WO2022112125A1

WO2022112125A1 - Method for identifying a substance on a surface

Info

Publication number: WO2022112125A1
Application number: PCT/EP2021/082282
Authority: WO
Inventors: Arne Walter; Marian Kraus; Jonas Grzesiak; Frank DUSCHEK
Original assignee: Deutsches Zentrum für Luft- und Raumfahrt e.V.
Priority date: 2020-11-24
Filing date: 2021-11-19
Publication date: 2022-06-02
Also published as: DE102020131014A1

Abstract

The invention relates to a method, a detection device, and a computer program product for identifying a substance (50) on a surface (52) using a detection device (100) which emits a light beam (30, 32) onto the substance so as to cover at least some regions of the substance, receives a response signal (34) in a receiving unit (20), and processes received response signals (64, 66, 74, 76) in a classifier (70). At least the following steps are carried out: (i) training the classifier (70) using training data sets (60) of at least one substance (50) which can be found on different surfaces (52), wherein the training data sets (60) comprise received response signals (64, 66) of the substance (50) on the surfaces (52) and of the clean surfaces (52), said response signals being linked to the desired prediction of the respective substance (50); (ii) entering at least one measurement data set (78) comprising a received response signal (74, 76) of the substance (50) to be identified which can be found on a surface (52) and of the clean surface (52) in the classifier (70); and (iii) classifying the measurement data set (78) in the classifier (70) and predicting the substance (50).

Description

description

title

Method of identifying a substance on a surface

State of the art

The invention relates to a method for identifying a substance on a surface using a light-based detection device and a computer program product, a data processing system and an evaluation unit for identifying a substance on a surface using such a method and a data processing system for such a method.

The automatic identification of substances on surfaces based on the classification of their spectral signature can be carried out in real time with light-based methods for the detection and identification or classification of substances, mixtures of substances and microorganisms on surfaces. These methods are active, the signal is induced by a light source, usually a laser pulse, and can be carried out directly, i.e. without sampling and processing. The signal response of a substance is at best spectral, temporal or otherwise characteristic, so that the substance can be identified or at least classified based on these spectra.

Classification is a grouping of objects into classes (or groups). This classification is part of the input during training of the predictive model. Identification is a special case here: in classification, each substance is assigned its own class. The predictive model is what is commonly referred to as the "classifier". For an input, the model predicts a class according to the classification. This process of classifying objects in existing (trained) class systems is called classification

After all, a classifier is an instance that carries out the classification, but also a classification. In both cases, pre-processing of the data is usually necessary.

In the case of optical detection on surfaces, the signal contribution of the surface may have changed due to interaction with the substance and, together with the signal from the substance, results in a mixed signal with different compositions.

In the so-called LIF (laser-induced fluorescence) spectroscopy, the broad fluorescence signals from the substance and surface, possibly modified by absorption bands, are often in similar spectral ranges and often cannot be separated from one another. The measured LIF signal of the same substance can look completely different on different surfaces.

The automatic identification or differentiation of substances based on their fluorescence spectra can be implemented with classification algorithms such as decision trees, support vector machines or neural networks. Regardless of the algorithm used, training a predictive model of a classifier follows the same pattern. Measurement data is linked to the desired output of the model. A set of such data from different substances on different surfaces forms a training data set. The prediction model is thus generated and can then be used to make a prediction about the substance on which the spectrum is based from a mixed spectrum measured during use.

Based on the mixed substance/surface response signals, a model can only make plausible statements for substances on surfaces that are known to it from the training. unknown

Depending on the structure of the model, substances are classified as unknown or wrongly assigned to a known class, which in the best case is similar to the substance at hand. When assigning a known substance to an unknown surface, the quality of the prediction depends on the strength and signature of the signal contribution and the

Structure of the training data set and data preprocessing.

Disclosure of the Invention The object of the invention is to provide an improved method for

identify a substance on a surface by means of a light-based detection device.

A further object is to provide an improved detection device for identifying a substance on a surface with such a

specify procedure.

A further object is to specify a computer program product for carrying out such an improved method.

A further object is to specify a data processing system for such an improved method. A further object is to specify an evaluation unit for such a detection device. The objects are solved by the features of the independent claims. Favorable configurations and advantages of the invention result from the further claims, the description and the drawing.

According to one aspect of the invention, a method for identifying a substance on a surface using a light-based

Proposed detection device, wherein the detection device irradiates a light beam on the substance located on the surface at least partially covering and receives a response signal emanating from the substance and / or the surface in a receiving unit and the received response signal in a

classifier processed. The method comprises at least the steps of training the classifier with training data sets of at least one substance located on different surfaces, the training data sets receiving response signals from the substance on the surfaces and receiving response signals from the pure

include surfaces associated with the desired prediction of the respective substance; Entering at least one measurement data set with at least one received response signal of the substance to be identified and located on a surface and at least one received response signal of the clean surface in the classifier; and classifying the measurement data set in the classifier and predicting the substance. The substance can be present directly on the surface. However, it can also be rubbed into or drawn into the surface. The substance can comprise one component, but it can also have a mixture of two or more different components. In this case, according to the method according to the invention, not the individual

Components identified, but the substance as a mixture.

The light source can be in the form of a laser, for example, so that the incident light beam is a laser beam. A fluorescence signal, for example, can be received as a response signal. The method can thus advantageously be based on laser-induced fluorescence spectroscopy. In this case, the response signal can be received with spectral resolution and processed with the classifier. Other measuring methods can also be advantageously used in the method according to the invention.

In addition to so-called mixed response signals, i.e. response signals from the substance and the surface, for different substances on different surfaces

Response signals from uncontaminated surfaces recorded. When recording mixed response signals for training the classifier, a light beam is radiated onto the substance located on the surface so that the light beam at least partially covers both the substance and the surface. In addition, response signals from the clean surface are recorded. This represents a favorable procedure for training the classifier. When measuring the substance to be identified, it is advantageous to only partially irradiate the substance at least. The surface can, but does not have to, also be irradiated. Mixed outgoing response signals, which are picked up by the receiving unit, represent the objects that are combined into classes during classification. The classifier's prediction model can thus, for a mixed response signal input, predict a class according to the classification.

Substances on surfaces can be identified or classified, for example, by an artificial neural network.

To train the neural network, the mixed response signals are combined with the response signals from the clean surfaces to form an input for the classifier and linked to the desired classification result.

This linking of mixed response signals is used both for the training phase of the classifier and for the classification of the measurement data set of the substance to be identified. This leads to a significant improvement in the quality of substance detection on both known and unknown surfaces. A clear improvement can also be achieved with a pronounced filter effect, such as partial absorption of the surface signal by the substance.

The method for improving the detection of substances on surfaces can advantageously be used wherever desired or undesired substances have to be detected on known or unknown surfaces and there are also clean surface areas. Analogous to the substance detection of a pure substance on a surface, for example, two pure substances can also be detected, with the second substance then being treated in the method according to the invention instead of the surface. In this case, separate recording signals of the pure substances must be available separately for identification, im

In contrast to the determination of a mixture of several components of a substance.

The method can be used advantageously for different types of detection methods in which measurement data from the substance and background are superimposed.

An application using the method of laser-induced

Fluorescence spectroscopy, in which, for example, pulsed lasers with a wavelength of 266 nm and detection using a spectrometer and a photomultiplier array are used, represents an advantageous embodiment of the method.

Some possible application examples are, for example, detection of hazardous substances by emergency services (police, fire brigade, military),

Detection of mold on walls, detection of pathogens on plants or feed and food, for example in agriculture or food production. According to a favorable embodiment of the method, a method of machine learning can be used as the algorithm of the classifier. Various classification algorithms, such as decision trees, support vector machines or neural networks can be used to advantage in order to process combined training data sets from mixed response signals and response signals from the clean surfaces as well as combined measurement data sets of the substance to be identified from mixed response signals and response signals from the clean surfaces and the To predict the substance with great accuracy or at least to rule out other substances.

According to a favorable embodiment of the method, a neural network can be used as the algorithm of the classifier. Neural networks can be used particularly advantageously to process combined training data sets from mixed response signals and response signals from the clean surfaces as well as combined measurement data sets of the substance to be identified from mixed response signals and response signals from the clean surfaces and to predict the substance accordingly with great quality.

According to a favorable embodiment of the method, a fluorescence signal of the substance and the surface can be recorded as the outgoing response signal. Among the light-based methods for the detection and identification or classification of substances, mixtures of substances and microorganisms on surfaces in real time, fluorescence spectroscopy (LIF) is particularly suitable as an active and direct method without taking or preparing samples.

After a favorable embodiment of the method, the

Output receiving unit spectrally resolved response signals received. In particular, the receiving unit can be designed as a spectrometer. An optical spectrometer or a spectral analysis of the outgoing response signals are not absolutely necessary. The method can advantageously provide for the use of wavelength-selective input channels. A spectrometer with a detector supplies such spectrally resolved response signals. However, spectrally resolved response signals can also be implemented using photodiodes with filters.

According to a favorable embodiment of the method, a laser, in particular a pulsed laser, can be used as the at least one light source.

For LIF spectroscopy, a narrow-band light source such as an LED is sufficient. A laser has the advantage over an LED that the losses due to divergence on the way to the sample are low. Both lasers and LEDs can be operated in pulsed mode. This is not absolutely necessary for LIF, but with suitable detection it can improve the signal-to-noise ratio.

A pulsed light source is required for time-resolved signals, since the decay of the response signal is recorded after a time-defined excitation.

According to a favorable embodiment of the method, the light beam can be used with at least two wavelengths, in particular in the ultraviolet. In this way, information from different fluorophores can advantageously be included in the classification of the measured response signals. For example, wavelengths in the UV-C range and in the UV-A range, for example wavelengths of 266 nm and 355 nm, can advantageously be used. UV radiation in the UV-A range has a wavelength in the range from 315 nm to 380 nm. UV radiation in the UV-B range has a wavelength in the range from 280 nm to 315 nm. Radiation in the UV-C range has a wavelength of 100 nm to 280 nm.

The excitation wavelength required for the method according to the invention depends on the substance. Many substances have absorption bands in the UV range, including those that also absorb in the visible range. In systems that are used in biological detection, the wavelengths of 266 nm (UV-C) and 355 nm (UV-A) that are frequently used in biology and are readily available can be used to advantage. Alternatively, 280 nm (UV-B / UV-C) and 355 nm can also be used. After a favorable embodiment of the method, at least one

Part of the outgoing response signal can be recorded as a time signal. In this way, a temporal signature of the outgoing response signal can be recorded via the time signal. In this way, additional information can advantageously be included in the classification of the measured response signals.

According to a further aspect of the invention, a detection device for identifying a substance on a surface is proposed, comprising at least one light source which at least partially covers the substance

irradiation of a light beam onto the substance located on the surface, an optical receiving unit, which is designed to receive a response signal emanating from the substance and/or the surface, and a classification which is designed to process a received response signal of the outgoing response signal and which is designed for training with training data sets and for identifying a substance from measurement data sets.

In addition to mixed response signals from different substances on different surfaces, response signals from the non-contaminated surfaces are recorded. When recording mixed response signals, a light beam is radiated onto the substance located on the surface so that the light beam covers both the substance and the surface at least in certain areas. In addition, response signals from the clean surface are recorded. To train the classifier, the mixed response signals are combined with the response signals of the clean surfaces to form an input for the classifier and linked to the desired classification result. The classifier is used for both the training phase and the

Classification of the measurement data set of the substance to be identified is used. The combination of mixed response signals with response signals from the pure surface leads to a significant improvement in the quality of substance detection on both known and unknown surfaces.

According to a favorable embodiment of the detection device, the algorithm of the classifier can be based on a machine learning method. Various classification algorithms, such as so-called decision trees, usually referred to as decision trees, so-called support vector machines, usually referred to as support vector machines, or neural networks can be used to advantage to combine training data sets from mixed response signals and response signals of the pure surfaces as well as combined measurement data sets of the to process the substance to be identified from mixed response signals and response signals from the pure surfaces and to predict the substance accordingly with great accuracy or at least to exclude other substances.

According to a favorable embodiment of the detection device, an algorithm of the classifier can be designed as a neural network. Neural networks can be used particularly advantageously to process combined training data sets from mixed response signals and response signals from the clean surfaces as well as combined measurement data sets of the substance to be identified from mixed response signals and response signals from the clean surfaces and to predict the substance accordingly with great quality.

According to a favorable configuration of the detection device, the receiving unit can be designed to receive fluorescence signals. Among the light-based methods for the detection and identification or classification of substances, mixtures of substances and microorganisms on surfaces in real time, LIF is particularly suitable as an active and direct method. A spectrometer with a PMT (photo multiplier tube) array, for example, can be used as the receiving unit. Depending on the desired number of channels, however, a system of bandpass filters, a monochromator, a tunable optical filter, or the like can also be used.

According to an advantageous embodiment, the detection device can comprise at least two light sources of different wavelengths, which are designed to radiate a light beam onto the substance located on the surface. In particular, the

Light sources have wavelengths in the ultraviolet, in particular wavelengths of 266 nm and 355 nm. In this way, information from different fluorophores can advantageously be included in the classification of the measured response signals.

Alternatively, a light source, for example a laser, can also be used which emits at least two different wavelengths.

After a favorable embodiment of the detection device, a part of the outgoing response signal, in particular via a

Output mirror can be recorded as a beam splitter as a time signal. In this way, a temporal signature of the outgoing response signal can be recorded via the time signal. In this way, additional information can be advantageous in the

Classification of the measured response signals are included.

According to a favorable configuration of the detection device, the at least one light source can be designed as a laser, in particular as a pulsed laser. For LIF spectroscopy, a narrow-band light source such as an LED is sufficient. A laser has the advantage over an LED that the losses due to divergence on the way to the sample are low. Both lasers and LEDs can be operated in pulsed mode. This is not absolutely necessary for LIF, but with suitable detection it can improve the signal-to-noise ratio. A pulsed light source is required for time-resolved signals, since the decay of the response signal is recorded after a time-defined excitation

With suitable detection, pulsed light sources increase the signal to ambient light quality, but are not absolutely necessary for spectral detection. Pulsed light beams are required to evaluate time signals.

According to a further aspect of the invention, a computer program product is proposed for identifying a substance on a surface using a light-based detection device, the computer program product comprising at least one computer-readable storage medium which comprises program instructions which can be executed on a computer system and cause the computer system to perform a method as described above.

At least the following steps are carried out: training the classifier with training data sets of at least one substance located on different surfaces, the training data sets comprising received response signals of the substance on the surfaces and received response signals of the clean surfaces, which are linked to the desired prediction of the respective substance; Entering at least one measurement data set with at least one received response signal of the substance to be identified and located on a surface and at least one received response signal of the clean surface in the classifier; and classifying the measurement data set in the classifier and predicting the substance.

The computer program product advantageously serves to implement the method according to the invention. To avoid unnecessary repetition, reference is made to the description of the method.

According to a further aspect of the invention, a data processing system is proposed for executing a data processing program which is computer-readable

Program instructions includes a method for identifying a substance on a surface using a laser-based

Detection device, as described above to perform.

The data processing system is advantageously used to carry out the method according to the invention. To avoid unnecessary repetition, reference is made to the description of the method.

According to a further aspect of the invention, an evaluation unit with at least one classifier for identifying a substance on a surface using a light-based detection device is proposed.

The evaluation unit with the classifier can have hardware modules such as FPGAs (field programmable gate arrays) or components produced by 3D printing, on which the method for identifying a substance on a surface using a light-based detection device is advantageously implemented. It is possible to build the predictive model in hardware instead of as part of a software solution. A simple and realistic example is an FPGA, where the logic is implemented in physical elements. Such a hardware implementation directly links a received response signal as input with a prediction as output.

With such a hardware implementation, the method according to the invention can be implemented with a high processing speed of the received response signals.

Drawing Further advantages result from the following

drawing description. Exemplary embodiments of the invention are shown in the figures. The figures, the description and the claims contain numerous features in combination. The person skilled in the art will expediently also consider the features individually and combine them into further meaningful combinations.

Examples show:

1, 2 show a schematic representation of the measuring principle of the method according to the invention, in which a light beam is radiated onto a substance located on a surface, covering it at least in some areas, and the outgoing response signal is recorded; 3 shows a schematic structure of a light-based

Detection device for the method according to the invention; 4 shows a procedural model of the method according to the invention for identifying a substance on a surface; and FIG. 5 classification results for the identification of different substances on different surfaces.

Embodiments of the invention

In the figures, components of the same type or having the same effect are denoted by the same reference symbols. The figures only show examples and are not to be understood as limiting. Directional terminology used below with terms such as "left",

"Right", "above", "below", "before", "behind", "after" and the like only serves to improve understanding of the figures and should in no way represent a restriction of generality. The components and elements shown, their design and use can vary according to the considerations of a person skilled in the art and can be adapted to the respective applications.

A schematic representation of the measuring principle of the method according to the invention is shown in Figures 1 and 2, in which a light beam 30 is irradiated onto a substance 50 located on a surface 52, at least partially covering it, and the outgoing response signal 40, 42, 44 is recorded. 1 shows a plan view of a surface 52 with a substance 50 thereon, while FIG. 2 shows a longitudinal section through the arrangement, in which the outgoing

Response signal 40, 42, 44 is shown symbolically in the form of an arrow. 1 shows three surfaces illuminated by light beams 30, of which one light beam 30 impinges directly on the substance 50, while the other two light beams 30 each hit the edge of the substance 50 and part of the light beam 30 hits the surface 52. The light beam 30 excites the irradiated areas, for example

Delivery of fluorescence radiation, which can be seen as the outgoing response signal 40, 42, 44 in FIG.

When a light beam 30 irradiates the substance 50 on the surface 52, the signal response of the outgoing response signal 30 consists of the

Response of the surface 52 and the response of the substance 50 in the form of fluorescence radiation together, depending on the type and distribution of the substance 50 on the surface 52, the signal responses can interact with the substance 50 and / or the surface 52, which lead to a changed signal response can. The measured signal is ultimately made up of different components, whereby the ratios can vary greatly depending on the surface, substance and type of distribution. Figure 2 shows the outgoing response signal 40, 42, 44, with a signal

40 emanates directly from the substance 50, while a second signal 42 emanates as a mixed signal from the substance 50 and the surface 52, and a third signal 44 emanates only from the surface 52. The signals 40, 42, 44 of the outgoing response signal can, for example, advantageously by means of laser-induced

Fluorescence spectroscopy (LIF) can be recorded. However, other light-based methods can also be used. The incident light beam 30 can, for example, have at least two

Have wavelengths, especially in the ultraviolet in the UV-C and UV-A range, in order to have interactions between the substance that are as favorable as possible 50 and surface 52 to detect. In this case, for example, laser light with wavelengths of 266 nm and 355 nm can be used. Alternatively, for example, light with wavelengths of 280 nm and 355 nm can also be used.

According to the method according to the invention, the various recorded signals of the outgoing response signal 40, 42, 44 can be input into a classifier 70 in a suitable combination, as shown later in FIG.

Figure 3 shows a schematic structure of a light-based detection device 100 for the method according to the invention for identifying a substance 50 on a surface 52. The detection device 100 comprises at least one light source 10, 12, which at least partially covers the substance 50 for irradiating a light beam 30, 32 the substance 50 located on the surface 52 is formed. The detection device 100 further comprises an optical receiving unit 20 which is used to receive the substance 50 and/or the surface 52 emanating from it

response signal 34, and a classifier 70, which is designed to process a received response signal 64, 66, 74, 76 (recognizable in Figure 3) of the outgoing response signal 34, and whose prediction model is a product of training with training data sets 60 and, for Identifying a substance 50 from

Measurement data sets 78 is formed.

The light beam 30, 32 of the at least one light source 10, 12 is directed onto the substance 50 via deflection mirrors 14, 16.

The receiving unit 20 can be designed, for example, to analyze signals from laser-induced fluorescence spectroscopy. That The outgoing response signal 34 can be coupled into optical fibers, for example, via a collecting mirror 18 and can be received in a spectrally resolved manner in the receiving unit 20 . An algorithm 90 of the classifier 70 (FIG. 4) can advantageously be based on a machine learning method. In particular, a neural network can be used for this.

The illustrated in the embodiment in the figure second light source 10, 12 can optionally to increase the accuracy of the

prediction of the substance to be identified. The light beam 32 of the second light source 12 is coupled into the beam path 30 of the first light source 10 via a deflection mirror 22 and a dichroic mirror 24 .

The two light sources 10, 12 can advantageously have wavelengths in the ultraviolet, in particular wavelengths of 266 nm and 355 nm, i.e. in the UV-C range and in the UV-A range. A part 36 of the outgoing response signal 34 can optionally be recorded as a time signal, in particular via an output mirror 26 as a beam splitter. In addition, a temporal signature of the outgoing response signal can be determined for the optional improvement of the prediction when classifying the measurement data sets.

Classifier 70, whose mode of operation is described in detail in Figure 4, can be arranged in an evaluation unit 38 of detection device 100, which also records the data from receiving unit 20, and optionally data of a time signal which is determined in time signal evaluation unit 28. the

Evaluation unit 38 can, for example, also control the at least one light source 10, 12 during the measurement. The evaluation unit 38 can further advantageously as a data processing system for executing a

Be formed data processing program, which includes computer-readable program instructions to the method for

Carry out identification of a substance 50 on a surface 52 by means of the light-based detection device 100 .

On the data processing system can advantageously a computer program product for identifying a substance 50 at a

Surface 52 be implemented by means of a laser-based detection device 100, wherein the computer program product comprises at least one computer-readable storage medium, which includes program instructions that are executable on the computer system and cause the computer system to the inventive

carry out procedures.

However, the evaluation unit 38 with the classifier 70 can also have hardware modules such as FPGAs (field programmable gate arrays) or components produced by 3D printing, on which the method according to the invention is advantageously implemented. With such a hardware implementation, the method can be implemented with a high processing speed of the received response signals.

A possible embodiment of the method according to the invention for identifying a substance 50 on a surface 52 using a light-based detection device 100 is shown in FIG. A detection device 100 (FIG. 3) radiates a light beam 30, 32 onto the substance 50 located on the surface 52, at least partially covering it. One of Substance 50 and/or Der The response signal 34 emanating from the surface 52 is received in a receiving unit 20 and the received response signal 64, 66, 74, 76 is processed in a classifier 70.

According to the method, the classifier 70 is first trained with training data records 60 of a large number of different substances 50 (FIG. 3) located on different surfaces 52 .

In addition to mixed response signals from different substances 50 on different surfaces 52 (FIG. 3), response signals from the non-contaminated surfaces 52 are recorded. When recording training signals, the light beam 30 is radiated onto the substance 50 located on the surface 52 so that the light beam 30 at least partially covers both the substance 50 and the surface 52 (FIG. 3). In addition, outgoing response signals of the clean surface 52 are recorded.

An initial measurement to determine a substance to be identified only has to cover the substance, whether and how much surface is irradiated is practically irrelevant.

The training data sets 60 include response signals 64 from the substances 50 on the surfaces 52 and response signals 66 from the clean surfaces 52, which are combined to form an input 62 in the classifier 70.

To train the neural network, the input 62 is linked to the desired classification result 72 as an output 68 to form a training data set 60 . The training data records 60 are thus permanently linked to the desired prediction of the respective substance 50 (FIG. 3). The classifier 70 thus goes through a series of training runs until a desired prediction quality of the respective substance 50 is reached.

After a measurement process on a substance 50 to be determined on a known or unknown surface 52, the measurement data set 78, which includes at least one response signal 74 of the substance 50 to be identified and located on a surface 52 and at least one response signal 76 of the clean surface 52, is used as input 62 entered into the classifier 70.

The measurement data set 78 is then classified in the classifier 70 and a prediction 72 of the substance 50 is output as an output 68 .

A method of machine learning can advantageously be used as the algorithm 90 of the classifier 70 . Various classification algorithms, such as e.g. Decision Trees, Support Vector Machines or Neural Networks can be used advantageously in order to generate combined training data sets 60 from mixed outgoing response signals and outgoing response signals from the clean surfaces 52 as well as combined measurement data sets 78 for the substance to be identified 50 from mixed response signals and response signals from the to process clean surfaces 52 and to predict the substance 50 accordingly with great quality. In particular, a neural network can be used for this.

The combination of mixed response signals with response signals from the clean surface 52 is used both for the training phase of the classifier 70 and for the classification of the measurement data record 78 of the substance 50 to be identified. This leads to a significant improvement in the quality of the substance detection both on known and on unknown surfaces 52. A significant improvement can also be achieved with a pronounced filter effect, such as partial absorption of the surface signal with the substance 50.

FIG. 5 shows classification results for the identification of different substances on different surfaces. The method was applied with measurement data on 13 different substances on 13 different surfaces each.

Substances included: albumin, baker's yeast, baking soda, curcumin, aspen pollen, flour, NADFI, pyridoxine, riboflavin, soy ferment, tryptophan, B. subtilis (suspension), E. coli (suspension).

Surfaces included: countertop (varnished) 200, flartwood (unvarnished) 202, cardboard 204, composite wood I 206, pressboard (varnished) 208, faux leather 210, hardwood (varnished) 212, towel (fabric) 214, PVC (black) 216, composite wood II 218, plastic (white) 220, paper (bleached)

222, pressboard (painted white) 224.

100 measurements each were carried out at 3 different positions with different substance assignments. LIF spectra after excitation at a laser wavelength of 266 nm were used.

Two types of training data sets were generated from 80% of the measurement data. One type of training data set contained the channels of the mixed response signal and appended the channels of an associated surface measurement. The other type of training data set contained only the channels of the mixed response signal. To determine the quality For the detection of substances on unknown surfaces, a training data set was created for each surface that did not contain the respective surface. The remaining 20% of the measurement data were used to evaluate the procedure and were not part of the training.

The proportion of measurements with correctly assigned substances in the total number of measurements (total accuracy 80) for substance detection on the various surfaces is shown in the figure.

Respective columns 82 show results of classifications on known surface with a mixed response and surface response as input to the classifiers. Respective columns 84 show results of classification on known surface with only mixed response as input

Input to the classifiers The respective columns 86 show results of classifications on an unknown surface with a mixed response signal and a surface response signal as input to the classifiers. The respective columns 84 show results of classifications on a known surface with only a mixed response signal as input to the classifiers

Overall accuracies in the classification of the measurement data for measurements on the different surfaces 200 to 224 are shown in each case. Mean values of the accuracy 80 over all surfaces are shown as dot-dash lines for the different categories 82, 84, 86, 88. By supplementing the mixed spectra with the pure surface spectra as input for the classifier, an improvement in the quality of the mean overall accuracy 80 of 15.2 percentage points could be achieved for the prediction of known substances on known surfaces included in the training, with a mean value of 82 of 70.5% compared to the mean 84 of 55.3%. For known substances on unfamiliar surfaces, a mean improvement of 11.2 percentage points was achieved with a mean 86 of 55.5% compared to the mean 88 of 44.3%.

For this data set, the improvement in the quality of the prediction on known surfaces is approximately of the order of magnitude that is also achieved by equipment supplements, which are sometimes associated with considerable additional costs, such as the use of a second excitation wavelength or measurement of the temporal signature of the

fluorescence signals.

Alternatively or additionally, however, by using a second excitation wavelength or measuring the temporal signature of the fluorescence signals in the method according to the invention

The overall accuracy of the forecasts can be further increased.

The increase in performance with non-linear behavior, eg interactions of the flicker background signal with the substance, was shown separately on modeled data sets with pronounced interactions. Numeral 10 light source

12 light source

14 deflection mirrors 16 deflection mirrors

18 collection mirrors

20 receiving unit

22 deflection mirrors

24 in-coupling mirror 26 out-coupling mirror

28 time signal measurement unit

30 light beam

32 light beam

34 outgoing response signal 36 decoupled response signal

38 evaluation unit

40 response signal substance

42 mixed response signal

44 response signal surface 50 substance

52 surface

60 training data set

62 input

64 received response signal 66 received response signal

68 edition

70 classifiers

72 prediction substance

74 received response signal 76 received response signal

78 measurement data set

80 accuracy 82 mixed spectrum+surface spectrum, known surface

84 mixed spectrum, known surface

86 mixed spectrum+surface spectrum, unknown surface

88 mixed spectrum, unknown surface

90 algorithm

100 detection device

200 countertop (lacquered)

202 hardwood (unpainted)

204 cardboard

206 Composite Wood I

208 pressboard (varnished)

210 faux leather

212 hardwood (varnished)

214 towel (cloth)

216 PVC (black)

218 composite wood II2

220 plastic (white)

222 paper (bleached)

224 pressboard (white)

Claims

Expectations

1. A method for identifying a substance (50) on a surface (52) using a light-based detection device (100), the detection device (100) directing a light beam (30, 32) onto the substance (50) located on the surface (52) radiates in at least partially covering and receives a response signal (34) emanating from the substance (50) and/or the surface (52) in a receiving unit (20) and received response signals (64, 66, 74, 76) in a classifier (70) processed, the method comprising at least the steps

(i) Training the classifier (70) with training data records (60) of at least one substance (50) located on different surfaces (52), the training data records (60) containing response signals (64) received from the substance (50) on the surfaces (52 ), and received clean surface (52) response signals (66) associated with the desired prediction of the respective substance (50);

(ii) Entering at least one measurement data set (78) with at least one received response signal (74) of the substance (50) to be identified and located on a surface (52) and at least one received response signal (76) of the clean surface (52) in the classifier (70);

(iii) classifying the measurement data set (78) in the classifier (70) and predicting the substance (50).

2. The method according to claim 1, characterized in that a method of machine learning is used as the algorithm (90) of the classifier (70).

3. The method according to claim 1 or 2, characterized in that a neural network is used as the algorithm (90) of the classifier (70).

4. The method according to any one of the preceding claims, characterized in that a fluorescence signal of the substance and the surface is recorded as the outgoing response signal (34).

5. The method according to any one of the preceding claims, characterized in that the receiving unit (20) outputs spectrally resolved response signals received, in particular that the receiving unit (20) is designed as a spectrometer.

6. The method according to any one of the preceding claims, characterized in that a laser, in particular a pulsed laser, is used as the at least one light source (10, 12).

7. The method according to any one of the preceding claims, characterized in that the light beam (30, 32) is used with at least two wavelengths, in particular in the ultraviolet.

8. The method according to any one of the preceding claims, characterized in that at least part of the outgoing response signal (34) is recorded as a time signal.

9. Detection device (100) for identifying a substance (50) on a surface (52) using a method according to any one of the preceding claims, comprising at least one light source (10, 12) which for at least partially covering the substance (50) irradiating a light beam (30, 32) onto the substance (50) located on the surface (52), an optical receiving unit (20) which is designed to receive a response signal ( 34) is designed, and a classifier (70) which is designed to process a received response signal (64, 66, 74, 76) of the outgoing response signal (34), and which for training with training data sets (60) and for identifying a Substance (50) is formed from measurement data records (78).

10. Detection device according to claim 9, characterized in that an algorithm (90) of the classifier (70) is based on a machine learning method.

11. Detection device according to claim 9 or 10, characterized in that the algorithm (90) of the classifier (70) is designed as a neural network.

12. Detection device according to one of claims 9 to 11, characterized in that the receiving unit (20) is designed to receive fluorescence signals.

13. Detection device according to one of claims 9 to 12, characterized in that the receiving unit (20) is designed to output spectrally resolved response signals received, in particular that the receiving unit (20) is designed as a spectrometer.

14. Detection device according to one of claims 9 to 13, comprising at least two light sources (10, 12) of different wavelengths, which are designed to radiate a light beam (30, 32) onto the substance (50) located on the surface (52), in particular wherein the light sources (10, 12) have wavelengths in the ultraviolet, in particular have wavelengths of 266 nm and 355 nm.

15. Detection device according to one of claims 9 to 14, characterized in that a part (36) of the outgoing response signal (34) can be recorded as a time signal, in particular via a decoupling mirror (26) as a beam splitter.

16. Detection device according to one of claims 9 to 15, characterized in that the at least one light source (10, 12) is designed as a laser, in particular as a pulsed laser.

17. Computer program product for identifying a substance (50) on a surface (52) using a light-based detection device (100), wherein the computer program product comprises at least one computer-readable storage medium which comprises program instructions that can be executed on a computer system are and cause the computer system to carry out a method, in particular according to at least one of claims 1 to 8, wherein at least the steps are carried out (i) Training the classifier (70) with training data records (60) of at least one substance (50) located on different surfaces (52), the training data records (60) containing response signals (64) received from the substance (50) on the surfaces (52 ), and received response signals (66) from the clean surfaces (52) associated with the desired prediction of the respective substance (50);

18. Data processing system for executing a

Data processing program which comprises computer-readable program instructions in order to carry out a method for identifying a substance (50) on a surface (52) by means of a laser-based detection device (100), in particular according to at least one of claims 1 to 8.

19. Evaluation unit (38) with at least one classifier (70) for identifying a substance (50) on a surface (52) by means of a light-based detection device (100), in particular according to at least one of claims 9 to 16, with a method, in particular according to at least one of claims 1 to 8.