Fault detection and monitoring for electric pump motors

Abstract

A fault monitor for monitoring for faults in pumps comprises a receiver to receive data relating to performance of a pump being monitored; a converter to convert the received data into a visual repres

Description

The invention relates to a fault monitor for monitoring for faults in pumps. More specifically, the invention relates to a fault monitor that uses machine learning to improve fault monitoring. In the industrial world, downtime of machinery represents a significant loss of revenue and impacts on reliability of all manner of systems. In particular, pumps used in water management systems can have critical importance both to give access to potable water for the general public, as well as regulate safety for water management. Failure of such pumps may have significant consequences. It is therefore desirable to reduce faults or failures of such pumps, and to anticipate when faults or failures are likely to happen. According to an example, there is provided a fault monitor for monitoring for faults in pumps. The fault monitor may comprise: a receiver to receive data relating to performance of a pump being monitored. The fault monitor may further comprise a converter to convert the received data into a visual representation of the data. The fault monitor may further comprise a machine learning module to process the visual representation. The processing of the visual representation of the data may comprise receiving the visual representation, analysing the visual representation using a machine learning algorithm to identify patterns in the visual representation and outputting a fault indication, if the patterns in the visual representation indicate an occurrence or prediction of a fault in the pump performance. In one example, the pump is a centrifugal pump. Centrifugal pumps are generally very reliable and often implemented in water management infrastructure. Therefore, when faults or failures do occur, these can have significant consequences and can be very hard to predict. With the fault monitor as described above however, improvements in prediction, prevention and repair of centrifugal pumps can be improved. In one example, the machine learning algorithm includes a convolutional neural network. A convolutional neural network is a form of machine learning algorithm that is particularly well suited to visual data analysis. Therefore, the combination of a convolutional neural network carrying out the processing of the data and that data being represented by a visualisation, creates a synergistic improvement in the processing and identification of faults. In one example, the visual representation of the data is a Concordia pattern. A Concordia pattern representation of the data may improve the subsequent processing of the visual data and therefore allow for more accurate identification and prediction of faults. In one example, the pump is supplied with three-phase power and the converter is further configured to convert three-phase current data into two-phase current data. Cross correlating the three-phase current data to output two-phase current data allows for a different visual representation of the data and, as shown below, a visual representation format that may improve the fault detection analysis. In one example, the converter uses Park Transformation to convert three-phase current data into two-phase current data. Park Transformation is one way to convert the current data that presents a simple but effective conversion, avoiding excessive processing. In one example, the fault monitor further comprises a three-phase current probe to detect current and output data, indicating the current detected, to the receiver. The three-phase current probe may be a connected, or Internet of Things, device, operable to utilise wireless communication to output the data to the receiver. In one example, the fault monitor further comprises a memory to store at least one of the data relating to the performance of the pump, the visual representation, the identified patterns, and the output fault indication. For completeness, the memory may store data relating to the performance of the pump or the visual representation or the identified patterns or the output fault indication or a combination of two or more thereof. The memory may be any suitable storage device. The data may be stored such that performance of the pump, the visual representation, the identified patterns, and the output fault indication, associated with the same event or timeframe, may be associated with each other. According to an example, there is provided a fault monitoring system comprising the fault monitor according to any of the examples described above. The system may further comprise multiple three-phase current probes, an application server and a data processing server on which visual data analysis may be carried out. The current probes may be located in any suitable location to detect the current. The application server and / or the data processing server may be located locally or remotely, and may be part of the same device or part of separate devices. According to an example, there is provided a fault monitoring method. The fault monitoring method may comprise receiving data relating to performance of a pump being monitored. The method may further comprise converting the received data into a visual representation of the data. The method may further comprise processing the visual representation, by a machine learning module. The processing of the visual representation of the data may comprise receiving, as an input, the visual representation, analysing the visual representation using a machine learning algorithm to identify patterns in the visual representation, and outputting a fault indication, if the patterns in the visual representation indicate an occurrence or prediction of a fault in the pump performance. According to an example, there is provided a computer-readable medium storing instructions which, when carried out on a computer, cause the computer to perform the method described above. The invention will now be described, by way of example only, with reference to the drawings, in which: Figure 1 is simplified schematic of a fault monitor; Figure 2 is a simplified flowchart of a processing method; Figure 3 is a simplified schematic of a fault monitoring system; and Figure 4 is a simplified flowchart of a processing method. The following description presents an exemplary embodiment and, together with the drawings, serves to explain principles of the invention. However, the scope of the invention is not intended to be limited to the precise details of the embodiments, since variations will be apparent to a skilled person and are deemed also to be covered by the description. Terms for components used herein should be given a broad interpretation that also encompasses equivalent functions and features. In some cases, several alternative terms (synonyms) for structural features have been provided but such terms are not intended to be exhaustive. Descriptive terms should also be given the broadest possible interpretation; e.g. the term "comprising" as used in this specification means "consisting at least in part of' such that interpreting each statement in this specification that includes the term "comprising", features other than that or those prefaced by the term may also be present. Related terms such as "comprise" and "comprises" are to be interpreted in the same manner. Directional terms such as “vertical”, “horizontal”, “up”, “down”, “upper” and “lower” are used for convenience of explanation usually with reference to the illustrations and are not intended to be ultimately limiting if an equivalent function can be achieved with an alternative dimension and / or direction. As shown in Figure 1, there is provided a fault monitor 10 for monitoring for faults in pumps. The fault monitor 10 may comprise: a receiver 100 to receive data relating to performance of a pump being monitored. The fault monitor 10 may further comprise a converter 110 to convert the received data into a visual representation of the data. The fault monitor 10 may further comprise a machine learning module 120 to process the visual representation. The processing of the visual representation of the data may comprise receiving the visual representation, analysing the visual representation using a machine learning algorithm to identify patterns in the visual representation and outputting a fault indication, if the patterns in the visual representation indicate an occurrence or prediction of a fault in the pump performance. In one example, the pump is a centrifugal pump. Centrifugal pumps represent approximately 70% of all kinds of pumps and are ubiquitous in the industrial world. Centrifugal pumps are generally very reliable and often implemented in water management infrastructure. Although modern pumps can last for many years, their sudden failure can lead to undesirable disruptions, or even catastrophic failures, e.g. when it affects a water supply in hospitals. Therefore, when faults or failures do occur, these can have significant consequences and can be very hard to predict. With the fault monitor as described above however, improvements in prediction, prevention and repair of centrifugal pumps can be improved. This has spurred research into intelligent condition monitoring techniques using signal processing and machine learning methods to detect, diagnose and predict faults by monitoring patterns in vibration, pressure or current signature sensors, as described in these examples. In one example, the machine learning algorithm includes a convolutional neural network. A convolutional neural network is a form of machine learning algorithm that is particularly well suited to visual data analysis. Therefore, the combination of a convolutional neural network carrying out the processing of the data and that data being represented by a visualisation, creates a synergistic improvement in the processing and identification of faults. In particular when the visualisation takes a form described in the following paragraphs. In one example, the visual representation of the data is a Concordia pattern. A Concordia pattern representation of the data may improve the subsequent processing of the visual data and therefore allow for more accurate identification and prediction of faults. Changes in a Concordia pattern visualisation may be analysed and, using a convolutional neural network, changes in the visualised pattern may be quickly identified and the inferred potential pump fault quickly identified or even predicted. A Concordia pattern representing a system under non-faulty conditions may take the form of a circle, and deviations from a circle may indicate faults. In some cases, different types of faults may have predictable ways in which the pattern deviates from the circular pattern. Further, small deviations may aid prediction of faults before a device, such as a pump, actually stops functioning. Therefore, downtime of the system may be mitigated or avoided. In one example, the pump is supplied with three-phase power and the converter is further configured to convert three-phase current data into two-phase current data. Cross correlating the three-phase current data to output two-phase current data allows for a different visual representation of the data and, as shown below, a visual representation format that may improve the fault detection analysis. In one example, the converter uses Park Transformation to convert three-phase current data into two-phase current data. Park Transformation is one way to convert the current data that presents a simple but effective conversion, avoiding excessive processing. In one example, the fault monitor further comprises a three-phase current probe to detect current and output data, indicating the current detected, to the receiver. The three-phase current probe may be a connected, or Internet of Things, device, operable to utilise wireless communication to output the data to the receiver. Converting the data from three-phase to two-phase allows for the data to be plotted on a two-dimensional graph. Other forms of visual representation are envisaged. In one example, the fault monitor further comprises a memory to store at least one of the data relating to the performance of the pump, the visual representation, the identified patterns, and the output fault indication. For completeness, the memory may store data relating to the performance of the pump or the visual representation or the identified patterns or the output fault indication or a combination of two or more thereof. The memory may be any suitable storage device. The data may be stored such that performance of the pump, the visual representation, the identified patterns, and the output fault indication, associated with the same event or timeframe, may be associated with each other. According to an example, there is provided a fault monitoring system comprising the fault monitor according to any of the examples described above. The system may further comprise multiple three-phase current probes, an application server and a data processing server on which visual data analysis may be carried out. The current probes may be located in any suitable location to detect the current. The application server and / or the data processing server may be located locally or remotely, and may be part of the same device or part of separate devices. In some examples, fault monitoring systems may include one or more current probes. According to an example, there is provided a fault monitoring method as shown in figure 2. The fault monitoring method may comprise receiving S101 data relating to performance of a pump being monitored. The method may further comprise converting S102 the received data into a visual representation of the data. The method may further comprise processing S103 the visual representation, by a machine learning module. The processing of the visual representation of the data may comprise receiving, as an input, the visual representation, analysing the visual representation using a machine learning algorithm to identify patterns in the visual representation, and outputting a fault indication, if the patterns in the visual representation indicate an occurrence or prediction of a fault in the pump performance. According to an example, there is provided a computer-readable medium storing instructions which, when carried out on a computer, cause the computer to perform the method described above. In some examples, the computer-readable medium may be a non-transitory computer-readable medium. In a further, more detailed example, the inventors were able to detect centrifugal pump induction motor faults by using a dataset collected from several on-site pumps deployed in real conditions in collaboration with a pump maintenance company. The detection may be done by binary classification of visual features of DQ / Concordia patterns with residual networks. Besides using real datasets, the application of image detection / analysis to systematically solve a real-life problem in engineering domain, through an artificial intelligence-augmented fault detection processing, led to improved prediction and detection of faults in the pump system. The centrifugal pump is the workhorse of many industrial and domestic applications, including water supply and wastewater treatment. While modern pumps are reliable, their unexpected failures may jeopardise safety or lead to significant financial losses. Therefore, there is a need in early fault diagnosis, detection and predictive monitoring systems. In this example, a machine learning algorithm is used to reliably detect motor faults using motor current signature analysis using data obtained from deployed pumps. Current may for example be detected by simply attaching current clamps to power supply wires of the pumps. Faults detectable with the method described herein may include, but are not limited to, impeller damage, stator winding, excessive vibrations, cavitation, and / or bearing damage faults, along with any other common pump faults. In one example, the fault detection method may be based on DQ pattern analysis. Park Transformation may be used to convert three-phase measurements into two components named D and Q. The visualisation of these two components on a two-dimensional plot may produce a circular shape for a healthy motor / pumps, whereas any distortions from a circle may indicate a fault condition. The proposed approach uses a convolutional neural network algorithm to capture those distortions to detect faults. Extracted DQ patterns may then be fed into the convolutional neural network transfer learning to analyse for binary classification of different type of faults, i.e. indicating the presence or absence of different types of faults. DQ patterns are created by measuring the three-phase line current of a pump, and then transforming that into two-phase DQ currents. These DQ currents are then plotted as patterns (with D (Ampere) current values as the abscissa and Q (Ampere) current values as the ordinate on a standard two-dimensional graph). Testing has shown that the shapes of the obtained DQ pattern plots may be affected by motor faults, such as for example, impeller or blockage faults. For example, in testing a healthy pump produced either a circle or hexagon shape, which became a fan shape when a blockage occurred. However, a different pattern may be produced for different types and severities of faults, allowing a machine learning model to learn those patterns and then predict sooner and / or with greater accuracy the occurrence of a fault. DQ therefore presents an intuitive method to detect induction motor centrifugal pump faults. This may then be combined with smaller scale convolutional neural networks (CNN) that do not require high resources, which may be deployed to consumer grade hardware. As CNNs are popular tools for computer vision problems thanks to their computational efficiency and parameter sharing capability, they may be scaled to consumer grade devices and used to perform simple image pattern recognition on DQ patterns. Park T ransformation For the data transformation Park transformation may be employed. The transform can be used by converting three-phase motor current signature to a two-phase system that has two 8 components named D and Q in order to describe three-phase IM phenomena with Park’s vector. where Xd and xq are Park’s vector components. The components of these vectors are derived from the weighted three-phases and their subtractions from each other. Park transform is essentially conversion of three-phase motor current data to two components called d and q. DQ / Concordia patterns are the plots of d and q against each other. They have various (disturbed) shapes that may give us an indication of whether the sampled motor has a failure or not, based on the pattern / signature. The identification of DQ Patterns DQ has an ability to detect several motor faults. Given that it can be applied to three-phase induction motors that any system (e.g pumps, wind turbines, etc) can have, it has the potential to be used in a wide range of areas. Each fault can have a distinct DQ pattern. The shape of a healthy (not faulty) motors’ DQ data may be perfectly circular, in some examples. Aside from the ability to detect the shape of healthy motors, DQ pattern is reported to demonstrate the presence of impeller, blockage, BRB and short-circuit turn failures. For BRBs and / or short circuit between turns faults, for every turn, the tracked motor current signature analysis (MCSA) plot produces a slightly altered version of every cycle and cause an elliptical shape of Park transform to form. The DQ plot may not overlap the area it has already passed due to fault-induced phases. For faults including impeller faults and pipe blockages, the faults may gradually change the shape of the DQ plots. In one study, faults include (hand-valve) half pipe block, full pipe block, (artificially) damaged impeller surface, and 120 degree angled damaged impeller surface. According to the findings in resulting line plots, a healthy pump gives a line plot that is relatively hexagonal. However, this status changes when the aforementioned faults are introduced, respectively. It is observed that, for every increasingly severe fault iteration, the hexagonal shape may be lost and a fan shape constructed. As discussed earlier, this newly formed fan shape may not be perfect and like BRB faults, for every period, the plot may have slightly different radius and period length. This causes the fan shape to be disrupted and not uniform in parallel to increasing fault severity. In some examples, faults may be detected in association with a device, and the output may be a simple fault / no fault output. In other examples, the fault / no fault-type output may be enhanced with an indication of the type of fault for example. System Architecture In some examples, a system may comprise the fault monitor described above, along with one or more loT devices deployed at one or more sites to detect current, an application server and a database, as shown in figure 3. Each loT device may be located near the monitored device. In the example of a pump, an loT device may be located at or near the pump control panel, to measure current on three phases using current clamps, and then transmits the raw data to the application server over, for example, GSM / GPRS link or any other wireless or wired communication method. In some examples, the measurements may be performed at most every 60 mins whenever the device, in this case a pump, is active. Sampling rates may for example be 1500 Hz, 3000 Hz, or 4500 Hz. Finally, the application server may run a dashboard application operable to provide an intuitive user interface for users of the system, which may include sensor data visualization, alarm generation. The application server may also store / retrieve raw sensor data from a database, such as an online database. In one example, Google Big Query Database was used. Feature Extraction and Classification The fault detection process will now be described with reference to figure 4. Initially, the three-phase data may be retrieved from the loT device, the pump itself or a database with the dimensions A-4500-n, meaning amplitude, collected &concatenated three phase data points and total number of collected signatures, respectively. Given that the collected data in this example are separate but not aligned due to the capacity of the device, the function ip is used to align the three-phases. The original data shape collected is 1500 ■ 3 where 1500 is the total amount of data points sampled per phase and 3 is the total phase number. The aligning process (tp) is done via the calculation of cross correlation of each signal with respect to each other. Then, additionally, there may be an optimization step to find the minimum amount required to shift and cut the signal from their ends to align. In the example set out, all aligning sequences were calculated to minimize the data point loss during the aligning process (e.g. aligning all phases w.r.t first, second or third phase). The obtained end signals are aligned and the dimensions are A (x <1500) (3 n) where x is the total amount of data points per phase. During the aligning process, some data points may be cut to align the phases, so the real number of data point per phase is less than or equal to 1500, in this example. Given that the phases are now separated, the n is tripled. The number of data points being 1500 is chosen because of the device collection and submission capacity per reading. The number may be increased if device capacity permits, which may improve accuracy further. Next, aligned three-phase data may be taken and Park transformation (a.k.a the function ¢) may be applied to obtain DQ data that has (x <1500) - (2 ■ ri) dimensions where we have d and q component for every n. After the transform, the data may be plotted as a grid-less RGB image that will be pre-processed for the machine learning model. After pre-processing, the plotted DQ image is resized and center-cropped to appropriate model input dimensions (224-224-3), normalized, and then fed to model (in some examples the model may include Res Net-34). The final step is the machine learning (ML) detection. The fed processed DQ images may be passed through a consumer-grade, hardware friendly and state of the art residual network. The obtained model results may be binary called “faulty” and “nonfaulty”, for example. However, the final layer can be extended to hold many fault classes. In some examples, plotting the DQs and feeding them as images may be advantageous as the plot will not have axis values. In other words, the pumps with higher voltages will not cause any overfitting with their higher DQ values. This may help to eliminate the bias based on the largeness and purely focusing on the image features. Data Processing and Augmentation In some examples, the labels for classification may be “faulty” or “1” and “non-faulty” or ”0”. Other labels are however envisaged. In testing, a training to validation ratio was chosen of ~ 3:1 and faulty to non-faulty ratio of ~ 1:1. The validation and training datasets were prepared with 1500 and 3000 Hz sampling frequency. The testing dataset comprises two pumps which are sampled with three different sampling rate (1500, 3000 and 4500 Hz) to provide additional unseen sampling distribution class to the model’s performance. For each sampling rate, at least 50 signatures are collected to have a balanced dataset. The inventors took this approach to create varied distributions of unseen datasets to train robust models. Besides, with the varied “unseen” sampling frequency (4500 Hz), the model’s robustness to unseen conditions was also examined. There are other reasons which also encouraged us to choose a multiple sampling rate experimentation. Testing different sampling rates allowed us not only to conduct device reliability testing, but also to collect data of different distributions from varying conditions. Collecting data of one sampling rate and then downscaling it can have carried information from higher frequencies that may not be real under real life conditions where less capable devices may not even register such information. The inventors present the specifications of the training (Table 1), validation (Table 2) and testing (Table 3) dataset to show the variation of the dataset gathered. A pumps’ speed and voltage are provided in the columns to show the variety of data distributions used in training the model, as this may influence the fault development, like creating variety of features for the model to learn. Table 1: Training pump specifications ID Fault Pump Type Voltage (V) Speed (rpm) Tr Fl Stator Winding pump / motor type for Youngvl 415 1385 Tr F2 Stator Winding pump / motor type for Youngv2 415 1385 Tr F3 Excess vibr., Bearing dmg. pump / motor type for BenchM5 400 2815 Tr NF1 Non-Faulty pump / motor type for Hestonvl 415 1450 Tr NF2 Non-Faulty pump / motor type for Hestonv2 415 1450 Tr NF3 Non-Faulty pump / motor type for NAWvl 400 1450 Table 2: Validation pump specification ID Fault Type Motor Type Voltage (V) Speed (rpm) Va Fl Impeller dmg., excess vibr. pump / motor type for BenchMl 400 1420 VaNFl Non-Faulty pump / motor type for Romseyv2 400 1450 Va NF2 Non-Faulty pump / motor type for NAWv2 400 1450 Va NF3 Non-Faulty pump / motor type for BenchM4 400 2815 Va NF4 Non-Faulty pump / motor type for TeaHvl 415 1445 Va NF5 Non-Faulty pump / motor type for TeaHv2 415 1445 5 Table 3: Testing pump specification ID Fault Type Pump Type Voltage (V) Speed (rpm) Te F Cavitation Booster pump 425 2800 Te NF Non-Faulty Booster pump 415 2860 In some examples, data augmentation techniques may be used to increase the dataset size. In the example described, such techniques were used to increase the dataset six-fold. The total list of these augmentation techniques include for the training dataset are vertical 10 mirroring, horizontal mirroring, 90 degrees clockwise rotation, 180 degrees clockwise rotation, 270 degrees clockwise rotation. During the augmentation for example the original image may be taken, and an augmentation function may be applied, and then saved. This may then be repeated until an original image and its five other augmented version are present in the dataset. Given that the data used for validation dataset is real and sampled from different pump systems, the inventors present the precision of the testing, and also validation, datasets, recall and accuracy table in Tables 5 and 6. In this part, the results of the model’s performance compared against the unseen validation and testing dataset. Table 7 shows that the ResNet-34 model performed the best at higher frequency samplings (e.g. 4500 Hz) both in faulty and non-faulty pumps, despite not being trained on that frequency. For further demonstration of the model’s performance, we used the metrics of precision, recall, F1 -score to measure the success which can be seen in Table 5 and 6. The sampling frequency for the datasets for these experiments is set out above. These metrics are defined as: Precision = TP / (TP + FP), Recall = TP / (TP + FN), Accuracy = (TP + TN) / (TP + TN + FP + FN), F1-score = 2TP / (2TP + FP + FN). These metrics were chosen, because they are, along with accuracy metric, efficient to enough to demonstrate the model’s robustness against over-fitting. With the bigger than 80% F1 -score, the models are capable enough to produce reliable decisions. Table 4: rises for each sulMlataset where trammg dataset is aijgmeiUai. Dataset same Faulty Data Nwa-Omlty Data Total Train dig 034S Mhd 12036 Validation 370 MS ws Testing 107 107 214 Table 5: Validation preciskm, recall and accuracy table Table &Testing dataset precision, recall and .^curacy table Table 7: Testing dataset size and results for each. sampling rate and pump for ResNet-34 The application of convolutional neural network models on non-conventional signal data types is an unseen and unique method. It is shown that, despite its unorthodoxy, the model does get impressive results on the validation dataset and the testing dataset. Furthermore, 5 its relative performance on varying sampling rates also shows that it is a robust technique too. The description herein refers to embodiments with particular combinations of features, however, it is envisaged that further combinations and cross-combinations of compatible 10 steps or features between embodiments will be possible. Indeed, isolated features may function independently as an invention from other features and not necessarily require implementation as a complete combination. Features described in any of the following claims may be combined with any features of any other claims.

Claims

1. A fault monitor for monitoring for faults in pumps, the fault monitor comprising:a receiver to receive data relating to performance of a pump being monitored;a converter to convert the received data into a visual representation of the data; anda machine learning module to process the visual representation, whereinthe processing of the visual representation of the data includes:receiving, as an input, the visual representation;analysing the visual representation using a machine learning algorithm to identify patterns in the visual representation; andoutputting a fault indication, if the patterns in the visual representation indicate an occurrence or prediction of a fault in the pump performance.

2. The fault monitor of claim 1, wherein the pump is a centrifugal pump.

3. The fault monitor of any preceding claim, wherein the machine learning algorithmincludes a convolutional neural network.

4. The fault monitor of any preceding claim, wherein the visual representation of the data is a Concordia pattern.

5. The fault monitor of any preceding claim, wherein the pump is supplied with three-phase power; andthe converter is further configured to convert three-phase current data into two-phase current data.

6. The fault monitor of claim 5, wherein the converter uses Park Transformation to convert three-phase current data into two-phase current data.

7. The fault monitor of claim 5 or claim 6, further comprising:a three-phase current probe to detect current and output data, indicating the current detected, to the receiver.

8. The fault monitor of any preceding claim, further comprising:a memory to store at least one of the data relating to the performance of the pump, the visual representation, the identified patterns, and the output fault indication.

9. A fault monitoring system comprising the fault monitor of any preceding claim.

10. A fault monitoring method comprising:receiving data relating to performance of a pump being monitored;converting the received data into a visual representation of the data;processing the visual representation, by a machine learning module; whereinthe processing of the visual representation of the data includes:receiving, as an input, the visual representation;analysing the visual representation using a machine learning algorithm to identify patterns in the visual representation; andoutputting a fault indication, if the patterns in the visual representation indicate an occurrence or prediction of a fault in the pump performance.

11. A computer-readable medium storing instructions which, when carried out on a computer, cause the computer to perform the method of claim 10.AMENDMENTS TO THE CLAIMS HAVE BEEN FILED AS FOLLOWS:-02 03 26CLAIMS:

1. A fault monitor for monitoring for faults in pumps, the fault monitor comprising:a three-phase current probe to detect current and output data relating to performance 5 of a pump being monitored;a receiver to receive the data relating to performance of a pump being monitored;a converter to convert the received data into a visual representation of the data; anda machine learning module to process the visual representation, whereinthe processing of the visual representation of the data includes:10 receiving, as an input, the visual representation;analysing the visual representation using a machine learning algorithm to identify patterns in the visual representation; andoutputting a fault indication, if the patterns in the visual representation indicate an occurrence or prediction of a fault in the pump performance; wherein15 the machine learning algorithm includes a convolutional neural network.

2. The fault monitor of claim 1, wherein the pump is a centrifugal pump.

3. The fault monitor of any preceding claim, wherein the visual representation of the20 data is a Concordia pattern.

4. The fault monitor of any preceding claim, wherein the pump is supplied with three-phase power; andthe converter is further configured to convert three-phase current data into two-phase 25 current data.

5. The fault monitor of claim 4, wherein the converter uses Park Transformation to convert three-phase current data into two-phase current data.02 03 266. The fault monitor of any preceding claim, further comprising:a memory to store at least one of the data relating to the performance of the pump, the visual representation, the identified patterns, and the output fault indication.5 7. A fault monitoring system comprising the fault monitor of any preceding claim.

8. A fault monitoring method comprising:collecting data relating to performance of a pump being monitored using a three-phase current probe;10 receiving the data relating to performance of a pump being monitored;converting the received data into a visual representation of the data;processing the visual representation, by a machine learning module; whereinthe processing of the visual representation of the data includes:receiving, as an input, the visual representation;15 analysing the visual representation using a machine learning algorithm toidentify patterns in the visual representation; andoutputting a fault indication, if the patterns in the visual representation indicate an occurrence or prediction of a fault in the pump performance; whereinthe machine learning algorithm includes a convolutional neural network.

209. A computer-readable medium storing instructions which, when carried out on a computer, cause the computer to perform the method of claim 8.

Images