Gas sensing device for sensing gas in a gas mixture and method of operating the same

Abstract

The present disclosure relates to a gas sensing device for sensing a gas in a gas mixture and a method of operating the same. The gas sensing device comprises: a photoacoustic spectrometry device comprising a radiator configured to emit light and comprising a gas detection chamber, the gas detection chamber being configured to expose the gas mixture to the light, the photoacoustic spectrometry device comprising a microphone configured to detect sounds in the detection chamber caused by exposing the gas mixture to the light, and the photoacoustic spectrometry device being configured to generate a signal sample corresponding to the concentration of the gas in the gas mixture based on the sounds detected by the microphone; and a computing device configured to receive the signal samples, the computing device comprising a feature extraction block, the feature extraction block being configured to calculate representations of the signal samples, such that one of the representations is calculated for each signal sample, wherein each representation comprises one or more feature values, each of the one or more feature values ​​referring to a characteristic of the corresponding signal sample.

Description

Technical Field

[0001] Embodiments relate to a gas sensing device for sensing a gas in a gas mixture. Further embodiments relate to a method for operating such a gas sensing device. More specifically, the present disclosure deals with the estimation of the concentration of a gas by using a photoacoustic spectrometry device. Background Art

[0002] Several gas sensors with different operating mechanisms (eg MOS, infrared, electrochemical, etc.), different form factors (portable devices or fixed devices) and different target applications (medical, industrial, automotive, mobile phones, etc.) are known in the art.

[0003] Acoustic coupling represents a problem for these sensors that smartphones use to detect pressure changes that are proportional to the amount of concentration of the gas present in the measurement volume. This is the case for so-called "open" non-resonant photoacoustic gas sensors, where a microphone is placed in the measurement volume and the system operates at relatively low frequencies (<100 Hz).

[0004] So-called "closed" photoacoustic sensors experience less significant degradation due to acoustic noise because the microphone is hermetically enclosed in the detector housing and is therefore at least partially isolated from the external noise environment. Similarly, resonant open photoacoustic gas sensors (operating at very high frequencies and having a very narrow spectrum) are less susceptible to degradation due to acoustic noise. However, closed photoacoustic sensors and resonant open photoacoustic gas sensors are bulky, costly, and unsuitable for high-end applications requiring high levels of sensitivity, wide measurement ranges, and fast temporal responses. Summary of the invention

[0005] A gas sensing device for sensing a gas in a gas mixture is provided. The gas sensing device comprises:

[0006] a photoacoustic spectrometry apparatus, wherein the photoacoustic spectrometry apparatus comprises a radiator configured to emit light, wherein the photoacoustic spectrometry apparatus comprises a gas detection chamber configured to expose a gas mixture to the light, wherein the photoacoustic spectrometry apparatus comprises a microphone configured to detect sounds in the detection chamber, the sounds caused by the exposure of the gas mixture to the light, and wherein the photoacoustic spectrometry apparatus is configured to generate a signal sample corresponding to a concentration of a gas in the gas mixture based on the sounds detected by the microphone; and

[0007] A computing device configured to receive signal samples, wherein the computing device comprises a feature extraction block configured to compute representations for the signal samples such that a representation of the representations is computed for each of the signal samples, wherein each of the one or more feature values ​​refers to a characteristic of the corresponding signal sample,

[0008] wherein the computing device comprises a decision making block comprising a trained model-based algorithm block having a plurality of inputs and at least one output, wherein the decision making block comprises one or more training models for the trained model-based algorithm block, wherein each of the feature values ​​of one of the representations is input into one of the inputs of the trained model-based algorithm block such that each of the feature values ​​is fed into an individual input of the input, wherein the decision making block creates a sensing result based on an output value of at least one output of the trained model-based algorithm block, wherein the output value is created by using at least one of the one or more training models at the decision making block such that the output value depends on a signal sample of a photoacoustic spectrometry device.

[0009] The photoacoustic spectrometry device is configured to measure the effects of absorbed electromagnetic energy, in particular the effects of light on matter by means of acoustic detection. It can be used to accurately measure the concentration of various gases due to its ability to evaluate a sample of an object in situ. The photoacoustic spectrometry device can be of an open type or a closed type. It can be a resonant photoacoustic spectrometry device or a non-resonant photoacoustic spectrometry device.

[0010] The radiator can generally be configured to emit light of a specific wavelength that can only be absorbed by a specific gas, so that the specific gas can be detected. The radiator can include a laser or any other light source. In addition, the radiator can include a filter to filter the generated light in such a way that the wavelength of the emitted light corresponds to the desired specific wavelength. The specific wavelength can be in the range of visible light or in the range of infrared light. The emitted light can be coherent or incoherent.

[0011] The signal sample is a sequence of time-discrete signal values, wherein the signal value is output by one of the gas sensors.

[0012] The feature extraction block may be configured to transform the signal samples into a representation, wherein the representation is based on characteristics of the signal samples. Some of the eigenvalues ​​of the representation may correspond to an absolute value of a mean deviation of the corresponding signal samples, a phase of the mean deviation of the corresponding signal samples, an absolute value of a standard deviation of the corresponding signal samples, or a phase of a standard deviation of the corresponding signal samples.

[0013] The decision making block receives a representation during an operation phase, wherein the decision making block includes a trained model-based algorithm block and one or more training models for the trained model-based algorithm block.

[0014] The algorithm block based on the training model is a processing stage that enables machine learning. Machine learning is done in a pre-operational training stage, where the training model is developed by comparing the actual output values ​​of the algorithm block based on the training model with the expected output values ​​of the algorithm block based on the training model for the defined inputs of the algorithm stage based on the training model. The training model has a predefined structure, where the equalization of the predefined structure is done during the training stage. After the training stage is over, the training model includes the learning content. In the operational stage for generating sensing results, one or more training models from the training stage are used to process the representations from the feature extraction block.

[0015] During the training phase, multiple training models can be built and then stored at the decision making block. The training models can differ in structure and / or parameters. During the operation phase, the most appropriate training model can be selected according to the specific use case.

[0016] The trained model-based algorithm block includes multiple inputs and at least one output, wherein each feature value of one of the representations is input to one of the inputs of the trained model-based algorithm block, such that each of the feature values ​​is fed into an individual input of the input, wherein a decision making block creates a sensing result based on an output value of at least one output of the trained model-based algorithm block, wherein the output value is created by using at least one of the one or more training models at the decision making block, such that the output value depends on a signal sample of the photoacoustic spectrometry device.

[0017] The decision making block provides a decision on either the classification of the concentration of the gas detected by the photoacoustic spectrometry device or the continuous measurement of the concentration of the gas detected by the photoacoustic spectrometry device. In the first case, a training model trained as a classification algorithm is used, and the sensed result is an alphanumeric term, such as "high" or "low". In the latter case, a training model trained as a regression algorithm is used, and the sensed result is a physical quantity, such as "4% by volume".

[0018] The gas sensing device according to the present disclosure addresses the intrinsic susceptibility of acoustic noise of a photoacoustic spectrometry device. It uses a robust algorithm and detection mechanism that can cope with calibration inaccuracies, drift, and other similar effects reliably and over a wide operating range.

[0019] The proposed gas sensing device provides an end-to-end solution of a gas sensor based on a photoacoustic spectrometry device, which is versatile, widely applicable to multiple applications and use cases (outdoor, indoor, health check, etc.) and can be embedded in smart portable devices. In particular, the gas sensing device can be used for air quality monitoring. Specifically, an algorithm is used, which acts on continuous sensor readings, which exploits features in the sensor response and exhibits low complexity and limited memory requirements.

[0020] The gas sensing device can reflect real-world scenarios where, for example, there are gas mixtures that lead to cross-sensitivity in the sensor response. In addition, the gas sensing device only requires a short time to reach a stable response level.

[0021] Gas sensing devices have low material costs and use robust and economical concrete mechanisms to be embedded in mass-produced consumer electronics (such as mobile phones) while delivering good continuous predictive performance in complex real-world scenarios, and therefore have to deal with challenges related to the availability of limited and noisy datasets, imperfect initial calibration, gas mixtures with varying analyte concentrations, modeling errors, etc.

[0022] Relying on a nonlinear activation function, the gas sensing device according to the present disclosure overcomes the limitation of the state of the prior art sensor relying on a linear regression fitting method. The proposed method is able to model the nonlinear dependency between the measured metric and the parameter to be estimated. Therefore, the sensor performance is significantly improved.

[0023] The gas sensing device according to the present disclosure is suitable for a wide range of applications, including high-end applications requiring a high level of accuracy, a high level of sensitivity, a wide measurement range, and a fast time response.

[0024] According to an embodiment of the present disclosure, a computing device includes a pre-processing block, wherein the pre-processing block is configured to receive signal samples from a photoacoustic spectrometry device, wherein the pre-processing block is configured to generate a pre-processed signal sample for each of the signal samples, and wherein the pre-processing block is configured to forward the pre-processed signal samples to a feature extraction block.

[0025] In such embodiments, the feature extraction block may compute representations for the signal samples by processing the pre-processed signal samples.

[0026] According to an embodiment of the present disclosure, the pre-processing block includes a noise suppression block, which is configured to suppress noise in the signal samples (especially noise caused by environmental noise) so that the pre-processed signal samples include noise-reduced signal samples with less noise than the corresponding signal samples.

[0027] By means of such features, the performance of the gas sensing device can be further improved.

[0028] According to an embodiment of the present disclosure, the algorithm block based on the training model includes a neural network using one or more training models and / or a random decision forest using one or more training models.

[0029] A random decision forest is a learning method for classification, regression, and other tasks that operates by building multiple decision trees at training time and outputting a class that is either the class (classification) or the mode of the average predictions (regression) of the individual trees.

[0030] Artificial neural networks are parameterized statistical models in which a number of nonlinear combinations are combined nonlinearly. Such systems "learn" to perform tasks by considering examples, usually without being programmed with rules for any specific task. Neural networks are based on a collection of connected nodes called artificial neurons. Each connection can transmit a signal from one artificial neuron to another. The artificial neuron that receives the signal can process it and subsequently signal additional artificial neurons connected to it. The structure or hyperparameters of the nodes of the neural network are predefined by the model, and the parameters of the connections are found by training the neural network. The structure and the corresponding parameters form a training model for the corresponding neural network.

[0031] The neural network may be, for example, a fully connected feed-forward neural network with one hidden layer.

[0032] According to an embodiment of the present disclosure, the pre-processing block comprises a domain transform block configured to transform the signal samples into a logarithmic frequency domain so as to obtain a logarithmic spectrum having a plurality of frequency bands for each of the signal samples,

[0033] The noise suppression block includes an algorithm block based on another training model, the algorithm block based on another training model having multiple inputs and multiple outputs,

[0034] wherein the noise suppression block comprises one or more further trained models for an algorithm block based on further trained models, wherein for each frequency band of the logarithmic spectrum of one of the signal samples, an amplitude value and a phase value are input to one of the inputs of the further algorithm block such that each of the amplitude values ​​and each of the phase values ​​is fed into an individual input of the input,

[0035] wherein the noise suppression block creates noise-reduced signal samples based on output values ​​of outputs of the algorithm blocks of the further training model,

[0036] wherein the output value of the algorithm block of the further training model is created by using at least one further training model of the one or more further training models at the noise suppression block, and

[0037] Each of the output values ​​of the algorithm block of the additional training model is a noise-reduced amplitude value or a noise-reduced phase value of a noise-reduced frequency band of a noise-reduced logarithmic spectrum of one of the preprocessed signal samples.

[0038] The transformation of the signal samples from the time domain to the logarithmic frequency domain results in a reduction in the amount of data that must be processed in the following processing blocks. In particular, it simplifies the processing of noise with different spectral characteristics.

[0039] The general remarks about the algorithmic block based on the training model of the training model and the decision making block are also valid for the algorithmic block based on the further training model of the further training model and the noise suppression block. However, the training model and the further training model must be established separately. In the latter case, during the training phase, the microphone output is used with and without acoustic noise. The algorithmic block of the further training model can be implemented as a denoising autoencoder. In particular, such an embodiment of the noise suppression block is most suitable for the case where the photoacoustic spectrometry device is of a closed type, because closed photoacoustic spectrometry devices usually include a mechanism for flowing a gas mixture into and out of the sensing chamber, which can generate acoustic noise.

[0040] According to an embodiment of the present disclosure, the algorithm block based on the additional training model includes an additional neural network using one or more additional training models and / or an additional random decision forest using one or more additional training models.

[0041] The further neural network may be, for example, a fully connected feed-forward neural network with one hidden layer.

[0042] According to an embodiment of the invention, the pre-processing block comprises a domain transform block configured for computing a logarithmic spectrum with a plurality of frequency bands for each signal sample.

[0043] The transformation of the signal samples from the time domain to the logarithmic frequency domain results in a reduction in the amount of data that has to be processed in the following processing blocks, in which case the decision making blocks do not include algorithm blocks based on an additional training model.

[0044] According to an embodiment of the present disclosure, the noise suppression block includes a bandpass filter or a lowpass filter such that the pre-processed signal samples are based on reduced-bandwidth signal samples having a lower bandwidth than corresponding signal samples.

[0045] A bandpass filter is a filter that passes frequencies within a certain frequency range and attenuates frequencies outside the frequency range. A lowpass filter is a filter that passes signals having frequencies below a selected cutoff frequency and attenuates signals having frequencies above the cutoff frequency.

[0046] In cases where it is not possible to access the microphone output without noise, or if separate training phases for training the model and further training the model are desired for complexity reasons, the use of frequency filters is preferred.

[0047] According to an embodiment of the present disclosure, a gas sensing device comprises one or more auxiliary sensors, wherein each of the auxiliary sensors is configured to generate an auxiliary signal sample corresponding to a physical quantity of a gas mixture;

[0048] One or more of the auxiliary sensors include

[0049] a temperature sensor for generating a first auxiliary signal sample corresponding to an auxiliary signal sample of the temperature of the gas mixture, and / or

[0050] a pressure sensor for generating a second auxiliary signal sample corresponding to the auxiliary signal sample of the pressure of the gas mixture, and / or

[0051] A humidity sensor is configured to generate a third auxiliary signal sample corresponding to the auxiliary signal sample of the humidity of the gas mixture.

[0052] By means of these features, the accuracy of the gas sensing device can be further increased.

[0053] According to an embodiment of the present disclosure, the decision making block is configured to select one or more selected training models from one or more training models based on auxiliary signal samples of one or more auxiliary sensors, wherein an output value of at least one output of the algorithm is created by using the one or more selected training models.

[0054] For example, the most appropriate training model may be selected based on humidity values ​​obtained from a humidity sensor, since humidity has an effect on the sensor response.

[0055] By means of these features, the accuracy of the gas sensing device can be further increased.

[0056] According to an embodiment of the present disclosure, the decision making block is configured to select one or more selected training models based on spectral information of the signal samples.

[0057] Spectral portions of the signal samples (e.g., very low frequencies) may be correlated to environmental parameters of the gas mixture (such as pressure, temperature, or humidity). Thus, such features may replace measurement data from auxiliary sensors, such that the need to have auxiliary sensors measuring such environmental parameters may be eliminated.

[0058] According to an embodiment of the present invention, the feature extraction block is configured to calculate the auxiliary representation of the auxiliary signal samples, such that for each of the auxiliary signal samples, one of the auxiliary representations is calculated, wherein each of the auxiliary representations comprises one or more auxiliary feature values, wherein each of the one or more auxiliary feature values ​​refers to a feature of the corresponding auxiliary signal sample; and

[0059] wherein each of the auxiliary feature values ​​of one of the auxiliary representations is input into an input of an input of an algorithm block based on a training model that is not used for the input feature value, such that the auxiliary feature values ​​of the auxiliary feature values ​​are fed into individual inputs of the input, wherein an output value of the algorithm block based on the training model is created such that the output value depends on the auxiliary signal sample.

[0060] Some of the represented auxiliary eigenvalues ​​may correspond to an absolute value of a mean deviation of corresponding auxiliary signal samples, a phase of a mean deviation of corresponding auxiliary signal samples, an absolute value of a standard deviation of corresponding auxiliary signal samples or a phase of a standard deviation of corresponding auxiliary signal samples.

[0061] By means of these features, the accuracy of the gas sensing device can be further increased.

[0062] According to an embodiment of the present disclosure, the detection chamber comprises one or more ventilation openings which are permanently opened during an operating phase of the gas sensing device.

[0063] The permanently open ventilation opening allows a permanent exchange between the gas mixture in the detection chamber and the gas environment mixture, so that changes in the concentration of the gas to be detected in the gas environment mixture can be detected without delay. A photoacoustic spectrometry device with such a ventilation opening can also be referred to as an open photoacoustic spectrometry device.

[0064] Another aspect of the invention relates to a method for operating a gas sensing device for sensing a gas in a gas mixture, wherein the gas sensing device comprises a photoacoustic spectrometry device, wherein the photoacoustic spectrometry device comprises a radiator configured for emitting light, wherein the photoacoustic spectrometry device comprises a gas detection chamber configured for exposing the gas mixture to the light, and wherein the photoacoustic spectrometry device comprises a microphone configured for detecting sounds in the detection chamber, the sounds being caused by exposing the gas mixture to the light, and wherein the gas sensing device comprises a computing device comprising a feature extraction block and a decision making block, wherein the decision making block comprises a trained model based algorithm block having a plurality of inputs and at least one output, wherein the decision making block comprises one or more training models for the trained model based algorithm block, wherein the method comprises the following steps:

[0065] generating, using a photoacoustic spectrometry device, a signal sample corresponding to a concentration of a gas in the gas mixture based on the sound detected by the microphone;

[0066] Receiving the signal sample using a computing device, wherein the computing device includes a feature extraction block and a decision making block, wherein the decision making block includes a trained model-based algorithm block having a plurality of inputs and at least one output, wherein the decision making block includes one or more trained models for the algorithm block;

[0067] computing representations for the signal samples using a feature extraction block such that a representation of the representations is computed for each of the signal samples, wherein each of the representations comprises one or more feature values, wherein each feature value of the one or more feature values ​​refers to a feature of the corresponding signal sample;

[0068] inputting each of the feature values ​​of one of the representations into one of the inputs of the trained model based algorithm block such that each of the feature values ​​is fed into an individual one of the inputs; and

[0069] A sensing result is created using a decision making block based on an output value of at least one output of the trained model-based algorithm block, wherein the output value is created by using at least one of the one or more training models at the decision making block such that the output value depends on a signal sample of the photoacoustic spectrometry device. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Embodiments of the present invention are described herein with reference to the accompanying drawings.

[0071] Figure 1 A schematic diagram showing a first embodiment of a gas sensing device according to the present disclosure is shown;

[0072] Figure 2 shows a schematic diagram of a second embodiment of a gas sensing device according to the present disclosure;

[0073] Figure 3 shows a schematic diagram of a first pre-processing block for a gas sensing device according to the present disclosure;

[0074] Figure 4 shows a schematic diagram of a second pre-processing block for a gas sensing device according to the present disclosure;

[0075] Figure 5 shows a schematic diagram of a third embodiment of a gas sensing device according to the present disclosure;

[0076] Figure 6A general method of training a training model for a gas sensing device according to the present disclosure is shown;

[0077] Figure 7 A general method of operating a gas sensing device according to the present disclosure is shown;

[0078] Figure 8 A more specific method of training a training model for a training model-based algorithm block for a gas sensing device according to the present invention is described;

[0079] Fig. 9 A more specific method of operating a gas sensing device according to the present disclosure is shown;

[0080] Fig.10 A general method of training an additional training model for an algorithm block based on an additional training model for a noise suppression block of a gas sensing device according to the present disclosure is shown;

[0081] Fig.11 A general method of operating a first pre-processing block for a gas sensing device according to the present disclosure is shown;

[0082] Fig.12 A general method of operating a second pre-processing block for a gas sensing device according to the present disclosure is shown;

[0083] Fig.13 shows errors generated by the presence of recorded speech at a gas sensing device according to the prior art;

[0084] Fig.14 shows errors generated by the presence of recorded music by Richard Strauss at a gas sensing device according to the prior art;

[0085] Fig.15 A training phase for computing a training model is shown; and

[0086] Fig.16 The operating stages of a gas sensing device according to the present disclosure are shown.

[0087] In the following description equal or equivalent elements or elements having equal or equivalent functionality are denoted by equal or equivalent reference numerals. DETAILED DESCRIPTION

[0088] In the following description, a number of details are set forth to provide a more thorough explanation of embodiments of the present invention. However, it will be apparent to those skilled in the art that embodiments of the present invention may be practiced without these specific details. In other examples, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the present invention. Additionally, the features of the different embodiments described below may be combined with each other unless otherwise specifically noted.

[0089] Figure 1 A schematic diagram of a first embodiment of a gas sensing device 1 for sensing a gas in a gas mixture MG according to the present disclosure is shown.

[0090] The gas sensing device 1 comprises:

[0091] a photoacoustic spectrometry device 2, wherein the photoacoustic spectrometry device 2 comprises a radiator 3 configured for emitting light CL, wherein the photoacoustic spectrometry device 2 comprises a gas detection chamber 4 configured for exposing the gas mixture MG to the light CL, wherein the photoacoustic spectrometry device 2 comprises a microphone 5 configured for detecting a sound SO in the gas detection chamber 4, the sound being caused by the exposure of the gas mixture MG to the light CL, and wherein the photoacoustic spectrometry device 1 is configured for generating a signal sample SIG corresponding to the concentration of the gas in the gas mixture MG based on the sound detected by the microphone 5; and

[0092] a computing device 6 configured for receiving signal samples SIG, wherein the computing device 6 comprises a feature extraction block 7 configured for calculating representations REP for the signal samples SIG, such that for each of the signal samples SIG a representation REP of the representations REP is calculated, wherein each of the representations REP comprises one or more feature values ​​FV, ​​wherein each of the one or more feature values ​​FV refers to a characteristic of the respective signal sample SIG,

[0093] Wherein the computing device 6 includes a decision making block 8, which includes a training model-based algorithm block 9 having multiple inputs 10 and at least one output 11, wherein the decision making block 8 includes one or more training models TM for the training model-based algorithm block 9, wherein each of the feature values ​​FV representing one of the REPs is input into one of the inputs 10 of the training model-based algorithm block 9, so that each of the feature values ​​FV is fed into an individual input 10 of the input 10, wherein the decision making block 8 creates a sensing result SR based on an output value OV of at least one output 11 of the training model-based algorithm block 9, wherein the output value OV is created by using at least one of the one or more training models TM at the decision making block 8, so that the output value OV depends on the signal sample SIG of the photoacoustic spectrometry device 2.

[0094] exist Figure 1 In the example of , each of the representations REP consists of three eigenvalues ​​FV1, FV2 and FV3. The eigenvalue FV1 is fed to the input 10.1, the eigenvalue FV2 is fed to the input 10.2 and the eigenvalue FV3 is fed to the input 10.3.

[0095] The output value OV available at the unique output 11 is used as the sensing result SR. In other embodiments, the sensing result SR is derived from the output value OV by a post-processing block.

[0096] Another aspect of the present disclosure relates to a method for operating a gas sensing device 1 for sensing a gas in a gas mixture MG, wherein the gas sensing device 1 comprises a photoacoustic spectrometry device 2, wherein the photoacoustic spectrometry device 2 comprises a radiator 3 configured for emitting light CL, wherein the photoacoustic spectrometry device 2 comprises a gas detection chamber 4 configured for exposing the gas mixture MG to the light CL, and wherein the photoacoustic spectrometry device 2 comprises a microphone 5 configured for detecting a sound SO in the detection chamber 4 caused by exposing the gas mixture MG to the light CL, and wherein the gas sensing device 1 comprises a computing device 6 comprising a feature extraction block 7 and a decision making block 8, wherein the decision making block 8 comprises a trained model based algorithm block 9 having a plurality of inputs 10 and at least one output 11, wherein the decision making block 8 comprises one or more training models TM for the trained model based algorithm block 9, wherein the method comprises the following steps:

[0097] Based on the sound SO detected by the microphone 5, the photoacoustic spectrometry device 2 is used to generate a signal sample SIG corresponding to the concentration of the gas in the gas mixture MG;

[0098] Using a computing device 6 to receive a signal sample SIG;

[0099] using a feature extraction block 7 to calculate a representation REP for the signal samples SIG such that for each of the signal samples SIG one of the representations REP is calculated, wherein each of the representations REP comprises one or more feature values ​​FV, ​​wherein each of the one or more feature values ​​FV refers to a characteristic of the respective signal sample SIG;

[0100] inputting each of the feature values ​​FV representing one of the REPs into one of the inputs 10 of the trained model based algorithm block 9 such that each of the feature values ​​FV is fed into an individual one of the inputs 10; and

[0101] The sensing result SR is created using the decision making block 8 based on the output value OV of at least one output 11 of the algorithm block 8 of the training model, wherein the output value OV is created by using at least one of the one or more training models TM at the decision making block 8, so that the output value OV depends on the signal sample SIG of the photoacoustic spectrometry device 2.

[0102] Figure 2 A schematic diagram of a second embodiment of a gas sensing device 1 according to the present disclosure is shown. The second embodiment is based on the first embodiment. In the following, only additional features are explained.

[0103] According to an embodiment of the present disclosure, the computing device 6 includes a pre-processing block 12, wherein the pre-processing block 12 is configured to receive a signal sample SIG from the photoacoustic spectrometry device 2, wherein the pre-processing block 12 is configured to generate a pre-processed signal sample PSIG for each signal sample SIG, and wherein the pre-processing block 12 is configured to forward the pre-processed signal sample PSIG to the feature extraction block 7.

[0104] In such an embodiment, each of the signal samples SIG is converted into a preprocessed signal sample PSIG by the preprocessing block 12. Each of the preprocessed signal samples PSIG is fed to a feature extraction block 7 which extracts feature values ​​FV1, FV2 and FV3 which refer to characteristics of the corresponding signal sample SIG.

[0105] Figure 3 A schematic diagram of a first pre-processing block 12 for a gas sensing device 1 according to the present disclosure is shown.

[0106] According to an embodiment of the present disclosure, the pre-processing block 12 comprises a noise suppression block 13 configured for suppressing noise NOI in the signal samples SIG so that the pre-processed signal samples PSIG comprise noise reduced signal samples NSIG having less noise NOI than corresponding signal samples SIG.

[0107] According to an embodiment of the present disclosure, the pre-processing block comprises a domain transform block 14 configured to transform the signal samples SIG into a logarithmic frequency domain so as to obtain a logarithmic spectrum LOS having a plurality of frequency bands FB for each of the signal samples SIG,

[0108] The noise suppression block 13 includes an algorithm block 15 based on another training model. The algorithm block 15 based on another training model has multiple inputs 16 and multiple outputs 17.

[0109] wherein the noise suppression block 13 comprises one or more further training models FTM for an algorithm block 15 based on further training models, wherein for each frequency band FB of the logarithmic spectrum LOS of one of the signal samples SIG, an amplitude value AV and a phase value PV are input to one of the inputs 16 of the further algorithm block 15, whereby each of the amplitude values ​​AV and each of the phase values ​​PV are fed into a respective input of the input 16,

[0110] wherein the noise suppression block 13 creates noise reduced signal samples NSIG based on the output values ​​NOV of the output 17 of the algorithm block 15 of the further training model,

[0111] wherein the output value NOV of the further trained model based algorithm block 15 is created by using at least one further trained model of the one or more further trained models FTM at the noise suppression block 15, and

[0112] Therein each of the output values ​​NOV of the further trained model based algorithm block 15 is a noise reduced amplitude value NAV or a noise reduced phase value NPV of a noise reduced frequency band NFB of a noise reduced logarithmic spectrum NLOS of one of the preprocessed signal samples PSIG.

[0113] The logarithmic spectrum LOS of each of the signal samples SIG including the noise NOI includes n frequency bands FB1 to FBn. For better overview, only the frequency bands FB1 and FBn are shown. The frequency band FB1 includes an amplitude value AV1 and a phase value PV1, and the frequency band FBn includes an amplitude value AVn and a phase value PVn.

[0114] The algorithm block 15 based on the further training model comprises m inputs 16 and m outputs 17. The amplitude value AV1 is fed to the input 16.1, the phase value PV1 is fed to the input 16.2, the amplitude value AVn is fed to the input 16m-1 and the phase value AVn is fed to the input 16m.

[0115] The noise reduction output value NOV1 of output 17.1 is a noise reduction amplitude value NAV1 of noise reduction frequency band NFB1, and the noise reduction output value NOV2 of output 17.2 is a noise reduction phase value NPV1 of noise reduction frequency band NFB1. The noise reduction output value NOVm-1 of output 17m-1 is a noise reduction amplitude value NAVn of noise reduction frequency band NFBn, and the noise reduction output value NOVm of output 17m is a noise reduction phase value NPVn of noise reduction frequency band NFBn. The value NAV1, the value NPV1, the value NAVn and the value NPVn are part of the noise reduction logarithmic spectrum NLOS, which is the noise reduction signal sample NSIG.

[0116] Figure 4 A schematic diagram of a second pre-processing block 12 for a gas sensing device 1 according to the present disclosure is shown.

[0117] According to an embodiment of the present disclosure, the pre-processing block 12 comprises a domain transform block 14 configured for calculating a logarithmic spectrum LOS having a plurality of frequency bands FB for each of the signal samples SIG.

[0118] According to an embodiment of the present disclosure, the noise suppression block 13 includes a bandpass filter 18 or a lowpass filter 18 such that the preprocessed signal samples PSIG are based on reduced-bandwidth signal samples BSIG having a lower bandwidth than corresponding signal samples SIG.

[0119] exist Figure 4 In the example of , each of the signal samples SIG is filtered by a bandpass filter 18 or a lowpass filter 18 to create bandwidth-reduced signal samples BSIG, which are transformed into a noise-reduced logarithmic spectrum NLOS by a domain transformation block 14, which is a noise-reduced signal sample NSIG. In other embodiments, the positions of blocks 14 and 14 may be swapped.

[0120] Figure 5 A schematic diagram of a third embodiment of a gas sensing device 1 according to the present disclosure is shown. The third embodiment is based on the first embodiment. In the following, only additional features are explained.

[0121] According to an embodiment of the present disclosure, the gas sensing device 1 comprises one or more auxiliary sensors 19, wherein each auxiliary sensor 19 is configured for generating an auxiliary signal sample ASIG corresponding to a physical quantity of the gas mixture MG;

[0122] One or more auxiliary sensors 19 include

[0123] a temperature sensor 19 for generating a first auxiliary signal sample ASIG corresponding to the temperature of the gas mixture MG, and / or

[0124] a pressure sensor 19 for generating a second auxiliary signal sample ASIG corresponding to the pressure of the gas mixture MG, and / or

[0125] A humidity sensor 19 for generating a third auxiliary signal sample ASIG which corresponds to the humidity of the gas mixture MG.

[0126] According to an embodiment of the present disclosure, the decision making block 8 is configured to select one or more selected training models TM from one or more training models TM based on auxiliary signal samples ASIG of one or more auxiliary sensors 19, wherein an output value OV of at least one output of the algorithm block 9 based on the training model is created by using the one or more selected training models TM.

[0127] According to an embodiment of the present disclosure, the decision making block 8 is configured to select one or more selected training models TM based on the spectral information of the signal samples SIG.

[0128] According to an embodiment of the present disclosure, the feature extraction block 7 is configured to calculate an auxiliary representation AREP for the auxiliary signal samples ASIG, such that for each of the auxiliary signal samples ASIG, one of the auxiliary representations AREP is calculated, wherein each of the auxiliary representations AREP comprises one or more auxiliary feature values ​​AFV, wherein each of the one or more auxiliary feature values ​​AFV refers to a characteristic of the corresponding auxiliary signal sample ASIG; and

[0129] Therein, each of the auxiliary feature values ​​AFV of one of the auxiliary representations AREP is input into one of the inputs 10 of the algorithm block 8 based on the training model, and the algorithm block 8 based on the training model is not used for the input feature value FW, so that each of the auxiliary feature values ​​AFV is fed into an individual input 10 of the input 10, so that an output value OV of the algorithm block 8 based on the training model is created, so that the output value OV depends on the auxiliary signal sample ASIG.

[0130] According to an embodiment of the present disclosure, the training model-based algorithm block 8 includes a neural network using one or more training models TM and / or a random decision forest using one or more training models TM.

[0131] According to an embodiment of the present disclosure, the detection chamber 4 comprises one or more ventilation openings 20 which are permanently opened during the operating phase of the gas sensing device 1 .

[0132] exist Figure 5 In the example, the auxiliary sensor 19 (which may be a temperature sensor 19, a pressure sensor 19 or a humidity sensor 19) is used to generate auxiliary signal samples ASIG, which are fed to the feature extraction block 7. The feature extraction block 7 establishes an auxiliary representation AREP for each of the auxiliary signal samples ASIG, wherein each of the auxiliary representations AREP includes auxiliary feature values ​​AFV1, AFV2 and AFV3. The auxiliary feature value AFV1 is fed to the input 10.4 of the algorithm block 9 based on the training model, the auxiliary feature value AFV2 is fed to the input 10.5 of the algorithm block 9 based on the training model, and the auxiliary feature value AFV3 is fed to the input 10.6 of the algorithm block 9 based on the training model. Therefore, the output value OV of the output 11 as the sensing result SR depends on the feature values ​​FV1, FV2, FV3 and the auxiliary feature values ​​AF1, AFV2, AFV3.

[0133] The gas detection chamber 4 comprises a ventilation opening 20.1 and a ventilation opening 20.2. It is obvious that a gas detection chamber for rapid detection comprises openings of more or less tightness.

[0134] Figure 6 A general method of training a training model TM for a gas sensing device 1 according to the present disclosure is shown.

[0135] exist Figure 6 In the training phase shown, signal samples SIG stored in the measurement database 21 and target output values ​​TOV stored in the label database 22 are given to a processing unit 23, such as a GPU or CPU, to optimize the parameters of some selected training models TM. The "other information" input corresponds to auxiliary signal samples ASIG of other environmental sensors (such as temperature, pressure or humidity), which enables the selected training model TM to be selected from a pool of training models TM. Alternatively, various "other information" inputs can be used as additional features of the same model. It is important that the measurement database 21 used for training covers the desired concentration range of gases, different correlated noise realizations and different environmental conditions.

[0136] Figure 7A general method of operating a gas sensing device 1 according to the present disclosure is shown.

[0137] according to Figure 6 The established training model TM is stored in the computing device 6, such as Figure 7 As shown, the concentration of the gas is predicted in real time based on the signal samples SIG from the photoacoustic spectrometry device 2 and based on the auxiliary signal samples ASIG from the auxiliary sensor 19. A plurality of training models TM may be made available and then, depending on the specific application or environmental conditions, the most appropriate training model TM may be selected during the operation phase.

[0138] Figure 8 A more specific method for training a training model TM of a training model based algorithm block 9 for a gas sensing device 1 according to the present invention is shown.

[0139] The input signal samples SIG of the training phase are first preprocessed at the preprocessing block 12: the sum of the photoacoustic signal as a function of time pa(t), the thermoacoustic signal as a function of time ta(t) and the noise NOI as a function of time n(t), in particular denoised or filtered according to the chosen embodiment, converted to the logarithmic frequency domain at the preprocessing block 12 to simplify the processing of the noise NOI with different spectral characteristics, and fed as preprocessed signal samples PSIG to the feature extraction block 7, which creates the representation REP. Now, the processing unit can generate a training model TM based on the representation REP, the target output value TOV and the auxiliary signal samples ASIG.

[0140] Fig. 9 A more specific method of operating the gas sensing device 1 according to the present disclosure is shown.

[0141] The input signal samples SIG of the operation phase (the sum of the photoacoustic signal as a function of time pa(t), the thermoacoustic signal as a function of time ta(t) and the noise NOI as a function of time n(t)) are first preprocessed at the preprocessing block 12, in particular denoised or filtered according to the selected embodiment, converted to the logarithmic frequency domain at the preprocessing block 12 to simplify the processing of the noise NOI with different spectral characteristics, and fed as preprocessed signal samples PSIG to the feature extraction block 7, which creates the representation REP. Now, using the training model TM, the representation REP and the auxiliary signal samples ASIG, the gas prediction SR can be calculated by the algorithm block 9 based on the training model.

[0142] The training model based algorithm block 9 may be implemented as a neural network 9 having a finite number of nodes, ie neurons and hidden layers.

[0143] Fig.10A general method for a further training model TM for training a further training model based algorithm block 15 for a noise suppression block 13 of a gas sensing device 1 according to the present disclosure is shown.

[0144] In some embodiments, the noise suppression block 13 comprises an algorithm block 15 based on a further training model and a further training model FTM, which needs to be trained separately from the training model TM for the algorithm block 9 based on the training model. The further training model FTM is established by a further processing unit 24, which utilizes the output SIG of the microphone 5 with and without noise as a label. This is a preferred choice for closed photoacoustic spectrometry devices 2 including valves or similar devices (for controlling the flow of gas into or out of the gas detection chamber 4), or for cases with better SNR, where good performance can be achieved by limited complexity of the noise suppression stage, which is conveniently implemented as a denoising autoencoder (lower SNR usually means more stages in the denoising autoencoder).

[0145] Fig.11 A general method of operating the first pre-processing block 12 for a gas sensing device 1 according to the present disclosure is shown.

[0146] The pre-processing block 12 includes a domain transformation block 14 and a noise suppression block 13, the noise suppression block 13 including an algorithm block 15 based on another training model and a Fig.10 Another training model FTM established as shown in .

[0147] Fig.12 A general method of operating the second pre-processing block 12 for a gas sensing device 1 according to the present disclosure is shown.

[0148] The pre-processing block 12 comprises a domain transformation block 14 and a noise suppression block 13, which comprises a bandpass or low-pass filter 18. A part of the noise suppression task can be implemented together with the concentration estimation of the gas by a trained model-based algorithm block 9. This is a preferred option if the designer does not have access to the noise-free output of the microphone 5, or if two separate training stages (one for noise suppression and one for prediction of the concentration of the gas) are not desirable for complexity reasons.

[0149] Fig.13 The error generated by the presence of recorded speech at a gas sensing device 1 according to the prior art is shown.

[0150] An exemplary performance of a prior art device with standard metric calibration and linear calibration algorithms without noise suppression is shown. The error generated by a speech recording is shown here.

[0151] Fig.14 The error generated by the presence of recorded music of Richard Strauss at a gas sensing device 1 according to the prior art is shown.

[0152] Shown here are the errors generated by a music recording with higher power in the low frequency range. Errors of up to 5000 ppm are observed.

[0153] Fig.15 The training phase for computing the training model TM is shown.

[0154] Fig.16 The operating stages of a gas sensing device 1 according to the present disclosure are shown.

[0155] The performance of the proposed device for different noise realizations (speech and Strauss music) is shown. While the original open photoacoustic spectrometry device 2 remains unchanged, a machine learning based approach is used to suppress the acoustic noise. A simple modification of the processing chain leads to obtaining an average error of less than 100 ppm.

[0156] In case of closed photoacoustic spectrometry device 2, even better performance is expected when noise suppression algorithms are introduced.

[0157] The elements of this disclosure are:

[0158] A noise suppression mechanism for a photoacoustic gas sensor operating in the presence of different noise realizations. The mechanism overcomes current sensor limitations without hardware changes in the sensing element.

[0159] - A related mechanism for concentration estimation of gases, which can map selected measurements (and processed versions thereof) into a ppm output.

[0160] By judiciously exposing the model to various noise and environmental conditions under the training phase conditions, a robust and compact prediction model is obtained that is able to handle residual acoustic noise from the previous phase as well as varying sensed environmental conditions (such as temperature or pressure changes).

[0161] The extracted model can be embedded on the device and applied to sensor (microphone) output samples measured in real time during the inference process.

[0162] Although the present invention has been described with reference to illustrative embodiments, this description is not intended to be interpreted in a limiting sense. Various modifications and combinations of the illustrative embodiments and other embodiments of the present invention will be apparent to those skilled in the art after reference to the description. Therefore, the appended claims are intended to cover any such modifications or embodiments.

Claims

1. A gas sensing device for sensing a gas in a gas mixture (MG), the gas sensing device (1) comprising: A photoacoustic spectrometry device (2), wherein the photoacoustic spectrometry device (2) comprises a radiator (3) configured to emit light (CL), wherein the photoacoustic spectrometry device (2) comprises a gas detection chamber (4), the gas detection chamber (4) being configured to expose the gas mixture (MG) to the light (CL), wherein the photoacoustic spectrometry device (2) comprises a microphone (5) configured to detect sound (SO) in the gas detection chamber (4), the sound being caused by exposing the gas mixture (MG) to the light (CL), and wherein the photoacoustic spectrometry device (1) is configured to generate a signal sample (SIG) corresponding to the concentration of the gas in the gas mixture (MG) based on the sound (SO) detected by the microphone (5); as well as A computing device (6) configured to receive the signal samples (SIG), wherein the computing device (6) comprises a feature extraction block (7), wherein the feature extraction block (7) is configured to calculate representations (REP) for the signal samples (SIG), such that one of the representations (REP) is calculated for each of the signal samples (SIG), wherein each of the representations (REP) comprises one or more eigenvalues ​​(FV), wherein each of the one or more eigenvalues ​​(FV) refers to a characteristic of the corresponding signal sample (SIG), wherein some of the eigenvalues ​​of the representations correspond to an absolute value of a mean deviation of the corresponding signal samples, a phase of the mean deviation of the corresponding signal samples, an absolute value of a standard deviation of the corresponding signal samples, or a phase of a standard deviation of the corresponding signal samples, wherein the computing device (6) comprises a decision making block (8), the decision making block (8) comprising a trained model-based algorithm block (9) having a plurality of inputs (10) and at least one output (11), wherein the decision making block (8) comprises one or more training models (TM) for the trained model-based algorithm block (9), wherein each of the feature values ​​(FV) of one of the representations (REP) is input to one of the inputs (10) of the trained model-based algorithm block (9), such that Each of the feature values ​​(FV) is fed into an individual one of the inputs (10), wherein the decision making block (8) creates a sensing result (SR) based on an output value (OV) of the at least one output (11) of the trained model based algorithm block (9), wherein the output value (OV) is created by using at least one of the one or more trained models (TM) at the decision making block (8) such that the output value (OV) depends on the signal sample (SIG) of the photoacoustic spectrometry device (2).

2. The gas sensing device according to claim 1, wherein the trained model based algorithm block (9) comprises a neural network using the one or more trained models (TM) and / or a random decision forest using the one or more trained models (TM).

3. The gas sensing device according to claim 1 or 2, wherein the computing device (6) comprises a pre-processing block (12), wherein the pre-processing block (12) is configured for receiving the signal samples (SIG) from the photoacoustic spectrometry device (2), wherein the pre-processing block (12) is configured for generating a pre-processed signal sample (PSIG) for each of the signal samples (SIG), and wherein the pre-processing block (12) is configured for forwarding the pre-processed signal samples (PSIG) to the feature extraction block (7).

4. A gas sensing device according to claim 3, wherein the pre-processing block (12) comprises a noise suppression block (13) configured for suppressing noise (NOI) in the signal samples (SIG) such that the pre-processed signal samples (PSIG) comprise noise reduced signal samples (NSIG) having less noise (NOI) than the corresponding signal samples (SIG).

5. The gas sensing device according to claim 4, wherein the pre-processing block comprises a domain transformation block (14) configured for transforming the signal samples (SIG) into a logarithmic frequency domain so as to obtain a logarithmic spectrum (LOS) having a plurality of frequency bands (FB) for each of the signal samples (SIG), The noise suppression block (13) comprises an algorithm block (15) based on another training model, wherein the algorithm block based on another training model has a plurality of inputs (16) and a plurality of outputs (17), wherein the noise suppression block (13) comprises one or more further trained models (FTM) for the further trained model based algorithm block (15), wherein for each frequency band (FB) of the logarithmic spectrum (LOS) of one of the signal samples (SIG), an amplitude value (AV) and a phase value (PV) are input to one of the inputs (16) of the further algorithm block (15), such that each of the amplitude values ​​(AV) and each of the phase values ​​(PV) are fed to an individual input of the input (16), wherein the noise suppression block (13) creates the noise reduced signal samples (NSIG) based on the output value (NOV) of the output (17) of the further trained model based algorithm block (15), wherein the output value (NOV) of the further trained model based algorithm block (15) is created by using at least one further trained model (FTM) of the one or more further trained models (FTM) at the noise suppression block (13), and wherein each of the output values ​​(NOV) of the further trained model based algorithm block (15) is a noise reduced amplitude value (NAV) or a noise reduced phase value (NPV) of a noise reduced frequency band (NFB) of a noise reduced logarithmic spectrum (NLOS) of one of the preprocessed signal samples (PSIG).

6. The gas sensing device according to claim 4, wherein the pre-processing block (12) comprises a domain transformation block (14) configured for calculating a logarithmic spectrum (LOS) having a plurality of frequency bands (FB) for each of the signal samples (SIG).

7. A gas sensing device according to any one of claims 4 to 6, wherein the noise suppression block (13) comprises a bandpass filter or a lowpass filter (18) such that the preprocessed signal samples (PSIG) are based on reduced-bandwidth signal samples (BSIG) having a lower bandwidth than the corresponding signal samples (SIG).

8. The gas sensing device according to claim 1, wherein the gas sensing device (1) comprises one or more auxiliary sensors (19), wherein each of the auxiliary sensors (19) is configured for generating an auxiliary signal sample (ASIG) corresponding to a physical quantity of the gas mixture (MG); wherein the one or more auxiliary sensors (19) include a temperature sensor (19) for generating a first auxiliary signal sample (ASIG) of said auxiliary signal samples (ASIG) corresponding to the temperature of said gas mixture (MG), and / or a pressure sensor (19) for generating a second auxiliary signal sample (ASIG) corresponding to said auxiliary signal sample (ASIG) of the pressure of said gas mixture (MG), and / or A humidity sensor (19) is provided for generating a third auxiliary signal sample (ASIG) of the auxiliary signal sample (ASIG) corresponding to the humidity of the gas mixture (MG).

9. The gas sensing device according to claim 8, wherein the decision making block (8) is configured for selecting one or more selected training models (TM) from the one or more training models (TM) based on the auxiliary signal samples (ASIG) of the one or more auxiliary sensors (19), wherein the output value (OV) of the at least one output of the training model based algorithm block (9) is created by using the one or more selected training models (TM).

10. The gas sensing device according to claim 1, wherein the decision making block (8) is configured to select the one or more selected training models (TM) based on spectral information of the signal samples (SIG).

11. The gas sensing device according to claim 8, wherein the feature extraction block (7) is configured for calculating auxiliary representations (AREP) for the auxiliary signal samples (ASIG), such that one of the auxiliary representations (AREP) is calculated for each of the auxiliary signal samples (ASIG), wherein each of the auxiliary representations (AREP) comprises one or more auxiliary feature values ​​(AFV), wherein each of the one or more auxiliary feature values ​​(AFV) refers to a characteristic of the corresponding auxiliary signal sample (ASIG); and wherein each of the auxiliary feature values ​​(AFV) of one of the auxiliary representations (AREP) is input to one of the inputs (10) of the trained model-based algorithm block (9), the trained model-based algorithm block (9) not being used for input feature values ​​(FW), such that each of the auxiliary feature values ​​(AFV) is fed to an individual one of the inputs (10), wherein the output value (OV) of the trained model-based algorithm block (9) is created such that the output value (OV) depends on the auxiliary signal samples (ASIG).

12. The gas sensing device according to claim 1, wherein the detection chamber (4) comprises one or more ventilation openings (20) which are permanently opened during an operating phase of the gas sensing device (1).

13. A method for operating a gas sensing device for sensing a gas in a gas mixture (MG), wherein the gas sensing device (1) comprises a photoacoustic spectrometry device (2), wherein the photoacoustic spectrometry device (2) comprises a radiator (3) configured to emit light (CL), wherein the photoacoustic spectrometry device (2) comprises a gas detection chamber (4), the gas detection chamber (4) being configured to expose the gas mixture (MG) to the light (CL), and wherein the photoacoustic spectrometry device (2) comprises a microphone (5) configured to detect sound (SO) in the detection chamber (4). ), the sound (SO) is induced by exposing the gas mixture (MG) to the light (CL), and wherein the gas sensing device (1) comprises a computing device (6), the computing device (6) comprising a feature extraction block (7) and a decision making block (8), wherein the decision making block (8) comprises a trained model-based algorithm block (9), the trained model-based algorithm block (9) having a plurality of inputs (10) and at least one output (11), wherein the decision making block (8) comprises one or more training models (TM) for the trained model-based algorithm block (9), wherein the method comprises the following steps: generating, using the photoacoustic spectrometry device (2), a signal sample (SIG) corresponding to the concentration of the gas in the gas mixture (MG) based on the sound (SO) detected by the microphone (5); using said computing device (6) to receive said signal sample (SIG); using the feature extraction block (7) to calculate representations (REP) for the signal samples (SIG), such that for each of the signal samples (SIG) one of the representations (REP) is calculated, wherein each of the representations (REP) comprises one or more eigenvalues ​​(FV), wherein each of the one or more eigenvalues ​​(FV) refers to a characteristic of the corresponding signal sample (SIG), wherein some of the eigenvalues ​​of the representations correspond to an absolute value of a mean deviation of the corresponding signal samples, a phase of a mean deviation of the corresponding signal samples, an absolute value of a standard deviation of the corresponding signal samples, or a phase of a standard deviation of the corresponding signal samples; inputting each of the feature values ​​(FV) of one of the representations (REP) into one of the inputs (10) of the trained model-based algorithm block (9) such that each of the feature values ​​(FV) is fed into an individual one of the inputs (10); and The decision making block (8) is used to create a sensing result (SR) based on an output value (OV) of the at least one output (11) of the trained model based algorithm block (9), wherein the output value (OV) is created by using at least one training model (TM) of the one or more training models (TM) at the decision making block (8), such that the output value (OV) depends on the signal sample (SIG) of the photoacoustic spectrometry device (2).

Images