial neuron to another. An artificial neuron that receives a signal can process it and then signal additional artificial neurons connected to it. A model predefines the structure of the nodes or the hyperparameters of a neural network and the parameters of the connections are found by training the neural network. Structure and the corresponding parameters form a trained model for the respective neural network.
A fully connected artificial neural network is an artificial neural network in which every neuron in one layer is connected to every neuron in the next layer. A gated recurrent unit is recurrent neural network using gating mechanisms.
According to embodiments of the disclosure the discriminative embedding network processor comprises a plurality of first gated recurrent units and a discriminative loss computation processor, which are configured for optimizing parameters, in particular weights and/or offsets, of the first trained model.
According to embodiments of the disclosure the third trained model based algorithm processor is implemented as a third artificial neural network, in particular as a second gated recurrent unit followed by a fully connected artificial neural network.
According to embodiments of the disclosure the classification processor is configured for preventing the quantification processor from creating sensing results, in case the classification processor is unable to create one of the classification value for one of the sets of the embedded feature values.
In a further aspect of the disclosure, a method for operating a gas sensing device for sensing one or more gases in a mixture of gases, which comprises one or more chemo-resistive gas sensors is disclosed. The method comprises the steps of:
using each of the gas sensors for generating signal samples corresponding to concentrations of the one or more gases in the mixture of gases;
using one or more heat sources for heating each of the gas sensors according to one or more temperature profiles;
using a preprocessing processor for receiving the signal samples from each of the gas sensors and for preprocessing the received signal samples in order to generate preprocessed signal samples for each of the gas sensors;
using a feature extraction processor for receiving the preprocessed signal samples and for extracting one or more feature values from the received preprocessed signal samples of each of the gas sensors based on characteristics of the received preprocessed signal samples of the respective gas sensor;
using a discriminative embedding network processor for receiving the sets of feature values and for creating for each of the sets of feature values a set of embedded feature values, wherein the discriminative embedding network processor comprises a first trained model based algorithm processor and a first trained model for the first trained model based algorithm processor, wherein the first trained model is configured for applying a loss function using discriminate weights to the sets of feature values in order to create the sets of embedded features values;
using a classification processor for receiving the sets of embedded feature values and for creating a classification value for each set of the embedded feature values, wherein the classification value indicates a class of the mixture of gases, wherein the classification processor comprises a second trained model based algorithm processor and a second trained model for the second trained model based algorithm processor, wherein the sets of embedded feature values are fed to an input of the second trained model based algorithm processor, wherein the classification values are provided at an output of the second trained model based algorithm processor; and
using a quantification processor configured for receiving the sets of embedded feature values, for receiving the classification values, and for creating for each of the gases a sensing result for each of the sets of embedded feature values, wherein the quantification processor comprises a third trained model based algorithm processor and a plurality of third trained models for the third trained model based algorithm processor, wherein the sets of embedded feature values are fed to an input of the third trained model based algorithm processor, wherein the sensing result are provided at an output of the third trained model based algorithm processor, wherein one third trained model of the plurality of third trained models is selected for creating the sensing results based on the classification values.
Preferred embodiments of the invention are subsequently discussed with respect to the accompanying drawings, in which:
Equal or equivalent elements or elements with equal or equivalent functionality are denoted in the following description by equal or equivalent reference numerals.
In the following description, a plurality of details is set forth to provide a more thorough explanation of embodiments of the present disclosure. However, it will be apparent to those skilled in the art that embodiments of the present disclosure may be practiced without these specific details. In other instances, well-known structures and devices are shown in block diagram form rather than in detail in order to avoid obscuring embodiments of the present disclosure. In addition, features of the different embodiments described hereinafter may be combined with each other, unless specifically noted otherwise.
The gas sensing device 1 is configured for sensing one or more gases in a mixture of gases. The gas sensing device 1 comprises:
one or more chemo-resistive gas sensors 2, wherein each of the gas sensors 2 is configured for generating signal samples SIG corresponding to concentrations of the one or more gases in the mixture of gases;
one or more heat sources 3, wherein the one or more heat sources 3 are controlled in such way that the gas sensors 2 are each heated according to one or more temperature profiles FTP, STP;
a preprocessing processor 4 configured for receiving the signal samples SIG from each of the gas sensors 2 and for preprocessing the received signal samples SIG in order to generate preprocessed signal samples PSS for each of the gas sensors 2;
a feature extraction processor 5 configured for receiving the preprocessed signal samples PSS and for extracting a set of feature values FV from each of the received preprocessed signal samples PSS of the gas sensors 2 based on characteristics of the received preprocessed signal samples PSS of the gas sensors 2;
a discriminative embedding network processor 6 configured for receiving the sets of feature values FV and for creating for each of the sets of feature values FV a set of embedded feature values EV, wherein the discriminative embedding network processor 6 comprises a first trained model based algorithm processor 7 and a first trained model 8 for the first trained model based algorithm processor 7, wherein the first trained model 8 is configured for applying a loss function using discriminate weights to the sets of feature values FV in order to create the sets of embedded features values EV;
a classification processor 9 configured for receiving the sets of embedded feature values EV and for creating a classification value CV for each set of the embedded feature values EV, wherein the classification value CV indicates a class of the mixture of gases, wherein the classification processor 9 comprises a second trained model based algorithm processor 10 and a second trained model ii for the second trained model based algorithm processor 10, wherein the sets of embedded feature values EV are fed to an input 12 of the second trained model based algorithm processor 10, wherein the classification values CV are provided at an output 13 of the second trained model based algorithm processor 10; and a quantification processor 14 configured for receiving the sets of embedded feature values EV and the classification values CV, wherein the quantification processor 14 is configured for creating for each of the gases a sensing result SR for each of the sets of embedded feature values EV, wherein the quantification processor 14 comprises a third trained model based algorithm processor 15 and a plurality of third trained models 16 for the third trained model based algorithm processor 15, wherein the sets of embedded feature values EV are fed to an input 17 of the third trained model based algorithm processor 15, wherein the sensing result SR are provided at an output 18 of the third trained model based algorithm processor 15, wherein one third trained model 16 of the plurality of third trained models 16 is selected for creating the sensing results SR based on the classification values CV.
According to embodiments of the disclosure the preprocessing processor 4 is configured for executing a baseline calibration algorithm for the signal samples SIG received from the gas sensors 2.
According to embodiments of the disclosure the preprocessing processor 4 is configured for executing a filtering algorithm for the signal samples SIG received from the gas sensors 2.
According to embodiments of the disclosure the feature extraction processor 5 is configured for extracting from the received preprocessed signal samples PSS a normalized sensor sensitivity as one of the feature values FV for each of the gas sensors 2.
According to embodiments of the disclosure the feature extraction processor 5 is configured for extracting from the received preprocessed signal samples PSS a slope of one of the preprocessed signal samples PSS as one of the feature values FV for each of the gas sensors 2.
According to embodiments of the disclosure the feature extraction processor 5 is configured for extracting from the received preprocessed signal samples PSS for each of the gas sensors 2 a time correlation between a first of the preprocessed signal samples PSS of the respective gas sensor 2 and a second preprocessed signal sample PSS of the respective gas sensor 2 as one of the feature values FV for the respective gas sensor 2.
According to embodiments of the disclosure the feature extraction processor 5 is configured for extracting from the received preprocessed signal samples PSS for each of the gas sensors 2 a spatial correlation between one of the preprocessed signal samples PSS of the respective gas sensor 2 and one of the preprocessed signal sample PSS of another of the gas sensors 2 as one of the feature values FV for the respective gas sensor 2.
According to embodiments of the disclosure the classification processor 9 is configured for preventing the quantification processor 14 from creating sensing results, in case the classification processor 9 is unable to create one of the classification value CV for one of the sets of the embedded feature values EV.
In a further aspect, the disclosure refers to a method for operating a gas sensing device 1 for sensing one or more gases in a mixture of gases, wherein the gas sensing device 1 comprises one or more chemo-resistive gas sensors 2, wherein the method comprises the steps of:
using each of the gas sensors 2 for generating signal samples SIG corresponding to concentrations of the one or more gases in the mixture of gases;
using one or more heat sources 3 for heating each of the gas sensors 2 according to one or more temperature profiles FTP, STP;
using a preprocessing processor 4 for receiving the signal samples SIG from each of the gas sensors 2 and for preprocessing the received signal samples SIG in order to generate preprocessed signal samples PSS for each of the gas sensors 2;
using a feature extraction processor 5 for receiving the preprocessed signal samples PSS and for extracting one or more feature values FV from the received preprocessed signal samples PSS of each of the gas sensors 2 based on characteristics of the received preprocessed signal samples PSS of the respective gas sensor 2;
using a discriminative embedding network processor 6 for receiving the sets of feature values FV and for creating for each of the sets of feature values FV a set of embedded feature values EV, wherein the discriminative embedding network processor 6 comprises a first trained model based algorithm processor 7 and a first trained model 8 for the first trained model based algorithm processor 7, wherein the first trained model 8 is configured for applying a loss function using discriminate weights to the sets of feature values FV in order to create the sets of embedded features values EV;
using a classification processor 9 for receiving the sets of embedded feature values EV and for creating a classification value CV for each set of the embedded feature values EV, wherein the classification value CV indicates a class of the mixture of gases, wherein the classification processor 9 comprises a second trained model based algorithm processor 10 and a second trained model ii for the second trained model based algorithm processor 10, wherein the sets of embedded feature values EV are fed to an input 12 of the second trained model based algorithm processor 10, wherein the classification values CV are provided at an output 13 of the second trained model based algorithm processor 10; and
using a quantification processor 14 configured for receiving the sets of embedded feature values EV, for receiving the classification values CV, and for creating for each of the gases a sensing result SR for each of the sets of embedded feature values EV, wherein the quantification processor 14 comprises a third trained model based algorithm processor 15 and a plurality of third trained models 16 for the third trained model based algorithm processor 15, wherein the sets of embedded feature values EV are fed to an input 17 of the third trained model based algorithm processor 15, wherein the sensing result SR are provided at an output 18 of the third trained model based algorithm processor 15, wherein one third trained model 16 of the plurality of third trained models 16 is selected for creating the sensing results SR based on the classification values CV.
After the preprocessing processor 4 and feature extraction processor 5 a discriminative embedding processor 6 is introduced which embeds the sets of feature values FV into a new space where the embedded feature values EV have better separability. The discriminative embedding processor 6 may be implemented as a neural network with a contrastive ‘distance-learning’ loss function such that, after transformation, embedded feature values EV belonging to the same gas or gas mixture types are closer to each other and those belonging to different gas types or gas mixtures are further apart. An example is provided in
It is sufficient to train the discriminative embedding network processor 6 on a selection of representative mixtures of gases. Afterwards, the discriminative embedding network processor 6 shall be capable of discriminating also unseen gases or use-cases.
Depending on the target application of the product, certain target gases or mixtures of gases have to be identified. For example, it was observed that, in a mixture, nitrogen dioxide is masked by ozone and, as such, nitrogen dioxide cannot—and shall not—be estimated in the presence of ozone. In this specific case, we need to identify three classes: specifically, nitrogen dioxide only, air, or ozone dominated gas mixture. Thresholds may be used to identify the different categories. For instance, ‘nitrogen dioxide only’ means nitrogen dioxide >1 ppb and ozone <10 ppb, ‘ozone-mixture’ corresponds to ozone >=10 ppb and ‘Air’ means ozone <10 ppb and nitrogen dioxide <1 ppb, and so on.
For other products/applications, a mixture of gases can still come into play (for example, a mixture of ozone and carbon monoxide) and, as such, this class and respective weights will have to be enabled at the classification processor 9.
Thanks to the initial discriminative feature embedding, the classification processor 9 can quite reliably identify, for each set of embedded feature values EV over time, the relevant class or scenario and, as such, recommend a certain third trained model 16 (regressor model) to the quantification processor 14, which eventually quantifies the concentration of the gases of interest for the corresponding time sample.
Furthermore, the classification processor 9 can also be equipped with an additional output, which detects an unknown gas for which then no concentration shall be estimated since no third trained model 16 is available for it. A possible implementation is shown in
The quantification processor 14 may apply O3-mixture and NO2-only weights from a specific third trained model 16 to the incoming sets of embedded feature values EV. In other applications, where the gas sensing device 1 is selective to more gases and more than one gas concentration has to be estimated at the same time, then gas mixture weights from another third trained model 16 will be applied.
Thanks to the initial discriminative feature embedding and to the presence of the classification processor 9, which provides the quantification processor 14 with additional information on the specific scenario, the quantification processor 14 can be greatly simplified and one can resort to simpler processing steps. A possible implementation is shown in
According to embodiments of the disclosure the first trained model based algorithm processor 6 is implemented as a first artificial neural network.
According to embodiments of the disclosure the discriminative embedding network processor 6 comprises a plurality of first gated recurrent units 19 and a discriminative loss computation processor 20, which are configured for optimizing parameters, in particular weights and/or offsets, of the first trained model 8.
The discriminative embedding network 6 may be implemented as a triplet gated recurrent unit network with triplet loss defined as
Loss=max(D+−D−α,0.0)
where
D
+=Σ(GRU(F)−GRU(F+))2
D
−=Σ(GRU(F)−GRU(F−))2
and α is a parameter such that 0<α<1. The training strategy is shown in
The recurrent nature of the triplet approach above achieves superior performance when dealing with time series data from low cost gas sensors 2. The difficulty here is that feature values FV from the past need to be included, since instantaneous feature values FV are not sufficient to correctly identify and learn the dynamics of the signal samples SIG.
Furthermore, thanks to the discriminative embedding network processor 6 in
According to embodiments of the disclosure the second trained model based algorithm processor 13 is implemented as a second artificial neural network, in particular as a fully connected artificial neural network.
According to embodiments of the disclosure the second trained model based algorithm processor 10 is implemented as an incremental linear discriminant analysis processor (not shown).
In its simplest form, the classification processor 9 could be implemented as a fully connected (FC) neural network with weights adjusted to the target output of the gas sensing device 1. As illustrated in
P
cl(t)=softmax(WciFem(t)+bcl)
For each temporal set of embedded feature values EV, the classification processor 9 assigns a value Pcl(t) of the classification value CV and passes this information to the quantification processor 14. For example, Pcl (t) could be chosen out of the set ‘NO2-only’, O3-Mixture’ or ‘Air’. Alternatively, a linear discriminant analysis classifier could also replace the fully connected layer in
According to embodiments of the disclosure the third trained model based algorithm processor 15 is implemented as a third artificial neural network, in particular as a second gated recurrent unit 21 followed by a fully connected artificial neural network 22.
Based on the recommendation of the classification processor 9, the appropriate pretrained set of weights and offsets, Wrg and brg, may be selected for second gated recurrent unit 21 and the second gated recurrent unit 21, applied to the same values Fem(t) of the embedded feature values EV used at the classification processor 9 and one or multiple gas predictions, for example in ppb/ppm, are finally delivered as illustrated in
If an unknown gas is found or if the classification processor 9 is not sure about its own output decision, which can be determined by the differences among class probabilities not exceeding a certain threshold, then the quantification processor 14 output may be put on standby.
Thanks to the improved embedded representation, simpler 1-D (single gas) weights/filters can be applied at the regressor as an alternative to complex 2D weights/filters learnt in gas mixture scenarios.
The upper chart shows the values “Pred NO2” of the sensing results SRI for a first gas, which is in this example nitrogen oxide, and the true values “True NO2” for the first gas over time.
The chart in the middle shows the values “Pred O3” of the sensing results SR2 for a second gas, which is in this example ozone, and the true values “True O3” for the second gas over time.
The lower chart shows the values “prediction” of the classification values CV and the true values “label” of the scenario over time.
According to embodiments of the disclosure a number of the chemo-resistive gas sensors 2 is greater than one, wherein at least some of the chemo-resistive gas sensors 2 have different sensitivities towards one or more of the gases.
Each sensor 2.1, 2.2, 2.3 and 2.4 in the array is heated by a heat source 3, whose temperature is being pulsed between first temperature T1 during a recovery phase and a second temperature T2 during sense phase (see Figure ii). In other embodiments, the sensors 2.1, 2.2, 2.3 and 2.4 in the array are heated by a plurality of heat sources 3. For example, each of the sensors 2.1, 2.2, 2.3 and 2.4 could be heated individually by one heat source of the plurality of the heat sources. The result of these controlled temperature oscillations is a more dynamic behavior of the signal samples SIG1, SIG2, SIG3, SIG4 as shown in
Several implementations of temperature pulsing mechanism are possible. For example, the temperature modulation could be the same for all sensors 2.1, 2.2, 2.3 and 2.4 or different in order to better exploit the different functionalizations of the base material and to improve gas separability. Similarly, multiple heater controls can be used (one for each sensor 2.1, 2.2, 2.3 and 2.4) or, alternatively, a single heater control in time division multiplexing with different applied voltages so as to obtain sensor specific temperature values.
The sensors 2.1, 2.2, 2.3 and 2.4 form a multi-gas sensor array, where a base material consisting of graphene is functionalized with different chemicals (e.g., Pd, Pt, and MnO2) for dissimilar selectivity. The interaction between graphene sheets and absorbed gas analytes would influence the electronic structure of the material, resulting in altered charge carrier concentrations and changed electrical conductance. Meanwhile, due to different sensitivity towards various gas molecules resistances of the sensors 2.1, 2.2, 2.3 and 2.4 also change in disparate patterns, making it possible to analyze complicated gas mixtures with one single sensor array.
According to embodiments of the disclosure the one or more gas sensors 2 are alternately operated in recovery phases RP and in sense phases SP;
wherein the one or more heat sources 3 are controlled in such way that the gas sensors 2 are each heated according to one or more first temperature profiles FTP of the one or more temperature profiles FTP, STP during the recovery phases RP and according to one or more second temperature profiles STP of the one or more temperature profiles FTP, STP during the sense phases SP, wherein for each of the gas sensors 2 a maximum temperature of the respective first temperature profile FTP is higher than a maximum temperature of the respective second temperature profile STP.
In the particular example of
Although some aspects have been described in the context of an apparatus, it is clear that these aspects also represent a description of the corresponding method, where a block or device corresponds to a method step or a feature of a method step. Analogously, aspects described in the context of a method step also represent a description of a corresponding block or item or feature of a corresponding apparatus.
The above described is merely illustrative, and it is understood that modifications and variations of the arrangements and the details described herein will be apparent to others skilled in the art. It is the intent, therefore, to be limited only by the scope of the impending claims and not by the specific details presented by way of description and explanation above.
| Number | Date | Country | Kind |
|---|---|---|---|
| 21202164 | Oct 2021 | EP | regional |
