Monitoring of transformer oil by neural-network-based analysis of NMR spectra

Abstract

There is disclosed herein a method to monitor transformer insulating liquid. The method comprises obtaining, from an NMR sensor arranged in a transformer insulating liquid tank, an NMR spectrum for the transformer insulating liquid. The NMR spectrum is processed using a trained an artificial neural network (ANN) to thereby obtain information about the transformer insulating liquid. Based on the obtained information about the transformer insulating liquid, a control assessment is generated for an aspect of the transformer.

Description

NEURAL-NETWORK-BASED ANALYSIS OF TRANSFORMER OILTechnical Field

[0001] The present disclosure relates to the analysis of transformer oil. More particularly, the present disclosure relates to a neural-network-based analysis of transformer oil, and includes a method for training a neural network to analyze transformer oil, a neural network trained according to such a method, and the use of such a trained neural network to analyze transformer oil. The present disclosure further relates to a computing device configured to perform such an analysis of transformer oil, and a system comprising sensors that feed information sensed from transformer oil to such a computing device.Background

[0002] During a transformer's operation, transformer oils experience both electrical and mechanical stresses. Additionally, contamination occurs due to chemical interactions with windings and solid insulation, often catalyzed by high operating temperatures. Over time, the original chemical properties of the oil gradually change, rendering it less effective for its intended purpose. As a result, large transformers and electrical equipment undergo periodic testing to assess their electrical and chemical properties, ensuring their suitability for continued use. In some cases, oil condition can be enhanced through filtration and treatment.

[0003] Some transformer oil tests can be carried out in the field, using portable test apparatus, but other tests, such as a test for dissolved gas, normally require a sample to be sent to a laboratory. Traditional analysis usually takes many weeks, starting with manual sampling, which needs to be shipped to the analysis lab, where it will be tested according to each of a number of lab schedules.Summary of the Present Disclosure

[0004] It is realized as a part of the present disclosure that transformer oil can be tested ‘online’ and without human intervention, at far smaller time intervals. In particular, it is realized as a part of the present disclosure that all pertinent information for transformer insulating liquid, and making operational decisions on the basis of such analysis, can be derived from nuclear magnetic resonance (NMR) spectra. Transformer insulating liquid is also referred to as ‘transformer oil’ or simply ‘oil’ herein, although it will be appreciated that the transformer insulating liquid may not comprise oil in some cases.

[0005] NMR spectra are already used in the field of transformer oil analysis, but currently performed ‘offline’, i.e. , taking samples to a laboratory. Furthermore, NMR techniques are currently only used to obtain a subset of oil parameters. According to aspects of the present disclosure, an NMR sensor can be installed into the transformer oil tank to provide regular spectra from the transformer oil, with this regular information being used to provide all the information that would otherwise have come from laborious and protracted testing methods.

[0006] Therefore, aspects of the present disclosure provide a significantly improved technique for analyzing transformer oil, and thus contribute to the improved maintenance of transformers. With improved maintenance, the transformers may operate more efficiently. As used herein, ‘transformers’ may be power and distribution transformers for power grids, with medium-voltage or high-voltage ratings.

[0007] NMR spectroscopy is extremely useful for identification and analysis of organic compounds. The principle on which this form of spectroscopy is based is simple. The nuclei of many kinds of atoms act like tiny magnets and tend to become aligned in a magnetic field. NMR spectroscopy measures the energy required to change the alignment of magnetic nuclei in a magnetic field.

[0008] The most common procedures focus on identifying the NMR spectra of hydrogen (1H) and carbon (13C). Depending on the molecular structure, the NMR response for these two “protons” will be different, generating a kind of ‘fingerprint’ for each molecule.

[0009] When a sample of the transformer insulating liquid, potentially containing several different contaminants, is analyzed, a resultant NMR spectrum containing dozens of such footprints overlaid will be obtained. A challenge is to correctly identify and quantify the substances present in that mixture.

[0010] Advanced algorithms support such interpretation of the NMR spectra, but, in almost all cases, the decision is highly dependent on the operator’s judgment. The conventional techniques for NMR spectra interpretation use the identified hydroxyl and carbonyl groups to populate a list of combinations of molecules in different proportions that will fit the full spectrum of substances, minimizing the total error. Such a procedure relies on the operator’s expert assessment to select the most appropriate combination of substances, which may not be the best mathematical answer.

[0011] Furthermore, small fluctuations in the readings may lead to big changes in the result. The best “fitting” may include molecules not present in the transformer or lead to variations on the content of the compounds not reflecting any realistic physical process.

[0012] Aspects of the present disclosure are therefore directed at providing a machine-learning-based approach to analyzing NMR spectra of transformer insulating liquid / oil to thereby obtain online information about the oil. Put another way, aspects of the present disclosure relate to the provision of an artificial neural network (ANN), which may also be referred to as a ‘model’.

[0013] A challenge of building an online sensor in this way is to incorporate ‘expert judgment’ in the model. To remove human intervention, it may not be viable to display a number of options to an operator. Indeed, the transformer-specific variations of the oil further complicate the ability of anoperator to make fully informed judgements, and hence the operator is often not an expert in the particular system under analysis. Thus, according to the present approach, no specialized worker should be required to perform an interpretation of the result.

[0014] On top of this, the expected results may not (only) be content of each compound, but (also) the final effect upon the transformer typical assessment parameters. A result may be presented to human operators in the form of a list of transformer assessment parameters such as insulating liquid breakdown voltage, dielectric dissipation factor, neutralization index, interfacial tension, water content, combustible and non-combustible dissolved gases content, oxidation of the insulating liquid, etc. Additionally or alternatively, the constituent compounds may be directly provided to a control process, which may have ANN-based learning to implement controls directly based on various compound profiles derived from an NMR spectrum.

[0015] Hence, according to a particular aspect of the present disclosure, there is provided a method to monitor transformer insulating liquid.

[0016] The method comprises obtaining, from an NMR sensor arranged in a transformer insulating liquid tank, an NMR spectrum for the transformer insulating liquid. The method may further comprise generating, using the NMR sensor, the NMR spectrum from a sample of the transformer insulating liquid. That is, the same sensor device may be configured with onboard compute to both generate the NMR spectrum and process the same. In other examples, the NMR sensor may be in communication with a separate device which receives the NMR spectrum from the NMR sensor.

[0017] The method further comprises processing the NMR spectrum using a trained an artificial neural network (ANN) to thereby obtain information about the transformer insulating liquid. Preferably, the ANN is a physics- informed neural network (PINN) having physics-based constraints built in, which substantially reduces the time and samples required to train the ANN. PINNs aim to learn the underlying physics of a system from data. Real-world physical systems can be highly complex, involving intricate interactions andnonlinear behavior. Capturing all relevant physical laws and boundary conditions accurately in a neural network architecture can be challenging. The synergies between the purely numerical algorithm (the neural network) and a physical model bring substantial advantages.

[0018] Another relevant advantage of the PINNs is the capacity to ensure numerical stability during training and avoid divergence or instability. The tracking history of identified compounds and the limited materials and process present in a transformer are example enforced physical constraints for the model.

[0019] Based on the obtained information about the transformer insulating liquid, a control assessment is generated for an aspect of the transformer. This control assessment may be any suitable assessment, such as the generation of a monitored parameter (e.g., moisture) for comparison with a set threshold, or the generation of an assessment as to whether intervention (such as maintenance) is required for the transformer, for example a replacement of the transformer insulating liquid.

[0020] By using an NMR sensor with an ANN to monitor the transformer insulating liquid, all or most of the parameters required can be generated with a single pipeline. Hence, the need to manually extract and process samples using a laboratory is obviated.

[0021] Moreover, the transformer insulating liquid can be monitored more frequently for changes. Preferably, at least the step of obtaining the NMR spectrum is performed at regular or dynamic intervals. Thus, greater information about the transformer insulating liquid can be obtained, which may be further utilized to re-train the ANN. Accordingly, the down-time for maintenance or repair of the transformer can be substantially reduced. Furthermore, the construction of the insulating liquid tank may be simplified, if regular access to the transformer insulating liquid is no longer required.

[0022] Processing the NMR spectrum using the trained ANN may comprise providing, as input to the ANN, the NMR spectrum for the transformer insulating liquid, and receiving, as output from the ANN, anestimated proportion of at least one constituent compound in the transformer insulating liquid and / or at least one transformer parameter for the transformer insulating liquid.

[0023] The at least one transformer parameter for the transformer insulating liquid may be derived from the estimated proportion of at least one constituent compound in the transformer insulating liquid and / or output with the estimated proportion of at least one constituent compound in the transformer insulating liquid. The transformer parameter(s) may include one or more of water content, presence of acid(s), dielectric dissipation factor (DDF) or power factor (PF), interfacial tension (IFT), oxidation inhibitor content, and breakdown voltage (BDV). This list may be non-exhaustive.

[0024] That is, the method according to the present disclosure may measure the amount of water present in the fluid. While this may seem to be a direct result from an NMR procedure, the variation of solubility with temperature and the possibility of having free water mixed with the liquid may require developing some additional correlations. Moreover, the traditional laboratory test used to evaluate the acidity of the insulation liquid was a titration using KOH (potassium hydroxide). The typical parameter for transformer assessment measures the amount of KOH (potassium hydroxide) required to neutralize the oil, in milligrams of KOH per gram of liquid, the “neutralization index” or “acid value”. Each identified acid and its content will lead to a different contribution to the “neutralization index”.

[0025] The DDF / PF values are essentially measurements of the content of polar contaminants in liquid. They may correlate the heat generated by the fluid when an electrical field is applied to the intensity of the applied field. While holding some similarities with the DDF, the IFT is an early indicator of the presence of polar contaminants.

[0026] The oxidation inhibitor content measures the content of some specific additives used to increase the oxidation stability of the insulating liquids. Furthermore, the laboratory test for BDV typically uses a pair of standard electrodes positioned 2mm or 2.5mm from each other to check therequired voltage for generating an electrical discharge between them. It is influenced by most contaminants, especially water, and by the presence of dissolved gases and particles.

[0027] The realization that some or all of these parameters can be derived from the NMR spectrum thus obviates the need for diverse testing procedures. The method according to the present disclosure therefore investigates the insulating liquid more deeply and accurately than the typical laboratory tests used for the transformers assessment that, at the same time, also delivers the results according to the traditional reference parameters for operators or established control systems, which may control based on these parameters instead of the presence / absence of specific constituent components.

[0028] According to a further aspect of the present disclosure, there of provided a system for performing the method according to the previously described aspect. The system comprises one or more NMR sensors arranged in an insulating liquid tank of a transformer, and a computing device. The computing device is configured to receive output NMR spectra from the one or more sensors and process the NMR spectra with the ANN to thereby obtain an estimate of a proportion of a constituent compound in the transformer insulating liquid and / or an estimate of a transformer parameter for the transformer insulating liquid.

[0029] The NMR sensor(s) and the computing device may be integrally formed in the same embedded system, or the computing device may be arranged remotely, e.g., implemented in a server environment such as a cloud computing arrangement.

[0030] According to yet a further aspect of the present disclosure, there is provided a method for training the ANN. The method comprises providing a plurality of training samples to the ANN, wherein each training sample comprises an NMR spectrum of transformer insulating liquid corresponding to convoluted spectra of constituent compounds in the transformer insulating liquid. Preferably, these NMR spectra are obtained using a similar or thesame NMR sensor as that which will be deployed in the system.

[0031] The training method further comprises processing the convoluted NMR spectrum using the ANN to estimate a proportion of at least one constituent compound in the transformer insulating liquid. The ANN is then trained by comparing the estimated proportion of the at least one constituent compound to a known proportion of said at least one constituent compound.

[0032] Additionally or alternatively, the method comprises processing the convoluted NMR spectrum using the ANN to estimate a transformer parameter for the transformer insulating liquid, and training the ANN by comparing the estimated transformer parameter to a known transformer parameter for the transformer insulating liquid. In such an example, the estimated proportion of the at least one constituent compound may be an intermediate derivation of the ANN before the parameter is estimated. Alternatively, the parameters may be output alongside the estimated proportion(s) of constituent compounds.

[0033] During training, the ANN may learn to associate specific spectral patterns with both compound proportions and transformer parameters. This dual learning process helps the ANN generalize better to new, unseen spectra, improving its overall performance. Knowing the proportions of individual compounds helps the ANN make more accurate predictions about the transformer oil parameters. This step ensures that the final output (e.g., moisture level, breakdown voltage) is based on a detailed understanding of the oil’s composition.

[0034] Processing the convoluted NMR spectrum using the ANN to estimate a proportion of at least one constituent compound in the transformer insulating liquid may comprise providing the NMR spectrum as a dataset to a plurality of input nodes in an input layer of the ANN, wherein each input node is provided with a respective subset of the dataset. The dataset may be a list of numbers (e.g., as comma separated values (CSV) or another format) representing each bin in the frequency range, e.g., corresponding to 1 Hz, 10Hz, or a different bin width. As an alternative, the dataset may be preprocessed to thereby reduce its size, by representing each spectrum as a list of tuples with (peak height, peak location), in descending order of peak height or in order of peak location. Such preprocessing may allow for a threshold peak height to be defined, such that characteristic aspects of the spectrum can be captured in a minimal amount of data.

[0035] The processing may further involve providing each output of the plurality of input nodes to one or more intermediate layers of the ANN, wherein each intermediate layer comprises a plurality of intermediate nodes, wherein each intermediate node of each intermediate layer has one or more associated weight functions.

[0036] The intermediate layers may include a convolutional layer, and may fully connected layer(s), such that the ANN is a convolutional neural network, such as a deep learning network. The ANN may instead be a support vector machine, or take some other suitable architecture, depending on the accuracy required for the implementation and the available data quality of the NMR spectrum inputs.

[0037] The processing may then comprise providing each output of the plurality of intermediate nodes of a last intermediate later to a plurality of output nodes in an output layer of the ANN, wherein each output node is associated with a corresponding constituent compound of the transformer insulating liquid. The output layer may comprise a SoftMax output layer for output nodes associated with constituent compounds of the transformer insulating liquid.

[0038] In some examples, the training of the ANN may enable probabilistic outputs for the presence of particular compounds to be read as proportions, such that a 0.01 % probability is equated to 100 parts per million (ppm) etc. In such cases, if the ANN is trained to recognize all possible constituents, then the outputs can be normalized to one (1 ), i.e., unity, so as to further simplify the training of the ANN.

[0039] In other examples, a first ANN may determine the probability ofthe presence of a particular compound, and a second ANN may determine the amount (e.g., in ppm) of the compound present, based on the determined probability.

[0040] Processing the convoluted NMR spectrum using the ANN to estimate a transformer parameter for the transformer insulating liquid may similarly comprise steps of: providing the NMR spectrum as a dataset to a plurality of input nodes in an input layer of the ANN, wherein each input node is provided with a respective subset of the dataset, providing each output of the plurality of input nodes to one or more intermediate layers of the ANN, wherein each intermediate layer comprises a plurality of intermediate nodes, wherein each intermediate node of each intermediate layer has one or more associated weight functions, and providing each output of the plurality of intermediate nodes of a last intermediate later to a plurality of output nodes in an output layer of the ANN, wherein each output node is associated with a corresponding transformer parameter for the transformer insulating liquid.

[0041] At least in a first round of training for the ANN, each training sample may be derived from a controlled sample of transformer insulating liquid. Thus, the known proportion of said at least one constituent compound can correspond to a predetermined proportion of the constituent compound in the controlled sample of transformer insulating liquid.

[0042] Based on a list of expected contaminants, a few hundreds of solutions may be prepared with known contents of each compound (for example, new insulating liquid with 100 ppm of Hydrogen dissolved). The NMR sensor may be used to analyze the solution multiple times, generating an expected spectrum with some standard deviations (error functions). The prepared solution may then be subjected to normal laboratory tests, measuring the transformer parameters. The resultant matrix can then be used for at least the first training of the neural network, allowing the prediction of the NMR spectra and the transformer parameters for any mix of the expectedcontaminants (included in a library).

[0043] The training may further comprise providing one or more NMR spectra for individual compounds with the plurality of training samples, wherein the one or more NMR spectra for individual compounds are chosen from a library of compounds expected to be constituent compounds in the transformer insulating liquid. For example, the ‘pure’ spectra of H2, CH4, H2O, etc. may be provided as further inputs, thereby enabling the ANN to compare the representative peaks of these spectra with the convoluted spectrum.

[0044] Providing the individual spectra of components in this away preserves the configuration of the ANN - in other words, the inputs are maintained the same (e.g., complex spectra from the NMR, and individual spectra of components) - and the dynamic variation can be in the spectrum coming from the NMR sensor.

[0045] According to further examples, the training method comprises providing a plurality of further training samples to the ANN, wherein each further training sample comprises an NMR spectrum of transformer insulating liquid corresponding to convoluted spectra of constituent compounds in the transformer insulating liquid.

[0046] Here, as above, the convoluted NMR spectrum is processed using the ANN to estimate a transformer parameter for the transformer insulating liquid, and the ANN is trained by comparing the estimated transformer parameter to a known transformer parameter for the transformer insulating liquid.

[0047] However, in this case, which may represent a subsequent training of the ANN, each training sample is derived from a sample of transformer insulating liquid from a(n actual operational) transformer. Hence, the known transformer parameter may correspond to a measured transformer parameter for the sample of transformer insulating liquid from the transformer, e.g., measured using laboratory tests.

[0048] Comparing the estimated proportion of the at least oneconstituent compound to a known proportion of said at least one constituent compound, as mentioned above, may comprise, for example, applying a loss function proportional to the squared difference between an output estimated proportion of the at least one constituent compound from the ANN and the known proportion of said at least one constituent compound.

[0049] Similarly, comparing the estimated transformer parameter to a known transformer parameter for the transformer insulating liquid may comprise applying a loss function proportional to the squared difference between an output estimated transformer parameter from the ANN and the known transformer parameter.

[0050] The training method may further comprise providing, in association with a (each) training sample, a transformer operating parameter, a transformer insulating liquid temperature, and / or a proportion of paper to transformer insulating liquid, associated with the (each) training sample. Thus, the ANN may learn to associate static, dynamic and operatic characteristics of the transformer system with changes in the transformer insulating liquid, which may serve as further constraints.

[0051] That is, a further training cycle may take place during the operational application of the ANN when the system is installed in a transformer system. As mentioned above, such a training cycle may include, as inputs, the NMR spectra from the analyzed oil sample, a library including the “NMR footprint” of the investigated substances (frequencies and amplitudes), transformer ratings (MVA, kVs, application, etc.), temperatures and loading, and / or estimation of the paper-to-oil proportion.

[0052] The results are the content of each substance and the associated transformer parameters. Before accepting the result, the system may compare each substance’s content's time-domain variation and transformer parameters. This verification considers the expected physical behavior when a model is available.

[0053] As mentioned above, the ANN may be a physics informed neural network (PINN), having at least one constraint based on a physicalproperty of the transformer insulating liquid. Thus, the PINN may comprise a time-variance constraint corresponding to an expected time variation in a proportion of a constituent compound and / or a transformer parameter.

[0054] As another example, the water content in the fluid may change with temperature due to the absorption of water by the insulating paper. Curves such as the “Piper Chart” can be associated with transient models for water diffusion to estimate balance of water between paper and oil. The fluctuation of the water content in the fluid may not be caused by the ingress or consumption of water, but rather by the change in the balance of water between the paper and the fluid. This may serve as a further constraint for the PINN to account for.

[0055] Other examples of “physically informed” aspects of the network are a typical degradation processes of a transformer, based on the loading and temperature (e.g., a modified Arrhenius model), and / or a correlation between the content of acids and the neutralization index. It may also include a “bias factor” for accounting for random contamination or unexpected substances.

[0056] An accepted result may thus become a new training point, and the ANN is preferably automatically retrained based on the latest data. This continuous evolution of the model (unsupervised training) will tailor the ANN to the particularities of each transformer’s typical behavior. The result is a more reliable prediction system, indicating abnormal behaviors more effectively.

[0057] Further aspects of the present disclosure provide an ANN trained according to the above-described training method, and a computing device having such an ANN stored thereon.

[0058] It will be appreciated that the advantages provided by the monitoring method and system, some of which are described above, and others of which will be appreciated from the present disclosure as a whole, will also be conferred by training an ANN for the limited purpose of performing the monitoring method, and hence by the ANN as such.Brief Description of the Drawings

[0059] One or more embodiments will be described, by way of example only, and with reference to the following figures, in which:

[0060] Figure 1 schematically shows a prior art approach to monitoring transformer insulating liquid;

[0061] Figure 2 schematically shows an approach to monitoring transformer insulating liquid according to aspects of the present disclosure;

[0062] Figure 3 schematically illustrates deconvolution of a convoluted NMR spectrum;

[0063] Figure 4 illustrates a method to monitor transformer insulating liquid according to an aspect of the present disclosure;

[0064] Figure 5 schematically shows an approach to training an artificial neural network for monitoring transformer insulating liquid according to aspects of the present disclosure; and

[0065] Figure 6 illustrates a method for training an ANN for monitoring transformer insulating liquid according to an aspect of the present disclosure.Detailed Description

[0066] The present disclosure is described in the following by way of a number of illustrative examples. It will be appreciated that these examples are provided for illustration and explanation only and are not intended to be limiting on the scope of the disclosure.

[0067] Figure 1A schematically shows a system 100 comprising a power and distribution transformer 102, hereinafter referred to simply as ‘transformer 102’ immersed in oil 104, where the oil 104 is an example of a transformer insulating liquid. The oil 104 is held in a tank 106.

[0068] According to a prior art approach to monitoring the oil 104, an operator performs laboratory tests, as generally referred to with the numeral 108. The tests 108 may include, for example, a test for acids / acidity. Thetraditional laboratory test used to evaluate the acidity of the oil 104 is a titration using KOH (potassium hydroxide). The typical parameter for transformer assessment measures the amount of KOH (potassium hydroxide) required to neutralize the oil, in milligrams of KOH per gram of liquid, the “neutralization index” or “acid value”.

[0069] Another test may be for the breakdown voltage (BDV), using a pair of standard electrodes positioned 2mm or 2.5mm from each other to check the required voltage for generating an electrical discharge between them. BDV is influenced by most contaminants, especially water, and by the presence of dissolved gases and particles.

[0070] An operator may be required to extract samples for sending to a laboratory. Hence, the time taken to obtain information about the oil 104 may be in the order of weeks. The extraction of a sample of the oil 104 from the tank 106 may further risk contamination from sampling tools. Moreover, by the time a result is returned, the conditions of the system 100 may have changed. Hence, any decisions based on the result may be outdated.

[0071] To provide more accurate and more frequent monitoring of transformer insulating liquid, the present approach (as schematically illustrated in figure 2) proposes providing an NMR sensor 210 in the tank 206 of the transformer system 200. It is noted that reference numerals corresponding to those used for components in figure 1 , incremented by 100, may correspond to those components, and hence they are not introduced in detail again.

[0072] The NMR sensor 210 is shown submerged entirely submerged in the oil 204. However, it will be appreciated that the NMR sensor 210 in other examples may not be in the tank 206 with the oil 204 but may instead be out of the oil 104 and have a fluid communication system for communication oil 204 to the NMR sensor 210.

[0073] At regular or dynamic intervals (e.g., every 15 minutes), a sample of oil 204 is processed by the NMR sensor 210 to provide a convoluted NMR spectrum 212 for the oil 204. The NMR spectrum 212 maycomprise a series of peaks of various amplitudes, corresponding to the chemical makeup of the oil, wherein the peaks may be shifted or otherwise morphed as a result of further factors such as the temperature of the oil 204.

[0074] As shown schematically in figure 2, the NMR spectrum is then provided to an artificial neural network (ANN) 214 for processing. In this example, the ANN is a physics-informed neural network (PINN 214) having a plurality of physical constraints factored in. For example, the PINN 214 may be provided with constraints based on known chemical decompositions or reactions of compounds present in the transformer system 200.

[0075] As well as the NMR spectrum 212, the PINN 214 may be provided with further information regarding the transformer system 200. For example, the PINN 214 may be provided with a transformer operating parameter (voltage, power, etc.), the temperature of the oil 204, and / or a(n estimated) proportion of paper-to-oil. The water content in the oil 204 may change with temperature due to the absorption of water by the insulating paper. Curves such as the “Piper Chart” can be associated with transient models for water diffusion to estimate balance of water between paper and oil 204. The fluctuation of the water content in the oil 204 may not be caused by the ingress or consumption of water, but rather by the change in the balance of water between the paper and the oil 204.

[0076] The output 216 of the PINN 214 may comprise an estimated proportion of one or more constituent compounds in the oil 204. For example, the estimated proportion of H2, CP , H2O, and other compounds - preferably all expected constituent compounds - may be output from the PINN 214, e.g., in parts per million (ppm). The output 216 of the PINN 214 may comprise an estimated parameter for the oil 204, such as BDV, which is typically measured using electrodes. It will be understood that a parameter such as BDV is directly influenced by the chemical content of the oil 204. Hence, it is realized as a part of the present disclosure that BDV (and other parameters) can be estimated from the NMR spectrum 212, which is representative of the chemical makeup of the oil 204. Parameters such as BDV may be derivedfrom the estimated proportions of chemical constituents output from the PINN 214, or these parameters may be output alongside such estimated proportions.

[0077] The output 216 is used to generate a control assessment, which may broadly be considered as some assessment of the status of the transformer oil 204, where there may be a desire to keep properties of the oil 204 within certain limits to preserve functionality of the transformer system 200. In an example, if a transformer parameter exceeds some limit, an alert may be generated calling for the oil 204 to be cleaned or replaced, or for a change in operation of the transformer 202.

[0078] An illustration of a deconvolution of a convoluted NMR spectrum 3000 is shown in figure 3. The convoluted spectrum 3000, which may be considered as a ‘spectrum of a mixture’ is the obtained reading from an NMR sensor, and it can be dissociated / deconvoluted based on known spectra 3100a, 3100b, 3100c of (possible) constituent compounds.

[0079] In this case shown in figure 3, the NMR system would indicate that the mixture contains a percentage of compound ‘a’, having spectrum 3100a, and compound ‘b’, having spectrum 3100b, and it does not contain compound ‘c’, having spectrum 3100c. While the position of the peaks is mostly associated with the types of1H and13C found (the so-called ‘footprint’ of each substance), the amplitude may indicate the concentration.

[0080] Advanced algorithms support such interpretation of the NMR spectra 3000, but, in almost all cases, the decision is highly dependent on the operator’s judgment. The conventional techniques for NMR spectra interpretation use the identified hydroxyl and carbonyl groups to populate a list of combinations of molecules in different proportions that will fit the full spectrum of substances, minimizing the total error. Such a procedure relies on the operator’s expert assessment to select the most appropriate combination of substances, which may not be the best mathematical answer.

[0081] Small fluctuations in the readings may lead to big changes in the result. The best ‘fitting’ may include molecules not present in the transformeror lead to variations on the content of the compounds not reflecting any realistic physical process.

[0082] The challenge of building an online NMR sensor, such as the NMR sensor 210 discussed in relation to figure 2, is to incorporate the ‘expert judgment’ in the PINN model 214.

[0083] It is not possible in this example to display a number of options to the operator, which may often not be an expert in the particular transformer system 200. Preferably, no specialized worker should be required to perform an interpretation of the result.

[0084] Hence, as discussed in connection with the example of figure 2, the NMR sensor 210 automatically analyzes a sample of the oil 204 every pre-determined interval, e.g., 15 minutes, 30 minutes, 1 hour or another interval that can be configured by the operator. The result to be presented to the operators can comprise a list of transformer assessment parameters such as insulating liquid breakdown voltage, dielectric dissipation factor, neutralization index, interfacial tension, water content, combustible and noncombustible dissolved gases content, oxidation of the insulating liquid, etc.

[0085] These parameters may be based on the constituent compounds in the oil 204. Alternatively, only the estimated proportions of constituent compounds may be output from the PINN 14 and another system, such as another PINN or a support vector machine, or the like, may derive the parameters from the constituent compounds, depending on the implementation.

[0086] Hence, more generally speaking, and as illustrated in figure 4, a method 400 to monitor transformer insulating liquid is provided. The method 400 comprises steps of obtaining 410, from an NMR sensor arranged in a transformer insulating liquid tank, an NMR spectrum for the transformer insulating liquid. Then, the method 400 comprises processing 420 the NMR spectrum using a trained an artificial neural network (ANN) to thereby obtain information about the transformer insulating liquid. Finally, and based on the obtained information about the transformer insulating liquid, the method 400comprises generating 430 a control assessment for an aspect of the transformer.

[0087] Although these steps are shown in a linear sequence, it will be appreciated that one or more steps may be performed cyclically and / or in parallel with each other, depending on the particular implementation.

[0088] In order to achieve such an objective, the ANN should be suitably trained using one or more training methods known to those in the art.

[0089] In an example, schematically represented in figure 5, a system 500 uses a variation of a PINN 514 with multiple steps of supervised training and continuous unsupervised training (machine learning). The first step is shown in figure 5.

[0090] The first training step comprises training of the ANN 514 using laboratory-prepared samples. Based on a list of expected contaminants, a few hundreds of solutions are prepared with known contents of each compound (for example, new insulating liquid with 100 ppm of H2 dissolved).

[0091] The NMR sensor is used to analyze the solution multiple times, generating an expected spectrum 512 with some standard deviations (error functions). The prepared solution is subjected to normal laboratory tests 508, measuring transformer parameters.

[0092] The resultant matrix is the first training of the neural network 514, allowing the prediction of the NMR spectra and the transformer parameters 515 for any mix of the expected contaminants (included in a library).

[0093] Training is performed using a loss function, where the output parameters 515 are compared with the laboratory tests 508. The loss function is proportional to the squared difference between the outputs 515 of the ANN 514 and the known content of the compound (known because the sample was prepared in a known way), and each individually measured parameter from the laboratory testing 508 (BDV, DDF, etc.).

[0094] Some input data 513 for the ANN 514 are ‘fixed’ in this firsttraining, such as the fluid temperature and transformer ratings (MVA, kV, number of phases, application, etc.).

[0095] The second training step comprises retraining of the ANN 514 using samples of new transformers. A few dozen samples from transformers, essentially of unaged oil, are used for retraining the ANN 514. Additional input data 513 such as the fluid temperature and the paper-to-oil ratio, which were fixed in the first training step, will be adjusted in this retraining phase of the ANN 514.

[0096] In this training step, the loss function is proportional to the squared difference between the outputs 515 of the ANN 514 and each individually measured parameter in the liquid from laboratory tests 508. While the compounds' content is unknown during this training step, the input data 513 of the ANN 514 that was previously fixed, such as the fluid temperature and transformer information, are now variables. Hence, the second training step enables the ANN 514 to estimate transformer parameters from an NMR spectrum 512 more robustly with real-life variables provided in the input data 513.

[0097] Figure 5 shows a supervised training process. However, the third training step comprises unsupervised learning. In the third training step, the developed ANN is applied for the identifying the compounds / contaminants based on the obtained spectrum of NMR.

[0098] In the third training step, the inputs comprise the NMR spectra from the analyzed oil sample, a library including the “NMR footprint” of the investigated substances (frequencies and amplitudes), transformer ratings (MVA, kVs, application, etc.), temperatures and loading, and estimation of the paper-to-oil proportion.

[0099] The results are the content of each substance and the associated transformer parameters. Before accepting the result, the system compares each substance’s content's time-domain variation and transformer parameters. This verification considers the expected physical behavior when a model is available.

[0100] Other examples of ‘physically informed’ aspects of the network comprise a typical degradation process of a transformer, based on the loading and temperature (i.e. , a modified Arrhenius model), and / or a correlation between the content of acids and the neutralization index.

[0101] The continuing unsupervised learning of the ANN allows the ANN to specialize based on the particularities of the transformer system, which may substantially evolve over time to become unlike any other transformer system. Viewed from one perspective, it can be said that a generalized transformer system model cannot be accurately derived. Hence, aspects of the present disclosure advantageously enable the differences between systems to be learned and factored into estimations.

[0102] Occasional lab testing may be performed for the oil from the operational transformer system, to provide data for supervised learning, thereby ensuring that the accuracy of the ANN is maintained.

[0103] In another example, multiple ANNs may be trained and used. The first ANN may provide, as output, the probabilities of having a given parameter / component, e.g., 90% likely to have H2 present in the complex spectrum coming from the NMR sensor. A second ANN may then quantitatively identify individual parameters / com pounds (e.g., 100 ppm H2, given the 90% probability).

[0104] Knowing that a given parameter is present in the insulating liquid with a given level of uncertainty, a second and specialized ANN is trained to identify the amount of, for example, H2 present in the insulating liquid, given the data utilized before to train the first NN (insulating liquid type, temperature, etc.) provided with the convoluted NMR spectrum as input again and the further input, resulting from the output of the first ANN, indicating the probability of H2.

[0105] Having a large number of different cases, and knowing the amount of H2 in in the insulating liquid, the second ANN will learn to quantify amounts. It can be noted, in the context of this example, that the actual value of H2, measured in the lab, is not an input to this secondary ANN butcompared with the output, allowing for the loss function to be proportional to the squared difference between the actual and the estimated H2 level by the second ANN.

[0106] The same will apply to the other parameters / compounds, meaning that there can be provided as many specialized ANNs as the number of parameters / compounds that are to be quantified. Such a two-ANN approach may simplify training.

[0107] It will be appreciated that the training may be performed in any suitable manner, preferably supervised. A general training method 600 for training the ANN is illustrated in figure 6.

[0108] The method 600 comprises providing 610 a plurality of training samples to the ANN, wherein each training sample comprises an NMR spectrum of transformer insulating liquid corresponding to convoluted spectra of constituent compounds in the transformer insulating liquid.

[0109] The method 600 further comprises processing 620 the convoluted NMR spectrum using the ANN to estimate a proportion of at least one constituent compound in the transformer insulating liquid (and / or an estimated parameter for the transformer insulating liquid).

[0110] The method 600 then comprises training 630 the ANN by comparing the estimated proportion of the at least one constituent compound to a known proportion of said at least one constituent compound (e.g., applying a loss function, as discussed above).

[0111] The present disclosure may be better understood through appreciation of the following numbered clauses:1 . A method to monitor transformer insulating liquid, comprising: obtaining, from an NMR sensor arranged in a transformer insulating liquid tank, an NMR spectrum for the transformer insulating liquid; processing the NMR spectrum using a trained an artificial neural network (ANN) to thereby obtain information about the transformer insulatingliquid; and based on the obtained information about the transformer insulating liquid, generating a control assessment for an aspect of the transformer.2. The method according to clause 1 , further comprising: generating, using the NMR sensor, the NMR spectrum from a sample of the transformer insulating liquid.3. The method according to clause 1 or clause 2, wherein processing the NMR spectrum using the trained ANN comprises: providing, as input to the ANN, the NMR spectrum for the transformer insulating liquid, and receiving, as output from the ANN, an estimated proportion of at least one constituent compound in the transformer insulating liquid and / or at least one transformer parameter for the transformer insulating liquid.4. The method according to clause 3, wherein the at least one transformer parameter for the transformer insulating liquid is derived from the estimated proportion of at least one constituent compound in the transformer insulating liquid and / or output with the estimated proportion of at least one constituent compound in the transformer insulating liquid.5. The method according to any preceding clause, wherein at least the step of obtaining the NMR spectrum is performed at regular or dynamic intervals.6. A system for performing the method according to any preceding clause, comprising: one or more NMR sensors arranged in an insulating liquid tank of a transformer; and a computing device configured to receive output NMR spectra from the one or more sensors and process the NMR spectra with the ANN to thereby obtain an estimate of a proportion of a constituent compound in the transformer insulating liquid and / or an estimate of a transformer parameter forthe transformer insulating liquid.7. A method for training the ANN according to any preceding clause, comprising: providing a plurality of training samples to the ANN; wherein each training sample comprises an NMR spectrum of transformer insulating liquid corresponding to convoluted spectra of constituent compounds in the transformer insulating liquid; processing the convoluted NMR spectrum using the ANN to estimate a proportion of at least one constituent compound in the transformer insulating liquid; and training the ANN by comparing the estimated proportion of the at least one constituent compound to a known proportion of said at least one constituent compound.8. The method according to clause 7, further comprising: processing the convoluted NMR spectrum using the ANN to estimate a transformer parameter for the transformer insulating liquid; and training the ANN by comparing the estimated transformer parameter to a known transformer parameter for the transformer insulating liquid.9. The method according to clause 7 or clause 8, wherein processing the convoluted NMR spectrum using the ANN to estimate a proportion of at least one constituent compound in the transformer insulating liquid comprises: providing the NMR spectrum as a dataset to a plurality of input nodes in an input layer of the ANN, wherein each input node is provided with a respective subset of the dataset; providing each output of the plurality of input nodes to one or more intermediate layers of the ANN, wherein each intermediate layer comprises a plurality of intermediate nodes, wherein each intermediate node of each intermediate layer has one or more associated weight functions; and providing each output of the plurality of intermediate nodes of a lastintermediate later to a plurality of output nodes in an output layer of the ANN, wherein each output node is associated with a corresponding constituent compound of the transformer insulating liquid.10. The method according to clause 9, wherein the output layer comprises a SoftMax output layer for output nodes associated with constituent compounds of the transformer insulating liquid.11 .The method according to any of clause 8 to 10, wherein processing the convoluted NMR spectrum using the ANN to estimate a transformer parameter for the transformer insulating liquid comprises: providing the NMR spectrum as a dataset to a plurality of input nodes in an input layer of the ANN, wherein each input node is provided with a respective subset of the dataset; providing each output of the plurality of input nodes to one or more intermediate layers of the ANN, wherein each intermediate layer comprises a plurality of intermediate nodes, wherein each intermediate node of each intermediate layer has one or more associated weight functions; and providing each output of the plurality of intermediate nodes of a last intermediate later to a plurality of output nodes in an output layer of the ANN, wherein each output node is associated with a corresponding transformer parameter for the transformer insulating liquid.12. The method according to any of clauses 7 to 11 , wherein each training sample is derived from a controlled sample of transformer insulating liquid, and wherein the known proportion of said at least one constituent compound corresponds to a predetermined proportion of the constituent compound in the controlled sample of transformer insulating liquid.13. The method according to any of clauses 7 to 12, further comprising: providing a plurality of further training samples to the ANN; wherein each further training sample comprises an NMR spectrum of transformer insulating liquid corresponding to convoluted spectra ofconstituent compounds in the transformer insulating liquid; processing the convoluted NMR spectrum using the ANN to estimate a transformer parameter for the transformer insulating liquid; and training the ANN by comparing the estimated transformer parameter to a known transformer parameter for the transformer insulating liquid; wherein each training sample is derived from a sample of transformer insulating liquid from a transformer, and wherein the known transformer parameter corresponds to a measured transformer parameter for the sample of transformer insulating liquid from the transformer.14. The method according to any of clauses 7 to 13, wherein comparing the estimated proportion of the at least one constituent compound to a known proportion of said at least one constituent compound comprises: applying a loss function proportional to the squared difference between an output estimated proportion of the at least one constituent compound from the ANN and the known proportion of said at least one constituent compound.15. The method according to any of clauses 7 to 14, wherein comparing the estimated transformer parameter to a known transformer parameter for the transformer insulating liquid comprises: applying a loss function proportional to the squared difference between an output estimated transformer parameter from the ANN and the known transformer parameter.16. The method according to any of clauses 7 to 15, further comprising: providing, in association with a training sample: a transformer operating parameter; a transformer insulating liquid temperature; and / or a proportion of paper to transformer insulating liquid, associated with the training sample.17. The method according to any of clauses 7 to 16, further comprising:providing one or more NMR spectra for individual compounds with the plurality of training samples; wherein the one or more NMR spectra for individual compounds are chosen from a library of compounds expected to be constituent compounds in the transformer insulating liquid.18. An artificial neural network (ANN) trained according to the method of any of clauses 7 to 17.19. The ANN according to clause 18, wherein the ANN is a physics informed neural network (PINN), having at least one constraint based on a physical property of the transformer insulating liquid.20. The ANN according to clause 19, wherein the PINN comprises a timevariance constraint corresponding to an expected time variation in a proportion of a constituent compound and / or a transformer parameter over time.21 .A computing device having the ANN according to any of clauses 18 to 20 stored thereon.

[0112] While the present disclosure is susceptible to various modifications and alternative forms, specific embodiments are shown and described by way of example in relation to the drawings, with a view to clearly explaining the various advantageous aspects of the present disclosure. It should be understood, however, that the detailed description herein and the drawings attached hereto are not intended to limit the disclosure to the particular form disclosed. Rather, the intention is to cover all modifications, equivalents, and alternatives falling within the scope of the following claims.

Claims

28P A T E N T C L A I M S1 . A method to monitor transformer insulating liquid, comprising: obtaining, from an NMR sensor arranged in a transformer insulating liquid tank, an NMR spectrum for the transformer insulating liquid; processing the NMR spectrum using a trained artificial neural network (ANN) to thereby obtain information about the transformer insulating liquid; and based on the obtained information about the transformer insulating liquid, generating a control assessment for an aspect of the transformer.

2. The method according to claim 1 , further comprising: generating, using the NMR sensor, the NMR spectrum from a sample of the transformer insulating liquid.

3. The method according to claim 1 or claim 2, wherein processing the NMR spectrum using the trained ANN comprises: providing, as input to the ANN, the NMR spectrum for the transformer insulating liquid, and receiving, as output from the ANN, an estimated proportion of at least one constituent compound in the transformer insulating liquid and / or at least one transformer parameter for the transformer insulating liquid.

4. The method according to claim 3, wherein the at least one transformer parameter for the transformer insulating liquid is derived from the estimated proportion of at least one constituent compound in the transformer insulating liquid and / or output with the estimated proportion of at least one constituent compound in the transformer insulating liquid.

5. The method according to any preceding claim, wherein at least the step of obtaining the NMR spectrum is performed at regular or dynamic intervals.

6. A system for performing the method according to any preceding claim, comprising: one or more NMR sensors arranged in an insulating liquid tank of a transformer; and a computing device configured to receive output NMR spectra from the one or more sensors and process the NMR spectra with the ANN to thereby obtain an estimate of a proportion of a constituent compound in the transformer insulating liquid and / or an estimate of a transformer parameter for the transformer insulating liquid.

7. A method for training the ANN according to any preceding claim, comprising: providing a plurality of training samples to the ANN; wherein each training sample comprises an NMR spectrum of transformer insulating liquid corresponding to convoluted spectra of constituent compounds in the transformer insulating liquid; processing the convoluted NMR spectrum using the ANN to estimate a proportion of at least one constituent compound in the transformer insulating liquid; and training the ANN by comparing the estimated proportion of the at least one constituent compound to a known proportion of said at least one constituent compound.

8. The method according to claim 7, further comprising: processing the convoluted NMR spectrum using the ANN to estimate a transformer parameter for the transformer insulating liquid; and training the ANN by comparing the estimated transformer parameter to a known transformer parameter for the transformer insulating liquid.

9. The method according to claim 7 or claim 8, wherein processing the convoluted NMR spectrum using the ANN to estimate a proportion of at least one constituent compound in the transformer insulating liquid comprises:providing the NMR spectrum as a dataset to a plurality of input nodes in an input layer of the ANN, wherein each input node is provided with a respective subset of the dataset; providing each output of the plurality of input nodes to one or more intermediate layers of the ANN, wherein each intermediate layer comprises a plurality of intermediate nodes, wherein each intermediate node of each intermediate layer has one or more associated weight functions; and providing each output of the plurality of intermediate nodes of a last intermediate later to a plurality of output nodes in an output layer of the ANN, wherein each output node is associated with a corresponding constituent compound of the transformer insulating liquid.

10. The method according to claim 9, wherein the output layer comprises a SoftMax output layer for output nodes associated with constituent compounds of the transformer insulating liquid.11 . The method according to any of claims 8 to 10, wherein processing the convoluted NMR spectrum using the ANN to estimate a transformer parameter for the transformer insulating liquid comprises: providing the NMR spectrum as a dataset to a plurality of input nodes in an input layer of the ANN, wherein each input node is provided with a respective subset of the dataset; providing each output of the plurality of input nodes to one or more intermediate layers of the ANN, wherein each intermediate layer comprises a plurality of intermediate nodes, wherein each intermediate node of each intermediate layer has one or more associated weight functions; and providing each output of the plurality of intermediate nodes of a last intermediate later to a plurality of output nodes in an output layer of the ANN, wherein each output node is associated with a corresponding transformer parameter for the transformer insulating liquid.

12. The method according to any of claims 7 to 11 , wherein each trainingsample is derived from a controlled sample of transformer insulating liquid, and wherein the known proportion of said at least one constituent compound corresponds to a predetermined proportion of the constituent compound in the controlled sample of transformer insulating liquid.

13. The method according to any of claims 7 to 12, further comprising: providing a plurality of further training samples to the ANN; wherein each further training sample comprises an NMR spectrum of transformer insulating liquid corresponding to convoluted spectra of constituent compounds in the transformer insulating liquid; processing the convoluted NMR spectrum using the ANN to estimate a transformer parameter for the transformer insulating liquid; and training the ANN by comparing the estimated transformer parameter to a known transformer parameter for the transformer insulating liquid; wherein each training sample is derived from a sample of transformer insulating liquid from a transformer, and wherein the known transformer parameter corresponds to a measured transformer parameter for the sample of transformer insulating liquid from the transformer.

14. The method according to any of claims 7 to 13, wherein comparing the estimated proportion of the at least one constituent compound to a known proportion of said at least one constituent compound comprises: applying a loss function proportional to the squared difference between an output estimated proportion of the at least one constituent compound from the ANN and the known proportion of said at least one constituent compound.

15. The method according to any of claims 7 to 14, wherein comparing the estimated transformer parameter to a known transformer parameter for the transformer insulating liquid comprises: applying a loss function proportional to the squared difference between an output estimated transformer parameter from the ANN and the known transformer parameter.3216. The method according to any of claims 7 to 15, further comprising: providing, in association with a training sample: a transformer operating parameter; a transformer insulating liquid temperature; and / or a proportion of paper to transformer insulating liquid, associated with the training sample.

17. The method according to any of claims 7 to 16, further comprising: providing one or more NMR spectra for individual compounds with the plurality of training samples; wherein the one or more NMR spectra for individual compounds are chosen from a library of compounds expected to be constituent compounds in the transformer insulating liquid.

18. An artificial neural network (ANN) trained according to the method of any of claims 7 to 17.

19. The ANN according to claim 18, wherein the ANN is a physics informed neural network (PINN), having at least one constraint based on a physical property of the transformer insulating liquid.

20. The ANN according to claim 19, wherein the PINN comprises a timevariance constraint corresponding to an expected time variation in a proportion of a constituent compound and / or a transformer parameter over time.21 . A computing device having the ANN according to any of claims 18 to20 stored thereon.

Images