Method and monitoring device for monitoring the operating state of a device

Abstract

Method for monitoring an operating state of a device with a rotating component, wherein a structure-borne sound signal of the device is measured and a spectrum (X) of the structure-borne sound signal or an envelope of the structure-borne sound signal is determined, characterized by the fact that to determine the operating state of the device, the shape of the spectrum (X) is quantitatively recorded, whereby at least one spectral audio characteristic (M1, M2) of the determined spectrum (X) is determined, where the spectral audio characteristic is the spectral flatness, the spectral skewness, the spectral crest factor, the spectral centroid, the spectral dispersion, the spectral decrease, the spectral slope or the spectral roll-off.

Description

[0001] The invention relates to a method for monitoring the operating state of a device. Furthermore, the invention relates to a monitoring device for monitoring the operating state of a device.

[0002] Such monitoring methods and systems are used primarily to monitor equipment with rotating components, for example, to detect faults in rolling bearings or gears at an early stage. This enables timely planning of maintenance measures and reduces the likelihood of unplanned downtime.

[0003] Damage to devices with rotating components can often be identified by the fact that the operating noise becomes louder or more jerky. To detect fault conditions, a structure-borne sound signal is therefore usually measured, and certain characteristics are derived from this structure-borne sound signal. Typically, characteristics are derived here that describe the signal power or the signal shape, such as the root mean square value, the peak value, the crest factor or the kurtosis. However, it has been found that the early detection of fault conditions based on these characteristics is often not satisfactory, so that fault conditions are detected too late or only when an unplanned downtime of the device can no longer be avoided. This impairs the availability of the device.

[0004] Ulrich Klein's book, Vibration Diagnostic Assessment of Machinery and Systems (2nd edition, Düsseldorf: Stahleisen GmbH, 2000), describes a method for monitoring the operating condition of a device with a rotating component.

[0005] Against this background, the object is to provide a method and a monitoring device for monitoring an operating state of a device, which enables increased availability of the monitored device.

[0006] The object is achieved by a method for monitoring an operating state of a device, in particular a device with a rotating component, wherein a structure-borne sound signal of the device is measured, a spectrum of the structure-borne sound signal or an envelope of the structure-borne sound signal is determined, and at least one spectral audio feature of the determined spectrum is determined to determine an operating state of the device.

[0007] In the method according to the invention, a spectrum of the structure-borne sound signal or a spectrum of an envelope of the structure-borne sound signal is determined to determine the operating state of the device. The shape of this spectrum characterizes the structure-borne sound signal independently of the volume and pitch. In devices with rotating components, the volume and / or pitch of the structure-borne sound signal is often dependent on the rotational speed of the rotating component. In order to quantitatively record the shape of the spectrum, the method determines at least one spectral audio feature of the previously determined spectrum. Thus, the operating state can be determined independently of the volume, pitch and / or rotational speed. Damage-related changes in the structure-borne sound signal can be detected earlier based on the shape of the spectrum, i.e. on the basis of the determined spectral audio features.This enables timely planning of maintenance measures and reduces the likelihood of unplanned downtimes, thus increasing the availability of the device.

[0008] The structure-borne sound signal is preferably measured using a structure-borne sound sensor. The structure-borne sound sensor is preferably arranged on a housing of the device so that it can measure vibrations of the housing, in particular a housing surface. The structure-borne sound sensor can measure acceleration and / or velocity and / or vibration displacement and derive a structure-borne sound signal from this.

[0009] A spectrum within the meaning of the invention is understood to be a frequency spectrum, in particular an amplitude spectrum, of the structure-borne sound signal or the envelope of the structure-borne sound signal. The spectrum is preferably generated by a Fourier transformation of the structure-borne sound signal or its envelope. Particularly preferably, the spectrum is generated by a discrete Fourier transformation of the structure-borne sound signal or its envelope, so that the spectrum is discrete and can be further processed digitally without analog-to-digital conversion.

[0010] The inventive embodiment of the method provides that the spectral audio feature is the spectral flatness, the spectral skewness, the spectral crest factor, the spectral centroid, the spectral spread, the spectral decrease, the spectral slope, or the spectral roll-off. These spectral audio features are defined as described below. X(k,n) denotes the (discrete) spectrum. k is the frequency line index, and n denotes the block index. K stands for the block length of the structure-borne sound signal, which corresponds to the number of samples in a block. The structure-borne sound signal preferably comprises exactly one block. With σ |X| is the standard deviation of the magnitude spectrum and with µ |X| the mean value of the magnitude spectrum.

[0011] Spectral flatness (English: spectral flatness) v SF is the ratio of the geometric mean and the arithmetic mean of the magnitude spectrum. The spectral flatness v SF is calculated according to equation 1. vSF(n)=∏K=0K2−1|X(k,n)|K22K⋅∑k=0K2−1|X(k,n)|

[0012] A very flat spectrum indicates a structure-borne sound signal with low contrast or a high noise content. A high value indicates a flat spectrum, while a low value indicates a non-flat spectrum.

[0013] Spectral skewness v SSK describes the symmetry of the spectral magnitude values ​​of the spectrum around the arithmetic mean µ |X| of the spectrum and is defined by equation 2. vSSK(n)=2∑k=0K2−1(|X(k,n)|−μ|x|)3K⋅σ|X|3

[0014] Positive values ​​of v SSKindicates an asymmetric distribution with a center of gravity to the left of the center, and negative values ​​indicate an asymmetric distribution with a center of gravity to the right of the center. If v SSK equals zero, the spectrum is symmetrical.

[0015] The spectral crest factor v Tsc describes the ratio of the maximum of the magnitude spectrum to the sum of the magnitude spectrum. It is defined by Equation 3. vTSC(n)=max0≤k≤K2−1|X(k,n)|⋅∑k=0K2−1|X(k,n)|

[0016] The value of the spectral crest factor is in the range 2K≤vTSC(n)≤1. Low values ​​indicate a flat spectrum and high values ​​indicate a sinusoidal spectrum.

[0017] The spectral centroid (vsc) defines the center of the spectral energy of the structure-borne sound signal. It is defined as the frequency-weighted sum of the spectrum, normalized by its unweighted sum, see Equation 4. vSC(n)=∑k=0K2−1k⋅|X(k,n)|2∑k=0K2−1|X(k,n)|2

[0018] The result of vsc is a frequency line index in the range 0≤vSC(n)≤K2−1. High vsc values ​​indicate a large proportion of high frequencies. A low value indicates a large proportion of low frequencies.

[0019] The spectral spread (vss) can be understood as the standard deviation of the power spectrum around the spectral center. It is defined by Equation 5. vSS(n)=∑k=0K2−1(k−vSC(n))2⋅|X(k,n)|2∑k=0K2−1|X(k,n)|2

[0020] The result of vsc is a frequency line index in the range 0≤vSS(n)≤K4. Low values ​​indicate a concentration of spectral energy in a specific frequency range.

[0021] The spectral decrease v SD is an approximate value for the steepness of the decay of a spectrum envelope over frequency. The spectral decrease V SD is defined by equation 6. vSD(n)=∑k=1K2−11k(|X(k,n)|−|X(0,n)|)∑k=1K2−1|X(k,n)|

[0022] The result of v SD is in the range v SD (n) ≤ 1.

[0023] The spectral slope V SSI is a measure of the steepness of the spectrum shape. It is calculated as a linear approximation of the magnitude spectrum according to Equation 7. VSSl(n)=∑k=0K2−1(k−μk)(|X(k,n)|−μ|X|)∑k=0K2−1(k−μk)2

[0024] Finally, the spectral roll-off (English spectral roll off) v SRthe frequency bin below which a given percentage p of the total sum of the magnitudes of the structure-borne sound spectrum lies, see equation 8.

[0025] Equation 8: vSR(n)=i|∑k=0i|X(k,n)|=p⋅∑k=0K2−1|X(k,n)|

[0026] Typical values ​​for p are 0.85 (corresponding to 85%) or 0.95 (corresponding to 95%). Low values ​​indicate that the spectrum contains only insignificant high-frequency components.

[0027] According to an advantageous embodiment of the method, the determined spectral audio feature is compared with a predefined threshold value, and if the threshold value is exceeded or undershot, a fault condition of the device is detected. Preferably, an acoustic and / or visual warning can be generated if a fault condition is detected. It is particularly advantageous if the determined spectral audio feature is compared with a predefined threshold value, and a fault condition is only detected if the threshold value is repeatedly exceeded or undershot. This can prevent individual events from triggering a false alarm.

[0028] According to an advantageous embodiment of the invention, several spectral audio features of the determined spectrum are determined to determine an operating state of the device. This can further improve the early detection of damage-related changes. This is because the determination of several spectral audio features enables a multivariate analysis of the determined spectral audio features, in which relationships between two or more spectral features can also be monitored. The spectral audio features are preferably selected from a list comprising spectral flatness, spectral skewness, spectral crest factor, spectral center of gravity, spectral dispersion, spectral decrease, spectral slope, and spectral roll-off. Preferably, three, four, five, six, seven, or eight spectral audio features of the spectrum are determined.

[0029] A further advantageous embodiment of the method is one in which, to determine the operating state of the device, a time-domain characteristic value of the structure-borne sound signal is additionally determined, preferably the zero-crossing rate or the predictivity ratio. Preferably, several time-domain characteristic values ​​of the structure-borne sound signal are determined, in particular the zero-crossing rate and the predictivity ratio of the structure-borne sound signal.

[0030] The zero crossing rate (vzc) is the number of sign changes in consecutive blocks of samples of the structure-borne sound signal. The zero crossing rate (vzc) is calculated according to Equation 9, where the sign function is defined according to Equation 10 and x(i - 1) = 0 is used for initialization if x(i - 1) does not exist. Here, i represents s (n) for the initial sample and i e(n) for the final sample. The larger the zero-crossing rate vzc, the more high-frequency components are contained in the structure-borne sound signal. vZC(n)=12⋅K∑i=is(n)ie(n)|sign[x(i)]−sign[x(i−1)]| sign[x(k)]={1,if x(i)>00,if x(i)=0−1,if x(i)<0

[0031] The predictivity ratio v Tp is a measure of how well the structure-borne sound signal can be predicted by a linear prediction or a zero-order prediction. Each sample i is compared using a previous sample and a prediction coefficient b j calculated according to equation 11. The predictivity ratio v Tp is calculated according to equation 12. Here, i s (n) for the initial sample and i e (n) for the final sample. x^(i)=∑j=1Obj⋅x(i−j) vTp(n)=∑i=is(n)ie(n)(x(i)−x^(i))2∑i=is(n)ie(n)x2(k)

[0032] According to an advantageous embodiment of the invention, in a training phase in which the device is in a predetermined good state, one, preferably several, spectral audio features are determined and stored. According to the invention, the spectral audio features are one or more spectral audio features selected from the list comprising spectral flatness, spectral skewness, spectral crest factor, spectral centroid, spectral scatter, spectral decline, spectral slope, and spectral roll-off. Preferably, three, four, five, six, seven, or eight spectral audio features of the spectrum are stored. Preferably, scatter ranges of the spectral audio features are learned and stored during the training phase. Preferably, a time-domain characteristic value of the structure-borne sound signal is additionally determined and stored, in particular the zero-crossing rate and / or the predictivity ratio of the structure-borne sound signal.

[0033] In this context, it is advantageous if, in a monitoring phase following the training phase, one, preferably several, spectral audio features are compared with the respective spectral audio feature stored in the training phase. By comparing the spectral audio features during the monitoring phase with the spectral audio features of the good state learned in the training phase, the degree of change compared to the good state can be quantified. Preferably, a time-domain characteristic value of the structure-borne sound signal is additionally compared with the respective time-domain characteristic value stored in the training phase, in particular the zero-crossing rate and / or the predictivity ratio of the structure-borne sound signal.

[0034] Preferably, a measure of the deviation of the operating state of the device from its normal state is determined. Particularly preferably, a fault condition is detected when the determined measure of the deviation of the operating state from its normal state exceeds a predetermined threshold. Upon detection of a fault condition, for example, an acoustic and / or visual warning can be generated and / or an emergency shutdown of the device can occur.

[0035] An advantageous embodiment of the method provides that in a training phase, in which the device is in a predetermined good state, several spectral audio features are determined and stored. In a monitoring phase following the training phase, these several spectral audio features are determined, each determination of these spectral audio features defining a further point in the multidimensional feature space, wherein a Mahalanobis distance is calculated for a point determined in the monitoring phase. Preferably, in the training phase, a time-domain characteristic value of the structure-borne sound signal is additionally determined and stored. In the monitoring phase, this time-domain characteristic value is determined, and the determination of the spectral audio features and the time-domain characteristic value defines a further point.The zero-crossing rate and / or the predictivity ratio of the structure-borne sound signal are used as time-domain values. The Mahalanobis distance of a point determined during the monitoring phase is a measure of this point's distance from the good state. The Mahalanobis distance can be determined according to Equation 13, where x denotes a feature vector of the spectral audio features and, if applicable, the additional time-domain parameters of the structure-borne sound signal, µ denotes the mean value, and S denotes the covariance matrix. r=(x−μ)TS−1(x−μ)

[0036] The covariance matrix S and the mean value µ are determined during the training phase based on the spectral audio features and, if applicable, the additional time-domain characteristics of the structure-borne sound signal and represent the distribution of the features for the known good condition of the machine. The feature vector x is calculated from the current structure-borne sound signal and describes the position of the current measurement point in the feature space. The Mahalanobis distance r of the current measurement represents the Euclidean distance of the current measurement point from the center µ of the distribution of the training data, scaled to account for the variance of the training data.

[0037] According to an alternative, advantageous embodiment, it is provided that in a training phase in which the device is in a predetermined good state, a plurality of spectral audio features and preferably a time domain characteristic value are determined and stored, wherein each determination of these spectral audio features and preferably of the time domain characteristic value defines a point in a multi-dimensional feature space, wherein in a monitoring phase following the training phase, these plurality of spectral audio features and preferably of the time domain characteristic value are determined, wherein each determination of these spectral audio features and preferably of the time domain characteristic value defines a further point in the multi-dimensional feature space, wherein for a point determined in the monitoring phase, a local density and an average density in the vicinity of the point are calculated.To determine the points in the feature space, in addition to the spectral audio features, time-domain parameters of the structure-borne sound signal, in particular the zero-crossing rate and / or the predictivity ratio of the structure-borne sound signal, can be used. Based on the local density and the mean density, the ratio of the mean density in the vicinity of a point to the local density at this point can be calculated as the kNN score. This results in values ​​close to 1 for points that lie within the range of the training data set. As the distance of the points from the training data set increases, the local density at the respective point decreases and the kNN score increases. This enables improved detection of damage-related error states for training data sets that have a more complex shape in the feature space.

[0038] To achieve the object mentioned at the outset, a monitoring device for monitoring an operating state of a device, in particular a device with a rotating component, is further proposed, comprising a structure-borne sound sensor for measuring a structure-borne sound signal of the device, and comprising an evaluation device which is configured to determine a spectrum of the structure-borne sound signal or an envelope of the structure-borne sound signal and to determine at least one spectral audio feature of the determined spectrum in order to determine an operating state of the device.

[0039] The monitoring device offers the same advantages as those already described in connection with the method according to the invention.

[0040] The advantageous features and embodiments explained in connection with the method can also be used in the monitoring device alone or in combination.

[0041] Further details and advantages of the invention will be explained below with reference to the exemplary embodiment illustrated in the drawings. Herein: Fig. 1 an exemplary temporal course of an effective value of a structure-borne sound signal; Fig. 2 a temporal progression of the crest factor of the structure-borne sound signal from Fig. 1; Fig. 3 a spectrum of an envelope of a structure-borne sound signal measured in a good condition; Fig. 4 a spectrum of an envelope of the structure-borne sound signal from Fig. 3 in an error state; Fig. 5 shows a further exemplary temporal progression of an effective value of a structure-borne sound signal and a result of a condition detection; Fig. 6 a representation of measured values ​​of a first spectral audio feature; Fig. 7 shows a representation of measured values ​​of a second spectral audio feature; Fig. 8 a representation of a feature space of the measured values ​​from Fig. 6 and Fig. 7; Fig. 9 a representation of a feature space with an ellipse drawn for a Mahalanobis distance r = 3; Fig. 10 a representation of the Mahalanobis distance of the measuring points from Fig. 9; Fig. 11 another representation of a feature space; and Fig. 12 a representation of the kNN score of the measurement points from Fig. 11.

[0042] In the Fig. Figure 1 shows an example of a time course of an effective value E of a structure-borne sound signal from a device designed as a wind turbine. The effective value was measured once a day over a period of more than two years. Both daily and seasonal fluctuations are evident. Starting in mid-2014, an increase in the effective value E is observed, which suggests a change in the operating state of the device. However, the fluctuation of the effective value E makes early detection difficult. Fig. Figure 2 shows a representation of the peak factor S of the same structure-borne sound signal. It can be seen that the damage in 2014 had no discernible influence on the peak factor S. Consequently, the fault condition can only be detected relatively late or not at all using the state-of-the-art methods that monitor the effective value E or the peak factor S of a structure-borne sound signal.

[0043] In the Fig. Figure 3 shows a spectrum of an envelope of a structure-borne sound signal of a device in good condition. Fig. Figure 4 shows a spectrum of the envelope of the structure-borne sound signal in a fault condition corresponding to the onset of damage. In the good condition according to Fig. 3, a harmonic structure can be seen in the envelope spectrum, which can be traced back to a gear pump adjacent to the monitored device. As shown in Fig. 4, with the onset of the defect, additional excitation caused by the damage masks the gear pump signal, and the envelope spectrum shows a clearly visible change. State-of-the-art monitoring of the RMS value would not have detected the damage at this point.

[0044] For monitoring purposes, a method for monitoring an operating state of the device is therefore used, wherein a structure-borne sound signal of the device is measured, wherein a spectrum of the structure-borne sound signal or an envelope of the structure-borne sound signal is determined, and several spectral audio features of the determined spectrum are determined to determine an operating state of the device.

[0045] Numerous spectral audio features have been developed for audio signal processing that can be used to classify and segment musical pieces. These features serve to describe the shape of the spectrum. The shape of the spectrum is closely linked to the timbre, which is typical of an instrument, for example. Timbre characterizes the signal independently of volume and pitch. In machines with rotating components, many components of structure-borne sound vary in power and frequency, corresponding to volume and pitch, with the speed. These audio features are therefore well suited for characterizing structure-borne sound signals, with the advantage of being low dependence on operating conditions.

[0046] Several of the following spectral audio features are used as spectral audio features to characterize the structure-borne sound signals: spectral flatness, spectral skewness, spectral crest factor, spectral centroid, spectral spread, spectral decrease, spectral slope and spectral roll-off.

[0047] In addition, the monitoring procedure determines one or two time-domain parameters of the structure-borne sound signal. These time-domain parameters are preferably the zero-crossing rate and / or the predictivity ratio.

[0048] The spectral audio characteristics or the time-domain characteristics can be compared with a predefined threshold value, whereby a fault condition of the device is detected if the threshold value is exceeded or undershot. Furthermore, it is possible for the respective determined spectral audio characteristic or the time-domain characteristic to be compared with a predefined threshold value, and a fault condition is only detected if the threshold value is repeatedly exceeded or undershot. This can prevent individual events from triggering a false alarm. Such univariate monitoring will be described below using the illustration in Fig. 5. The upper section shows an example of the time course of an effective value E of a structure-borne sound signal, with a threshold value T being indicated. In the present case, a single detection of exceedance of the threshold value T results in an error condition F=1, see the lower section of the Fig. 5.

[0049] Preferably, however, a multivariate detection of the operating state of the device is carried out. In this case, first, in a training phase in which the device is in a predetermined good state, several of the aforementioned spectral audio features are determined and stored. For example, two, three, four, five, six, seven, or eight of these spectral audio features of the spectrum can be determined and stored. In addition, the zero-crossing rate and / or the predictivity ratio are preferably determined and stored. In a monitoring phase following the training phase, the same spectral audio features and, if applicable, the zero-crossing rate and / or the predictivity ratio are then determined and compared with the data determined in the training phase.

[0050] In Fig. 6 shows several measured values ​​N of a first feature M1. Fig. Figure 7 shows a corresponding representation of several measured values ​​N of a second feature M2. The features M1, M2 can be the aforementioned spectral audio features and / or the zero-crossing rate and / or the predictivity ratio. Fig. Figure 8 shows the second feature M2 plotted against the first feature M1. The measurement data contains two anomalous measurement points a and b. In addition, boundaries are drawn at a distance of three standard deviations from the sample center. In the univariate analysis of feature 1, the anomalous measurement points cannot be identified because both points are unremarkable with regard to feature 1 with regard to the entire scatter range of feature 1. In the middle graph, measurement point a can be identified as an anomaly because this point lies significantly outside the scatter range of the remaining data. In the feature space representation (right), it is clearly visible that the first feature M1 and the second feature M2 are correlated with each other. Furthermore, both measurement points a and b can be easily identified visually as anomalies. This fact can be used by multivariate methods to improve condition monitoring.For this purpose, the scatter ranges of different features are learned in a training phase, taking into account their correlations for the good state of the monitored machine. By comparing the features of a current measurement with the learned good state, the degree of abnormality of the current measurement can be quantitatively assessed in the feature space and used for monitoring. Two examples of this approach are presented below.

[0051] Based on the illustrations in Fig. 9 and Fig. 10, the multivariate monitoring of the operating condition of a device based on monitoring the Mahalanobis distance r will be described below. Monitoring the Mahalanobis distance r (see Equation 13) is a comparatively simple method for learning the normal state and detecting anomalies. Calculating the Mahalanobis distance requires knowledge of the feature vector x, the mean µ, and the covariance matrix S. The covariance matrix S and the mean µ are determined in the training phase using the features, for example, M1 and M2, of the structure-borne sound signals and represent the distribution of the features for the known good condition of the machine. Feature vector x is calculated in the monitoring phase from the current structure-borne sound signal and describes the position of the current measurement point in the feature space.The Mahalanobis distance r of the current measurement represents the Euclidean distance of the current measurement point from the center µ of the distribution of the training data, scaled taking into account the scatter of the training data.

[0052] In Fig. Figure 9 illustrates the procedure assuming a multivariate normal distribution as the training data set and two exemplary abnormal measurement points a and b. The dashed line in Fig. Figure 9 represents an ellipse on which all points with a Mahalanobis distance of r = 3 from the center µ of the training data are located. In the one-dimensional case, this would correspond to a distance of three standard deviations from the mean. The graph in Fig. Figure 10 shows that the anomalous measurement points a and b can be clearly distinguished from the training data using the Mahalanobis distance r. If the distribution of the training data can be sufficiently well described by a multidimensional generalized ellipse, the Mahalanobis distance r can be used to summarize a large number of features and characterize the anomalous nature of a measurement point. The time course of this characteristic value can be monitored for limit violations and used to detect anomalous machine conditions.

[0053] An alternative method for multivariate monitoring is described below using the illustrations in Fig. 11 and Fig. 12. The previously described distance-based method based on the Mahalanobis distance r takes into account the distribution of the training data summarized by the mean and the covariance matrix. For training data sets that have a more complex shape in the feature space, anomaly detection can be improved by taking local properties of the distribution into account. For this purpose, a simplified method based on the k-nearest neighbor (kNN) approach is proposed, based on the local outlier factor (LOF). For a point P, the local density can be determined as the inverse of the mean distance to its k nearest neighbors in the training data set. Analogously, the mean density in the vicinity of point P can be calculated as the mean of the local densities of its k nearest neighbors N 1..kcan be determined. Using the ratio of the average density in the vicinity of point P to the local density at point P as the kNN score, values ​​close to 1 result for measurement points P that lie within the range of the training data set. As the distance of point P from the training data set increases, the local density at this point decreases, and the kNN score increases.

[0054] Fig. 11 and Fig. 12 illustrate the method using the joint distribution of two features of a real training data set and two conspicuous measurement points a and b. In the feature space ( Fig. 11), the conspicuous points a and b are visually clearly distinguishable from the training data. The kNN score for the training data ( Fig. 12) fluctuates around the value 1. The score values ​​for points a and b differ significantly from the training data. The boundary shown at a score value of approximately 4 corresponds to the course of the dashed line in the feature space ( Fig. 11). In areas with high training data density, the decision boundary runs closer to the training data set. In areas with low data density and consequently less information about the boundary of the good condition, the boundary runs at a greater distance. The density-based approach thus better accounts for the specific distribution of the training data set. Analogous to the Mahalanobis distance, the time course of the kNN score can also be monitored for threshold violations to detect abnormal machine conditions.

[0055] In the methods explained above for monitoring the operating state of a device, in particular a device with a rotating component, a structure-borne sound signal of the device is measured, and a spectrum of the structure-borne sound signal or an envelope of the structure-borne sound signal is determined. Furthermore, at least one spectral audio feature of the determined spectrum is determined to determine the operating state of the device.The method can be carried out with a monitoring device for monitoring an operating state of a device, in particular a device with a rotating component, with a structure-borne sound sensor for measuring a structure-borne sound signal of the device and with an evaluation device which is designed to determine a spectrum of the structure-borne sound signal or an envelope of the structure-borne sound signal and to determine at least one spectral audio feature of the determined spectrum in order to determine an operating state of the device. List of reference symbols a, b measuring points E Effective value f frequency F Error condition M1, M2 feature N Sample number S crest factor t time T threshold X Spectrum

Claims

[1] Method for monitoring an operating state of a device with a rotating component, wherein a structure-borne sound signal of the device is measured and a spectrum (X) of the structure-borne sound signal or an envelope of the structure-borne sound signal is determined, characterized by , that to determine the operating state of the device, the shape of the spectrum (X) is quantitatively recorded, wherein at least one spectral audio feature (M1, M2) of the determined spectrum (X) is determined, where the spectral audio characteristic is spectral flatness, spectral skewness, spectral crest factor, spectral centroid, spectral dispersion, spectral decrease, spectral slope or spectral roll-off. [2] Method according to claim 1, characterized by that, in order to determine the operating state of the device, several spectral audio features (M1, M2) of the determined spectrum (X) are determined. [3] Method according to one of the preceding claims, characterized by that in order to determine the operating state of the device, a time-domain characteristic value of the structure-borne sound signal is additionally determined, preferably the zero-crossing rate and / or the predictivity ratio of the structure-borne sound signal. [4] Method according to one of the preceding claims, characterized by that in a training phase in which the device is in a predetermined good condition, one or more spectral audio features are determined and stored, wherein preferably a time domain characteristic value of the structure-borne sound signal is additionally determined and stored. [5] Method according to claim 4, characterized bythat in a monitoring phase following the training phase, one or more spectral audio features (M1, M2) are compared with the respective spectral audio feature stored in the training phase, wherein preferably additionally a time domain characteristic value of the structure-borne sound signal is compared with the respective time domain characteristic value stored in the training phase. [6] Method according to claim 5, characterized by that a measure of the deviation of the operating state of the device from its good state is determined, in particular wherein a fault state is detected if the measure of the deviation exceeds a predetermined threshold value. [7] Method according to one of the preceding claims characterized byin that in a training phase in which the device is in a predetermined good state, a plurality of spectral audio features, and preferably a time-domain characteristic value, are determined and stored, wherein in a monitoring phase following the training phase, these plurality of spectral audio features (M1, M2) and preferably the time-domain characteristic value are determined, wherein each determination of these spectral audio features (M1, M2) and preferably of the time-domain characteristic value defines a further point in the multi-dimensional feature space, wherein a Mahalanobis distance is calculated for a point determined in the monitoring phase. [8] Method according to one of claims 1 to 6, characterized byin that in a training phase in which the device is in a predetermined good state, a plurality of spectral audio features (M1, M2) and preferably a time-domain characteristic value are determined and stored, wherein each determination of these spectral audio features and preferably of the time-domain characteristic value defines a point in a multi-dimensional feature space, wherein in a monitoring phase following the training phase, these plurality of spectral audio features (M1, M2) and preferably of the time-domain characteristic value are determined, wherein each determination of these spectral audio features (M1, M2) and preferably of the time-domain characteristic value defines a further point in the multi-dimensional feature space, wherein for a point determined in the monitoring phase, a local density and an average density in the vicinity of the point are calculated. [9] Monitoring device for monitoring an operating state of a device, in particular a device with a rotating component, with a structure-borne sound sensor for measuring a structure-borne sound signal of the device, characterized by an evaluation device which is designed to determine a spectrum (X) of the structure-borne sound signal or an envelope of the structure-borne sound signal and to quantitatively record the shape of the spectrum (X) in order to determine the operating state of the device, wherein at least one spectral audio feature (M1, M2) of the determined spectrum (X) is to be determined, where the spectral audio characteristic is spectral flatness, spectral skewness, spectral crest factor, spectral centroid, spectral dispersion, spectral decrease, spectral slope or spectral roll-off.

Images