BEARING

Field of the Invention

The present invention relates to a method and a device for determining operational properties of a rotating machine. In particular, the present invention relates to a method and device for determining vibrational properties of a bearing in a rotating machine.

Technical Background

Bearings and in particular rolling bearings are used in rotating machines for a wide range of applications. Rolling bearings provide low friction rotation by means of the rolling elements arranged between the outer and inner race of the bearing.

In order to improve the availability, up-time and life span of a rotating machine, the bearing properties are often carefully monitored as there is a need for efficient analysis tools that enable early warning of possible bearing failure.

There is a wide spectrum of methods used for condition monitoring. In particular, several different approaches are used for fault detection; such as limit checking, trend checking, signal models, process models, multivariate data analysis, and heuristic knowledge-based methods. In addition there are different approaches for fault diagnosis, which can be roughly divided in classification methods and heuristics, and probabilistic inference methods. Limit checking is a commonly used approach because the interpretation is straightforward and it can be implemented with or without advanced models, but the results depend in a critical way on the characteristics of the signals and the methods used to extract information from the signals or models. False alarms and low sensitivity are common problems when simplistic methods are used. Model-based fault detection and diagnosis methods are commonly used and functions well when an accurate model that remains valid over time can be engineered, which however means that experts need to be involved in the development and deployment of such methods. A difficulty with that approach is that the solution is application-specific and can be invalidated if system components change, which limits the general applicability of the approach.

Fault detection and diagnosis systems for rotating machinery often include vibration analysis, and sometimes also acoustic-emission (AE) analysis. Frequency analysis can enable early detection of low-amplitude high-frequency signal components that are associated with emerging faults. However, such methods discard time-domain patterns in the signal and have limited sensitivity to broadband and stochastic patterns. Several components are required to monitor these modalities, including the sensor (typically of piezoelectric type), preamplifier, signal-processing system and fault detection and diagnosis methods. The signal-processing system can be a

microcontroller, FPGA, computer, oscilloscope, or an application-specific integrated circuit (ASIC). The pre-processed signal is analyzed in an automated fashion with methods like those described above, or manually by humans.

Moreover, in existing condition monitoring systems, there is a need for data reduction, in particular in embedded and wireless sensor systems. Some condition monitoring signals require a high sampling rate, which is

problematic for (low-cost) embedded and wireless systems using

conventional technology. For example, vibration and acoustic emission signals require sampling rates on the order of several kHz up to the MHz range.

Machines, sensors and the interface between the two are also subject to complex environmental conditions, operational variations and changes over time. As a consequence, the modeling of the machine is a task that can be both expensive and difficult to perform and account for with conventional technology. In addition, some phenomena and failure modes are complex and difficult to model. Accordingly, there is a need for a technology that requires a minimum of manual configuration and application-specific development. Summary of the Invention

In view of the above-mentioned desired properties of condition monitoring for a rotating machine, and the above-mentioned and other drawbacks of the prior art, it is an object of the present invention to provide an improved method for determining operational properties of a bearing in a rotating machine, and a device configured to perform such a method.

According to a first aspect of the invention, it is therefore provided a method for condition monitoring of a bearing in a rotating machine, the method comprising; receiving vibration data or acoustic emission data from the rotating machine; identifying time-domain features in the data using a self- learning algorithm to form a set of features, each feature being represented by a unique waveform, until a converged set of features is reached; storing the converged set of features as a steady-state set of features; in a

monitoring state, identifying time-domain features in the data through the self learning algorithm and updating the set of features based on the features indentified in the monitoring state; comparing the updated set of features to the steady-state set of features; and if the updated set of features is different from the steady-state set of features, determining that the conditions of the rotating machine has changed.

To receive vibration or acoustic emission data should in the initial state be understood as receiving data for a period of time at least such that a converging set of features is reached. In the monitoring state, vibration or acoustic emission data is received so that an updated set of features can be determined. Data may for example be received continuously. Vibrations and acoustic emissions are highly interrelated as vibrations in a component may give rise to acoustic emissions.

The rotating machine may be any rotating arrangement that generates detectable vibrations or acoustic emission during operation. In particular, the rotating machine may comprise one or more bearings and the vibration or acoustic emission data received may be related to one or more specific bearings within the rotating machine. A time-domain feature may also be referred to as an atom, where each identified feature, atom or time-domain waveform represents a particular vibrational or acoustic-emission property of the rotating machine.

Examples of the mathematical basis for self-learning algorithms are well known in the scientific literature and readily available to the skilled person. Accordingly, the present invention is not limited to any specific mathematical model or particular algorithm.

The present invention is based on the realization that by using a self- learning algorithm to characterize a rotating machine through a set of time- domain features, a change in the condition of the rotating machine may be identified by observing if the set of features change. The automated detection of novel features in a signal can enable preventive actions and improved maintenance planning in a broad range of applications. Furthermore, through the claimed method, changes can be detected without the need for fault modeling. Moreover, by indentifying a change in the set of features, instead of comparing each stored feature individually to an updated feature, also the addition or removal of features in the set can readily be identified. The method comprises an initial state in which a steady-state set of features is identified, and a monitoring state where the rotating machine is monitored and features are identified to determine if vibrational or acoustic-emission properties are changing over time. In the monitoring state, an updated set of features is formed which is compared to the steady-state set of features. The updated set of features is typically continuously updated and compared to the steady state set of features.

According to one embodiment of the invention, it is determined that the updated set of features is different from the converged steady-state set of features if at least one of the features has changed in frequency, shape, and/or amplitude, and if the change is larger than a predetermined threshold value. Accordingly, a change in any of the aforementioned parameters larger than a predetermined threshold value may be the result of a change of the condition of a component of the rotating machine. If a detected difference is smaller than the predetermined threshold value, it may be determined that the difference is the result of noise or that the difference is within the range of known measurement uncertainties.

In one embodiment of the invention, it is determined that the set of features is different from the steady-state set of features if at least one feature has been added to or removed from the steady-state set of features. The addition of a feature to the steady-state set of features is indicating that properties of the rotating system have changed. Likewise, the removal of a feature is an indication of change, and further analysis may be performed to determine if the change is a potential cause for failure.

According to one embodiment of the invention, the method may comprise generating an output to a host system only when a difference between the updated set of features and the steady-state set of features is identified. Thereby, where the method is used in a centralized and/or remote condition monitoring system, communication between a sensing device utilizing the method and a host system only takes place when a change is identified, thereby significantly reducing the data rate compared to

conventional monitoring systems. By reducing the data rate, i.e. the amount of information that is communicated, the energy consumption of a monitoring device can be kept at a minimum.

According to one embodiment of the invention, the vibration or acoustic-emission data may advantageously be received in the form of an analog signal, the method further comprising the step of converting the analog signal to a digital signal.

In one embodiment of the invention, the self-learning algorithm may advantageously start from a randomized initial state. By using an

unsupervised learning methodology based on a randomized initial state, only a minimum of configuration is required prior to use, and the method may be applied to a wide range of applications without any adaptation or prior knowledge of the application at hand.

According to one embodiment of the invention, the method may further comprise determining a cross correlation between each feature in the steady state set of features and a corresponding feature in the updated set of features.

According to one embodiment of the invention, the method may further comprise determining a signal-to-noise ratio (SNR) between each feature in the updated set of features and said received signal.

According to a second aspect of the invention, there is provided a condition monitoring device comprising; an analog input configured to receive a signal indicative of vibrational or acoustic-emission properties of a rotating machine; an amplifier configured to amplify the signal; a low pass filter configured to block frequencies of the signal higher than a predetermined cutoff frequency; an analog to digital converter configured to sample the analog signal to form a digital signal; a control unit configured to identify vibrational features in the signal using a self-learning algorithm to form a set of features, each feature being represented by a unique waveform, until a converged set of features is reached; a storage unit configured to store the converged set of features as a steady-state set of features; wherein the control unit is further configured to continuously identify time-domain features in the signal and to update the set of features, a change detector configured to compare the updated set of features to the steady-state set of features; and if the set of features is different from the steady-state set of features, determine that the conditions of the rotating machine has changed.

The device integrates an amplifier with a general-purpose signal processing and condition-monitoring system for change detection in terms of learned additive features. Thereby, the signal from a vibration sensor or acoustic sensor is fed directly to the device and no further data processing or communication is required during normal operation. Because the device is general and not designed for one particular condition monitoring application the design can be refined and optimized (in FPGA's) and then realized in the form of an ASIC, which is power efficient and suitable for embedded and wireless application.

The adaptive feature detection device models the received signal in terms of waveforms with compact support that is learned from the signal using an unsupervised statistical optimization algorithm. The device adapts the shape and length of the waveforms, herein also referred to as features where each feature is represented by a unique waveform, to a specific signal from a specific rotating machine. The signal is modeled as a linear

superposition of features and noise. In particular, the device enables unsupervised learning of features present in a signal, detection of changes of the features, detection of new features in the signal, and detection of removal of features in the signal. Moreover, if desired, identified features may be requested by and communicated to a host system for further analysis.

The device reports significant changes to the set of learned features or residual of the signal model. The device does not generate any output data during normal operation, i.e. when the input signal corresponds to a superposition of the historically learned features comprised in the steady- state set of features, and noise. Nor does it require any further information processing or human intervention in that case. Therefore, the device can significantly reduce the data rate that needs to be communicated and analyzed. Because the device can adapt the features to a particular signal and machine, the condition monitoring process can potentially be simplified. Given a set of features, the signal model is similar to a matched filter.

Therefore, the signal to noise ratio is high in the presence of additive noise. Thus, new features in the signal can potentially be detected before they could have been detected with standard methods like spectral analysis. In summary, the unsupervised learning, high data reduction and high signal to noise ratio that are enabled by the device can be useful for early detection of faults in rotating machines.

Moreover, the device can simplify and enable continuous condition monitoring of rotating machines as it enables wireless condition monitoring with minimal bandwidth requirements.

In one embodiment of the invention, the low pass filter of the condition- monitoring device may advantageously comprise an analog low pass filter, an analog-to-digital (AD) converter, and a digital low pass filter. By combining an analog filter with a digital filter, the advantages of both types of filters are achieved. Moreover, by arranging the analog filter prior to the AD-converter, the cut-off frequency of the analog filter may be set with respect to the sampling frequency of the AD-converter such that the signal does not comprise frequencies too high to be accurately sampled.

There is also provided a condition monitoring arrangement comprising a condition monitoring device as described above and a sensor configured to detect vibrations of a rotating machine and to convert the vibrations into an electrical signal.

Furthermore, the condition monitoring arrangement may in one embodiment comprise an acoustic emission sensor configured to detect acoustic emissions from a rotating machine and to convert the acoustic emissions into an electrical signal. The electrical signal representing the vibrations or acoustic emissions can then be coupled to the condition- monitoring device.

Further features of, and advantages with, the present invention will become apparent when studying the appended claims and the following description. The skilled person realize that different features of the present invention may be combined to create embodiments other than those described in the following, without departing from the scope of the present invention.

Brief Description of the Drawings

These and other aspects of the present invention will now be described in more detail with reference to the appended drawings showing an example embodiment of the invention, wherein:

Fig. 1 is a flow chart outlining the general steps according to an embodiment of the invention;

Fig. 2 schematically illustrates a system according to an embodiment of the invention; and

Fig. 3 schematically illustrates exemplary features in an embodiment of the invention. Detailed Description of Preferred Embodiments of the Invention

The present invention will now be described more fully hereinafter with reference to the accompanying drawings, in which currently preferred embodiments of the invention are shown. This invention may, however, be embodied in many different forms and should not be construed as limited to the embodiments set forth herein; rather, these embodiments are provided for thoroughness and completeness, and fully convey the scope of the invention to the skilled person. Like reference characters refer to like elements throughout.

Fig. 1 is a flow chart outlining the general method steps of the invention steps, the method illustrated in Fig. 1 will be discussed with reference to Fig. 2 schematically illustrating a condition-monitoring device 200 capable of performing the method.

In a first step 102, data is received continuously in the form of an electrical signal representing analog acoustic emissions or vibrations from the rotating machine through an analog input 202. The acoustic emissions or vibrations are converted into an electronic signal by a sensing device (not shown) arranged in connection with the rotating machine and provided from the sensor to the analog input 202. The sensing device may be any suitable sensing device such as a piezoelectric of MEMS-based sensor. In cases where the properties of the sensing device are unknown, an amplifier 204 is preferably included to ensure that the amplitude of the signal is sufficient for further treatment of the signal. Accordingly, a condition-monitoring device 200 comprising an amplifier 204 may be used with different types of sensing devices. The input signal is low-pass filtered in an analog low-pass filter 206 in order to avoid aliasing effects. Next, the analog signal is converted to a digital signal in an analog-to-digital converter 208 (ADC). Alternatively, or in combination with the analog low-pass filter 206, a digital low-pass filter may be arranged after the ADC. Thereby, the well known properties of the two types of filters can be combined to provide the required filtering of the signal. A clock 210 synchronizes the digital signal processing and a buffer 210 is used to facilitate further signal processing after AD conversion. In some applications where acoustic emission is received it may be advantageous to use a band-pass filter to filter out lower-frequency vibration signals which may otherwise dominate over the much lower-amplitude acoustic emission at higher frequencies.

Next 104, time-domain features from the vibration or acoustic emission signal are identified using a self-learning algorithm implemented in a signal decomposition block 214 communicating with a feature optimization block 216. Based on the result of signal decomposition, the gain of the amplifier 204 may be controlled via a gain control module 218. The signal is modeled as a linear superposition of noise and features. In an initial state, features are identified until the number and shape of identified features converge to form a steady-state set of features. Each feature corresponds to a waveform having a finite length. The steady-state set of features is stored 106 to enable comparison with later identified features. Identification of the steady-state set of features is typically done in an initial state where it is known that the rotating machine is operating under normal conditions.

In a monitoring state, the vibration or acoustic-emission data is continuously monitored and signal decomposition and feature optimization is performed to identify 108 features and to update the set of features. The updated set of features is continuously compared 1 10 with the stored steady- state set of features to determine if the updated set of features is different from the steady-state set of features.

If it is determined that the set of features has changed, in the change detector 220, this may be reported as an event at the device output 222. Thereby, communication with a host system is only required when it has been observed that a change of properties of the rotating machine has occurred. From the nature of the reported feature, or set of features, it is often possible to determine the nature of the defect, for example based on empirical data or models.

While the input signal is sampled continuously, the signal

decomposition and feature optimization algorithm works with a signal window that is at least twice as long as the longest feature identified in the steady- state set of features. A longer window is preferable to avoid edge effects.

The sampling frequency of the ADC 208 is based on the sensing device connected to the input 202. Using a conventional vibration sensor the sampling frequency may be on the order of a few tens of kHz whereas for an acoustic emission sensor the sampling frequency may be on the order of several MHz. For a rotating machine operating at low rotational velocities, such as a wind mill, a sampling frequency of a few hundreds of Hz may be sufficient. However, the method as such is not limited to any particular frequencies, but is instead generally applicable to in principle any signal comprising information about the vibrational or acoustic-emission properties of a rotating machine.

A discussion outlining the mathematical methodology for unsupervised feature learning and detection can be found in del Campo et al. "FPGA Prototype Of Machine Learning Analog- To-Feature Converter For Event- Based Succinct Representation Of Signals" IEEE International Workshop on Machine Learning for Signal Processing (MLSP2013), September 22-25, 2013, Southampton, UK, by the inventors. In principle, the algorithm is based on sparse coding with a learned dictionary for feature extraction. During the initial stage when the features are identified, features are detected and matched to the received signal so that only a noise-like residual remains.

Moreover, determining if there is a difference between the updated set of features and the steady state set of features may involve observing the cross-correlation between corresponding features as a measure of the change of each feature per time unit. After convergence of the features, i.e. after the steady-state set of features had been identified and the method is in the monitoring state, only a very small change is expected during normal operation and a cross-correlation value should thus be close to one. The determination that the set of features has changed may thus be based on the observed cross-correlation value where a suitably selected threshold value lower than one may indicate a change of features. If a defect is introduced which alters the properties of an existing feature, the cross-correlation value for the relevant feature will decrease rapidly. The rapid decrease can be readily detected and may be reported as an event.

Alternatively, or in combination, the respective signal-to-noise ratio (SNR) between each feature and the full signal may be determined and monitored. Also here, a change in one of the features will result in a rapid change of the SNR, which may be detected.

In Fig. 2, the analog input 202 and event output terminals 222 are externally accessible, whereas the remaining blocks of the condition monitoring device 200 may be integrated in a single electronic circuit or package for example in the form of an ASIC or FPGA.

Fig. 3 schematically illustrates exemplary identified features

represented by waveforms, where the upper waveform 302 is an impulse-like waveform and the lower waveform 304 is a harmonic waveform. There is in principal no limitation in the number of features in the set of features. An impulse-like waveform detected in the vibration data is typically the result of a defect bearing where a damaged race or ball in the bearing typically gives rise to an impulse having a high frequency (in comparison to the operating frequency of the rotating machine) and an exponential decay in amplitude. Similar features may be found for defect cogs and the like. During modeling of the features, an impulse waveform representing a feature grows in length until a steady-state condition is reached, for example when the amplitude of the tail of the waveform is below a predetermined threshold value.

The harmonic waveform 304 is typically related to the operating frequency and vibrational modes of the rotating machine during normal operation and is found in all rotating machines. Both the fundamental frequency and higher order harmonics may be observed. Imbalance or misalignment of components in the rotating machine, such as in a shaft, may be observed as an increase in amplitude in one or more of the harmonic features.

Even though the invention has been described with reference to specific exemplifying embodiments thereof, many different alterations, modifications and the like will become apparent for those skilled in the art. Also, it should be noted that parts of the system may be omitted,

interchanged or arranged in various ways, the conditioning monitoring device yet being able to perform the functionality of the present invention.

Additionally, variations to the disclosed embodiments can be understood and effected by the skilled person in practicing the claimed invention, from a study of the drawings, the disclosure, and the appended claims. In the claims, the word "comprising" does not exclude other elements or steps, and the indefinite article "a" or "an" does not exclude a plurality. The mere fact that certain measures are recited in mutually different dependent claims does not indicate that a combination of these measured cannot be used to advantage.