Method and apparatus for generating a winding faults detection model and for detecting winding faults in stator windings of an electrical machine

Abstract

A method for generating a winding fault detection model for detecting a winding fault in stator windings of an electrical machine comprises: - collecting batches of samples of three-phase voltage and current signals, - for one or more load conditions, pre-processing a first set of batches to generate, for each batch in the first set of batches, a first set of features comprising sequence components associated with the three-phase current and voltage signals, and a zero-sequence admittance and a negative-sequence impedance of the electrical machine obtained from said sequence components, - calculating learned values of the zero-sequence admittance and negative-sequence impedance, - calculating, for each batch in a second set of batches, a second set of features comprising an inter-turn and a phase-to-ground fault severity index based on the learned values, - training a non-supervised anomaly detection model to detect winding faults, based on the second sets of features, - storing the anomaly detection model for subsequent inference.

Description

METHOD AND APPARATUS FOR GENERATING A WINDING FAULTS DETECTION MODEL AND FOR DETECTING WINDING FAULTS IN STATOR WINDINGS OF AN ELECTRICAL MACHINE

[0001] TECHNICAL FIELD

[0002] Various example embodiments relate generally to methods and apparatus for detecting winding faults in stator windings of electrical machines, for example motors or generators or transformers, supplied with a three-phase voltage signal drawing a three-phase current signal through the stator windings.

[0003] BACKGROUND

[0004] The most frequent failures in such electrical machinery are stator winding faults.

[0005] Stator winding faults may be caused by the gradual deterioration of windings insulation due to e;g. a combination of electromechanical-force-induced vibrations, high dv / dt voltage surges, thermal overload...

[0006] Stator winding faults may be of different types, essentially three types known as phase-to-ground fault, inter-turn fault and phase-to-phase fault. Phase-to-ground fault happens when the insulation of a winding degrades to the point that a phase conductor comes into contact with the casing of the electrical machine or the ground. Inter-turn fault occurs when the insulation between turns of the same winding (within a single phase) fails. This causes shortcircuiting between these turns, leading to a partial short circuit within the phase itself. Phase- to-phase fault happens when the insulation between two phases breaks down, causing them to short-circuit. All such faults alter the impedance and the current distribution in the affected phases, resulting in uneven impedance across the phases, disrupting the balance between the three phases (the magnitudes of the three phases are no longer equal and / or the phases are no longer 120 degrees apart). Such asymmetry is accompanied by thermal stresses that result in deteriorating the reliability of the electrical machine. It causes various operational inefficiencies and potential failures. Addressing these faults promptly through detection and corrective actions is crucial for maintaining the health and efficiency of industrial electrical machinery.

[0007] Efficient detection and early identification of winding faults in alternative current electrical machinery is still a subject of research.

[0008] SUMMARY

[0009] In a first aspect, a computer implemented method for generating a winding fault detection model is disclosed. This detection model enable detecting a winding fault in stator windings of an electrical machine supplied with a three-phase voltage signal drawing a three- phase current signal through the stator windings,. It comprises:- a collecting step for collecting a plurality of batches of samples of the three-phase voltagesignal and the three-phase current signal,- for one or more load conditions, a pre-processing step for pre-processing a first set of batches to generate, for each batch in the first set of batches, a first set of features comprising at least sequence components associated with the three-phase current and voltage signals, and a zero-sequence admittance and a negative-sequence impedance of the electrical machine obtained from said sequence components,- a first calculating step for calculating learned values of the zero-sequence admittance and negative-sequence impedance of the electrical machine, based on the first sets of features,- a second calculating step for calculating, for each batch in a second set of batches, a second set of features comprising at least an inter-turn fault severity index and a phase-to-ground fault severity index based on the learned values of the zero-sequence admittance and negativesequence impedance of the electrical machine,- a training step for training a non-supervised anomaly detection model to detect winding faults, based on the second sets of features,- a storing step for storing the anomaly detection model for subsequent inference.

[0010] In an embodiment, the first set of batches comprises a predetermined number of batches.

[0011] In an embodiment, the pre-processing step comprises a low pass filtering of the collected batches.

[0012] In an embodiment, the pre-processing step comprises filtering out transients in the collected batches.

[0013] In an embodiment, the pre-processing step comprises removing batches collected before initial zero crossing.

[0014] In an embodiment, the first calculating step comprises applying a recursive least square to adaptively obtain the learned values of the zero-sequence admittance and negativesequence impedance of the electrical machine.

[0015] In an embodiment, the electrical machine has a given capacity and the first set of features further comprises a load percentage representative of the load condition associated with the batch being processed.

[0016] In an embodiment, the sequence components comprise a zero-sequence current component and a negative-sequence voltage component, and the second calculating step comprises:- using the learned values of the zero-sequence admittance to determine a noise factor of the zero-sequence current component and obtain the phase-to-ground fault severity index from the noise factor of the zero-sequence current component,- using the learned values of the negative-sequence impedance to determine a noise factor ofthe negative-sequence voltage component and obtain the inter-turn fault severity index from the noise factor of the negative-sequence voltage component.

[0017] In an embodiment, the sequence components further comprise a negativesequence current component and a zero-sequence voltage component, and the second set of features further comprises a magnitude of the negative-sequence current component and a magnitude of the zero-sequence voltage component.

[0018] In an embodiment, the anomaly detection model is an isolation forest model.

[0019] In a second aspect a computer-implemented method is disclosed for detecting winding faults in an electrical machine, by running inferences of a winding fault detection model obtained through the method disclosed in the first aspect.

[0020] In a third aspect, an apparatus is disclosed comprising at least one processor and at least one memory storing instructions that, when executed by the at least one processor, cause the apparatus at least to perform the method described above in relation to the first and the second aspects.

[0021] With the above-described methods, a non-supervised model, for example an isolation forest model, is used to learn the normal trends of a given healthy electrical machine and establish a stable performance baseline of the machine. A comprehensive understanding of normal operating behavior of the electrical machine is established by extracting features from the three-phase current and voltage signals during the model generation phase. Once trained, the detection model is used to pinpoint any deviations from the established stable performance and report that a winding fault happened. For example, an alert can be raised after a certain number of anomalies are detected. The detection model proposed herein dynamically learns the behavior of each electrical machine specifically, and then detects the fault during inference, based on new collected samples. This allows higher precision compared with known winding fault detection methods.

[0022] Depending on the application, the electrical machine may or may not experience variations in load. The proposed method and apparatus consider load variation when building the detection model by collecting and pre-processing batches of samples of the three-phase current and voltage signals for several load conditions. In other words, the features used to train the detection model are obtained for a plurality of load conditions when the load is variable. Further, the load percentage, which is indicative of the load condition, is generated as part of the pre-processing step and used when calculating the learned values of the zerosequence admittance and negative-sequence impedance of the electrical machine. This allows getting a detection model which is load independent.

[0023] The disclosed method and apparatus are not dependent on the type of sensors being used. Any sensors enabling collection of samples of the three-phase current and voltagesignals can be used.

[0024] The disclosed method and apparatus allow detecting any type of winding fault e.g. inter-turn, phase-to-ground, phase-to-phase faults. For example, this can be used to alert maintenance for immediate action. This not only enhances the reliability and efficiency of the operations of the electrical machine but also minimizes downtime.

[0025] BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Example embodiments will become more fully understood from the detailed description given herein below and the accompanying drawings, which are given by way of illustration only and thus are not limiting of this disclosure.

[0027] FIG.1. is a block diagram of system comprising an electrical machine as disclosed herein.

[0028] FIG.2 is a flow chart of a method for generating a winding fault detection model for detecting a winding fault in stator windings.

[0029] FIG.3 is a detailed view of the pre-processing step of the method described with reference to FIG.2.

[0030] FIG.4 is a detailed view of the second calculating step and the training step of the method described with reference to FIG.2.

[0031] FIG.5 is a flow chart of a method for detecting winding faults in an electrical machine based on the detection model generated according to the method described with reference to FIG.2 to FIG.4.

[0032] DETAILED DESCRIPTION

[0033] Various example embodiments will now be described more fully with reference to the accompanying drawings in which some example embodiments are shown.

[0034] Detailed example embodiments are disclosed herein. However, specific structural and functional details disclosed herein are merely representative for purposes of describing example embodiments. The example embodiments may, however, be embodied in many alternate forms and should not be construed as limited to only the embodiments set forth herein. Accordingly, while example embodiments are capable of various modifications and alternative forms, the embodiments are shown by way of example in the drawings and will be described herein in detail. It should be understood, however, that there is no intent to limit example embodiments to the particular forms disclosed.

[0035] The embodiments disclosed herein relate to a method and apparatus for detecting stator winding faults in three-phase alternative current (AC) electrical machines.

[0036] FIG. 1 is a schematic block representation of a 3-phase system comprising a three-phase input 110 which supplies, through a distribution circuit 120, a three-phase voltage signal to an electrical machine 130, drawing a three-phase current into the stator windings ofthe electrical machine. The system of FIG.1 further comprises an anomaly detection circuit 140 which receives samples of the 3-phases voltage and current signals captured by sensors 150. The anomaly detection circuit 140 comprises means for performing a method for generating a winding faults detection model and / or means for performing a method for detecting winding faults in an electrical machine as disclosed herein. For example, the means comprise at least one processor and at least one memory storing instructions that, when executed by the at least one processor, cause the anomaly detection circuit to perform a method for generating a winding faults detection model and / or a method for detecting winding faults in an electrical machine as disclosed herein.

[0037] A method for generating a winding fault detection model will now be described with reference to FIG.2. The model generation method begins with a collecting step 210, where a plurality of batches of samples of the three-phase voltage signal and the three-phase current signal are collected through sensors 150. These data are used to establish the normal operating behavior of the machine.

[0038] A pre-processing step 220 is performed for a first set of batches comprising a predetermine number of batches Q for P load conditions with Q and P being integers greater than or equal to 1. If the load is stable, there is only one load condition so that P=1 and the first set of batches comprises N=Q batches. When the load is variable, multiple load conditions are considered: a load segmentation is applied, meaning that Q buckets of P batches are pre- processed, each bucket corresponding to a different load condition.

[0039] The first set of batches is pre-processed to generate a first set of features which comprise sequence components associated with the three-phase current and voltage signals, and a zero-sequence admittance and a negative-sequence impedance of the electrical machine obtained from said sequence components.

[0040] Sequence components are a set of orthogonal components used to analyze unbalanced conditions in three-phase electrical systems. They are derived from the symmetrical components theory which states that the three-phase voltage and current can be mathematically reduced to three distinct sets of symmetrical sequence components referred to as positive, negative and zero-sequence components. The positive-sequence components represent the symmetrical part of the three-phase system and have equal magnitudes and 120-degree phase differences. The negative-sequence components represent the unsymmetrical part caused by an unbalanced load or fault in one phase and have equal magnitudes and 120-degree phase differences. Negative-sequence components are indicative of unbalance in the system. The zero-sequence components represent the common-mode part of the three-phase system, such as the neutral current or ground fault current. Zero-sequence components are typically present in systems with ground faults.

[0041] Admittance is a measure of how easily a circuit allows current to flow. Zero-sequence admittance refers to the admittance calculated using the zero-sequence components of the voltage and current signals. The analysis of the zero-sequence admittance provides insights into how effectively the machine’s neutral is grounded. For example, it can be used to detect phase-to-ground winding faults.

[0042] Impedance is the inverse of admittance. It measures the opposition that a circuit presents to the flow of alternating current. Negative-sequence impedance is calculated using the negative-sequence components of the voltage and current signals. It provides a measure of the opposition to current flow under unbalanced conditions. For example, it can be used to detect inter-turn winding faults.

[0043] As will be further described below with reference to FIG.3, the pre-processing step prepares the data in the first set of batches and performs the extraction of the first set of features based on the prepared data. Data preparation ensures that the data are suitable for the following steps of the method.

[0044] After the pre-processing step 220, a first calculating step 230 occurs. In this first calculating step 230, learned values of the zero-sequence admittance and the negativesequence impedance of the electrical machine are calculated based on the first sets of features generated at step 220 using a recursive least square method or an alternative method. The learned values represent the operating parameters of the electrical machine under normal conditions, which will be used at step 240 to calculate fault severity indexes used for training the anomaly detection model.

[0045] At step 240, a second set of features is calculated, for a second set of batches, from the learned values of the positive and negative admittance and the negative-sequence impedance. The second set of batches may or may not overlap with the first set of batches. The second set of features comprises an inter-turn fault severity index and a phase-to-ground fault severity index.

[0046] At step 250, when enough values of inter-turn and phase-to-ground fault severity indexes have been obtained, a non-supervised anomaly detection model is trained to detect winding faults based on the second set of features. The anomaly detection model, once trained, can identify patterns in the data that deviate from the established baseline, which may indicate the presence of a winding fault. The trained model is stored at step 260 for subsequent inference. Advantageously, the non-supervised anomaly detection model is an isolation forest model. This is not limitative: other non-supervised anomaly detection model known in the art can be used as well.

[0047] FIG.3 is a detailed view of an exemplary embodiment of the pre-processing step220 of the method described with reference to FIG.2. It shall be understood that some of the steps described in relation to FIG.3 can be omitted depending on the specific implementation.

[0048] In the embodiment depicted with reference to FIG.3, the pre-processing step 220 includes several steps intended for cleaning the data used to generate the first set of features. At step 320 a low pass filtering is applied to the three phase voltage data when the machine starter is a Variable-Frequency Drive (VFD) starter. Otherwise, when the machine starter is a Direct On-Line (DOL) started, step 320 is omitted.

[0049] Indeed, VFDs control the speed of an induction machine by varying the frequency and voltage of the power supplied to the machine. This is typically achieved using pulse-width modulation (PWM) techniques, which can introduce high-frequency switching noise and harmonics into the voltage signal. These high-frequency components can distort the voltage waveform and affect the accuracy of subsequent analysis making it difficult to accurately extract features such as fundamental and sequence components. Low pass filtering helps to remove these unwanted high-frequency components, ensuring that the voltage signal used for analysis is clean and representative of the actual operating conditions of the machine.

[0050] Then at step 332, the supply and the load of the machine are evaluated based on the fundamental component of the voltage and current signals. For example, a zerocrossing detector is used to calculate the supply frequency variation per cycle. And the effective value of the voltage and current signals is being evaluated for each complete cycle of the waveform by calculating the voltage and current RMS per cycle (Root Mean Square - the RMS is a measure used to quantify the magnitude of a varying Alternative Current (AC) voltage or current signal over one complete cycle of the waveform, representing the effective value of the waveform, which corresponds to the equivalent Direct Current (DC) value that would deliver the same power to the load).

[0051] The voltage and current RMS are used to assess the power quality delivered to the electrical machine. Consistent RMS values indicate stable power quality, while significant variations may suggest issues such as transient disturbances. The RMS values are also used to detect changes in the load conditions. For instance, if the load of the machine changes due to varying operational demands, this will be reflected in the RMS values. A stable load will show consistent RMS values, whereas a variable load will show fluctuations.

[0052] At step 334, transients in the data are filtered out. For example, transients are detected when the following conditions are met:- standard deviation of the supply frequency is above 5%, or- voltage standard deviation is below 5%, or- current standard deviation is below 8%.

[0053] Transients in the three-phase voltage and current signals can occur due to sudden changes in the load or other disturbances. By filtering out these transients, the preprocessing step enhances the stability and reliability of the subsequent analysis, ensuring thattransient events do not lead to false detections of winding faults.

[0054] When the prerequisites for transient in the data are met, the data are processed at step 340. Advantageously, at step 340, incomplete cycles before the first zero-crossing are eliminated. This action is taken to ensure that the data used for feature extraction and model training is consistent and begins at a standardized point in the waveform cycle. By eliminating incomplete cycles before the first zero crossing, the method avoids potential discrepancies that could arise from partial data, thereby improving the accuracy of the winding fault detection model.

[0055] The pre-processed data, now devoid of high-frequency noise, transients, and incomplete cycles, is ready for the extraction of the first set of features, which will be used to calculate the learned values of the zero-sequence admittance and the negative-sequence impedance of the electrical machine.

[0056] At step 350, the features which are part of the first set of features are extracted from the data. For example, step 350 comprises the following steps, executed for each batch of samples:- calculating at least part of the sequence components from the fundamental component of the signals;- calculating a power factor and a load percentage;- calculating real and imaginary parts of the voltage and current sequence components;- calculating initial values of the zero-sequence admittance and the negative-sequence impedance of the electrical machine;- calculating the magnitude and phase angle of the zero-sequence admittance and the negative-sequence impedance;- storing the results in a data store.

[0057] The power factor is a measure of how effectively electrical power is being transmitted to the load. It is obtained as the cosine of the phase angle between the voltage and current fundamentals.

[0058] The load percentage indicates how much of machine capacity is currently being used compared with its maximum capacity. It provides an indication of the current load condition. For example, it can be calculated as follows: voltage RMS * current RMS * power factorLoad factor = - - - - - * 100 horsepower * 745.7 where the horsepower is a known parameter indicating the maximum sustained mechanical output power the machine can deliver under specific operating conditions and 745.7 is the universal constant to convert the horse power to watt.

[0059] Optionally other features can be calculated at step 350. For example, it is possible to calculate the positive and / or the negative-sequence impedance.

[0060] When the calculations made at step 332 show that the load condition is variable, at step 360, it is checked whether a predetermined number of buckets of batches has been processed, each bucket corresponding to a different load condition. The indication of the load condition associated with a given batch can be derived from the load percentage calculated at step 350. This process, referred to as load segmentation, ensures that the features in the first set of features have been obtained with different load conditions, so that the first set of features constitute a good representation of the normal operational behavior of the electrical machine even when the load profile of the machine is variable.

[0061] When there is only one load condition or when the predetermined number of buckets has been processed, the method continues with the next step 230 which will be further described below with reference to FIG.4. Otherwise, the method returns to step 210 so that additional samples can be collected for additional buckets.

[0062] FIG.4 is a detailed view of the first and the second calculating steps of the method described with reference to FIG.2. The first calculating step 230 comprises a step 410 for checking if a predetermined number of batches N has been collected. When the number of collected batches equal the predetermined number of batches (N), at step 420, the features of the first sets of features are retrieved. And at step 430, learned values of the positive admittance, the negative admittance, and the negative-sequence impedance of the electrical machine are determined by applying a recursive least square method to adaptively refine the initial values obtained at step 350. The recursive least square method enables account for changes in the electrical machine's behavior over time, ensuring that the learned values accurately represent the machine's operating conditions.

[0063] When the number of collected batches is different from the predetermined number of batches (N), at step 440, it is checked if the number of collected batches is higher than the predetermined number of batches (N). When it is higher, the method continues at step 450. When it is lower, the method returns to step 210 to collect additional data.

[0064] At step 450 and 460 respectively, the learned values are utilized to determine noise factors for the zero-sequence current component and the negative-sequence voltage component. The noise factors are indicative of deviations from the expected behavior of the electrical machine and are used to calculate the fault severity indexes. Specifically, the learned value of the zero-sequence admittance is used to determine the noise factor of the zerosequence current component. From this noise factor, the phase-to-ground fault severity index is obtained. Similarly, the learned value of the negative-sequence impedance is used to determine the noise factor of the negative sequence voltage component, from which the interturn fault severity index is calculated.

[0065] More specifically, the noise factor of the zero-sequence current componentrepresents the deviation of the zero-sequence current component from its expected value under normal operating conditions. The noise factor of the zero-sequence current component is calculated by comparing the measured zero-sequence current with the expected zerosequence current, which is derived from the learned value of the zero-sequence admittance of the electrical machine. As a significant deviation in the zero-sequence current component can indicate a phase-to-ground fault, this noise factor is used to obtain the phase-to-ground fault severity index.

[0066] And The noise factor of the negative-sequence voltage component represents the deviation of the negative-sequence voltage component from its expected value under normal operating conditions. The negative-sequence voltage component is indicative of unbalanced voltages, which can be caused by inter-turn faults or other asymmetrical conditions in the electrical machine. It is calculated by comparing the measured negative-sequence voltage with the expected negative-sequence voltage, which is derived from the learned values of the negative-sequence impedance of the electrical machine. As a significant deviation in the negative-sequence voltage component can indicate an inter-turn fault, this noise factor is used to obtain the inter-turn fault severity index.

[0067] The phase-to-ground fault severity index and the inter-turn fault severity index are part of the second set of features that are used to train the anomaly detection model. These indexes provide a quantitative measure of the severity of potential winding fault, which is crucial for the accurate detection of such faults during the inference phase.

[0068] Other features can be used in addition to the fault severity indexes for improved accuracy. For example, it is possible to use the magnitude of the negative-sequence current component and / or the magnitude of the zero-sequence voltage component.

[0069] The training step 250 is initiated when enough second sets of features have been collected. This is checked at step 470. If the number of second sets of features is found to be insufficient, the method returns to step 440 for obtaining additional second sets of features.

[0070] The anomaly detection model, once trained, at inference time, can detect winding faults by analyzing new (inference) batches of samples (step 510 in FIG.5), extracting the second set of features corresponding to the inference batches (step 520 in FIG.5), and running the trained model with the extracted second set of features as input (step530 in FIG.5). The trained model can identify patterns indicative of winding faults, which deviate from the normal behavior known from the model. Deviations are flagged as potential winding faults (step 540 and 550 in FIG.5), enabling timely intervention and maintenance to prevent further damage to the electrical machine.

[0071] Although aspects herein have been described with reference to particularembodiments, it is to be understood that these embodiments are merely illustrative of the principles and applications of the present disclosure. It is therefore to be understood that numerous modifications can be made to the illustrative embodiments and that other arrangements can be devised without departing from the spirit and scope of the disclosure as determined based upon the claims and any equivalents thereof.

[0072] For example, the data disclosed herein may be stored in various types of data structures which may be accessed and manipulated by a programmable processor (e.g., CPU or FPGA) that is implemented using software, hardware, or combination thereof.

[0073] It should be appreciated by those skilled in the art that any block diagrams herein represent conceptual views of illustrative circuitry embodying the principles of the disclosure. Similarly, it will be appreciated that any flow charts, flow diagrams, state transition diagrams, and the like represent various processes which may be substantially implemented by circuitry.

[0074] Each described function, engine, block, step can be implemented in hardware, software, firmware, middleware, microcode, or any suitable combination thereof. If implemented in software, the functions, engines, blocks of the block diagrams and / or flowchart illustrations can be implemented by computer program instructions I software code, which may be stored or transmitted over a computer-readable medium, or loaded onto a general purpose computer, special purpose computer or other programmable processing apparatus and I or system to produce a machine, such that the computer program instructions or software code which execute on the computer or other programmable processing apparatus, create the means for implementing the functions described herein.

[0075] In the present description, block denoted as "means configured to perform ..." (a certain function) shall be understood as functional blocks comprising circuitry that is adapted for performing or configured to perform a certain function. A means being configured to perform a certain function does, hence, not imply that such means necessarily is performing said function (at a given time instant). Moreover, any entity described herein as "means", may correspond to or be implemented as "one or more modules", "one or more devices", "one or more units", etc. When provided by a processor, the functions may be provided by a single dedicated processor, by a single shared processor, or by a plurality of individual processors, some of which may be shared. Moreover, explicit use of the term "processor" or "controller" should not be construed to refer exclusively to hardware capable of executing software, and may implicitly include, without limitation, digital signal processor (DSP) hardware, network processor, application specific integrated circuit (ASIC), field programmable gate array (FPGA), read only memory (ROM) for storing software, random access memory (RAM), and non-volatile storage. Other hardware, conventional or custom, may also be included. Their function may be carried out through the operation of program logic, through dedicated logic,through the interaction of program control and dedicated logic, or even manually, the particular technique being selectable by the implementer as more specifically understood from the context.

[0076] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting. As used herein, the singular forms "a," "an," and "the," are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms "comprises," "comprising," "includes," and / or "including," when used herein, specify the presence of stated features, integers, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0077] As used herein, the term "and / or," includes any and all combinations of one or more of the associated listed items.

[0078] Benefits, other advantages, and solutions to problems have been described above with regard to specific embodiments of the disclosure. However, the benefits, advantages, solutions to problems, and any element(s) that may cause or result in such benefits, advantages, or solutions, or cause such benefits, advantages, or solutions to become more pronounced are not to be construed as a critical, required, or essential feature or element of any or all the claims.

Claims

CLAIMS1. A computer implemented method for generating a winding fault detection model for detecting a winding fault in stator windings of an electrical machine supplied with a three- phase voltage signal drawing a three-phase current signal through the stator windings, the method comprising:- a collecting step for collecting a plurality of batches of samples of the three-phase voltage signal and the three-phase current signal,- for one or more load conditions, a pre-processing step for pre-processing a first set of batches to generate, for each batch in the first set of batches, a first set of features comprising at least sequence components associated with the three-phase current and voltage signals, and a zero-sequence admittance and a negative-sequence impedance of the electrical machine obtained from said sequence components,- a first calculating step for calculating learned values of the zero-sequence admittance and negative-sequence impedance of the electrical machine, based on the first sets of features,- a second calculating step for calculating, for each batch in a second set of batches, a second set of features comprising at least an inter-turn fault severity index and a phase-to- ground fault severity index based on the learned values of the zero-sequence admittance and negative-sequence impedance of the electrical machine,- a training step for training a non-supervised anomaly detection model to detect winding faults, based on the second sets of features,- a storing step for storing the anomaly detection model for subsequent inference.

2. A method as claimed in claim 1, wherein the first set of batches comprises a predetermined number of batches.

3. A method as claimed in claim 1 or 2, wherein the pre-processing step comprises a low pass filtering of the collected batches.

4. A method as claimed in any of claims 1 to 3, wherein the pre-processing step comprises filtering out transients in the collected batches.

5. A method as claimed in any of claims 1 to 4, wherein the pre-processing step comprises removing batches collected before initial zero crossing.

6. A method as claimed in any of claims 1 to 5, wherein the first calculating step comprises applying a recursive least square to adaptively obtain the learned values of the zerosequence admittance and negative-sequence impedance of the electrical machine.

7. A method as claimed in any of claims 1 to 6, wherein the electrical machine has a given capacity and the first set of features further comprises a load percentage representative of the load condition associated with the batch being processed.

8. A method as claimed in any of claims 1 to 7, wherein the sequence components comprise a zero-sequence current component and a negative-sequence voltage component, and the second calculating step comprises:- using the learned values of the zero-sequence admittance to determine a noise factor of the zero-sequence current component and obtain the phase-to-ground fault severity index from the noise factor of the zero-sequence current component,- using the learned values of the negative-sequence impedance to determine a noise factor of the negative-sequence voltage component and obtain the inter-turn fault severity index from the noise factor of the negative-sequence voltage component.

9. A method as claimed in any of claims 1 to 8, wherein the sequence components further comprise a negative-sequence current component and a zero-sequence voltage component, and the second set of features further comprises a magnitude of the negative-sequence current component and a magnitude of the zero-sequence voltage component.

10. A method as claimed in any of claims 1 to 9, wherein the anomaly detection model is an isolation forest model.

11. A computer implemented method for detecting winding faults in an electrical machine, by running inferences of a winding fault detection model obtained through the method claimed in any of claims 1 to 10.

12. An apparatus comprising at least one processor and at least one memory storing instructions that, when executed by the at least one processor, cause the apparatus at least to perform the method as claimed in any of claims 1 to 11.

Images