FIELD OF INVENTION

The present invention relates generally to rotating electrical machines, such as motors and generators, and more particularly to fault detection of electrical machines .

BACKGROUND

The history of fault diagnosis and protection of

electrical machines is as archaic as the machines themselves. However nowadays, condition monitoring of electrical machines has become increasingly essential. It plays a very important role for safe operation and helps to avoid heavy production losses when installed in production lines. Solutions of the prior art have concentrated on sensing specific failure modes in stator, rotor and bearing. All of the presently available techniques require the user to have some expertise in order to distinguish a normal operating condition from a potential failure state. One known method for monitoring electrical machine is known as motor current signature analysis (MCSA) , which

utilises the results of spectral analysis of the stator current. This technique is based on Fourier analysis. In MCSA, the monitoring method is based on the behaviour of the current at the side band associate with the fault.

For that, intimate knowledge of the machine construction is required.

It is further known that when the load torque varies with rotor position, the current will contain spectral components, which coincide with those caused by the fault condition. The torque oscillation results in stator currents that can obscure, and often overshadow, those produced by the fault condition. This technique is not also able to discriminate in many cases among different faults.

Fourier analysis is very useful for many applications where the signals are stationary. However, it is not appropriate for analysing a signal that has a transitory characteristic such as drifts, abrupt changes and

frequency trends. To overcome this problem, Fourier analysis has been adapted to analyse small sections of the signal in time, this technique is known as Short Time Fast Fourier Transform (STFFT) . STFFT represents a sort of compromise between time and frequency based views of a signal and it provides information about both. However, the precision is determined by the size of the window.

SUMMARY

An object of the present invention is to provide a solution to fault detection which can detect several different fault conditions.

According to a first aspect of the invention, it is presented a method for detecting faults in a rotating electrical machine. The method comprises the steps of: selecting at least one frequency band to analyse, each frequency band being a frequency band of a measured entity to analyse; obtaining a plurality of magnitude measurements over time, making up a frequency band series, for each of the at least one frequency bands; and evaluating the frequency band series to determine the presence or absence of a plurality of different fault conditions. In one embodiment, there may be a plurality of frequency bands.

The step of obtaining a plurality of magnitude

measurements may utilise discrete wavelet transform. The step of evaluating may use a weight function to evaluate the frequency band series.

The step of evaluating may use a different weight

function for different time periods. If there are a plurality of frequency bands that are analysed, each of the weight functions evaluates a plurality of frequency bands .

The measured entity may be the stator current of the rotating electrical machine.

The measured entity may be a physical vibration of the rotating electrical machine.

The measurements may be taken during a phase when

transients are expected.

The phase may be a start-up phase of the rotating

electrical machine or a shut-down phase of the rotating electrical machine.

The fault conditions may be at least two conditions selected from the group consisting of a broken rotor bar, static air gap eccentricities, dynamic air gap

eccentricities, opening of stator coil, short circuit of stator coil, abnormal connection in stator winding, cracked end rings, bent shaft, short circuit in rotor field windings, bearing failures, and gearbox failures. The at least one frequency bands may all be sub-bands within the range of 0 to 5 kHz.

A second aspect of the invention is an apparatus for detecting faults in a rotating electrical machine. The apparatus comprises: a frequency band selector arranged to select at least one frequency band to analyse, each frequency band being a frequency band of a measured entity to analyse a measurer arranged to obtain a

plurality of magnitude measurements over time, making up a frequency band series, for each of the at least one frequency bands; and an evaluator arranged to evaluate the frequency band series to determine the presence or absence of a plurality of different fault conditions.

The apparatus may further comprise a current measuring device arranged to measure a stator current of the rotating electrical machine.

The apparatus may further comprise a vibration sensor arranged to measure a vibration of the rotating

electrical machine. A third aspect of the invention is a computer program for a fault detecting apparatus, the computer program

comprising computer program code which, when run on the media recording device, causes the media recording device to perform the method according to the first aspect. A fourth aspect of the invention is a computer program product comprising a computer program according to the third aspect and a computer readable means on which the computer program is stored. In general the invention is based on the following.

First, a signal, e.g. stator current or mechanical vibration, of an electrical machine is measured. This signal is divided into frequency bands, in a similar way to band pass filters, e.g. using discrete wavelet transform. By analysing the evolution over time of these frequency bands, a plurality of fault conditions can be identified; i.e. not only one fault condition as known in the art. For example, the frequency bands can be 0-8, 8- 16, 16-32, 32-64, 64-128, 128-256, and 256-512 Hz. For example, eccentricity and/or broken rotor bars can be identified this way. The analysis can for example be performed during start-up of the machine.

Hence, in this invented method, the frequency bands can be evaluated as they progress during a transient (start or stop), e.g. using the wavelet analysis. Since the frequency bands due to the different root causes evolve differently, this method allows discriminating with high certainty. The solution has been found to be able to detect e.g. broken rotor bars and eccentricity faults and distinguishing it with respect to other failures and load torque oscillations.

It is to be noted that any feature of the first, second third and fourth aspects may, where appropriate, be applied to any other aspect.

Generally, all terms used in the claims are to be

interpreted according to their ordinary meaning in the technical field, unless explicitly defined otherwise herein. All references to "a/an/the element, apparatus, component, means, step, etc." are to be interpreted openly as referring to at least one instance of the element, apparatus, component, means, step, etc., unless explicitly stated otherwise. The steps of any method disclosed herein do not have to be performed in the exact order disclosed, unless explicitly stated. BRIEF DESCRIPTION OF DRAWINGS

The invention is now described, by way of example, with reference to the accompanying drawings, in which:

Fig Ia is a schematic diagram showing an environment where the present invention could be applied using stator current measurements,

Fig Ib is a schematic diagram showing an environment where the present invention could be applied using vibration measurements,

Fig 2 is a schematic diagram showing modules of the fault detecting apparatus of Figs Ia and/or b,

Fig 3 is a flow chart illustrating the use of the fault detecting apparatus of Figs Ia and/or b,

Fig 4 is a graph of a discrete wavelet decomposition of a starting current of a healthy motor, Fig 5 is a graph of a discrete wavelet decomposition of a starting current of a motor with one broken rotor bar,

Fig 6 is a graph of a discrete wavelet decomposition of a starting current of a motor with 10 per cent

eccentricity, Fig 7 is a graph of a discrete wavelet decomposition of a starting current of a motor with two broken bars and at the same time load torque oscillations, Fig 8 is a diagram illustrating a sub-band coding algorithm of a discrete wavelet transform,

Fig 9 is a graph illustrating discrete wavelet transform filtering performed with the Mallat algorithm, Fig 10 shows one example of a computer program product comprising computer readable means,

Fig 11 is a schematic graph of sidebands of a starting current of a motor in an embodiment,

Fig 12 is a graph showing rotor slip and left side band component frequency of a starting current of a motor with one broken rotor bar, and

I Fig 133- is a graph showing rotor slip and left side band component frequency of a starting current of a motor with 10 per cent dynamic eccentricity. DETAILED DESCRIPTION OF PREFERRED EMBODIMENTS

The present invention will now be described more fully hereinafter with reference to the accompanying drawings, in which certain 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 by way of example so that this disclosure will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art. Like numbers refer to like elements throughout the description.

Fig Ia is a schematic diagram showing an environment where the present invention could be applied to stator current measurements. As known in the art per se, a rotating electrical machine 1, such as a motor or

generator, comprises a stator 3 and a rotor 2. These are only displayed schematically here; the actual

configuration of the rotor and stator typically varies from what is shown. Nevertheless, the stator 3 is

stationary and the rotor 2 rotates. A current to/from the stator flows through a cable and is measured with a current measuring device 4, such as an ammeter. The ammeter provides a measurement signal to a fault

detecting apparatus 5, also known as fault detector. The fault detector can detect a plurality of faults of the rotating electrical machine 1.

Fig Ib is a schematic diagram showing an environment where the present invention could be applied to vibration measurements. Here, a vibration sensor 6 measures

vibrations of the stator 3. The vibration sensor can for example comprise an accelerometer, a velocity sensor, a displacement sensor. Optionally, the vibration sensor can send measurement values wirelessly to the fault detecting apparatus 5. The vibration sensor thus provides a

measured entity in form of a measurement of physical vibration to the fault detecting apparatus 5.

It may be possible to combine both types of measurements, i.e. stator current and vibration. Alternatively or additionally, flux measurements can be used, e.g.

obtained using a search coil.

Embodiments herein are based on the realisation that sideband components vary in frequency over time in a way which is characteristic for any faults that may be present . Fig 11 is a schematic graph of sidebands of a starting current of a motor. The graph shows peaks in frequency domain diagram of the current signal. Only peaks are shown, as indicated by arrows, not the entire graph.

Moreover, only the first left and right sidebands, along with the main frequency are shown. A main frequency is indicated by a relatively large arrow 61. A left first sideband, indicated by arrow 60, is a peak at a frequency fi which is the peak which is closest in frequency to the main frequency, which is below the main frequency. A first right sideband is indicated by arrow 62.

Fig 12 is a graph showing rotor slip and left side band component frequency of a starting current of a motor with one broken rotor bar. The upper graph 70 shows the rotor slip over time, while the lower graph 72 shows the frequency of the left sideband 60 (Fig 11) when the motor with a broken bar is started. Significantly, the left side band frequency drops all the way to zero before it returns to a steady state at around 45 Hz. Fig 13 is a graph showing rotor slip and left side band component frequency of a starting current of a motor with 10 per cent dynamic eccentricity. The upper graph 74 shows the rotor slip over time, while the lower graph 76 shows the frequency of the left sideband 60 (Fig 11) when the motor with eccentricity issues is started.

Significantly, the left side band frequency remains at a frequency of about 28 Hz, which is far from the main frequency (here 50Hz) .

Embodiments of the present invention are directed to utilising the characteristics of the left sideband which differ for different motor problems. This is done by examining lower frequencies of the current in the

starting transient. In particular, the time when the motor is at about half its operating speed is shows a large variation between various conditions of the motor. Fig 2 is a schematic diagram showing modules of the fault detecting apparatus 5 of Figs Ia and/or b. The various modules 7-9 can be implemented by means of software and/or hardware. It is also to be noted that the modules may share some hardware components such as controllers and memory 10. A controller (not explicitly shown but can be used to implement part or all of the modules) is provided using any suitable central processing unit

(CPU) , microcontroller, digital signal processor (DSP) , etc., capable of executing software instructions stored in a memory 10. The memory 10 can be any combination of read and write memory (RAM) and read only memory (ROM) . The memory 10 also comprises persistent storage. The persistent memory can be any single one or combination of magnetic memory, optical memory, or solid state memory. A frequency band selector 7 selects the frequency bands to analyse. A measurer 8 receives the measurement data to analyse, e.g. vibration data or stator current data. An evaluator 9 evaluates the measured data for the selected frequency bands to evaluate whether one or more faults have occurred in the rotating electrical machine 1, as will be explained in more detail below.

The fault detecting apparatus 5 may be connected to other systems (not shown) for further handling when faults are detected. For example, the fault detecting apparatus can be connected to a monitoring and alarm system, or it can be arranged to autonomously stop the rotating electrical machine 1. The fault detecting apparatus 5 may further be provided with a user interface (not shown), e.g.

comprising a display and a keypad or keyboard, mouse or trackball, etc. Optionally, a speaker can also be

comprised in the fault detecting apparatus 5 to allow generation of an audible alarm when faults are detected.

The fault detecting apparatus 5 can be implemented using a general purpose computer such as a personal computer with appropriate input for the measurements, analog and/or digital. Fig 3 is a flow chart illustrating the use of the fault detecting apparatus of Figs Ia and/or b, in a method to detect faults of the rotating electrical machine 1.

In one embodiment, the method is processed during the start-up phase of the rotating electrical machine.

Alternatively or additionally, the method can be run during a shut-down phase. The method can be started automatically when monitored machine is turned on, or using signals from a control system such as a motoro control system. Optionally, the method can be run

periodically or when transients are detected.

In an initial select frequency bands step 20, suitable frequency bands are selected. This selection can be suitably configured, whereby this step, for example, can select frequency bands 0-8 Hz, 8-16 Hz, 16-32 Hz, 32-64 Hz, 64-128 Hz, 128-256 Hz, and 256 - 512 Hz. This selection is just an example and other suitable frequency bands can be selected. In one embodiment, only frequency bands under 5 kHz are selected.

In an obtain magnitude measurements step 22, a plurality of magnitude measurements are obtained over time for each of the frequency bands, thus making up a frequency band series. This can be implemented using discrete wavelet transform. In other words, the measurement signal is received and the frequency components of the signal are extracted according to the frequency bands of the

previous step.

The discrete wavelet transform decomposition provides a set of wavelet signals (approximation and details) . Each one of those signals contains the time evolution of the components within the original measurement signal that are included within its corresponding frequency band, according to the band expressions shown above. The analysis of those signals can allow the detection of some patterns caused by the evolution of the components associated with the fault. For the analyses performed,

Daubechies wavelets can be employed, although other types of wavelet families, such as Morlet or biorthogonal, also provide satisfactory results. The use of such a high- order wavelet is justified by the decrease in the overlap between bands.

In an evaluate signal for faults step 24, the frequency bands are analysed to detect any of a plurality of potential faults. The details of this evaluation will be explained in more detail with reference to Figs 4-7 below.

In a conditional any faults step 26, the process ends if there are no faults. However, if there are faults, the process continues to a react to faults step 28.

In the react to faults step 28, the process reacts to the detected fault or faults. For example, one or more corresponding alarm signals can be sent to an operation management system. Alternatively or additionally, an alert or alarm can be displayed on the fault detecting apparatus 5. Optionally, an audible signal can be

generated. Once the reaction has been processed, the process ends.

The method is started again whenever suitable fault detection should be performed.

Figs 4 to 8 are graphs of a discrete wavelet

decompositions of a starting current of a motor in various conditions. The starting current was obtained from simulations of a motor with details as follows:

Rated Power 35 kW

Rated frequency 100 Hz

Voltage 400 V

Rated current 64 A

Connection Star

Number of pole pairs 2

Number of stator slots 48

Number of rotor bars 40

The topmost graph in Figs 4-8 shows the stator current s during the start-up phase. The graphs below that show wavelet components at frequencies indicated on the right hand side of the graphs and are denoted aio, dio, dg .. d ₄ , according to denotations used when wavelet theory is explained in more detail with reference to Figs 8 and 9.

Firstly, Fig 4 is a graph of a discrete wavelet

decomposition of a starting current of a healthy motor. In this case that we have a motor fed with a 50 Hz signal, which is most notably seen in the decomposition in ds (sub-band from 32 to 64 Hz) . The lower frequency sub-bands (e.g. aio, dio and dg) become steady after initial signals subside, when we do not have induction motor faults. However, when there are faults, such as broken rotor bars, the frequency content of these sub-bands changes abruptly, even for incipient fault cases. For example Fig. 5 shows the decomposition for the case of one broken bar for the same motor. Here, as indicated by ellipse 30, the dio band, 8-16 Hz has a significantly different appearance compared to the healthy motor. The signal is stronger and lasts longer than in the healthy motor.

The lower frequency band harmonic due to the broken bar evolves, first the signal is in ds, and then after the first 0.3 s it appears in dg, after the signals continues to the detail signal dio, then, to the approximation signal aio, after 0.5 s, it appears in dio again. After that, it moves to dg and to the main signal sub-band (ds) when the transient disappears. Consequently, the fault of the broken rotor can be detected by analysing any one or more of these frequency bands. For optimal detection accuracy, all of the affected frequency bands can

optionally be analysed.

The graphs of Fig 6 illustrate the case of an

eccentricity problem, here 10 per cent of dynamic

eccentricity. Compared to the healthy motor illustrated in Fig 4, the low frequency component appears first in ds, then after 0.4 s moves to dg, where continues in steady state, because the frequency is about 25 Hz. For the case of a superior harmonic, this can be seen at the very beginning is in signals d ₇ and ds, and in the steady state it appears in d ₇ . This is a superior harmonic of a frequency of around 75Hz.

Fig 7 is a graph of a discrete wavelet decomposition of a starting current of a motor with two broken bars and at the same time load torque oscillations occurring at low frequencies (from 50 to 15 Hz) . It is clear that the patterns produced by the broken bars, see ellipse 34, are distinguishable from the pattern due to the torque oscillations, see ellipse 36. It has thus been shown that with the inventive method and apparatus presented herein, individual faults can be detected and identified, event when a plurality of faults are present.

While the example above shows the use of fault detection for a motor, the same principle can be used for fault detection of a generator. Furthermore, while the example above shows the use of fault detection during a start-up phase, the same principle can be used for fault detection during shut-down or other transient phase. Discrete wavelet transform will now be explained in some more detail.

For many years, Fast Fourier Transform (FFT) has been used for signal processing of the stator current, as it has been found to be suitable for the study of a wide range of signals. Nevertheless, it only allows the extraction of the frequency content of a signal,

eliminating the information concerning time-localization of the frequency components. Short Time Fast Fourier Transform (STFFT) is better in this aspect, but implies some constraints regarding the selection of the optimum window size for data analysis.

To overcome the previous shortcomings, the wavelet theory is here used as a tool for analysing signals with

frequency spectrum varying in time. It allows a time- localization of the frequency components occurring within the signal, being able to extract their time evolution. This property enables the detection of characteristic patterns within the evolution of those components, which can be related to the occurrence of certain phenomena.

The discrete wavelet transform performs the decomposition of a sampled signal s (t) (si, S2, S3 ... S _n ) onto an approximation signal at a certain decomposition level n (a _n (t) ) and n detail signals (d _D (t) with j varying from 1 to n) . Each signal is the product of the corresponding coefficients (approximation coefficients for a _n and wavelet coefficients for d _D ) and the scaling function or the wavelet function at each level, respectively. The signal then can be approximated to s(t) = ∑a; ^{r ■} < (t) +∑∑ JS/ ^• ψl it) = a _n + d _n + . . . + O ₁ ( 1 )

2 =1 J = I 2 = 1

Where α", β _∑ ^J are the scaling and wavelet coefficients, respectively, φ"(t),ψ^(t) are the scaling function at level n and wavelet function at level j , respectively, and n is the decomposition level. a _n is the approximation signal at level n and d _D is the detail signal at level j . The practical procedure for the application of discrete wavelet transform is known as Mallat's algorithm or sub- band coding algorithm; the approximation signal behaves as a low-pass filter whereas each wavelet signal behaves as a pass-band filter, extracting the time evolution of the components of the original signal included within its corresponding frequency band.

Figure 8 shows the sub-band coding algorithm regarding the coefficients of the transform at the different levels in one example. Beside the length of those coefficients, frequency content at each level is shown, considering an original signal with n=512 samples and sampling frequency of f _s . The sampling frequency is typically between 1 and 20 kHz, depending on the main frequency and hardware and software capabilities. It is shown how the original sampled signal S [n] is passed firstly through a half-band high-pass filter g[n] and a low pass filter h[n] .

According to Nyquist criterion for sampling, a down- sampling by two can be performed, obtaining, for

successive levels, half the number of samples of the previous level. These coefficients, multiplied by the scaling function or the wavelet function at each level (which depends on the selected mother wavelet) give the approximation and detail signals at the different levels. The analysis of these signals can be used in the

embodiments herein.

If f _s is the sampling rate used for capturing s (t) , the detail d _D contains the information concerning the signal components whose frequencies are included in the interval [2 ^~<D+1) f _s , 2 ^~J f _s ] . The approximation signal a _n includes the low-frequency components of the signal, belonging to the interval [0, 2 ^"(n+1) f _s ] .

This filtering process performed in a discrete wavelet transform illustrated in Fig 9. However, as can be seen in Fig 9, there is a certain overlap between bands due to the non-ideal characteristic of the wavelet filters, raising the problem of aliasing between bands. That is, when discrete wavelet transform sub-bands are sub-sampled by a factor of two, according to the Mallat algorithm, the Nyquist criterion is violated and frequency

components above or below the cut-off frequency of the filter are aliased into the wrong sub-band. This

phenomenon causes what is known as shift-variance. The number of decomposition levels (n _d ) is related to the sampling frequency of the signal being analysed (f _s ) . In order to get an approximation signal containing

frequencies below the supply frequency, or main frequency f, the number of decomposition levels that has to be considered is given by:

^ ₌ int( ^l0g(f - ^{/ f)} ) (2,

Log(2)

Based on the discrete wavelet transform decomposition, different algorithms can subsequently be implemented to automatically detect the motor condition. This can for example be done using a weight function of the

approximation signal. Alternatively, this can be achieved by simply using the approximation coefficients. One particular example is to use an appropriately configured weight function, based on weighing the approximation and details signal in the neighbourhood of the moment in time when the motor velocity reaches half of the steady state speed. This point is selected because when the lower side band component due to broken bars reach the lower

frequency, this can be seen in the approximation signal (aio) , (see Fig. 5). Also see the discussion above with reference to Figs 12 and 13.

The number of levels in the discrete wavelet transform decomposition depends on the sampling frequency. One example is to select 10 levels of decomposition, in order to have at least three sub-bands below the sub-band that contain the fundamental frequency (50 or 60 Hz) . From the calculated weight and taking into account the rms value of the current in the steady state, five indicators are calculated for five different time periods. Each of the indicators are a result of a particular weight function for that time period, which takes into consideration particular frequency patterns for the various fault conditions. In one embodiment there are different sets of weight functions to evaluate different fault conditions, wherein each set of weight functions is adapted to detect a particular fault condition . The analysis of these indicators determined the motor condition. Fig. 6 shows a block diagram of the detection algorithm. The preset values are calculated from the machine in the healthy condition .

Fig 10 shows one example of a computer program product comprising computer readable means 50. On this computer readable means 50, a computer program can be stored, which computer program can cause a computer to execute the method according to embodiments described herein. In this example, the computer program product is an optical disc, such as a CD (compact disc) , a DVD (digital

versatile disc) or a blu-ray disc. The computer readable means can also be solid state memory, such as flash memory or a software package distributed over a network, such as the Internet. The computer readable means can hold a computer program for methods for the fault

detecting apparatus 5.

While the embodiments above mainly relate to the fault conditions of broken rotor bars and eccentricity, the same principle can be used to detect a long range of faults, e.g. broken rotor bar, static air gap

eccentricities, dynamic air gap eccentricities, opening of stator coil, short circuit of stator coil, abnormal connection in stator winding, cracked end rings, bent shaft, short circuit in rotor field windings, bearing failures, and gearbox failures.

The invention has mainly been described above with reference to a few embodiments. However, as is readily appreciated by a person skilled in the art, other

embodiments than the ones disclosed above are equally possible within the scope of the invention, as defined by the appended patent claims.