DE102013013551A1 - Method for detecting significantly changed signals against a background of dominant disturbances - Google Patents

Abstract

The method according to the invention of significantly changed signals on devices with periodically repeating, uniform or uniformly accelerated movements comprises the following steps:
- Recording the signals of the device by means of a sensor
- Determining the frequencies of individual amplitudes of the signals
- Calculation of the probabilities of individual amplitudes
Calculation of the entropy values, in particular the Shannon entropy, the Kolmogorov-Sinai entropy and / or the Renyi entropy from the probabilities
- Comparison of the entropy of the current time unit with the entropy of previous time units and / or with the entropy of reference data

Description

The invention generally relates to a method and a device which continuously records measured variables (hereinafter: "signals") and can recognize, evaluate and signal changes in the signal properties. The aim is to detect or predict degradations, disturbances or damage of devices with periodically or uniformly or even accelerated moving parts.

[State of the art]

A variety of individual methods are known, which allow for special applications to deduce from measured signals of various kinds on operating and wear conditions of moving parts:
It is possible, on the basis of the acoustic frequency spectrum (structure-borne noise data), which was measured on a test bench of devices with periodically repeating movements, for example, for sliding bearings, to recognize whether a so-called stick-slip effect is taking place during the test, the Represents malfunction of the bearing. A typical example of this effect is a squeaky door hinge.

It is known that the development of such an effect can already be detected in advance for the diagnosis of bearings, gears and other devices by means of structure-borne noise.

There are a number of established methods in this area. These can be roughly divided into two categories:

1. Statistical methods

It is generally known that damage in the initial stage, for example in rolling bearings, manifests itself by the fact that the curvature of the distribution density of the bearing vibrations changes, which can be detected by the so-called kurtosis. The kurtosis is a simple statistical index and therefore an analysis accessible.

2. Methods of spectral analysis

Here, the measured vibration or sound signal is first subjected to a so-called Fourier transform to determine its frequency spectrum. This can then be compared, for example, with a reference spectrum or examined for changes in frequencies that can be determined in advance due to bearing or gear geometry.

However, these methods are specially adapted to their environment and often largely "blinded" to changes outside the considered set of frequencies.

3. Decompositions into orthogonal and other components

If several signals recorded in parallel are available in the form of time series, these can be decomposed by suitable decomposition into the same number of statistically independent component time series.

Various concepts of statistical independence can be used here for further subdivision into principal component analysis, in which the components are only decorrelated, and methods with higher requirements for independent component analysis ("Independent Component Analysis") ,

All of these methods have specific disadvantages in the current application:
The kurtosis is known to be affected by noise, damage, but it shows little robust and is statistically difficult to evaluate, as their distribution is sensitive to non-specific changes. The spectral analysis usually makes strong assumptions about the operating environment, is very computationally intensive and more or less blind to weak statistical signals with unstable phasing.

There are few asymptotic results about the statistical properties of principal components, beyond which little is known, and for stronger concepts of statistical independence, the distributions of the components themselves represent variables that must be adapted depending on certain constraints, so here hardly any general statements are possible.

From the DE 199 43 689 A1 For example, a method and apparatus for monitoring and / or diagnosing moving machinery and / or machine parts is known. The aim of such measurement and evaluation is to detect as early as possible damage to a machine or a machine part such as a rolling bearing in order to replace this before its total failure if necessary.

[Task]

The object of the present invention is to avoid the disadvantages of the prior art.

The invention should be as universally applicable and therefore may make only a few assumptions about the respective operating environment. In addition, it should also be able to take into account weak statistical signals and detect changes in the signal even against a background of dominant disturbances. In this case, both direct measurement signals from machines and systems and sensors used therein as well as indirect, resulting from preprocessing in control units signals are evaluated or converted so that better than before a clear statement about the actual presence of damage or interference is possible.

Finally, this method should be able to be carried out so quickly that suddenly occurring damage and / or malfunctions can be detected without delay.

[Solution of the task]

The solution of these objects results from the features of the main claim, while advantageous embodiments and modifications of the invention are the dependent claims.

• – Aufnahme der Signale (bspw. akustische, elektrisch, optisch) der Vorrichtung mittels eines Sensors, das bedeutet: Aufnahme einer Menge von Signalamplituden
• – Ermittlung der Häufigkeiten einzelner Amplituden der Signale
• – Berechnung der Wahrscheinlichkeiten einzelner Amplituden
• – Berechnung der Entropiewerte, insbesondere der Shannon Entropie, der Kolmogorov-Sinai-Entropie und/oder der Renyi-Entropie aus den Wahrscheinlichkeiten
• – Vergleich der Entropie der aktuellen Zeiteinheit mit der Entropie von vorhergehenden Zeiteinheiten und/oder mit der Entropie von Referenzdaten

The task is solved by a method with the following properties:
A method of detecting significantly altered signals on devices having periodically repeating, uniform or uniformly accelerated motions, comprising the steps of:

• - Recording the signals (eg, acoustic, electrical, optical) of the device by means of a sensor, which means: recording a set of signal amplitudes
• - Determining the frequencies of individual amplitudes of the signals
• - Calculation of the probabilities of individual amplitudes
• Calculation of the entropy values, in particular the Shannon entropy, the Kolmogorov-Sinai entropy and / or the Renyi entropy from the probabilities
• - Comparison of the entropy of the current time unit with the entropy of previous time units and / or with the entropy of reference data

A device for detecting significantly different signals, which uses this method, has at least one signal input, which makes it possible to passively detect a signal and to supply it to a suitable further processing. This can be done, for example, by the analog-to-digital conversion of a voltage level at a high-impedance input.

In advantageous embodiments of the device, the device also has active elements in particular energy supply means to ensure the power supply of the sensors via constant voltage or constant current and / or receiving means to already elsewhere in digital form converted signals via suitable interfaces, such as wired serial, parallel or network -Interfaces, or even wireless optical (IrDA) or radio-based interfaces (Bluetooth, ultra-broadband, WLAN) to receive.

In a further embodiment, the device has an internal digital or analog calculating unit in order to further process the received signals.

For the processing of the signals, the device has at least one operating state.

In the so-called intervention mode, which represents the regular operating state, the statistical properties of the measured signals are compared with predefined and / or determined during operation of the device statistical properties (eg variance, mean and / or entropy) and possible significant deviations from determined this. The output of the data of the deviation to the outside and / or the information that a deviation is present through an optical, acoustic and / or electrical signal is referred to below as signaling.

In advantageous embodiments of the device, there is at least one further optional operating state, the so-called exploration operation. In this case, signals are usually not sent to the outside, but during this operating state, the statistical properties of all detected signals are first measured or calculated over a certain period of time and logged internally on an internal non-volatile memory. This serves to condition the device since comparative values for the intervention mode are generated here. Signal sections with significant deviations can also be output either in analog or in digital form via a suitable interface.

In further advantageous embodiments, the device has an optical and / or acoustic signal output means z. B. a separate screen to make the deviations directly representable.

For further signaling in intervention mode, the device has at least one signal output to provide the information of the significant deviations of one or more input signals from at least one, before the measurement or in the exploration mode defined statistical property (eg variance, mean and / or entropy) to send.

In further advantageous embodiments of the device, a more differentiated signaling is achieved by one of the following features or a combination of the following features:
The signaling sends the size of the threshold of significance of a statistical property, not just exceeding it. This happens, for example, by outputting a suitably scaled analog voltage value.

The signaling takes place in a direct output of voltage values via other digital interfaces, in order to communicate via external infrastructure z. B. to ensure a mobile network.

The signal data is either recorded on an internal non-volatile memory and / or transmitted in digital form via one of the above-mentioned interfaces.

To analyze the data, mathematical methods are used in both phases, which are described below.

In each method, the signal of the device (for example acoustic, optical or electrical) is detected in a first step. For this purpose, sensors such as vibration sensors, acceleration sensors, microphones, photosensors, voltage meters, ammeters or other suitable sensors are used.

For the entropy analysis, the data of the sensors is repeatedly recorded in blocks of the same size and an entropy measure is determined for each individual block.

The period over which the individual blocks extend is always the same and depends solely on the device to be monitored. It can move in the range of a few hundredths or tenths of a second when fast rotating parts are to be monitored, but for slow moving parts with a long period, the period could extend over several seconds or even longer periods of time. In principle, it is favorable to detect a number of 100-1000 data points per block in a resolution in which the individual measured values are not significantly influenced by converter noise, since this noise could artificially increase the entropy and thus possibly mask slight damage events.

The empirical probabilities for the occurrence of individual measured values are then determined from the measured values obtained and, in turn, the information content of the data. The information content of the signal determines its statistical significance. Here is an example: If only the same value is transmitted and thus its probability of occurrence equals 1, the information content of the signal is 0 and therefore minimal. If, on the other hand, a value occurs only once, it carries the maximum possible information.

The simplest information measure that can be determined from a block of N measured values with occurrence probability p _k is the so-called Shannon entropy S: = -Σp _k logp _k . If only one measured value occurs in a block, then S = 0, if all the characters are different, then the maximum value logN is reached.

If data from M sensors are available in parallel, they can be combined into tuples (x ₁ , ..., x _M ) and also from this an entropy can be calculated using the occurrence probabilities p _{k of} the blocks.

This entropy of the current time unit is compared with the entropy of previous time units and / or with the entropy of reference data. Reference data can be derived, for example, from the amplitudes of a device known as undamaged and / or interference-free. After this comparison, a status value is output.

Alternatively, the relative entropy can also be calculated as a distance measure between current measured values and reference measured values. In this case, the entropy with respect to the probability p _{t of} the measured value at time t and the reference value at time _{t is} calculated with probability q _t

The resulting entropy results in the so-called Kullback information, which describes, so to speak, the information gain associated with the new measurement.

für ein beliebiges β eine selektive Betonung oder Dämpfung des Beitrags niedriger Wahrscheinlichkeiten erlaubt oder Kolmogorov-Sinai-Entropien können auch zur Anwendung kommen und verlangen spezielle Parameteranpassungen.Other entropy measures like the Rényi entropy the

for any β, selective emphasis or attenuation of the contribution of lower probabilities is allowed, or Kolmogorov-Sinai entropies can also be applied and require special parameter adjustments.

The determined entropy now serves as a monitoring signal for whether the device is within a trouble-free operating range.

With a significant change in entropy, failure of the device, which is for example a bearing, is likely. But you get this information some time before the real failure. In this respect, there is still time to shorten maintenance periods.

• • den Mittelwert über mehrere Entropiewerte zu bilden, da sich die Verteilung einer solchen Summe einer Normalverteilung annähert, deren Parameter dann einer statistischen Analyse zugänglich sind,
• • bei zyklischen Abläufen eine Anzahl von Entropiewerten zu streichen, falls diese jeweils zum selben Zeitpunkt des Zyklus gestört werden und/oder
• • die tatsächliche Verteilungsfunktion der Entropie aus den empirisch gewonnenen Werten durch Bootstrapping oder Jackknifing zu approximieren und damit die Ergebnisqualität zu verbessern. Beim Jackknifing werden dabei nacheinander einzelne Messwerte verworfen und mit der so ausgedünnten Stichprobe wiederholt Werte für statistische Schätzer wie die Varianz berechnet, die sich dann im Idealfall einer t-Verteilung anpassen. Beim Bootstrapping werden aus den gewonnenen Messwerten durch wiederholtes Ziehen mit Zurücklegen neue Stichproben gewonnen, deren Parameter besser geschätzt werden können.

In a further embodiment, the entropy values are subjected to post-conditioning prior to the statistical analysis in order to ensure certain statistical properties or to eliminate disturbances caused by cyclical machine processes. Post-conditioning are

• • to form the mean over several entropy values, since the distribution of such a sum approaches a normal distribution, the parameters of which are then accessible to a statistical analysis,
• • to delete a number of entropy values in cyclical processes, if they are disturbed at the same time of the cycle and / or
• • to approximate the actual distribution function of entropy from the empirically obtained values by bootstrapping or jackknifing, thereby improving the quality of the results. In the case of jackknifing, individual measured values are rejected one after the other and with the thus thinned out sample values for statistical estimators such as the variance are calculated, which then ideally adapt to a t-distribution. During bootstrapping, new samples are obtained from the measured values obtained by repeated dragging with traversing, the parameters of which can be better estimated.

• • Erstens eine Mittelwertbildung der Messwerte, so dass sich schon vor der Bestimmung der empirischen Entropie annährend eine Normalverteilung eingestellt hat. Da für die Normalverteilung und verschiedene andere Verteilungen die Entropie analytisch bekannt ist, erfolgt damit ein Vergleich der empirisch ermittelten Werte mit dem analytischen Sollwert
• • Zweitens das Glätten des Datenstroms durch Filterung, insbesondere das Herausfiltern von im Vorfeld ermittelten Störfrequenzen
• • Drittens das Durchführen einer Fouriertransformation des Signals und Bestimmung der Entropie für ausgewählte Frequenzbereiche

In a further embodiment, the measured values are subjected to preconditioning in order to ensure certain statistical properties or to eliminate disturbances due to cyclical machine sequences. Preconditioning includes various measures that are taken individually or in combination. These are:

• • First, an averaging of the measured values, so that even before the determination of the empirical entropy, a normal distribution has been established. Since the entropy is analytically known for the normal distribution and various other distributions, a comparison of the empirically determined values with the analytical setpoint value is thus carried out
• Secondly, the smoothing of the data stream by filtering, in particular the filtering out of previously determined interference frequencies
• Third, performing a Fourier transform of the signal and determining the entropy for selected frequency ranges

embodiments

1. Single chip solution

In an exemplary embodiment, the method is implemented on a chip having input and output interfaces. This chip calculates the individual entropy values for the data of the input signal and outputs them as a data block. In this case, by a command to the chip, the parameters of the entropies, such as the number of data points to be used per entropy value N, for example N = 128, and the desired entropy can be set, for example the Shannon entropy. If a data stream with a sampling frequency of 48,000 Hz is now present, the chip calculates an entropy value s for every 128 data points and outputs it. The output is in this example with a frequency of 48,000 / 128 = 375 Hz.

2. Computer with Microform factor (NATRIX)

It is also possible to implement the method in an independent measuring and evaluation unit, which can be integrated into both existing and new systems. This computer in the form of a microformat as NumericAl Test, Reporting and Information boX (short NATRIX) has several analogue and digital interfaces that can be used to read in signals from several sensors and to output the corresponding evaluation via analog and digital output interfaces. The individual parameters (number of measured values per entropy N and parameters of the Rényi entropy as well as the type of entropy evaluation (Shannon, Rényi, Kolmogorov-Sinai or entropies or Kullback information respectively with or without post- or preconditioning) become for example set or adapted via a web interface of NATRIX.

The output is either directly as a data array of the entropy values, or as an individual signal that can be recognized by the control unit of the system in which the NATRIX is used, e.g. B. rectangular signal with the value 0 detected as a signal for no defect and 1 as a signal for detected defect.

In an exemplary embodiment, acquiring data having a resolution of, for example, 16 bits and a sampling frequency of 1 kHz is nothing other than detecting the data stream of a signal source that is 1000 times one of 65536 per second. Sends sign.

It will now be analyzed, for example, for each second, which characters were sent how often and this information is summarized in a Entropiemaß. This describes vividly speaking, the measure of "disorder" in this string. If the same sign is sent 1000 times in one second, the maximum order and the entropy is 0, if 1000 different characters are sent, then maximum disorder prevails and the entropy reaches a maximum value, for example log (1000).

In addition to the entropy described here, Shannon's entropy, there are various entropy measures, including, for example, the order of the characters, and since the measurements are naturally ordered, the entropy of the differences of adjacent values is analyzed and / or the measurement data is changed to others Way preconditioned. Over time, samples of entropy values are collected, which in turn are subjected to various statistical analyzes.

[Illustration legends and reference list]

In the following description, further aspects and embodiments of the present invention are disclosed. Furthermore, reference is made to the accompanying drawings which form a part hereof, and in which is shown by way of illustration one or more examples in which the invention may be practiced. This disclosure of the invention is not intended to limit the features or main elements of the invention to a specific embodiment. Rather, the various elements, aspects and features disclosed in the embodiments may be combined by one of ordinary skill in the art in various ways to achieve one or more advantages of the present invention. It should be understood that other embodiments may be utilized and structural or logical changes may be made without departing from the scope of the present invention. Like reference numerals designate corresponding like parts.

1 shows the entropy analysis (1a) and the probability analysis (1b) for a buzzing sound, which builds up slowly over a period of 15 minutes against a background of white noise. Shortly before the time of 600 seconds, the probability of equal variance decreases to just over 0% compared to earlier times, while the probability of a larger mean value rises to just under 100%.

2 shows the entropy analysis (2a) and the probability analysis (2b) for a patter noise, which slowly builds up over a period of 15 minutes against a backdrop of white noise. Shortly after 450 seconds, the probability of equal variance decreases to nearly 0% compared to earlier times, while the probability of a larger average increases to almost 100%.

3 shows the Entropieanalysen of 4 Hubfahrten a lifting table, each with a load, z. B. in the form of a product lifts. The current of the motor which moved the lift table was analyzed. These lift trips lasted 0.75 seconds each and then come to rest. In 3 out of the 4 lift runs (solid line), the trip runs smoothly, while in the 4th lift (dotted line) the load is not optimally placed on the lift table. Even at the first entropy value (> 0.5), a clear deviation can be seen in comparison to the normal running stroke (entropy value almost 0).

4 shows the entropy analysis of the angular velocity of a smooth, electric, wave. In one application it is known that undesired operating states of a machine lead to short deviations of this angular velocity. The angular velocity can be monitored by scanning the shaft with a magnetoresistive sensor which induces an eddy current through an integrated permanent magnet in the moving shaft. If the angular velocity of the shaft changes, the eddy currents and thus the sample value supplied by the sensor also change.

However, the output value supplied by the sensor can be additionally subjected to a more or less intense noise by various disturbing influences, such as amplifier noise or the grain structure in the structure of the shaft material. Nevertheless, with the entropy method detection of the deviation is possible even with strong noise.

5 shows the flow of the process in a flow chart

First, the signal of the device is picked up by means of a measuring sensor. This is the data entry into the system. Prior to further processing, it is optionally possible to precondition the data by averaging the measured values, filtering out previously determined interference frequencies and / or performing a Fourier transformation of the signal.

From this data follows the entropy analysis. This includes the determination of the frequencies of individual amplitudes and, consequently, the calculation of the probabilities of individual amplitudes. From these probabilities, the information content and thus the entropy values, in particular the Shannon entropy, the Kolmogorov-Sinai entropy and / or the Renyi entropy of the data are calculated.

Optionally, a further conditioning of this entropy data in the form of an averaging over several entropy values and / or approximation of the actual distribution function of the entropy from the empirically obtained values by bootstrapping or jackknifing can now be carried out.

Finally, the entropy of the current time unit is compared with the entropy of previous time units and / or with the entropy of reference data (setpoint value). This represents the evaluation of the data. Subsequently, a state value is output. This can be, for example, a binary value: 0 = entropy below critical value, machine can continue running. 1 = entropy over critical value machine must stop. But also a multi-level state value is possible.

6 shows the integration of NATRIX into an existing system.

The NumericAl Test, Reporting and Information boX (short NATRIX) has several analog and digital interfaces. It serves to detect the sensor data, to carry out the method according to the invention and to forward the state value to a control unit.

Via NATRIX, the process steps are output, i. H. the input of the entropy parameters, the reference values and also of termination conditions in which the device should be switched off with periodically recurring, uniform or uniformly accelerated movements (eg motor, pump, etc.).

Furthermore, the Natrix also serves for the external communication of the monitoring system and the storage of the data in exploration operation.

QUOTES INCLUDE IN THE DESCRIPTION

This list of the documents listed by the applicant has been generated automatically and is included solely for the better information of the reader. The list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions.

Cited patent literature

• DE 19943689 A1 [0012]

Claims (4)

Method for determining significantly changed signals on devices with periodically recurring, uniform or uniformly accelerated movements, comprising the following steps: - Recording a set of signal amplitudes of the device by means of a sensor - Determination of the frequencies of individual amplitudes - Calculation of the probabilities of individual amplitudes Calculation of the entropy values, in particular the Shannon entropy, the Kolmogorov-Sinai entropy and / or the Renyi entropy from the probabilities - Comparison of the entropy of the current time unit with the entropy of previous time units and / or with the entropy of reference data. Method for detecting significantly different signals according to Claim 1, wherein the entropy values before the statistical analysis are subjected to postconditioning in the form of averaging over several entropy values and / or approximation of the actual distribution function of the entropy from the empirically obtained values by bootstrapping or jackknifing. A method for detecting newly occurring signals according to claim 1 or 2, wherein the values of the amplitude values of the device subject to preconditioning, in particular averaging of the measured values, filtering out previously determined noise frequencies and / or performing a Fourier transform of the signal determining the entropy for selected frequency ranges become. Device for detecting newly occurring signals, characterized in that it comprises a method according to one of claims 1 to 3.

Images