Method and apparatus for determining the signal-to-noise ratio of a signal

Abstract

A signal (125) comprises a superposition of a useful signal (115) and a noise signal (120). A method (200) for determining a signal-to-noise ratio (600) of such a signal (125) comprises steps of determining a frequency spectrum of the signal (125); detecting a predetermined useful signal (115) in the frequency spectrum of the signal (125); determining a first, detection-relevant, and a second, detection-irrelevant section of the frequency spectrum of the signal (125); and determining the signal-to-noise ratio (600) with respect to a first power of the signal (125) at the first section and a second power of the signal (125) at the second section.

Description

[0001] The present invention relates to the detection of a signal. In particular, the invention relates to the determination of a signal's signal-to-noise ratio.

[0002] The state of the art is shown in DE 102 30 685 A1, DE 10 2005 017 833 A1 and DE 10 2017 129 168 A1.

[0003] An acoustic signal in the vicinity of a motor vehicle can be evaluated in order to control the vehicle. For example, the sound of a siren in the surrounding area can be detected and the distance and direction of the sound source can be determined. The siren can be used by a special vehicle with special rights to request other vehicles to give it the right to pass or to allow it to pass. In order to clear the way for a special vehicle approaching from behind, it can be advisable to stop the vehicle or move to the side. If the special vehicle is even further away, this could cause a traffic jam that would hinder the special vehicle. However, the distance cannot be accurately determined based on volume alone, as this can be subject to several influencing factors.

[0004] A classifier for detecting a predetermined wanted signal in a captured signal typically provides some kind of probability or confidence value, which, especially in neural networks, is close to 1 or 0. This value does not yet indicate the reliability of the result. Especially in audio analysis, classification is difficult when the signal-to-noise ratio (SNR) is poor. Some methods calculate a spectrogram and search for patterns within it. If the pattern barely stands out from the background noise, detection is difficult.

[0005] A specific result can be estimated more accurately based on the SNR. However, known techniques for determining the SNR are complex or inaccurate. One object underlying the present invention is to provide an improved technique for determining the SNR of a signal that comprises a superposition of a wanted signal and a noise signal, in which a predetermined wanted signal is to be detected.

[0006] The invention solves this problem by means of the subject matter of the independent claims. Subclaims specify preferred embodiments.

[0007] A signal comprises a superposition of a wanted signal and a noise signal. According to a first aspect of the present invention, a method for determining a signal-to-noise ratio of such a signal comprises the steps of determining a frequency spectrum of the signal; detecting a predetermined wanted signal in the frequency spectrum; determining a first section of the frequency spectrum relevant for detection and a second section irrelevant for detection; and determining the signal-to-noise ratio with respect to a first power of the signal at the first section and a second power of the signal at the second section.

[0008] It was recognized that the spectrogram not only allows for the detection of the predetermined wanted signal, but also for the determination of the SNR. Features that a classifier uses to identify the predetermined wanted signal typically contribute a large portion of the wanted signal, so that in sections where these features are located, the wanted signal can account for a large portion of the overall signal. For these sections, the power of the wanted signal can be determined as the first power. The second power of the noise signal can be determined for other sections.

[0009] Sections that make a predetermined minimum contribution to signal detection can be determined as relevant. Sections that are not determined as relevant can be determined as irrelevant. The distinction between relevant and irrelevant can be made in different ways, with the method used being able to be adjusted in each case to shift the distinction in favor of relevant or irrelevant sections.

[0010] The frequency spectrum can generally be represented in any way that allows further processing, for example, as a vector or a matrix. Such a general representation can also easily be created as a graphical representation, which can be called a spectrogram. A matrix can be mapped onto an image surface, and values ​​of the individual matrix elements can specify color or brightness values ​​of picture elements (pixels). In one embodiment, further processing steps are performed on the graphical representation. For example, image processing or image recognition methods can be applied to identify the desired signal in the graphical representation.

[0011] Preferably, the frequency spectrum represents a temporal progression of the signal, which can be displayed as a two-dimensional spectrogram. The representation can depict three parameters in different dimensions. A first parameter comprises a frequency range, a second parameter a signal power, and a third parameter the time.

[0012] Preferably, a spectrogram of the signal is determined using a short-time Fourier transformation (STFT). A predetermined time range can be selected, and a Fourier transformation can be applied within this range. Preferably, a Fast Fourier Transformation (FFT) is performed, which can be efficiently executed on a digital processing device.

[0013] In a first variant, an isometric representation can be selected, in which the frequency, for example, is plotted horizontally and the power vertically, resulting in a bar graph for a given point in time. Bar graphs for more recent points in time can be isometrically offset, for example, to the top right. To visualize a current sampling, older bar graphs can be moved to the top right (or back), and new bar graphs can be inserted at the bottom left (or front). This representation uses two colors and can be processed in monochrome.

[0014] In a second variant, for example, the frequency range is plotted vertically and the time horizontally. The signal power can be represented as the color of a predetermined color scale. Such a representation can be referred to as a false-color representation. The color scale is usually spanned between easily distinguishable colors, for example, from red to blue or from yellow to violet. A running scan can be animated by moving the representation horizontally, with measurements being added at one edge and removed at the other. This representation can provide a good overview of a selected time window, and previous measurements cannot obscure each other in perspective. Graphic processing can be performed using conventional, color-capable graphic methods.However, this representation can make it more difficult, at least for a human observer, to visually recognize small differences between sections.

[0015] Other representation forms are also possible. In this case, an advantage of the selected representation can be exploited in determining the SNR and / or detecting the predetermined wanted signal. Reference is preferably made herein, without limiting generality, to the second described variant. A section mentioned herein can correspond to an area with uniform information in the representation. The representation can be rasterized into image points (pixels), and a smallest possible section can correspond to one or more image elements.

[0016] For or as part of the recognition process, the relevance of all segments can be determined. A segment whose relevance reaches a predetermined threshold can be determined as relevant in the sense described above. A segment whose relevance lies below the threshold can be determined as irrelevant in the sense described above. The threshold can be selected to ensure a sufficiently reliable determination of the SNR. The threshold can be adjusted to the assumed relevance of the classifier used to recognize the predetermined wanted signal. This allows different classifiers to use differently defined relevances.

[0017] The threshold can comprise a predetermined proportion of the maximum relevance achieved by a section. In other words, the sections can be sorted according to their assigned relevancy. Sections whose relevancy exceeds the threshold can be used to determine the first performance. The threshold can be determined absolutely or relatively. In one example, all sections whose relevancy is at least 90% of a maximum possible relevance can be used. Conversely, the 10% of all sections with the highest achieved relevancy can also be used to determine the first performance.

[0018] In a further embodiment, the performance of all sections is weighted with their assigned relevances. The first performance can then be determined based on the weighted performance. The relevances of the sections can be used as a kind of mask to primarily consider the performance of those sections whose contribution to the detection of the wanted signal is greatest. A performance with respect to multiple sections can be determined, in particular, as a root mean square (RMS).

[0019] Method according to one of the preceding claims, wherein the detection is performed using a convolutional neural network (CNN). A CNN is a well-understood machine learning technique and can be used to detect the predetermined signal. For this purpose, the CNN can be trained to detect the predetermined signal in numerous runs. Training data can include samples of the signal under different circumstances. For example, if it is a predetermined acoustic signal, the signal can be detected from different directions, different distances, at different transmission volumes, and / or at different driving speeds of the signal source and / or the motor vehicle.In addition, the signal can be detected under various background noises, which may include, for example, traffic noise, construction noise, music or other noises encountered in the area of ​​a moving motor vehicle.

[0020] The relevance of a section for recognition can be determined by interpolating a neuron's activation to the signal. If, for example, a CNN is used to detect or classify the predetermined signal, a local feature can be determined within it during recognition. An active neuron can calculate activity "near" an associated section. The internal activity of a neuron can correspond to the relevance of an associated section during recognition. The activity can be interpolated to the input signal to determine the associated section.

[0021] In another embodiment, the relevance is determined with reference to a class activation map (CAM). The class activation map can be used to understand how an artificial neural network makes a decision. The CAM is preferably determined with reference to a processing of a graphical representation of the frequency spectrum by a neural network. In particular, a gradient-weighted class activation map of the graphical representation can be created, which can provide a relevance degree for each section of the representation. Various program packages are freely available for this purpose, for example "grad-cam", "torch-cam" or "Score-CAM". The relevance degree can be displayed graphically in false colors; the result usually has the same dimensions as the signal and can also be referred to as a heat map.

[0022] Relevance can also be determined using layer-wise relevance propagation (LRP). This allows regions in the input data that had a positive influence on the classification result to be highlighted. This can exploit certain properties of a multilayer neural network.

[0023] The classifier can implement any machine learning technique, and the determination of the relevance of a section to recognition can also be determined using another suitable Explainable Artificial Intelligence technique.

[0024] Sections that belong to the wanted signal can each be replaced by an interpolation with adjacent sections. The second power can be determined with respect to a resulting representation. A section can be counted as part of the wanted signal if its associated relevance for detecting the predetermined signal is sufficiently high. Adjacent sections that do not belong to the wanted signal can be used for interpolation, effectively removing a relevant part of the representation.

[0025] The interpolation can be performed using a suitable interpolation function. In practice, any predetermined, smooth function can be applied to replace the value of a section belonging to the wanted signal. The function can be chosen as typical depending on the wanted signal to be recognized. The error in calculating the interpolation fit to a section can be weighted according to the assigned relevance. For example, the error can be multiplied by (1 - R), where R is the determined relevance of a section. A large value of R corresponds to high relevance, and a low value to low relevance. Typically, the value range for R is between 0 and 1.

[0026] The difference between the target signal and the background spectrogram can then be calculated individually for each pixel and then averaged based on the level.

[0027] The signal-to-noise ratio can be determined as the quotient of the first and second power levels. The power of the wanted signal can be determined as a single first numerical value, and that of the noise signal as a single second numerical value, and the numerical values ​​can be divided accordingly. The numerical values ​​are preferably averages of the values ​​assigned to the respective sections and can preferably be determined as root mean square (RMS).

[0028] In another embodiment, a difference between the power of the wanted signal and the power of the noise signal can be determined for each section of the representation. The signal-to-noise ratio can then be determined based on a level-based averaging of the differences.

[0029] The technique proposed herein can also be used with a classifier capable of recognizing multiple predetermined signals. If the signal is an acoustic signal, particularly related to vehicle traffic, the classifier can, for example, recognize useful signals from a horn, a bell, a police whistle, a siren (follow-up tone signal), a yelp signal, a wail signal, or another special signal. Using the described technique, a signal-to-noise ratio can be determined for each detected useful signal.

[0030] To improve the determination, a section that lies in a frequency range that is atypical for the predetermined desired signal can be removed from the determination. The value of the determined section can be replaced by an interpolation with respect to values ​​of adjacent sections, as described herein.

[0031] In a preferred development of the described technology, a motor vehicle can be controlled based on a detected wanted signal and the determined signal-to-noise ratio of a sampled acoustic signal. The signal can be detected, in particular, in the surroundings of the motor vehicle. The control can take place in real time, i.e., with a predetermined (small) maximum time delay between the sampling and a control decision or control intervention. The motor vehicle preferably comprises a motorcycle, a passenger car, a truck, or a bus. The control can take place within the framework of deterministic logic or by means of machine learning methods. It should be noted that the control does not necessarily have to trigger a control intervention in the motor vehicle immediately after the signal is detected and the SNR is determined.In some cases, a response to a detected signal can be intentionally delayed, for example, depending on the signal, the SNR, or a driving situation of the vehicle or a source of the signal. A driving situation can be determined, for example, based on a position, a speed, a direction, or an acceleration. The driving situation can also be determined based on the state of an on-board system, such as a drive motor.

[0032] According to a further aspect of the present invention, a device for determining a signal-to-noise ratio of a signal comprising a superposition of a useful signal and a noise signal comprises: a sensor for sampling the signal; and a processing device. The processing device is configured to determine a frequency spectrum of the signal; a predetermined useful signal in the frequency spectrum of the signal; to determine a first section of the frequency spectrum that is maximally relevant for detection and a second section that is minimally relevant for detection; and to determine the signal-to-noise ratio with respect to a first power of the signal at the first section and a second power of the signal at the second section.

[0033] Recognition can be performed using a classifier implemented using any technology. For example, the classifier can use machine learning methods; alternatively, a statistical approach or an adaptive filter is also conceivable. The classifier can be formed, emulated, or included in the processing device. Alternatively, a discrete classifier can also be used.

[0034] The processing device can be configured to carry out a method described herein in whole or in part. For this purpose, the processing device can be implemented electronically and comprise a programmable microcomputer or microcontroller, and the method can be in the form of a computer program product with program code means. The computer program product can also be stored on a computer-readable data carrier. Features or advantages of the method can be transferred to the device, or vice versa.

[0035] The invention will now be described in more detail with reference to the accompanying figures, in which: Fig. 1 a system; Fig. 2 a flow diagram of a process; Fig. 3 an example spectrogram; Fig. 4 an input signal for a classifier for a predetermined useful signal; Fig. 5 Relevances of sections of a spectrogram for the detection of a predetermined useful signal; and Fig. 6 represents a specific course of a signal-to-noise ratio of a useful signal in a signal.

[0036] Fig. 1 shows a system 100 comprising a motor vehicle 105 and an external sound source 110. The sound source 110 is assumed here to be a special vehicle, for example, an emergency vehicle of the police, rescue service, or fire department. The sound source 110 emits a useful acoustic signal 115, which is superimposed with a noise signal 120. The noise signal can originate from the sound source 110 or another source in the vicinity of the motor vehicle 105. A signal 125 can be detected on the motor vehicle 105, which comprises a superposition of the useful signal 115 and the noise signal 120.

[0037] On board the motor vehicle 105 is a device 130, which is preferably configured to detect the signal 125, recognize a predetermined wanted signal 115, and determine a signal-to-noise ratio. The device 130 comprises one or more sensors 135 and a processing device 140. A sensor 135 shown is designed as a microphone. A memory 145 is optionally provided. The device 130 preferably has an interface 150 to provide a processing result. Such a result can include an indication of a detected, predetermined wanted signal 115 and / or a signal-to-noise ratio determined in this regard. The motor vehicle 105 can be controlled based on such a result. For example, the motor vehicle 105 can slow down or drive further to the right in order to more easily allow a special vehicle 110 approaching from behind to pass.In one embodiment, the device 130 is also configured to provide a control signal to control the motor vehicle 105. The control can be performed, in particular, in the longitudinal direction and / or in the transverse direction.

[0038] The memory 145 is preferably configured to store information about one or more useful signals 115. Optionally, the memory can also temporarily or permanently store other information, for example, a detected signal 125 or a processing result.

[0039] Fig. 2 shows a flowchart of a method 200, which can preferably be carried out using the device 130 on board the motor vehicle 105. In a step 205, the signal 125 is acquired. A discrete-time sampling preferably takes place at a predetermined sampling rate, which can, for example, be approximately 15,000 Hz. In a step 210, a time period can be determined, which, for example, has a length of 0.1 s. Samples that fall within the determined time period can be combined.

[0040] In a step 215, a frequency spectrum can be calculated with respect to samples from the specific time period. An FFT can be used for this purpose. Together with the formation of the time period in step 210, this can correspond to the functionality of an STFT.

[0041] In a step 220, a graphical representation of the frequency spectrum can be created. Such a representation can also be called a spectrogram. A graphical representation can also be created that includes spectrograms of several time periods, as described below with reference to Fig. 3. Steps 205 to 220 can be executed continuously or cyclically in a loop, so that the spectrogram can be updated or expanded with incoming samples. Typically, the spectrogram is built until a predetermined duration is reached, for example, approximately 5 seconds, and then updated to represent data of the predetermined duration prior to the current time.

[0042] In a step 225, a predetermined wanted signal 115 can be detected in the signal 125 based on the graphical representation—or another selected representation. For this purpose, a classifier can be used that can detect one or more wanted signals 115. If multiple wanted signals 115 are detected, the following steps of the method 200 can be performed multiple times, each with reference to one of the wanted signals 115.

[0043] The frequency spectrum can be given in the form of a matrix divided into rows and columns into sections of equal size. Each section can be formed as a picture element, which can also be called a pixel. Several pixels can also be combined to form a section. In a step 230, sections that are uncharacteristic of the detected useful signal 115 can be removed from the signal 125. If, for example, a subsequent signal with two tones has been detected, a section that temporarily represents a third tone and originates, for example, from the horn of another road user, can be removed. During removal, a value assigned to the section can be replaced by a value that corresponds to values ​​of surrounding sections that are assigned to the noise signal 120.Such a replacement may involve interpolation or, if the section to be replaced is located at the edge of the representation, extrapolation.

[0044] A signal-to-noise ratio can be defined as the quotient of a first power of the wanted signal and a second power of the noise signal. To determine the quotient, the first power of the wanted signal 115 and the second power of the noise signal 120 can each be determined based on the representation of the frequency spectrum.

[0045] In a step 235, sections of the representation can be determined that are or were sufficiently relevant for the recognition of the wanted signal 115. Particularly if the wanted signal 115 was determined using a CNN, the relevance of the individual sections can be determined based on the activities of individual neurons during the recognition process. Alternatively, a gradient-weighted class activation map (grad CAM) or layer-wise relevance propagation (LRP) can be used to determine the relevance of the sections. The result can be provided in the form of a mask that has the same dimensions as a grid-like representation of the signal 125.

[0046] A section whose assigned relevance exceeds a predetermined threshold can be assigned to the wanted signal 115, while another section can be part of the noise signal 120. For this purpose, specific relevancy values ​​can be compared with the threshold. The values ​​of the sections can also be weighted by the relevancy values. Values ​​of the resulting sections of the wanted signal 115 can be averaged in a step 240 to form a single number that represents the first power. For this purpose, the values ​​can be calculated, for example, as a root mean square.

[0047] In a step 245, those sections of the representation that are not considered relevant for recognition can be determined. For this purpose, values ​​of those sections of the representation that are considered relevant can be ignored or replaced by values ​​determined based on values ​​of adjacent sections. For example, a value of a section can be determined based on an interpolation of opposite neighboring sections with respect to the section. For this purpose, a predetermined interpolation function can be used, which is optionally selected depending on the recognized useful signal 115.

[0048] With respect to the specific sections of the noise signal 120, a second power can be determined in a step 250, which represents a single number that represents the averaged power of the sections that are not part of the useful signal 115. For this purpose, a root mean square (RMS) can be calculated, for example, over the values ​​of the sections.

[0049] In a step 255, the signal-to-noise ratio can be determined by dividing the first power by the second power. The result can be expressed on a logarithmic scale, for example, in dB or dB(A). It should be noted that the determined SNR applies to the detected wanted signal 115 and in the selected time domain of the representation. The method 200 can be executed continuously, so that the SNR can change over time with incoming samples.

[0050] To determine how reliably the predetermined wanted signal 115 could be detected, it can be helpful to consider the SNR. For example, if the SNR is low, the determination reliability may be low, so a response to the detected wanted signal 115 should be carried out with caution.

[0051] Fig. Figure 3 shows an example spectrogram 300. A time is plotted horizontally, and a frequency is plotted vertically. The intensity of a frequency represented at a given time in signal 125 is represented by a false color. At time 12 s, a tone signal with a frequency of approximately 800 Hz and several harmonics is clearly visible. A little later, this signal is replaced by a tone signal with a frequency of approximately 500 Hz. This repetitive alternation can be visually recognized throughout the entire spectrogram 300. The underlying wanted signal 115 is an alternating signal with two tones, also known as a siren.

[0052] However, the wanted signal 115 is not clearly visible from the beginning, but only from approximately 9 s onward. Prior to this, the ratio between the power of the wanted signal 115 and the noise signal 120 is too small to allow a clear distinction. This can be seen from the fact that sections of the wanted signal 115 differ only slightly from sections of the noise signal 120.

[0053] Fig. 4 shows an input signal 400 for a classifier for a predetermined useful signal 115 in the time range 10 s to 12 s. A different color scale than in Fig. 3 is selected for the false color display. In the vertical direction, the same scale is used as in Fig. 3 used.

[0054] Fig. 5 shows relevances 500 of sections of a spectrogram 300 for the detection of a predetermined wanted signal 115, also in a false color representation. Horizontal and vertical dimensions of the representation correspond to those of Fig. 4. It can be seen that for the detection of the wanted signal 115, primarily those sections were relevant which lie in the frequency range below 1,000 Hz and have essentially the (vertical) bandwidth of one of the audio signals.

[0055] Fig. Figure 6 shows a specific signal-to-noise ratio 600 curve of a wanted signal 115 in a signal 125. The SNR 600 was determined by dividing, at different times, the first average power of the sections relevant for detection by the second average power of the sections not relevant for detection. It should be noted that the representation of SNR values ​​600 over the entire time range of the signal 125 in Fig.3. Before time 2 s, for example, the SNR 600 cannot be determined because the useful signal 115 has not yet been detected.

[0056] In one embodiment, the wanted signal 115 can be detected at a higher SNR 600 and then assumed for a range of a lower SNR 600, so that an SNR 600 can be determined even in a range in which the wanted signal 115 could not be directly detected. This may enable a retroactive SNR determination. Reference symbol 100 systems 105 Motor vehicle 110 Sound source 115 useful signal 120 noise signal 125 Signal 130 Device 135 Sensor, Microphone 140 processing facility 145 storage 150 interface 200 procedures 205 Capture signal 210 Select time period 215 Determine frequency spectrum 220 create graphical representation 225 detect predetermined wanted signal Remove 230 uncharacteristic sections 235 relevant sections determine 240 determine first performance Determine 245 remaining sections 250 second performance determine 255 Determine SNR 300 spectrograms 400 input signal 500 relevances 600 signal-to-noise ratio

Claims

[1] Method (200) for determining a signal-to-noise ratio (600) of a signal (125) comprising a superposition of a useful signal (115) and a noise signal (120); the method (200) comprising the following steps: - determining (215) a frequency spectrum of the signal (125); - detecting (225) a predetermined useful signal (115) of the frequency spectrum of the signal (125); - determining (235, 245) a first section of the frequency spectrum of the signal (125) that is relevant for the detection and a second section that is irrelevant for the detection; and - Determining (255) the signal-to-noise ratio (600) with respect to a first power of the signal (125) at the first section and a second power of the signal (125) at the second section. [2] The method (200) of claim 1, wherein the frequency spectrum represents a temporal progression of the signal (125). [3] The method (200) of claim 1 or 2, wherein the relevance of all sections for recognition is determined (235, 245); and wherein a section whose relevance reaches a predetermined threshold is determined as relevant; and a section whose relevance is below the threshold is determined as irrelevant. [4] The method (200) of claim 3, wherein the threshold comprises a predetermined proportion of a maximum relevance achieved by a section. [5] Method (200) according to one of the preceding claims, wherein performances of all sections are weighted with their relevance; and the first performance is determined on the basis of the weighted performances. [6] Method (200) according to one of the preceding claims, wherein the recognition is carried out by means of a convolutional neural network. [7] The method (200) of claim 6, wherein the relevance is determined with respect to an interpolation of an activation of a neuron to the signal (125). [8] Method (200) according to one of the preceding claims, wherein the relevance is determined with respect to a class activation map. [9] Method (200) according to one of the preceding claims, wherein the relevance is determined with respect to a layer-wise relevance propagation. [10] Method (200) according to one of the preceding claims, wherein sections belonging to the useful signal (115) are each replaced by an interpolation with respect to adjacent sections; and the second power is determined with respect to a resulting representation (300). [11] Method (200) according to one of claims 1 to 9, wherein sections belonging to the useful signal (115) are each replaced on the basis of a predetermined smooth function with respect to adjacent sections; and the second power is determined with respect to a resulting representation (300). [12] Method (200) according to claim 10 or 11, wherein the second power is determined on the basis of an averaging of the powers of the sections of the representation (300). [13] Method (200) according to claim 10 or 11, wherein for each section of the representation (300) a difference between the power of the useful signal (115) and the power of the noise signal (120) is determined; wherein the signal-to-noise ratio (600) is determined on the basis of an averaging of the differences. [14] Method (200) according to one of the preceding claims, wherein signal-to-noise ratios (600) are determined with respect to a plurality of predetermined useful signals (115). [15] Method (200) according to one of the preceding claims, wherein a portion lying in a frequency range that is atypical for the predetermined useful signal (115) is removed. [16] Method (200) according to one of the preceding claims, wherein the signal (125) comprises an acoustic signal (125). [17] Method (200) according to claim 16, further comprising controlling a motor vehicle (105) on the basis of a detected wanted signal (115) and the determined signal-to-noise ratio (600) of an acoustic signal (125) sampled in the environment of the motor vehicle (105). [18] Device (130) for determining a signal-to-noise ratio (600) of a signal (125) comprising a superposition of a useful signal (115) and a noise signal (120); the device comprising: - a sensor (135) for sampling the signal (125); and - a processing device (140) configured to determine a frequency spectrum of the signal (125); to detect a predetermined useful signal (115) in the frequency spectrum of the signal (125); to determine a first section of the frequency spectrum that is maximally relevant for detection and a second section that is minimally relevant for detection; and to determine the signal-to-noise ratio (600) with respect to a first power of the signal (125) at the first section and a second power of the signal (125) at the second section.

Images