Machine learning architecture for asset condition prediction

Abstract

A method for generating a machine learning (ML) prediction model for a monitored target utility asset is disclosed. The method includes obtaining, from one or more historical sensor reading databases, historical sensor reading data associated with a target utility asset, and filtering the historical sensor reading data to separate nonrelevant data that is abnormal with respect to the operation of the target utility asset from relevant historical sensor reading data. The method further includes clustering the relevant historical sensor reading data to identify historical sensor reading data that is related to other utility assets which have been identified as being similar to the target utility asset, and utilizing the clustered relevant historical sensor reading data to generate at least one ML prediction model that is configured to respectively determine at least one predicted future condition of the target utility asset.

Description

MACHINE LEARNING ARCHITECTURE FOR ASSET CONDITION PREDICTIONTECHNICAL FIELD

[0001] The present disclosure relates generally to machine learning (ML) and utility asset devices (e.g., electrical transformer devices), and more particularly to utilizing ML generated models to predict future conditions associated with one or more utility asset devices.BACKGROUND

[0002] State of the art monitoring solutions primarily focus on previous and past readings to determine the current condition and / or operational status of a monitored utility asset, such as a liquid immersed electrical transformer device. For example, dissolved gas analysis (DGA) is the one of the most valuable measures for evaluating the well-being of a liquid immersed transformer. Notably, DGA involves the evaluation of a wide variety of dissolved gases (e.g., particles per million (ppm) concentration) that are present in the insulating liquid (e.g., oil) contained in the transformer device. Based on the ratio and quantities of the dissolved gases, several techniques (e.g., Duval Triangles and Pentagons, Rogers Ratios, ETRA, etc.) have been previously employed to attempt to discern the underlying problems of a transformer device. For example, dissolved gas readings can be captured via sensors that detect the amount of gas in the insulating liquid of the transformer at the time of a reading. However, the system operator has no knowledge of future values of DGA, and current methodologies are limited only to the analysis of past datapoints as opposed to learning the transformer device’s behavior over time. Knowing future DGA values, or other like future utility asset values and / or conditions, effectively enables the system operator to perform predictive maintenance, thereby avoiding problems in the target utility asset that can become worse over time.SUMMARY

[0003] A method for generating a machine learning (ML) prediction model for a monitored target utility asset is disclosed. The method includes obtaining, from one or more historical sensor reading databases, historical sensor reading data associated with a target utility asset, and filtering the historical sensor reading data to separate nonrelevant data that is abnormal withrespect to the operation of the target utility asset from relevant historical sensor reading data. The method further includes clustering the relevant historical sensor reading data to identify historical sensor reading data that is related to other utility assets which have been identified as being similar to the target utility asset, and utilizing the clustered relevant historical sensor reading data to generate at least one ML prediction model that is configured to respectively determine at least one predicted future condition of the target utility asset.

[0004] According to one embodiment of the method, the at least one ML prediction model includes one or more of a dissolved gas analysis, DGA, prediction model, an asset load prediction model, and / or an asset condition prediction model.

[0005] According to one embodiment of the method, the historical sensor reading data includes i) historical sensor readings that are unique to the target utility asset, ii) historical sensor readings from utility assets that are similar in type to the target utility asset and are operating in the same utility entity as the target utility asset, and / or iii) historical sensor readings from utility assets associated with a plurality of utility entities classified as being similar to the target utility asset.

[0006] According to one embodiment, the method further includes utilizing at least one sensor device of the target utility asset to obtain sample reading data from the target utility asset.

[0007] According to one embodiment of the method, the sample reading data is provided to the one or more historical sensor reading databases and / or the at least one ML prediction model.

[0008] According to one embodiment of the method, the at least one ML prediction model is generated in a manner that is asynchronous and independent from the obtaining of the sample reading data from the target utility asset.

[0009] According to one embodiment of the method, the predicted future condition includes a discrete status prediction comprising a plurality of different condition levels.

[0010] According to one embodiment of the method, the predicted future condition is a set of two or more discrete values for ranking its condition from a normal class value to different levels of abnormality and / or severity class values.

[0011] According to one embodiment of the method, wherein the one or more historical sensor reading databases includes at least one large global database.

[0012] According to one embodiment, the method further includes utilizing the clustered relevant historical sensor reading data includes applying a moving window to one or more subsets of the historical sensor reading data to determine microtrends that are utilized to train and / or retrain the at least one ML prediction model.

[0013] According to one embodiment, the method further includes generating synthetic sensor reading data based on an analysis of the clustered relevant historical sensor reading data.

[0014] According to one embodiment of the method, the at least one ML prediction model is trained with the generated synthetic sensor reading data and the clustered relevant historical sensor reading data.

[0015] According to one embodiment of the method, the target utility asset is a liquid immersed transformer device and the sample reading data includes a sensor reading of a dissolved gas sample from an insulating liquid contained in the liquid immersed transformer device

[0016] According to one embodiment, the method further comprises presenting the at least one predicted future condition of the target utility asset via one or more graphical user interfaces.

[0017] According to one embodiment, the method further comprises utilizing the clustered relevant historical sensor reading data to generate at least one ML prediction model includes training and validating the at least one ML prediction model.

[0018] According to one embodiment of the method, the at least one ML prediction model is configured to determine the at least one predicted future condition of the target utility asset in response to receiving and processing current sensor reading data originating from the target utility asset.

[0019] According to one embodiment of the method, the historical sensor reading data includes trend related measurement data, data representative of a number of outlier values, and / or abnormal value data.

[0020] According to one embodiment, the method further comprising issuing a notification to at least one system operator of the at least one predicted condition to the target utility asset in response to determining that the at least one predicted future condition represents an abnormal result.

[0021] According to one embodiment of the method, the notification includes an audio and / or visual alert directed to a system operator.

[0022] The subject matter described herein for generating a machine learning prediction model for a monitored target utility asset may be implemented in hardware, software, firmware, or any combination thereof. As such, the terms “function,” “engine,” and / or “module” as used herein refer to hardware, software, and / or firmware for implementing the feature being described. In one exemplary implementation, the subject matter described herein may be implemented using a computer readable medium having stored thereon computer executable instructions that when executed by the processor of a computer control the computer to perform steps.

[0023] Exemplary computer readable media suitable for implementing the subject matter described herein include non-transitory computer readable media, such as disk memory devices,chip memory devices, programmable logic devices, and application specific integrated circuits. In addition, a computer readable medium that implements the subject matter described herein may be located on a single device or computing platform or may be distributed across multiple devices or computing platforms. At least a portion of any of the modules, predictive model instances, and / or engines described herein may be implemented in hardware.BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The accompanying drawings, which are included to provide a further understanding of the disclosure and are incorporated in and constitute a part of this application, illustrate certain non-limiting embodiments of inventive concepts. In the drawings:

[0025] Figure l is a block diagram depicting an example machine learning architecture used for predicting target asset conditions according to some embodiments;

[0026] Figure 2 is a block diagram depicting an example host device configured to generate and deploy machine learning models according to some embodiments;

[0027] Figure 3 is a flow chart diagram depicting an example method for predicting target asset conditions according to some embodiments;

[0028] Figure 4 is a diagram depicting an example screen display for providing continuous value prediction monitoring data according to some embodiments;

[0029] Figure 5 is a diagram depicting an example screen display for providing discrete value prediction monitoring data according to some embodiments; and

[0030] Figure 6 is a diagram depicting an example screen display for providing binary value prediction monitoring data according to some embodiments.DETAILED DESCRIPTION

[0031] Some of the embodiments contemplated herein will now be described more fully with reference to the accompanying drawings. Embodiments are provided by way of example to convey the scope of the subject matter to those skilled in the art, in which examples of embodiments of inventive concepts are shown. Inventive concepts may, however, be embodied in many different forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of present inventive concepts to those skilled in the art. It should also be noted that these embodiments are not mutually exclusive. Components from one embodiment may be tacitly assumed to be present / used in another embodiment.

[0032] The disclosed subject matter includes a system that is configured to continuously create one or more machine learning (ML) predictive models configured to predict future conditions and / or statuses of a target asset. As used herein, a target asset may refer to a utility asset device and / or system, such as a liquid filled electrical transformer, distribution lines, transmission lines, substations, shunt reactors, circuit breakers, batteries, switchgears, autotransformers, phaseshifters, high voltage direct current (HVDC) transformers, furnace transformers, and the like.

[0033] Examples of the predicted future conditions or statuses associated with the target asset may include dissolved gas content monitoring, load monitoring, moisture monitoring, insulation monitoring, temperature monitoring, and any other monitoring measure for the target asset. In some embodiments, the disclosed subject matter may also be configured to train and / or subsequently retrain an ML predictive model using data from one or more large databases of target asset data (e.g., transformer operational data) as described below. In addition to i) learning a particular transformer’ s behavior over time, ii) filtering fluctuations and outliers in measurements, and iii) applying hypothesis tests, the comparison of a target asset’s behavior with other similar units placed in a common cluster (e.g., sister units, family of similar units, etc.) is an important measure for minimizing errors and to differentiate typical from atypical behaviors and trends. Notably, an informed prediction based on monitored data may be the difference between recommending de-energizing a healthy target asset unit, and dealing with an unacceptable risk of failure in an unhealthy asset.

[0034] As indicated above, the disclosed subject matter pertains to the creation of one or more ML predictive models that predict future conditions and statuses of a target asset. For purposes of example, the present disclosure will describe the disclosed subject matter in the context of predicting future conditions for an electrical transformer device utilizing DGA monitoring. However, it is understood that this example target asset and / or specific monitoring does not limit the scope of the disclosed subject matter, which may also be applied to other target asset types and monitoring measures.

[0035] Notably, the prediction of asset conditions as related to the generation of gases in transformer devices is a very complex phenomenon that may or may not indicate eventual faults. For example, a transformer may operate at a given temperature, which triggers the formation of some gases produced by the insulating liquid itself (i.e., stray-gassing) and whose detected presence cannot be correlated to any electrical fault, dielectric fault, or thermal fault in the active part. Some gases (e.g., CO and CO2) may also be generated in the initial phase of operation of a particular transformer, thus being unrelated to any fault type. On other occasions, gases may be formed due to specificities of the loading profile r the electrical system where the asset is inserted (e.g., presence of harmonics) that can appear and disappear, hence such gases may alsonot be understood as an indication of a fault or abnormal behavior. Yet, the presence of other gases may be correlated to the loading condition of the transformer and demonstrate a strong correlation with operational procedures. There is also the possibility of insulating liquid (e.g., oil) contamination caused by maintenance or other devices, such as load-type changers. It is also well-known that some electrical transformers may operate at higher gas concentration levels (i.e., atypical levels) for an extended period of time, without failure or compromising operation. Finally, there is the issue of statistical reference. Ideally, users can create their own statistical references based on their own network experience. Levels found in international standards serve as useful references in the absence of knowledge of the user’s own data. With all that in mind, a machine learning solution has been developed to acquire sensor reading data from a large database(s) of target asset data that is sourced by different utilities and uses the sensor reading history of the sensor devices available in the target asset (e.g., transformer) to create ML predictive models that allow for the prediction of future DGA values for the transformer. Other sensor reading data that can be used for other condition predictions and / or value predictions include load data, insulating liquid quality parameter data, bushing data, and insulating liquid temperature data, and the like.

[0036] One example ML architecture used for predicting target asset conditions is depicted in Figure 1. Specifically, Figure 1 illustrates an exemplary architecture of system 100, which can be deployed to generate one or more ML predictive models configured to conduct predictions of target asset conditions. System 100 includes a target asset 110, a sensor reading database collection 104, a ML predictive model generation engine 106, a ML predictive model group 114, a condition prediction manager 130, a data analysis and monitoring engine 150, and at least one monitoring tool 115. Although Figure 1 is described in the context of a liquid-immersed transformer asset and the prediction of dissolved gas analysis (DGA) condition(s), other conditions and / or target assets deployed by utility entities (e.g., power companies) may be monitored in a similar manner without departing from the scope of the disclosed subject matter.

[0037] Referring to Figure 1, target asset 110 may include a utility asset (e.g., a liquid immersed electrical transformer) that has been deployed in a power grid system operated by a power utility company. Notably, target asset 110 may include one or more sensor devices that are configured to collect, measure, and / or communicate one or more monitored readings associated with the operational status of target asset 110. For example, target asset 110 may include a DGA monitoring sensor device that collects DGA data 161, a load measurement monitoring sensor device that collects load data 162 (e.g., the amount of electricity flowing through the transformer), a bushings monitoring sensor device that collects bushings data 163, atemperature monitoring sensor device that collects cooling and / or temperature data 164, an insulation monitoring sensor device that collects insulation data 165, and the like.

[0038] Using DGA data collection as an example, dissolved gas readings may be captured through sensors that detect the amount of gas in the insulating liquid at the time of reading (e.g., on-line reading). Alternatively, dissolved gas readings may be obtained through laboratorial analysis of samples of the insulating liquid. While the first method allows short intervals between readings, the second method (typically deemed as more reliable than the first method) takes days to be completed. When any abnormal behavior is identified, it is essential for the system to define the severity of the issue. Several aspects are to be considered within the concept of severity, including the likelihood of that behavior to result in a recommendation for taking the unit out of service and the timeframe for that to occur. Since there is no way for the asset operator / manager to anticipate the future statuses of DGA, such prediction depends strongly on the experience of the technician and includes a large percentage of guesswork.

[0039] In Figure 1, collected data 161-165 is graphically depicted as a collection of sensor data 108. Notably, sensor data 108 may be communicated to and / or collected by one or more databases (e.g., data source servers hosting and / or managing the collected sensor data) for storage and subsequent distribution to ML predictive model generation engine 106 and / or condition prediction manager 130.

[0040] In some embodiments, database collection 104 comprises a plurality of different historical data databases, including a target asset historical database 111, a utility company historical database 112, and a utility conglomerate historical database 113. Notably, these historical databases store historical sensor reading data that may be used by ML predictive model generation engine 106 to generate, train, and / or retrain ML predictive models (see description below). In some embodiments, each of historical databases 111-113 is hosted by a separate dedicated database host server. In other embodiments, a single database host server may be configured to host and / or store two or more of historical databases 111-113. In some embodiments, the historical databases 111-113 may be stored in a single large global historical database.

[0041] In some embodiments, target asset historical database I l l is configured to contain historical data related to the particular target asset (e.g., a specific transformer operating as an asset in the utility power grid system) that the utility entity is attempting to obtain a predicted condition. For example, target asset historical database 111 may contain past and previously collected DGA measurement readings, load readings, bushing readings, temperature readings, and / or insulation readings corresponding to a single specific target transformer device operating in the utility power grid.

[0042] Similarly, utility company historical database 112 is configured to store previous sensor readings made from other transformer devices and / or target assets deployed by a single / common company (i.e., the same utility entity that deploys the target asset being analyzed and / or monitored). For example, the data stored in utility company historical database 112 may include past and previously collected DGA measurement readings, load readings, bushing readings, temperature readings, and / or insulation readings corresponding to all the transformer devices that are operating in the utility entity’s power grid (e.g., utility entity’s ‘transformer fleet’). In addition, utility company historical database 112 may also be configured to store past and previously collected data that is associated with other target assets (e.g., substations, transmission lines, distribution lines, etc.) that operate in the utility’s power grid.

[0043] Likewise, utility conglomerate database 113 is configured to store previous readings made from a multitude of transformers and / or target assets deployed by a number of participating utility entities and / or companies (e.g., a SuperMinds database). For example, the data stored in utility conglomerate database 113 may include past and previously collected DGA measurement readings, load readings, bushing readings, and / or temperature readings corresponding to each of a plurality of transformer devices that are operating all of the power grids of the participating utility entities. In addition, utility conglomerate database 113 may also be configured to store past and previously collected data that is associated with other target assets (e.g., substations, transmission lines, distribution lines, etc.) that operate in all of the power grids of the participating utility entities. In some embodiments, the data stored and / or collected from utility conglomerate database 113 is conducted in an anonymous manner (i.e., data identifying the specific utility entities is not stored and / or collected).

[0044] In some embodiments, each of historical databases 111-113 may be configured to provide access to its respective historical asset data to a ML predictive model generation engine 106 (e.g., asset filtering engine 121 and / or asset clustering engine 122). In particular, the historical asset sensor reading data obtained from databases 111-113 may be used to create and / or train ML model algorithms that correspond to the various prediction models included in ML predictive model group 114. For example, Figure 1 includes ML predictive model engine 106 that is configured to execute one or more processes to create, train, retrain, and / or deploy the machine learning models for one or more different monitored parameters (e.g., DGA, load, temperature, etc.) associated with the target asset(s).

[0045] As indicated in Figure 1, ML predictive model generation engine 106 comprises a plurality of components and / or modules including an asset filtering engine 121, an asset clustering engine 122, a synthetic data generation engine 123, a data pre-processing engine 124, a ML predictive model training and validation engine 125, and a ML predictive modeldeployment engine 126. In some embodiments, ML predictive model generation engine 106 may comprise a plurality of algorithms that are hosted and / or executed on a local computing device or a remote cloud-based computing device.

[0046] In some embodiments, asset filtering engine 121 is configured to obtain and / or assess historical data related to target asset 110 from one or more of historical databases 111-113. For example, asset filtering engine 121 may be configured to query (e.g., submit a request for stored reading data to) each of historical databases 111-113 on a periodic basis, or in response to a realtime request by a system operator. In other embodiments, the historical databases 111-113 may be configured to transmit historical sensor reading data to the ML predictive model engine 106 (e.g., its asset filtering engine 121 and / or asset clustering engine 122) on a periodic basis.

[0047] Once the historical data readings are received from one or more of historical databases 111-113, asset filtering engine 121 may be configured to conduct a filtering process on the sensor reading data. For example, asset filtering engine 121 can be configured to identify reading data as being irrelevant, abnormal, and / or not useful, and thus should be filtered out (i.e., separated) from the remaining relevant historical sensor reading data. For example, reading data that was captured when the target asset (e.g., transformer) was de-energized (i.e., no electricity flowing through transformer), when one or more sensor devices were not functioning properly, or some other similar abnormal situation. Additionally, asset filtering engine 121 can also be configured to filter out reading data values by detecting and / or identifying abnormal extreme values (e.g., usual peak or very low values). Notably, such reading data values can be identified as being an anomaly and / or impossibly irregular, such as if a temperature reading indicates a sensor measurement of 1000 degrees Celsius. Similarly, asset filtering engine 121 can be configured to exclude and / or filter out data readings / points that are outliers, thereby allowing the remaining relevant sensor reading data to represent more accurately the operational status of the monitored target asset.

[0048] In some embodiments, after the sensor reading data is filtered by asset filtering engine 121 (as described above), the filtered data may be forwarded to and processed by asset clustering engine 122. Because there is a multitude of different target assets and / or transformers (e.g., by design and / or model number), the typical readings associated with one type of transformer can vary significantly when compared to readings of other types of transformers. As such, asset clustering engine 122 is configured to perform a clustering operation that attempts to determine which of the multitude of transformers indicated in the obtained historical reading data are most similar (e.g., same design, same model, same make, etc.) to the specific target transformer that the system is trying to monitor and / or predict a future condition. When performing the clustering operation, asset clustering engine 122 is configured to identify relevanttransformers, such that ML predictive model generation engine 106 will only utilize the (remaining) historical sensor reading data that corresponds to transformers that are the most similar to the target transformer asset. For example, engine 122 may cluster the data from transformers that largely resemble the behavior of the target asset 110 so that the ML predictive models are trained on data that can best predict the behavior of the target asset transformer. As such, this measure contributes toward ensuring that the system’s prediction of asset condition is as accurate as possible.

[0049] The number and type of signals coming from the transformer are dependent on the sensors available in the transformer or target asset. Accordingly, the clustering operation is adaptable to future changes in the transformer. For example, if a new sensor is installed in the transformer, the asset clustering engine 122 can adapt to transformers that similarly include that type of sensor. The new sensor data would allow the asset clustering engine 122 to consider the sensor reading data from the new sensor, thus improving the clustering results to a cluster of assets that is more similar to the target asset (with the newly added sensor).

[0050] In some embodiments, reading data from historical databases 111-113 may first be received by asset clustering engine 122 (i.e., instead of asset filtering engine 121). In such configurations, asset clustering engine 122 can be configured to conduct its clustering processing prior to the filtering actions executed by asset filtering engine 121. Notably, the scope of the disclosed subject matter is not limited by the order in which the asset clustering engine 122 and the asset filtering engine 121 perform their respective operations on the data from historical databases 111-113.

[0051] After conducting the filtering and clustering processing indicated above (regardless of order), ML predictive model generation engine 106 is configured to determine if the amount of sensor reading data is sufficient to create, train, and / or retrain ML predictive models. If ML predictive model generation engine 106 determines that additional reading data is required to construct (or train) an fully representative ML predictive model, it may utilize its synthetic data generation engine 123 to (optionally) generate realistic synthetic data that can be used in lieu of, or in addition to, real historical data readings from databases 111-113. For example, the synthetic data generation engine 123 may be configured to analyze the cluster that was created (e.g., the output of asset clustering engine 122) for the particular target transformer asset and subsequently synthetically generate additional new data based on the behavior of the analyzed cluster.

[0052] The next stage of processing by the ML predictive model generation engine 106 includes data pre-processing for training. In some embodiments, this operation may be conducted by data pre-processing engine 124. In some embodiments, data pre-processing engine 124 may be configured to conduce a number of pre-processing functions including, but notlimited to, transforming time series data into tabular data, conduct feature engineering, perform correlation analysis, cleaning and / or handling missing values, encoding categorical variables, and scaling numerical features, and the like. Further, data pre-processing engine 124 may also be configured to create a moving window for conducting microtrend analysis and / or perform all manners of data manipulation in order to create a ML predictive model. In some embodiments, the process of ML predictive model training is continuous and new models are generated periodically to reflect the latest behavior of the transformer.

[0053] After data pre-processing is conducted, ML predictive model generation engine 106 is configured to utilize model training engine 125 to train a ML predictive model with the pre- processed data output from data pre-processing engine 124. In some embodiments, modeling training engine 125 is configured to train one or more ML predictive models by dividing the available sensor reading data into two or more sets: a training set and a testing / validation set. The training set can be used by engine 125 to generate and teach a predictive model, while the testing / validation set is used to evaluate the model’s performance. In some embodiments, the ML predictive model being generated comprises an artificial neural network that is subjected to multiple iterations of training. In each iteration, the predictive model sees the entire training dataset, adjusts its parameters, and tries to improve its predictive functionality.

[0054] In some embodiments, model training engine 125 may also be configured to conduct validation operations on a trained ML predictive model. Further, predictive model training engine 125 can be configured to constantly retrain one or more previously generated ML predictive models based on the new sensor data that ML predictive model generation engine 106 receives. Notably, the retrained ML predictive models are ultimately designated to replace the previously deployed ML predictive models (see below). In some embodiments, ML predictive model generation engine 106 is configured to perform the retraining of ML predictive models on a continuous basis.

[0055] After an ML predictive model is trained and validated, ML predictive model generation engine 106 is configured to utilize its model deployment engine 126 to deploy one or more ML predictive models. In some embodiments, the result of the deployment process is a set of ML predictive models (e.g., ML predictive model group 114) that is generated specifically for the target asset 110. For example, each deployed ML predictive model can predict a monitoring value and / or condition for the target asset. For example, Figure 1 illustrates a DGA predictive model 131, a load predictive model 132, and an insulation predictive model 133 (among others) that are deployed for target asset 110. Notably, predictive models 131-133 may be respectively configured to predict a DGA condition, a load condition, and a temperature condition that will occur at the target asset 110 at some predefined time period in the future. Although MLpredictive model generation engine 106 is continuously generating, training, and deploying ML predictive models, it is important to note that its operation (e.g., the operation collectively executed by engines 121-126) is conduct asynchronously and / or independently from the data collection operations associated with databases 111-113.

[0056] Once a ML predictive model is trained and deployed for the target transformer and is able to predict future condition values (e.g., DGA) for the target transformers, the predictive model training engine 125 may be configured to retrain such a predictive model after deployment. In some embodiments, the predictive model training engine can be configured to continuously learn the behavior of each specific target asset (e.g., transformer) by frequently retraining on micro trends of historic sensor reading data that is obtained from databases 111- 113.

[0057] Likewise, in a manner that is independent from (but possibly concurrent with) the ML predictive model generation and training processes executed by ML predictive model generation engine 106, the system 100 may be configured to conduct the actual condition prediction process using the recently gathered sensor data 108 from the target asset 110. For example, system 100 includes a condition prediction manager 130 that is configured to first receive current sensor reading data from the target asset 110 and subsequently perform preprocessing operations on said sensor reading data to match the input requirements of the currently deployed ML predictive models (e.g., models 131-133). More specifically, the sensor reading data is pre-processed by condition prediction manager 130 to comply with the expectations of the ML predictive model(s) 131-133. By providing the pre-processed data as input, each of the deployed ML predictive models can generate a predictive future value (e.g., a ‘diagnosis prediction output’) corresponding to the target asset’s future operation. For example, Figure 1 illustrates i) a future DGA diagnosis prediction 141 that is output from DGA predictive model 131, ii) a future load diagnosis prediction 142 that is output from load predictive model 132, and iii) a future insulation diagnosis prediction 143 that is output from insulation predictive model 133. Notably, ML predictive models 131-133 may be hosted and executed on a local computing device or a remote cloud-based computing device.

[0058] In some embodiments, one or more of these future diagnosis predictions 141-143 are provided to a data analysis and monitoring engine 150. Further, data analysis and monitoring engine 150 may be configured to also receive recent sensor data 108 from the target asset 110. Notably, data analysis and monitoring engine 150 can utilize the generated predictive future value(s) associated with diagnosis predictions 141-143 to predict and / or identify a future condition that will be exhibited by the target asset 110.

[0059] In some embodiments, the predictive future values associated with diagnosis predictions 141-143 are subsequently used together with current readings (e.g., collection of sensor data 108) from the target asset 110 to run the data analysis and proprietary monitoring algorithms (e.g., engine 150). Notably, data analysis and monitoring engine 150 may be configured to i) receive a diagnosis prediction output from a deployed predictive model and ii) receive current sensor reading data from a target asset sensor device (via condition prediction manager 130) and conduct an analysis to determine a predictive condition for a particular monitored operational metric (e.g., DGA, load, temperature, etc.) for a specific target asset. For example, data analysis and monitoring engine 150 may utilize the received data to perform any one or more of: i) a Continuous Value Prediction, ii) a Discrete Status Prediction, iii) Binary Value Prediction, iv) Time Estimation for reaching a Discrete Status Condition, and v) Probabilistic indication of the Future Condition. Examples of such predictions are described below.

[0060] In some embodiments, data analysis and monitoring engine 150 is configured to provide the predicted conditions to a system operator in a number of ways, including but not limited to, providing notifications, alerts, and / or information (e.g., visual and / or audio alarms, flags, etc.). Notably, data analysis and monitoring engine 150 may be configured to forward the predicted conditions for monitored target asset 110 and / or associated recommendations to system operator via monitoring tools 115. In some embodiments, system monitoring tools may include various screen displays that are visually presented to a system operator (e.g., via a computer screen or panel). Examples of the screen displays provided by system monitoring tools 115 are depicted in Figures 4-6 and discussed further below. Ultimately, such predictions produced by the disclosed subject matter is expected to allow a target transformer asset owner / operator to anticipate issues, thereby performing predictive maintenance to prevent a catastrophic failure. For example, data analysis and monitoring engine 150 may provide recommendation remedial actions to monitoring tools 115, such as a recommendation / notification for executing specific diagnostic test, de-energizing the target asset, shedding electric load from the target asset, or any like user instruction or feedback.

[0061] In some embodiments, data analysis and monitoring engine 150 may be hosted locally on an computing device with ML predictive model generation engine 106 and predictive models 131-33. In other embodiments, data analysis and monitoring engine 150 may be a cloudbased implantation that is hosted by a remote computing device.

[0062] Examples of Usage

[0063] Each of the ML predictive models described above produces an output (e.g., a predicted value of the model), which serves as information the data analysis and monitoringengine 150 uses to derive a predicted condition or status of the target asset. Data analysis and monitoring engine 150 be configured to derive predicted values in a number of ways, each of which can be graphically displayed using the monitoring tools 115 (e.g., Lumada Asset Performance Management). Examples of the various types of prediction techniques that may be conducted by data analysis and monitoring engine 150 from data output from an ML predictive model include, but are not limited to, i) Continuous Value Prediction, ii) Discrete Status Prediction, iii) Binary Value Prediction, iv) Time Estimation for reaching a Discrete Status Condition, and v) Probabilistic indication of the Future Condition.

[0064] Using the practical example pertaining to the DGA predictive model, one possible predicted value that can be produced from the model with regard to ‘Continuous Value Prediction’ includes an estimation of gas trends, such as the most likely ppm level for each gas and the like. In some embodiments, continuous value prediction may be used by engine 150 to estimate gas trends (e.g., the most likely ppm level for each gas). Such trends may be graphically presented to a user via screen displays. For example and referring to Figure 4, screen display 400 provides a user interface display which may be presented on a user’s computer screen. Notably, screen display 400 presents graphical plots 401-403, each of which depicts past values and a future predicted value (e.g., predicted values 411-413) associated with a single monitored gas. In particular, graphical plot 401 depicts data related to acetylene gas, graphical plot 402 depicts data related to ethylene gas, and graphical plot 403 depicts data related to ethane gas.

[0065] Similarly, another possible predicted value that can be produced from a generated predictive model with regard to ‘Discrete Status Prediction’ includes the analysis of the data with respect to a plurality of condition levels (e.g., Statuses 1, 2, 3, and 4). In some embodiments, ‘Statuses 1, 2, 3, and 4’ are established by using predefined thresholds. For example, data analysis and monitoring engine 150 may derive a predicted value that is then compared to (or cross-referenced with) a plurality of predefined ranges associated with each of Status 1, 2, 3, and 4. For example, in the context of a DGA predictive model, ‘ Status 1 ’ may be defined as all determined (predictive) gas concentration values (e.g., ppm) being below the 90thpercentile according to IEEE C57.104-2019 (see Table 1 below), while ‘Status 2’ can be defined as any determined predictive gas concentration value exceeding the 90thpercentile, but still lower than the 95thpercentile (and / or exceeding a first predefined rate of increase threshold / value ).‘ Status 3’ can be defined as any predictive gas concentration value between the 95thand 98thpercentile (and / or exceeding a second predefined rate of increase threshold / value), and ‘Status 4’ may be defined as any predictive value above the 98thpercentile (and / or exceeding a third predefined rate of increase threshold / value).Table 1—80* percentile gas concentrations as a function ofO2 / N2 ratio and age in pUL (ppm)Table 1. from IEEE C57.104-2019 used to classify DGA readings into Normal and Abnormal values

[0066] Such condition levels may be graphically presented to a user via screen displays. For example and referring to Figure 5, screen display 500 provides a user interface display which may be presented on a user’s computer screen. Notably, screen display 500 presents graphical plots 501-503, each of which depicts past and current values associated with a single monitored gas. In particular, graphical plot 501 depicts data related to acetylene gas, graphical plot 502 depicts data related to ethylene gas, and graphical plot 503 depicts data related to ethane gas. Each of plots 501-503 depicts four different tiers or zones, which correspond to each of Conditions 1-4. For example, plot 503 illustrates a tier 551 that corresponds to a Condition 1, a tier 552 that corresponds to a Condition 2, a tier 553 that corresponds to a Condition 3, and a tier 554 that corresponds to a Condition 4. An example of prediction is displayed in plot 501, in the presented text positioned below the gas name (‘Acetylene’), where it is predicted that the next reading will fall within Condition 2. These conditions reflect the current status of the transformer in terms of where the target asset measurement currently reads for each gas, as compared to the percentile levels of a large corpus of assets.

[0067] Yet another possible predicted value that can be produced from a generated predictive model with regard to ‘Binary Value Prediction’ includes the analysis of the data with respect to two condition statuses. In some embodiments, a ‘typical’ condition and ‘atypical / abnormal’ condition are established by using predefined thresholds. For example, data analysis and monitoring engine 150 may derive a predicted value that is then compared to (or cross-referenced with) two predefined ranges associated with each of the typical condition and the atypical condition. For example, in the context of a DGA predictive model, ‘typicalcondition’ may be defined as any determined (predictive) gas concentration value (e.g., ppm) that is equal to and / or below the 90thpercentile according to IEEE C57.104-2019, while ‘atypical’ can be defined as any determined predictive gas concentration value that is above the 90thpercentile.

[0068] Such binary value prediction levels may be graphically presented to a user via screen displays. For example and referring to Figure 6, screen display 600 provides a user interface display which may be presented on a user’s computer screen. Notably, screen display 500 presents graphical plots 601-603, each of which depicts past and current values associated with a single monitored gas. In particular, graphical plot 601 depicts data related to acetylene gas, graphical plot 602 depicts data related to ethylene gas, and graphical plot 603 depicts data related to ethane gas. Like Figure 5, each of plots 601-603 depicts four different tiers or zones.However, when utilizing binary value prediction, only the typical and atypical conditions are of concern. As such, plot 601 contains / displays text of the prediction, which is positioned below the gas title (i.e., ‘Acetylene'). Notably, the prediction text indicates that the next predicted value will fall within the atypical zone. In this case, the atypical zone is defined by the highest dotted line in figure 601. Every dot below the highest dotted line would be considered to be within the typical zone.

[0069] In other embodiments, the screen displays 400-600 can be configured to display alerts, notifications, and / or predicted conditions. For example, data analysis and monitoring engine 150 may determine a predictive value that can be used to estimate an amount of time (e.g., hours, days, weeks, etc.) for a particular gas concentration to reach any one of “Conditions 1-4” or alternatively, derive a probabilistic indication of a future condition. In both of these scenarios, such information may be provided to a user via any one of screen displays 400-600. For example, the displayed information may include a “Prediction: Next sample will reach the atypical zone in 5 weeks” or “Prediction: Transformer will be diagnosed with partial discharge within 30 days” (e.g., see field 610 in Figure 6 as an example of prediction information).

[0070] Figure 2 is a block diagram of a host device 200 in accordance with various aspects described herein. As used herein, the host device 200 may be or comprise various combinations of hardware and / or software, including a standalone server, a blade server, a cloud-implemented server, a distributed server, a virtual machine, container, or processing resources in a server farm. The host device 200 may be configured to function as a machine learning predictive model generation device that is adapted to generate (e.g., manage components / engines 121-126) and / or execute ML predictive models (e.g., predictive models 131-133) as described herein. Host device 200 may also be configured to conduct the data analysis and monitoring operations (e.g., data analysis and monitoring engine 150) described herein.

[0071] The host device 200 includes processing circuitry 210 that is operatively coupled via a bus (not shown) to an input / output interface 214, a network interface 216, a power source (not shown), and a memory 212. Other components may be included in other embodiments.

[0072] In some embodiments, processing circuitry 210 may comprise one or more computer processors, processing circuits, device, and / or components including, but not limited to, a single core or multi-core central processing unit (CPU), a graphical processing unit (GPU), a microprocessor, a microcontroller, and the like.

[0073] Memory 212 may comprise any type of non-volatile memory (e.g., flash memory, read only memory (ROM), programmable read-only memory (PROM), EPROM and EEPROM memory, etc.) and / or volatile memory (e.g., dynamic random-access memory (DRAM), static random-access memory (SRAM), etc.). In some embodiments, memory 212 includes a ML predictive model generation manager 106 that comprises one or more computer programs including an asset filtering engine 121, an asset clustering engine 122, a synthetic data generation engine 123, a data pre-processing engine 124, a ML predictive model training and validation engine 125, and a ML predictive model deployment engine 126 (as also shown in Figure 1). Memory 212 may also be configured to include and / or store a plurality of predictive models, such as DGA predictive model 131, load predictive model 132, and insulation predictive model 133. Notably, engines 121-126 and predictive models 131-133, when stored in memory 212, may be executed by processing circuitry 210 to perform the functionalities described herein. Embodiments of the host device 200 may utilize only a subset or all of the components shown. The aforementioned engines and predictive models may be implemented in a container-based architecture.

[0074] Figure 3 is a flow chart diagram depicting an example method 300 for predicting target asset conditions according to some embodiments. In some embodiments, method 300 may be a process or algorithm (e.g., ML model generation manager 106 and / or one or more of engines 121-126) that is stored in memory and executed by processing circuitry of one or more computing devices. In block 301, the method 300 includes obtaining, from one or more historical sensor reading databases, historical sensor reading data associated with a target utility asset. In some embodiments, the ML predictive model generation engine 106 is configured to receive historical sensor reading data from at least one of a monitored asset historical database 111, a historical company database 112, and / or a conglomerate historical database 113.

[0075] In block 302, the method 300 includes filtering the historical sensor reading data to separate nonrelevant data that is abnormal with respect to the operation of the target utility asset from relevant historical sensor reading data. In some embodiments, ML predictive model generation engine 106 is configured to utilize its asset filtering engine 121 to filter the obtainedhistorical sensor reading data to exclude abnormal sensor reading data from the relevant historical sensor reading data.

[0076] In block 303, the method 300 includes clustering the relevant historical sensor reading data to identify historical sensor reading data that is related to other utility assets which have been identified as being similar to the target utility asset. As indicated above, the filtering and clustering steps of blocks 302 and 303 may be performed in any order. For example, some embodiments of the disclosed subject matter may perform the clustering step in block 303 prior to the filtering step of block 302. In some embodiments, ML predictive model generation engine 106 is configured to utilize its asset clustering engine 122 to cluster the relevant historical sensor readings into one or more groups / clusters.

[0077] In block 304, the method 300 includes utilizing the clustered relevant historical sensor reading data to generate at least one ML prediction model that is configured to respectively determine at least one predicted future condition of the target utility asset. In some embodiments, ML predictive model generation engine 106 is configured to process the clustered and filtered historical sensor reading data to generate one or more ML prediction models, each of which can determine a future condition of a target utility asset.

[0078] In certain embodiments, some or all of the functionality described herein may be provided by processing circuitry executing instructions stored on in memory, which in certain embodiments may be a computer program product in the form of a non-transitory computer- readable storage medium. In alternative embodiments, some or all of the functionality may be provided by the processing circuitry without executing instructions stored on a separate or discrete device-readable storage medium, such as in a hard-wired manner. In any of those particular embodiments, whether executing instructions stored on a non-transitory computer- readable storage medium or not, the processing circuitry can be configured to perform the described functionality. The benefits provided by such functionality are not limited to the processing circuitry alone or to other components of the computing device, but are enjoyed by the computing device as a whole, and / or by end users generally.

Claims

CLAIMS1. A method for generating a machine learning, ML, prediction model for a monitored target utility asset (110), the method comprising: obtaining (301), from one or more historical sensor reading databases (104), historical sensor reading data associated with a target utility asset (110); filtering (302) the historical sensor reading data to separate nonrelevant data that is abnormal with respect to the operation of the target utility asset (110) from relevant historical sensor reading data; clustering (303) the relevant historical sensor reading data to identify historical sensor reading data that is related to other utility assets which have been identified as being similar to the target utility asset; and utilizing (304) the clustered relevant historical sensor reading data to generate at least one ML prediction model (114) that is configured to respectively determine at least one predicted future condition of the target utility asset (110).

2. The method of claim 1 wherein the at least one ML prediction model (114) includes one or more of a dissolved gas analysis, DGA, prediction model (131), an asset load prediction model (132), and / or an asset condition prediction model (133).

3. The method of any of claims 1-2 wherein the historical sensor reading data includes i) historical sensor reading data (111) that is unique to the target utility asset (110), ii) historical sensor reading data (113) from utility assets that are similar in type to the target utility asset and are operating in the same utility entity as the target utility asset, and / or iii) historical sensor reading data (113) from utility assets associated with a plurality of utility entities classified as being similar to the target utility asset.

4. The method of any of claims 1-3 further comprising utilizing at least one sensor device of the target utility asset (110) to obtain sample reading data from the target utility asset.

5. The method of any of claims 1-4 wherein the sample reading data is provided to the one or more historical sensor reading databases (104) and / or the at least one ML prediction model(114).

6. The method of any of claims 1-5 wherein the at least one ML prediction model (114) is generated in a manner that is asynchronous and independent from the obtaining of the sample reading data from the target utility asset (110).

7. The method of any of claims 1-6 wherein the predicted future condition includes a discrete status prediction comprising a plurality of different condition levels.

8. The method of any of claims 1-7 wherein the predicted future condition is a set of two or more discrete values for ranking its condition from a normal class value to different levels of abnormality and / or severity class values.

9. The method of any of claims 1-8 wherein the one or more historical sensor reading databases (104) includes at least one large global database.

10. The method of any of claims 1-9 wherein utilizing the clustered relevant historical sensor reading data includes applying a moving window to one or more subsets of the historical sensor reading data to determine microtrends that are utilized to train and / or retrain the at least one ML prediction model.

11. The method of any of claims 1-10 further comprising generating synthetic sensor reading data based on an analysis of the clustered relevant historical sensor reading data.

12. The method of any of claims 1-11 wherein the at least one ML prediction model (114) is trained with the generated synthetic sensor reading data and the clustered relevant historical sensor reading data.

13. The method of any of claims 1-12 wherein the target utility asset (110) is a liquid immersed transformer device and the sample reading data includes a sensor reading of a dissolved gas sample from an insulating liquid contained in the liquid immersed transformer device.

14. The method of any of claims 1-13 further comprising presenting the at least one predicted future condition of the target utility asset (110) via one or more monitoring tools (112).

15. The method of any of claims 1-14 wherein utilizing the clustered relevant historical sensor reading data to generate at least one ML prediction model (114) includes training and validating the at least one ML prediction model (114).

16. The method of any of claims 1-15 further wherein the at least one ML prediction model is configured to determine the at least one predicted future condition of the target utility asset in response to receiving and processing current sensor reading data originating from the target utility asset.

17. The method of any of claims 1-16 wherein the historical sensor reading data includes trend related measurement data, data representative of a number of outlier values, and / or abnormal value data.

18. The method of any of claims 1-17 further comprising issuing a notification via one or more monitoring tools (112) to at least one system operator of the at least one predicted condition to the target utility asset in response to determining that the at least one predicted future condition represents an abnormal result.

19. The method of any of claims 1-18 wherein the notification includes an audio and / or visual alert directed to a system operator.

20. A computer program comprising program code (106) to be executed by processing circuitry (210) of a machine learning predictive model generation device, whereby execution of the program code causes the machine learning predictive model generation device (200) to perform operations comprising any of the operations of claims 1-19.

21. A non-transitory computer-readable medium having instructions stored therein that are executable by processing circuitry (201) of machine learning predictive model generation device (200) to cause the machine learning predictive model generation device to perform operations comprising any of the operations of claims 1-19.

22. A machine learning predictive model generation device (200) comprising: processing circuitry (210); and memory (212) coupled with the processing circuitry (210), wherein the memory (212)includes instructions that when executed by the processing circuitry (210) causes the machine learning predictive model generation device (200) to perform operations according to any of claims 1-19.

Images