Fault detection in synchronous machines

Abstract

To provide methods of fault detection in synchronous machines.SOLUTION: A method of fault detection in a synchronous machine includes using at least one sensor to determine parameters linked to the magnetic field generated within the synchronous machine, including parameters based on one or more of magnetic field strength, rotor current or voltage, stator current or voltage, and vibration. The sensor measurements are processed to identify data artefacts linked to the magnetic field, where the processing includes one or more signal processing techniques based on time, frequency, and both time and frequency. The output of the signal processing is analyzed in order to identify and categorize irregularities in the magnetic field that are indicative of a fault in the synchronous machine. The analyzing step comprises recognizing patterns in the processed sensor measurements, via use of computer aided pattern recognition techniques such as via machine learning algorithms.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a method for fault detection in a synchronous machine, as well as an associated fault detection system that can be combined with a synchronous machine, and a corresponding computer program product. [Background technology]

[0002] Synchronous machines, especially large synchronous machines such as hydroelectric generators, play a vital role in the generation of electricity. High reliability of power supply depends on these synchronous generators. Unplanned shutdowns of energy production centers (power plants) and production lines are one of the most serious concerns in the power industry. All large-scale energy production systems rely on large electrical machines, especially synchronous machines. Large synchronous machines are one of the most expensive pieces of equipment in power plants. Furthermore, their maintenance and repair are expensive, and if a fault causes a synchronous machine to be disconnected from the network, it leads to economic losses. Today, regular maintenance systems are used in most industries related to electrical machines. Summary of the Invention [Problem to be solved by the invention]

[0003] Synchronous machines can experience various types of electrical and mechanical faults. Mechanical faults can induce vibrations or unbalanced magnetic pull on certain parts or the entire generator. Mechanical faults can be caused by improper operating conditions, mechanical deterioration, or impediments during machine assembly. If the minimum air gap between the rotor core and the stator core fluctuates, the synchronous generator will experience what is known as an eccentricity fault. There are two types of eccentricity: static and dynamic. Static eccentricity occurs when the minimum length of the unevenly distributed air gap is fixed and spatially fixed. Static eccentricity is one of the common faults in synchronous generators. While small eccentricities do not cause damage to the machine, they should be detected early, before the rotor core rubs against the stator core and windings.

[0004] Generally, faults can occur due to external factors, such as a severe short circuit in the power grid, or internally due to a staged fault. Rapid response to faults is the responsibility of the synchronous machine's protection system. Typical protection systems use voltage and current profiles at the stator terminals or their associated extracted data. In the case of hydroelectric generators, hydroelectric power plants often have sophisticated protection systems consisting of overvoltage, overcurrent, and differential relays, as well as some subsystems and equipment that quickly respond to abnormal operating conditions and automatically trip. The purpose of the protection system is to ensure immediate and accurate disconnection of the production unit in the event of a severe fault in the machine or the power grid connected to the production unit. Based on measurement data, the protection system can disconnect the electric machine from the power system to protect it from fast, destructive faults. However, existing protection systems cannot detect staged faults within the machine that could lead to future severe failures. Therefore, it is beneficial to have a condition monitoring system that performs additional fault detection tasks. The present invention relates to a method and system used in this context. [Means for solving the problem]

[0005] Viewed from a first aspect, the present invention provides a method of fault detection in a synchronous machine, the method comprising: using at least one sensor to determine parameters related to the magnetic field generated within the synchronous machine, including parameters based on one or more of magnetic field strength, rotor current or voltage, stator current or voltage, and vibration; processing the sensor measurements to identify data artifacts related to the magnetic field, the processing including one or more signal processing techniques based on time, frequency, and both time and frequency; and analyzing the output of the signal processing to identify and classify irregularities in the magnetic field that are indicative of faults in the synchronous machine, the analysis including recognizing patterns in the processed sensor measurements through the use of computer-aided pattern recognition techniques, such as through machine learning algorithms.

[0006] This method allows for more effective fault detection of synchronous machines. The first aspect of the method relates to detecting faults within the synchronous machine itself, e.g., with respect to mechanical and / or electrical components of the synchronous machine typically found inside the machine. Faults within the machine are distinguished from faults external to the machine, e.g., with respect to the excitation system or external electrical network for magnetizing the machine. Faults in such components are typically difficult to identify and diagnose without invasive measurements and / or modifications to the machine. It would be advantageous to provide an effective non-invasive technique. This method is particularly advantageous for large synchronous machines, such as synchronous electric machines used as hydroelectric generators. While other fault detection methods exist, they do not focus on using magnetic fields to diagnose numerous problems with synchronous machines through the identification and classification of irregularities in the magnetic field. The method may be for detecting faults related to one or more of eccentricity, damper winding faults, and short circuits.

[0007] A synchronous machine is an AC electric machine whose shaft rotation, in steady state, is synchronized with the frequency of the generated voltage, or with the supply voltage if the machine acts as a motor. The rotation period of a synchronous machine is therefore equal to an integer or fractional AC cycle given by the number of pole pairs. Synchronous machines as described herein may include polyphase AC electromagnets on the machine's stator that generate a magnetic field that rotates in sync with the oscillations of the line current.

[0008] In this context, it is important to understand the difference between synchronous and induction machines. Synchronous machines do not rely on current induction to create the rotor's magnetic field, so they rotate at a speed locked to the line frequency. In contrast, induction machines require slip; the rotor must rotate slightly slower than AC to induce current in the rotor windings. The method described here does not pertain to induction machines.

[0009] The method may include using signal processing techniques for frequencies higher than the line frequency. That is, the method does not work solely with signals within the power frequency range (typically 50 Hz or 60 Hz) and may in fact not be particularly relevant to signal processing techniques that use identifiable patterns within that frequency range. Thus, the method may involve operating at frequencies higher than (and optionally not including) the power frequency range, i.e., higher than 60 Hz. The method advantageously involves signal processing techniques based on frequencies that are superharmonics of the line frequency, typically frequencies well above the line frequency, which may be frequencies above 75 Hz, e.g., 75 Hz to 400 Hz, and / or may include frequencies in the kHz range and above. Thus, the frequencies used in the method of the first aspect may be frequencies above 1 kHz, and in particular may be superharmonics of the line frequency of a synchronous machine. Thus, analyzing the output may include recognizing patterns in the processed sensor measurements that include such frequencies, the patterns being identifiable in these larger frequency ranges. Signal processing techniques may include using sampling frequencies in the kHz range, such as 5 kHz or 10 kHz or higher sampling frequencies.

[0010] As mentioned above, the method can be used for fault detection in large synchronous machines, such as hydroelectric generators. In this context, a large synchronous machine is one that, when acting as a generator, can generate 100 kW to 500 MW or more of electricity. Hydroelectric generators are electrical machines used to generate electricity from flowing water and are typically salient pole synchronous generators. The fault detection method can be particularly useful for salient pole synchronous generators.

[0011] The fault detection method exploits irregularities in the magnetic field of a synchronous machine. This may be, for example, a transient magnetic field that occurs during the starting or stopping of the machine. Such transient magnetic fields may be distinguished from steady-state or fluctuating magnetic fields that occur during the ongoing use of a synchronous machine, such as when a hydroelectric generator is operating at a steady speed or is otherwise continuously used to generate electricity. It has been found that fault detection can be improved by identifying and classifying irregularities in such transient magnetic fields. Thus, the method may be used during periods when transient magnetic fields are present to find irregularities indicative of faults, and thus may involve fault detection during the starting of a synchronous machine. The inventors have discovered that some fault types may be best detected while transient magnetic fields are present, or may only be detected when the machine is in a transient state.

[0012] The method may involve using at least one existing sensor or at least one non-invasive sensor. This provides the various advantages described above in fault detection without the need for additional invasive sensors or modifications to the synchronous machine. This is beneficial for fault detection in existing machines. Alternatively, additional sensors may be provided compared to conventional large synchronous machines when appropriate, including when designing new machines with integrated fault detection capabilities. In particular, it may be beneficial to measure magnetic field signals more directly via sensors with elements invasively installed, such as in the air gap between the rotor and stator. In some examples, the method utilizes a Hall Effect sensor or search coil installed in the air gap.

[0013] If an existing sensor is used, the existing sensor may be a sensor already provided on the synchronous machine, such as one or more sensors used during operation of the synchronous machine to control the machine. For example, this may be a voltage or current sensor used to monitor the performance of the machine. The existing sensor may be a non-invasive sensor, i.e., a sensor mounted external to the machine and / or a sensor that performs measurements in a contactless manner, or a sensor integrated within the synchronous machine.

[0014] Alternatively or additionally, at least one non-invasive sensor may be used with the machine specifically for fault detection, such as a sensor provided as part of a fault detection system, in which case the sensor is temporarily deployed with the machine to perform the fault detection method. Non-invasive sensors that may be used include sensors that measure electrical properties such as voltage, current, resistance, or related parameters. This can be done via additional sensors or via existing sensors, such as voltage sensing on the potential transformer of a synchronous machine. Furthermore, an accelerometer may be used for vibration signal acquisition, which is deployed externally, i.e., non-invasively, to the machine.

[0015] The method may use multiple sensors, where multiple different measurements are taken by two or more sensors, including existing sensors and / or non-invasive sensors.

[0016] In some examples, the method uses only existing and / or non-invasive sensors. Thus, no new invasive sensors are used. For example, the method may not require any physical modifications to the synchronous machine to install or attach new sensors.

[0017] Advantageously, the method may not involve any data transfer from the moving parts of the rotor, and therefore may not require sensors located on the moving parts of the rotor and / or additional slip rings or other modifications for data transfer, which minimizes the impact of the fault detection method on the performance of the machine, as the moving parts of the machine are not disturbed by additional sensors.

[0018] The method can use voltage and / or current sensors, such as sensors provided in the potential transformer (PT) of a synchronous machine. Typically, there are one or more sensors for voltage and / or current in the potential transformer. This allows existing sensors to be used in the proposed method, and sensor measurements from the potential transformer are used to identify magnetic field irregularities, such as during transient magnetic fields during machine starting. Such measurements can be used to identify and classify faults, such as damper winding faults or eccentricity faults, for example.

[0019] The method may include determining a turning radius of a rotor of a synchronous machine and using the turning radius to identify and classify magnetic field irregularities. As described further below, the turning radius may be evaluated to determine various faults. In some cases, in embodiments, measurements from a sensor in a potential transformer may be used to determine the turning radius.

[0020] Processing the sensor measurements includes one or more signal processing techniques. Examples include time series data mining (TSDM), Fourier transform (FT), fast Fourier transform (FFT), Hilbert transform (HT), Hilbert-Huang transform (HHT), continuous wavelet transform (CWT), and discrete wavelet transform (DWT). These techniques may include using sampling frequencies in the kHz range, such as 5 kHz or 10 kHz or higher.

[0021] In some example implementations, the method may include using turning radius data, optionally obtained via time-series data mining, to detect damper winding faults, particularly to detect broken damper bars. This can be done by mapping the phase space for the induced voltage in the field winding of the synchronous machine. Advantageously, this voltage can be determined or measured via a sensor in the potential transformer. It has been found that when a damper bar breaks, the turning radius mapped based on the phase space of the induced voltage in the field winding of the synchronous machine differs from a machine without a broken damper bar. This difference allows pattern recognition techniques to automatically identify magnetic field irregularities indicative of a broken damper bar based on recognizing characteristic patterns in the phase space diagram. Essentially, the amplitude of the turning radius increases when the damper bar breaks. The method may include pattern recognition based on distinguishing (or matching) between measurements of a synchronous machine known to be good and measurements of a synchronous machine known to have a broken damper bar. It will be appreciated that a similar method may provide a way to detect other types of faults based on pattern recognition for measurements of synchronous machines known to have other faults.

[0022] Alternatively or additionally, in some exemplary implementations, the method may include using turning radius data, optionally obtained via time-series data mining, to detect eccentricity faults, advantageously including static eccentricity faults. This can be done by mapping the phase space for the induced voltage in the field winding of the synchronous machine. Similar to the damper winding fault detection described above, this voltage can be determined or measured via a sensor in the potential transformer. When detecting eccentricity faults, analyzing the output of the signal processing (e.g., from time-series data mining) may include determining a normalized turning radius with reference to the turning radius of a machine known to be good. The normalized turning radius may be defined as the difference between the normal turning radius and the measured (i.e., suspected fault) turning radius divided by the normal turning radius. This results in an index that enables automatic identification and classification of eccentricity faults by evaluating the value of the normalized turning radius (which is zero for a good machine).

[0023] The analyzing step includes recognizing patterns in the processed sensor measurements through the use of computer-aided pattern recognition techniques. Thus, the method may include comparing patterns found in the processed sensor measurements to patterns believed to be indicative of a fault in the synchronous machine. This may be a specific fault, such that pattern recognition may provide a clear diagnosis of the fault. Alternatively, this may be an indicator of a non-specific fault, requiring further investigation or consideration in reference to other measurements, including other processed sensor measurements (and optionally patterns therein), before the fault type can be diagnosed. The computer-aided pattern recognition technique may be based on a machine learning algorithm. For example, the pattern recognition may be based on a machine learning algorithm trained using a plurality of processed sensor measurements known to be associated with fault-free machines and a plurality of processed sensor measurements known to be associated with machines having faults. The machine learning process may include training using a plurality of processed sensor measurements known to be associated with a particular category of fault (e.g., eccentricity faults or damper winding faults, and short circuits) and / or a particular type of fault (e.g., damper winding faults in the form of broken damper bars).

[0024] Viewed from a second aspect, the present invention provides a fault detection system for fault detection in a synchronous machine, the fault detection system comprising: a data processing device for connecting to at least one sensor for receiving parameters relating to a magnetic field generated within the synchronous machine, the parameters including parameters based on one or more of a magnetic field strength, a rotor current or voltage, a stator current or voltage, and vibration; a data processing device, processing the sensor measurements to identify data artifacts related to the magnetic field, the processing including one or more signal processing techniques based on time, frequency, and both time and frequency; and analyzing the output of the signal processing to identify and classify irregularities in the magnetic field that are indicative of faults in the synchronous machine, the analysis including recognizing patterns in the processed sensor measurements through the use of computer-aided pattern recognition techniques, such as through machine learning algorithms.

[0025] The data processing device may be configured to perform the steps described above in relation to the first aspect and any of its features. The invention further extends to a synchronous machine, such as a large synchronous machine acting as a hydroelectric generator, including a fault detection system. The fault detection system may include structural and / or functional features as described above in relation to the first aspect and any of its features. The fault detection system is configured and used for detecting faults within the synchronous machine, which are distinguished from faults external to the machine, such as problems related to the excitation system or the external electrical network, as described above. The fault detection system is advantageously related to frequencies higher than line / power frequency, as described above.

[0026] For example, the fault detection system may include at least one existing sensor or at least one non-invasive sensor, or other sensors described above. This may be an existing and / or non-invasive sensor disposed on the synchronous machine to detect parameters based on one or more of magnetic field strength, rotor current or voltage, stator current or voltage, and vibration. The existing sensor may be a sensor already provided on the synchronous machine, such as one or more sensors provided for use during operation of the synchronous machine for machine control. For example, this may be a voltage or current sensor used to monitor machine performance. The existing sensor may be a non-invasive sensor, i.e., a sensor configured to be attached to the outside of the machine and / or a sensor that performs measurements in a non-contact manner, or a sensor integrated within the synchronous machine. Alternatively or additionally, the at least one non-invasive sensor may be used with the machine specifically for fault detection, such as a sensor provided as part of the fault detection system. In this case, this sensor is configured to be temporarily disposed with the machine to perform fault detection. The fault detection system may include multiple sensors, and multiple different measurements are provided to the data processing device by two or more sensors, including existing and / or non-invasive sensors.

[0027] In some cases, the fault detection system consists only of existing and / or non-invasive sensors on the synchronous machine, thus eliminating the need to install new sensors on the machine, especially new invasive sensors that would require physical modifications to the synchronous machine.

[0028] The fault detection system may be configured to use sensors without data transfer from the moving part of the rotor, and therefore there may be no sensors located on the moving part of the rotor, which minimizes the impact of the fault detection method on the performance of the machine, as the moving parts of the machine are not obstructed by additional sensors.

[0029] The data processing device may be any device suitably configured to perform the required method steps. The data processing device may be a dedicated processor of the fault detection system and, therefore, may include a suitable processor, such as a computer processor, with appropriate data input and output connections. Alternatively, the fault detection system may include a general-purpose computing device, such as a desktop computer, laptop computer, tablet, or smartphone, configured to perform the required method steps, either in its primary role or in a secondary role, such as through the use of a software application provided for installation on the general-purpose computing device.

[0030] The fault detection system may include a suitable interface for communication of the sensor signals to the data processing device, such as a wired or wireless system for transmission of the output signals from the sensors to the data processing device.

[0031] Viewed from a third aspect, the present invention provides a computer program product comprising instructions that, when executed in a fault detection system such as of the second aspect, are configured to cause a data processing device of the fault detection system to process sensor measurements to identify data artefacts related to the magnetic field, the processing comprising one or more signal processing techniques based on time, frequency, and both time and frequency, and to analyse output of the signal processing to identify and classify irregularities in the magnetic field that are indicative of faults in the synchronous machine, the analysis comprising recognizing patterns in the processed sensor measurements through the use of computer-aided pattern recognition techniques, such as through machine learning algorithms.

[0032] The computer program product may include instructions configured to configure a data processing apparatus to perform other steps as described above in relation to the method of the first aspect and any features thereof. The computer program product may be, for example, firmware or software configured for a data processing apparatus, such as firmware for a special purpose processor or software for a general purpose computing device.

[0033] Certain exemplary embodiments of the present invention will now be described, by way of example only, with reference to the accompanying drawings, in which: [Brief explanation of the drawings]

[0034] [Figure 1] 1 is a flowchart of a fault detection system. [Figure 2] A two-dimensional scheme of a simulated salient pole synchronous generator is shown. [Figure 3] The distribution position of the damper bars in the SPSG is shown. [Figure 4] Two plots are included comparing the induced voltage in the rotor field winding of the SPSG in normal and faulty cases for different numbers of BDBs. [Figure 5] FIG. 10 illustrates the phase space of healthy damper bars and one broken damper bar at the edge of an SPSG rotor pole. [Figure 6] Examples of locations of broken damper bars in an SPSG and how this affects the asymmetry of the air gap field are shown with reference to (a) six BDBs at each opposing pole, (b) two BDBs at the edges of two adjacent poles, (c) two BDBs at the edges of two opposing poles, and (d) two BDBs at two pole pitch distances. [Figure 7] The diagram on the left shows a simulated salient pole synchronous generator and the location of the Hall effect sensor installed in the air gap, and the diagram on the right shows the magnetic flux density of the right sensor under normal (and less than 20%) static eccentricity faults. [Figure 8]Pole diagrams show the simulated average magnetic flux density for each pole for a) the right sensor and b) the left sensor. [Figure 9] FIG. 10 is a diagram showing a detailed procedure for signal extraction in wavelet transform. [Figure 10] The wavelet coefficients of D7 values ​​for a normal machine and a machine with static eccentricity fault are shown. [Figure 11] 10 is a plot showing the air-gap flux density of an SPSG at no load with seven shorted turns. [Figure 12] 1 is a polar diagram of the average magnetic flux density of the SPSG poles under no load. [Figure 13] 1 shows the frequency spectrum of the radial magnetic flux density under no load. [Figure 14] The frequency spectrum of the magnetic flux density for the sum of the two sensors at no load is shown. [Figure 15] The experimental results of the radial magnetic flux density measurement under no-load conditions are shown. DETAILED DESCRIPTION OF THE INVENTION

[0035] As described above, the synchronous machine fault detection method can use various parameters related to the magnetic field generated within the synchronous machine, including parameters based on one or more of the following: magnetic field strength, rotor current or voltage, stator current or voltage, and vibration. These sensor measurements are processed to identify magnetic field-related data artifacts with the intent of finding patterns indicative of magnetic field irregularities to diagnose faults. The processing of the sensor measurements involves one or more signal processing techniques based on time, frequency, and both time and frequency. The frequencies of interest are higher than the line frequency of the synchronous machine and its harmonics. These may be frequencies greater than 75 Hz, e.g., 75 Hz to 400 Hz, and / or may include frequencies in the kHz range or higher. Typically, the sampling frequency used for this method is in the kHz range or higher. Figure 1 shows a flowchart of the major steps of an exemplary fault detection process.

[0036] As shown in the following examples, the proposal involves analyzing the output of signal processing to identify and classify magnetic field irregularities indicative of synchronous machine faults. It will be appreciated that this analysis may utilize computer-aided pattern recognition techniques to recognize patterns in the processed sensor measurements. Several possible examples are described below to illustrate possible ways to implement such a method for different types of faults.

[0037] Example 1 - Broken Damper Bar (BDB) Previously, research into BDB and end-ring faults in large salient-pole synchronous generators (SPSGs) has been limited due to the low statistical population of these faults compared to other types of faults. Damper bars are used for synchronization of SPSGs and rapid-response transients. Additionally, damper bars protect the rotor winding during short-circuit faults in the stator bars. BDB faults in synchronous machines can occur due to insufficient rigid connections between the damper bars and the end-rings. Even small breaks increase the current density in the adjacent bars of a BDB. BDB currents flow through the adjacent damper bars, which leads to excessive resistive losses and consequent temperature increases. Therefore, BDB faults create hot spots around the adjacent bars, which in turn increase the breakage rate of the adjacent bars.

[0038] A study of a pumped storage machine with a BDB fault has yielded some useful results. The electromagnetic torque required during machine starting is partially provided by the current through the damper bar. The magnetic flux density of the machine under fault conditions was studied. It was shown that a BDB fault causes an asymmetric magnetic field. In addition, the starting time of the machine under fault conditions was also investigated and proved to increase with the fault. However, the starting time is not an appropriate indicator for diagnosing a BDB fault, as many factors, such as misalignment, eccentricity fault, or load conditions, can affect the machine starting time.

[0039] The damper bars in a synchronous motor are used for synchronization and damping purposes until the rotor reaches synchronous speed. Therefore, the characteristics of the synchronous machine are similar to that of an induction motor before synchronization. Under steady-state conditions, the amplitude of the current passing through the bars is low, making BDB fault detection difficult.

[0040] This example proposes a novel method for detecting BDB faults in SPSGs during startup. The startup procedure involves first rotating the machine at its nominal speed, and then increasing the excitation current of the rotor field winding as a ramp function over a period of several seconds.

[0041] Modeling SPSG with BDB using FEM A two-dimensional (2-D) scheme of the SPSG simulated using the finite element method (FEM) is shown in Figure 2. Complete geometric and physical details of the simulated SPSG, such as the stator slots, rotor pole saliency, and damper bars, are considered. The nonlinear characteristics of the laminated magnetic core, vortex effects, are considered to simulate the SPSG under normal and BDB fault conditions. In this simulation, the SPSG is analyzed under synchronous speed with the rotor field current increasing from 0 to its nominal value. A transient analysis during the machine voltage ramp-up was performed to simulate this SPSG. In this FEM model, the equations of motion are considered to include the magnetic and coupled mechanical forces, and the electrical equations are considered to describe the rotor field supply.

[0042] Furthermore, saturation conditions, stator and rotor slot design have a significant impact on the fault signature, which must be considered in FEM modeling. The machine winding configuration and time and space harmonics due to the power supply play an important role in the fault detection procedure. The DC current supplied to the rotor field winding by the power electronics can cause special time harmonics in the air-gap magnetic field. Furthermore, fractional slot windings on the stator can also have significant side effects on the magnetic flux density and the resulting stator terminal voltage and load current.

[0043] In this model, a 100 kVA rated SPSG with stator and rotor laminations made of M-400 material is simulated. The specifications of an exemplary SPSG are shown in Table I. The simulated SPSG model has two layers of fractional slot stator windings and a rotor field winding powered by an ideal DC source. The BDB during SPSG startup is modeled using FEM. Figure 3 shows the distribution location of the damper bars in the SPSG. In the case of a fault, it is assumed that the damper bars are completely broken and the corresponding current is zero.

[0044] [Table 1]

[0045] Theoretical analysis of BDB failure Pulsating magnetic field due to a faulty damper bar The magnetic field in the air gap of an SPSG consists of magnetic fields from both the stator and rotor, in addition to the pulsating magnetic flux density from the damper windings. The magnetic flux density in the air gap during startup includes only the magnetic flux density of the rotor and damper bars. One well-known and practical method of fault diagnosis focuses on monitoring the air gap magnetic flux density [8]. To obtain this signal, Hall effect sensors or search coils attached to the stator teeth or slots must be utilized. During transient operation of a synchronous machine, time harmonics, in addition to space harmonics, induce voltages in the rotor damper bars. The damper bars are short-circuited by the end rings at both ends of the rotor poles. As a result, current can pass through the damper bars, generating a magnetic field in the air gap. The amplitude of this magnetic field varies significantly from transient to steady-state operation.

[0046] To analytically monitor the pulsating magnetic field during machine starting, the magnetomotive force (MMF) of the damper bar can be calculated as described below.

[0047]

number

[0048] where p is the number of pole pairs, α is the angle of the damper bar relative to a reference point in a 2-D plane in radians (Figure 4), and I j is the damper bar current, ω is the angular velocity, and ζ is the spatial harmonic number since the harmonics of the winding function are given by ζ=1±6n (n is an integer).

[0049] Pulsating magnetic flux density (B p ) in terms of the generated magnetomotive force of the damper loop at the effective length (l) of the machine pole is given by:

[0050]

number

[0051] In normal operation of an SPSG, the rotor magnetic field and the air-gap magnetic field from the damper bars are symmetric. According to the above equation, the current in the faulted bar passes through the adjacent bars, which increases the current density in the loop and causes local saturation. The asymmetric current distribution in the SPSG rotor bars results in an unbalanced magnetic field in the air-gap.

[0052] Induced voltage in field winding due to BDB fault The unbalanced magnetic field in the air gap due to the BDB fault induces a voltage in the rotor field winding. The total flux due to this distributed air gap flux density connecting the poles and damper bars is given by:

[0053]

number

[0054] α i and α j is the position of the rotor pole through which the magnetic flux link passes, so r r is the outer diameter of the rotor. According to Faraday's law, the induced voltage of the rotor field winding is as follows:

[0055]

number

[0056] where N is the number of turns in the rotor winding. Figure 4 shows the induced voltage in the field winding of an SPSG under normal and fault conditions. The induced voltage in the normal case is due to rotor and stator slot harmonics and the inherent asymmetry of the machine. A BDB fault increases the pulsating magnetic field which distorts the air-gap field, resulting in an increase in the amplitude of the induced voltage in the rotor field winding.

[0057] The amplitude of the induced voltage in the rotor field winding of an SPSG is directly related to the location and number of BDBs on the rotor pole. The current amplitude in the middle bar is smaller than that of the adjacent bars. The reluctance of the path through which the interlinkage flux reaches the stator core through the middle bar is smaller than that of other bars at the rotor pole edge. Therefore, the amplitude of the induced voltage in the excitation winding with a BDB at the rotor pole center must be smaller than that at the rotor pole edge. As can be seen in Figure 4, the amplitude of the induced voltage with three BDBs at the pole center is smaller than that with one BDB at the rotor pole edge.

[0058] Feature Extraction Feature extraction is an essential part of the fault detection procedure for electric machines. Appropriate metrics must be used to examine the most affected signals of an SPSG under a BDB fault. After examining the sensitivity of signals obtained through FE simulation to BDB faults, the induced voltage of the rotor field winding was selected because of its high fault sensitivity compared to other signals. Meanwhile, the air-gap magnetic field of a machine is perhaps the most reliable signal that can be used for various fault detection purposes. However, fault detection based on air-gap magnetic flux density is an invasive method that requires installing sensors inside the machine, which is almost impossible. Signal processing tools for fault detection in electric machines are also important. Signals acquired during transient periods are non-stationary, and therefore most signal processing tools, such as FFT, are not applicable.

[0059] The radius of rotation (RG) can be used as a suitable indicator to probe the trend of BDB faults in SPSGs. RG has a specific value for any number of broken damper bars. RG is based on a time series data mining (TSDM) approach. TSDM is applied to the induced voltage in the rotor field winding to detect hidden patterns due to BDB faults in SPSGs during transient conditions.

[0060] Time Series Data Mining Methods TSDM is a nonlinear signal processing approach based on a discrete stochastic model of a reconstructed phase space based on dynamical systems theory. It is proven that a metrically equivalent state space can be reproduced by a single sampled state variable. In other words, the dynamic invariants are also preserved in the reconstructed state space. To recover the state space of the SPSG, the induced voltage of the rotor field winding in two states, normal and faulty, is considered as the state variable. In other words, the acquired signals can reproduce a topologically equivalent state space similar to the original SPSG system in normal or faulty cases.

[0061] Two methods can be used to reconstruct the state space: time-delay embedding and differential embedding. Differential embedding is not practical for experimental results because the results have high-order derivatives that are sensitive to noise. Therefore, to find the invariants of the dynamic system, time-delay embedding is chosen, which converts scalar points into vector form. Assume that the time series of the induced voltage of the field winding is as follows:

[0062] e={e(j)-e(j-1),j=2,3,...,K}

[0063] where j is the time index and K is the number of sampled signals. The reconstructed state space (also called phase space) for j equals 10 for healthy and faulty SPSGs is shown in Figure 5. RG is used to quantify any change in the area of ​​the mass generated by the TSDM to distinguish between healthy and faulty states, and can be shown as follows:

[0064]

number

[0065] where l is the time delay in phase space, μ and μ l are the centers of rotation for each dimension.

[0066] When a BDB fault occurs, the RG amplitude increases. Because the machine has some inherent asymmetry, the RG amplitude was present even in the normal case. Figure 5 shows the phase space of a normal SPSG with one BDB and a faulted SPSG. According to Figure 5, the BDB fault increases the radius of the mass, as the radius is 0.1119 in the normal case and increases to 2.4307 when one damper bar breaks. It can be seen that when the damper bar at the center of the rotor pole (DB-4) breaks, the RG amplitude does not have a significant sensitivity, as the lowest current density passes through that bar.

[0067] Furthermore, the magnitude of the RG with three BDBs in the center of a pole is smaller than with one BDB at the edge because their current is lower. However, the increase in RG due to a fault depends on the location and number of BDBs. For example, the amount of current passing through the bars at the edge of a rotor pole is the largest, resulting in more pronounced fluctuations relative to the other bars. The location of the BDBs is an important factor that can modify the RG amplitude. If a BDB fault occurs on two edge bars of the same pole, the RG value is smaller than when there is one edge BDB. The pulsating field due to two BDBs at the edge of a pole is twice that of a single BDB. However, due to the symmetry of the fault, the pulsating field almost cancels out. However, there is still an asymmetry in the air-gap field that does not completely cancel out their effects, resulting in an RG index. The worst case scenario, which significantly increases the RG amplitude, is when two BDBs are adjacent to each other on the same pole. In this case, the current from the two BDBs passes through the third BDB, causing local saturation. As a result, the pulsating magnetic field, local saturation, and lack of magnetic field due to the BDB result in a strong unbalanced magnetic field, which induces large voltages in the rotor field winding.

[0068] Effect of location of broken damper bar on the proposed index The magnitude of RG depends on the asymmetry level of the air-gap magnetic field caused by the location of the BDB fault. Figure 6 shows the location of the BDB fault at different rotor poles. The third and fourth columns of Table II show the variation of the RG index with respect to the location of the BDB fault at different poles.

[0069] [Table 2]

[0070] In case (a), the RG amplitude increases by a factor of two compared to three BDBs on one pole. However, due to the lower current density in the middle bar, its value is not expected to increase with the number of BDBs. Case (b), with two BDBs on the two edges of adjacent rotor poles, is expected to have a higher degree of RG, but the amplitude is not larger compared to when there is one BDB on the edge of one pole. The magnetic flux density level varies around the circumference of one pole (north or south). In case (c), both rotor poles have the same magnetic flux density polarity, which increases the RG amplitude by twice that of the BDBs on the pole edges. In case (d), the amplitude does not increase, but only partially decreases, which can be explained based on case (b).

[0071] conclusion This example deals with detailed modeling, analytical study, and condition monitoring of an SPSG under a BDB fault. Analytical methods prove that the damper bars induce pulsating magnetic fields due to fluctuations in damper bar current as a result of a BDB fault, which can distort the air-gap magnetic flux density. This flux distortion induces electromotive forces in the rotor field winding, which has extreme sensitivity to BDB faults compared to non-invasive methods. To extract accurate features for diagnosing BDB faults, a time-series data mining method was applied to the induced voltage at the rotor field terminals. This index was shown to have high sensitivity to BDB faults. Furthermore, the effect of the number of bars and their location on the RG was investigated. It was shown that the RG amplitude increases with an increase in the number of BDBs. Furthermore, the magnitude of the RG can increase when BDBs are placed on poles with the same polarity.

[0072] Example 2 - Static eccentricity If the minimum air gap between the rotor core and the stator core fluctuates, the synchronous generator will experience a condition known as eccentricity fault. There are two types of eccentricity: static and dynamic. Static eccentricity occurs when the minimum length of the unevenly distributed air gap is a fixed length and spatially fixed. Static eccentricity is a common fault in synchronous generators. Even small eccentricities will not cause damage to the machine, but they should be detected early before the rotor core rubs against the stator core and windings.

[0073] Over the years, many methods have been explored to diagnose static eccentricity faults in synchronous generators. Fault detection based on non-invasive techniques such as stator terminal voltage or current, or machine parameters, have been used to detect eccentricity faults in synchronous generators. However, these methods are unable to detect faults early because the rotor topology may mask fault indicators in the aforementioned signals.

[0074] In this example, static eccentricity faults in salient pole synchronous generators are detected using the air gap magnetic flux density. A finite element approach is used to simulate the synchronous machine. The location and number of Hall effect sensors installed in the air gap of the machine are described. Wavelet transform is used as a processing tool to handle the magnetic field signals. A new index is introduced to detect static eccentricity faults. It is demonstrated that the proposed criterion index can accurately detect the normal or faulty state of the machine and the severity of the fault.

[0075] Finite element modeling: A 100 kVA salient pole synchronous generator has been modeled using two-dimensional finite element techniques. The synchronous machine has been modeled in two cases, namely, normal eccentricity and static eccentricity with different levels of severity. The synchronous generator is tested at no load, constant synchronous speed and rated current of the rotor field winding. The finite element modeling of the synchronous machine for the normal case is shown in Figure 7. The specifications of the synchronous generator are listed in Table III below.

[0076] [Table 3]

[0077] Static eccentricity measurement points: The magnetic flux density distribution does not vary within a full rotor rotation under static eccentricity faults. However, this is not true for rotor field winding inter-turn faults or dynamic eccentricity faults. Static eccentricity causes the magnetic flux density to vary with position, so the location of the magnetic flux density measurement points is important. Figure 7 shows the locations of the measurement points in the air gaps of two pairs of synchronous generators. Two measurement points are referred to as the right measurement point and the left measurement point, and the other two points are referred to as the top measurement point and the bottom measurement point (the locations of the sensors are shown in red circles). The locations of the top and bottom measurement points are assigned at a 90-degree angle to the left and right measurement points.

[0078] For each pair of measuring points to experience the same variation in flux density under normal or static eccentricity conditions, they must be located on opposite sides of the synchronous machine. However, the variation in flux density at each measuring point must not be exactly the same due to lagging flux phenomena.

[0079] The magnetic flux density measurement points must detect static eccentricity faults regardless of the fault orientation. When the measurement points are positioned perpendicular to the static eccentricity orientation, there is no significant variation in the air gap magnetic flux density. As a result, the measured magnetic fields in the direct and transverse axes are separated and can be used to detect the direction of the static eccentricity fault. However, with four measurement points distributed along the air gap in the above configuration, eccentricity faults must be detected at a single pair of measurement points, regardless of the fault orientation. The maximum angle between the measurement points and the static eccentricity orientation must not exceed 90 degrees. Figure 7 shows the variation in magnetic flux density at the correct measurement points for normal and static eccentricity faults of less than 20%.

[0080] In this example, static eccentricity was imposed along the positive x-axis, the y-axis, and 45 degrees with respect to the positive x-axis in all simulations. The air gap length increases at the right measurement point and decreases equally at the left measurement point. In other words, the amplitude of the magnetic flux density at the left measurement point under a given level of static eccentricity increases compared to the normal case. Figure 8 shows the polar diagrams of the average magnetic flux density for each pole for the right and left measurement points under normal and eccentricity fault conditions.

[0081] Signal Processing: The wavelet transform (WT) is a useful signal processing tool used in various fields, such as power system and electromechanical analysis. The wavelet transform exploits the temporal localization of different frequency components of a signal. Unlike traditional frequency-domain signal processing tools, the wavelet transform does not use a fixed-width window. The wavelet analysis function adjusts the time width of a given signal according to its frequency components, with lower frequencies in a wider window and higher frequencies in a narrower window. In other words, the wavelet transform can be used to handle signals with vibrations and localized impulses by decomposing high-frequency components into shorter time intervals and low-frequency components into longer time intervals.

[0082] In this example, we use Daubechies-8 as the mother wavelet. Higher-order wavelets, like D-8, have higher resolution to improve the quality of fault detection in electrical machines. Figure 9 shows the signal decomposition procedure using a discrete wavelet transform, where S is the input signal, and LPF and HPF are low-pass and high-pass filters. Prior to the wavelet transform, a given signal is divided into two halves, which are the inputs to the LPF and HPF. The output of the first-level LPF is separated into half of the frequency bandwidth. This procedure continues until the given signal is decomposed to a predefined value of that level. The sampling frequency in this example is 10 kHz, and based on the Nyquist theorem, the highest frequency a signal can contain is 5 kHz. As a result, the frequency bandwidth of the first level of the wavelet transform needs to be 5 to 2.5 kHz.

[0083] Fault detection: A wavelet transform (Daubechies-8) is applied to the magnetic field signal acquired using a Hall-effect sensor in the air gap. Level 7 of the wavelet transform shows better sensitivity to increasing static eccentricity compared to other wavelet levels in a salient-pole synchronous generator. Figure 10 presents the absolute values ​​of the detailed signal at level 7 (D7) coefficients for normal and static eccentricity levels less than 2.5%, 10%, and 20%. A comparison of the magnitude of the wavelet coefficients for static eccentricity cases under different severity levels shows that faults increase the vibration level at D7. This is due to air gap irregularities, which lead to an increase in the amplitude of subharmonics in the magnetic field. A novel reference index is proposed to quantify the value of D7 under different eccentricity fault levels.

[0084]

number

[0085] Here, the oscillation of the absolute value of the wavelet coefficient (D7) is defined per unit with respect to the average value of the air-gap magnetic field. According to Table VI, the value of the proposed index for the right sensor decreases, while the amplitude of the index for the left sensor increases. This is because, in the case of static eccentricity in the positive x-axis direction, the air-gap length on the right side increases, which leads to a decrease in the magnetic field in the air-gap. Therefore, the proposed feature should decrease. It can be seen that the reference index decreased from 0.6007 in the normal case to 0.5969, 0.5975, and 0.5597 in the cases of 2.5%, 10%, and 20% static eccentricity. The difference between the normal reference index and the 2.5% static eccentricity proves that the proposed feature can detect static eccentricity faults in their early stages.

[0086] [Table 4]

[0087] Conclusion: This example introduces a novel feature as a reference indicator for early detection of static eccentricity faults in salient-pole synchronous generators. The magnetic field acquired using Hall-effect sensors installed in the air gap of the synchronous machine is utilized, as it provides relevant information about irregularities due to faults in the air gap. A finite element approach is used to model the synchronous machine under normal and faulty conditions, taking into account all detailed geometric and material properties. Daubechies-8 is used as the mother wavelet to analyze the magnetic field under normal and faulty conditions. Its accuracy demonstrates the ability of this indicator to detect faults early. In the final version of this paper, results for eccentricity on two other axes (y-axis, at 45 degrees relative to the positive x-axis) will be added. A section on fault location detection will also be added. Furthermore, the simulations are validated by experimental results.

[0088] Example 3 - Short Circuit This example includes a detailed electromagnetic analysis of an SPSG under normal and incipient inter-turn short-circuit faults using FE methods. Based on the mean radial flux, polar diagram, sum of two sensor flux densities, and frequency spectrum monitoring, a procedure is proposed for short-circuit faults in the excitation winding. The effects of sensor location, sampling frequency, and data resampling are studied. Load effects on the proposed method are examined. The simulation results are verified by a custom-built 100kVA SPSG.

[0089] electromagnetic analysis Reliable fault diagnosis requires a sufficiently accurate modeling method. In this example, the time-stepping finite element method is used to simulate the SPSG. The modeling takes into account the detailed geometric complexity of the machine, such as the stator slots and rotor damper bars. In addition, it includes rotor pole saliency, the spatial distribution of the armature windings, and the nonlinearity of the core material. The specifications of the proposed SPSG are summarized in Table V below.

[0090] [Table 5]

[0091] Under a turn-to-turn short circuit fault, the total magnetomotive force of the faulted pole is reduced. As a result, the magnetic flux density of the working pole decreases, distorting the air-gap magnetic field. Two Hall-effect sensors are positioned in opposite directions on the stator teeth to measure the air-gap radial magnetic field (these two sensors are designated as the right and left points). The modeled SPSG has 14 poles, each with 35 turns. Turn-to-turn short circuit faults of different severity, ranging from 1 to 10 turns, are simulated. Figure 11 shows the variation of the air-gap magnetic field for one complete rotor revolution, with 7 of the 35 turns shorted in one pole. Because the faulted pole with reduced ampere turns experiences a reduced magnetomotive force, when the sensor sweeps the faulted pole, the magnetic flux density in the air-gap has a smaller amplitude than that of a healthy pole. By comparing the average magnetic flux density of each pole, turn-to-turn short circuit fault diagnosis is possible. However, fault detection is difficult when the number of shorted turns or the ratio of shorted turns to the total pole turns is low.

[0092] Figure 12 shows the average magnetic flux density of each pole in space as pole diagrams for a normal and a faulty synchronous generator. In a normal, ideal machine, the radial distance from the origin of the pole diagram to the average value of each pole is equal. The average value of the rotor pole field decreases with increasing number of shorted turns. The average magnetic flux density of the faulty pole relative to the average magnetic flux density of all poles for 1, 2, 3, 7, and 10 shorted turns is 99%, 97.9%, 96.8%, 93.7%, and 88.8%, respectively.

[0093] The frequency spectrum of the air-gap magnetic field in Figure 13 shows the rapid change in sideband harmonics as a result of an inter-turn short circuit fault in the excitation winding. The amplitude of the fault-related harmonics increased with increasing number of shorted turns. The exponential frequency identifiers are as follows:

[0094] f fault =f s ±kf r

[0095] In the formula, f s is the electrical frequency, f r is the rotor mechanical frequency, p is the number of pole pairs, and k is an integer. The plot in Figure 13 has the largest amplitude spike for a healthy machine, with smaller amplitude spikes overlaid, from top to bottom, at 10, 7, 3, 2, 1, and possibly normal. The most significant fault-related harmonic components in the frequency spectrum appear at frequencies lower than the fundamental harmonic. The amplitude of the 7.14 Hz component with one shorted turn is approximately 60.22 dB or approximately 1 mT. With two shorted turns, the amplitude increases to approximately 2 mT, and with 10 turns, the amplitude is approximately 10.3 mT. The magnitudes of all fault-related harmonics show a nearly linear increase in amplitude with increasing number of shorted turns.

[0096] The frequency spectrum of the sum of the magnetic fields of two Hall-effect sensors installed in the air gap can also reveal the machine condition. Under normal, balanced operating conditions, the two sensors, based on theory, experience identical fluctuations in magnetic flux density, summing to zero. In the case of a shorted turn in the field winding, the reduced magnetic flux density of the faulted pole causes a spike in the sum of the magnetic flux density each time the faulted pole passes one of the measurement points. Simulation results for a normal synchronous generator, as shown in Figure 14, show that the fundamental component and its odd multiples effectively cancel out in the frequency spectrum. Thus, the plot in Figure 14 shows that the components of the normal machine are canceled out, with the remaining spikes being overlaid by 10, 7, 3, 2, and 1 turns, from top to bottom. The frequency components appearing in the frequency spectrum are a direct result of the shorted turn in the field winding distorting the inherent magnetic symmetry of the normal machine. Additionally, the frequency spectrum contains fewer harmonics associated with the fault; this spectrum contains only the odd multiples. The total amplitude of the magnetic field spectrum is higher than the radial flux spectrum which makes fault detection easier.

[0097] Experimental Test Equipment A 100 kVA salient pole synchronous generator was used to verify the proposed theory. Apart from the size, the generator topology is similar to real hydroelectric generators in typical hydroelectric power plants such as those used in Norway. The experimental SPSG has 14 poles and the air gap length is 1.75 mm to achieve the appropriate synchronous reactance.

[0098] The absolute values ​​of the magnetic flux density for the faulted pole and the two rear poles passing through the right sensor are shown in Figure 15. The faulted pole responded as expected, with the magnetic flux density decreasing with increasing number of shorted turns. Thus, in the leftmost section of Figure 15, the highest line is the normal one, followed by lines moving downwards showing turns 1, 2, 3, 7, and 10, with 10 being the lowest line. The magnetic flux density for the adjacent pole with the opposite polarity also decreases slightly as the number of shorted turns increases. This is a result of the reduced magnetic flux tracing a path through the faulted pole to the adjacent pole of opposite polarity.

[0099] conclusion In this example, an FE method is used to study an inter-turn short circuit fault in the excitation winding of a salient-pole synchronous generator due to an air-gap magnetic field. The FE results are verified using a 100-kV custom-built SPSG. The proposed procedure for inter-turn short circuit fault detection is based on using two Hall-effect sensors with opposite orientations. The severity and location of the fault can be revealed by comparing the average magnetic field in the pole diagram, the sum of the two sensors' magnetic fields, and their frequency spectra. Experimental verification results are as follows: By analyzing the average magnetic field distribution or the pole diagram, multiple shorted turns could be immediately detected. A uniform increase in the amplitude of low-order harmonic components in the magnetic flux density spectrum could be observed. Monitoring the sum of the magnetic fluxes of the two sensors in a machine with only short-circuit faults can be an invaluable tool for diagnostic purposes, as the magnitude of the sum of the magnetic flux density spectrum responds better with increasing numbers of shorted turns. The measurement needs to be coordinated with an encoder to identify the location of the faulty pole.

[0100] Machine Learning The use of machine learning algorithms enhances the effectiveness of the above examples by providing automation and reproducibility for recognizing patterns in processed signals. With reference to Example 1, a machine learning pattern recognition system can be utilized to determine a match between a measured / calculated radius of rotation and a known radius of rotation pattern indicative of a broken damper bar. With reference to Example 2, a machine learning pattern recognition system can be used to evaluate data including criteria indices to identify indications of possible static eccentricity faults. With reference to Example 3, a machine learning system can be used to replace some or all of the human input associated with analyzing mean field distributions or polar diagrams.

[0101] The sensor types used in the above examples, as well as other sensor types discussed herein, may be used in combination, and the various signal processing steps and generation of indices may also be performed in combination. A combined signal processing system may perform all of the signal processing, and the resulting processed data may be automatically evaluated, such as via a machine learning system, to check for various types of faults.

Claims

1. 1. A method for fault detection in a synchronous machine, said method comprising: determining, using at least one sensor, parameters related to the magnetic field generated within the synchronous machine, including parameters based on one or more of magnetic field strength, rotor current or voltage, stator current or voltage, and vibration; processing the sensor measurements to identify data artifacts related to the magnetic field, the processing including one or more time, frequency, and both time and frequency based signal processing techniques; analyzing an output of the signal processing to identify and classify irregularities in the magnetic field that are indicative of faults in the synchronous machine, the analysis including recognizing patterns in the processed sensor measurements through the use of computer-aided pattern recognition techniques, such as through machine learning algorithms.

2. The method of claim 1 , wherein the method is for detecting faults associated with one or more of an eccentricity fault, a damper winding fault, and a short circuit fault.

3. 3. The method of claim 1 or 2, wherein the method is for fault detection in large synchronous machines such as hydroelectric generators.

4. 4. A method according to claim 1, 2 or 3, wherein the method is carried out during the presence of transient magnetic fields of the type that occur during start-up or stop-up of the machine.

5. 10. The method of any one of the preceding claims, wherein the sensors comprise at least one pre-existing sensor or at least one non-invasive sensor.

6. 10. A method as claimed in any one of the preceding claims, comprising using one or more pre-existing sensors, including one or more voltage or current sensors configured for use during operation of the synchronous machine for monitoring and / or control of the machine.

7. 10. A method according to any one of the preceding claims, wherein at least one non-invasive sensor is provided as part of a fault detection system, the sensor being temporarily located with the machine for carrying out the fault detection method.

8. 10. A method according to any one of the preceding claims, wherein the method does not require any physical modifications to the synchronous machine in order to place or install new sensors.

9. 10. A method according to any one of the preceding claims, wherein the sensor measurements do not include data transferred from a moving part of the rotor.

10. 10. A method as claimed in any one of the preceding claims, comprising using a sensor provided in a potential transformer of the synchronous machine.

11. 10. A method according to any one of the preceding claims, comprising determining a radius of rotation of the rotor of the synchronous machine and using the radius of rotation to identify and classify irregularities in the magnetic field.

12. 12. The method of claim 11, comprising obtaining turning radius data obtained via time series data mining, and using the turning radius data to detect damper winding or eccentricity faults.

13. 13. A method according to claim 11 or 12, comprising mapping a phase space for induced voltages in a field winding of the synchronous machine.

14. The method of claim 13 , wherein the induced voltage is determined or measured via a sensor in the potential transformer.

15. 15. A method according to any one of claims 11 to 14, comprising attempting to identify damper winding faults in the form of broken damper bars, said analysing step comprising comparing the measured turning radius with an equivalent measurement of the turning radius of a healthy synchronous machine.

16. 16. A method according to any one of claims 11 to 15, comprising attempting to identify eccentricity faults, and wherein the step of analysing the output of the signal processing comprises determining a normalised radius of rotation with reference to a radius of rotation of a machine known to be normal, the normalised radius of rotation being defined as the difference between the radius of rotation of a normal synchronous machine and the measured radius of rotation, this difference being divided by the normal radius of rotation.

17. 17. The method of claim 16, including automatic identification and classification of possible eccentricity faults by evaluating the normalized radius of gyration value against a threshold value.

18. 10. A method according to any one of the preceding claims, wherein the computer-aided pattern recognition technique is based on a machine learning algorithm trained using a plurality of processed sensor measurements known to be associated with fault-free machinery and a plurality of processed sensor measurements known to be associated with machinery having a fault.

19. 1. A fault detection system for fault detection in a synchronous machine, the fault detection system comprising: a data processing device for connecting to at least one sensor to receive parameters related to the magnetic field generated within the synchronous machine, the parameters including parameters based on one or more of magnetic field strength, rotor current or voltage, stator current or voltage, and vibration; The data processing device processing the sensor measurements to identify data artifacts related to the magnetic field, the processing including one or more time, frequency, and both time and frequency based signal processing techniques; and analyzing the output of the signal processing to identify and classify irregularities in the magnetic field that are indicative of faults in the synchronous machine, the analysis including recognizing patterns in the processed sensor measurements through the use of computer-aided pattern recognition techniques, such as through machine learning algorithms.

20. A fault detection system according to claim 19, wherein the data processing device is configured to carry out a method according to any one of claims 1 to 18.

21. 21. A fault detection system according to claim 19 or 20, comprising at least one conventional sensor and / or at least one non-invasive sensor.

22. 22. A fault detection system according to claim 19, 20 or 21, wherein the fault detection system uses sensors without data transfer from the moving part of the rotor.

23. A large synchronous machine acting as a hydroelectric generator, said synchronous machine including a fault detection system according to any one of claims 19 to 22.

24. A computer program product comprising instructions, which when executed in a fault detection system according to any one of claims 19 to 23, cause a data processing device of said fault detection system to: processing the sensor measurements to identify data artifacts related to the magnetic field, said processing including one or more time, frequency, and both time and frequency based signal processing techniques; and analyzing an output of the signal processing to identify and classify irregularities in the magnetic field that are indicative of faults in a synchronous machine, the analysis including recognizing patterns in the processed sensor measurements through the use of computer-aided pattern recognition techniques, such as through machine learning algorithms.

25. A computer program product according to claim 24, comprising instructions adapted to configure a data processing apparatus to perform the other steps of any one of claims 1 to 18.

Images