Method for generating signal processing logic, apparatus for controlling, monitoring, and / or analyzing a physical asset, and power system

Abstract

To generate the signal processing logic (42), machine learning model training is performed, which includes training one or more encoders and one or more decoders. At least one encoder of the trained machine learning model can be used to provide the signal processing logic (42) to a device (40) that executes the signal processing logic (42) to control, monitor, and / or analyze a physical asset.

Description

[Technical field]

[0001] FIELD OF THEINVENTION Embodiments of the invention relate to systems and methods for generating signal processing logic, devices for executing signal processing logic, and power systems. In particular, embodiments of the invention relate to devices and systems that operate to process signals containing measurements of properties of a physical asset or components connected to a physical asset into one or more features for performing control, monitoring, and / or analytical operations. [Background technology]

[0002] 2. Background of the Invention There are several applications where it is necessary to include a protection, monitoring, and / or control system. Examples of such applications include power system asset health monitoring, among others. Additional examples include power system protection, where protective devices are used to disconnect faulty portions of an electrical network, or process monitoring systems used to identify anomalous behavior in an industrial plant that may indicate a developing fault.

[0003] As the number of sensors deployed in modern utility or industrial systems continues to grow, the identification of features that depend on sensor measurements and are meaningful for control, monitoring, and / or analysis decisions is a considerable challenge. While human experts typically have a good understanding of how to design some features to be meaningful for control, monitoring, and / or analysis purposes, such techniques are prone to error and do not always yield particularly good features given the large number of sensor data available.

[0004] Computer-aided techniques have been used for control, monitoring, and / or analytical purposes in utility or industrial automation systems. For example, various methods can be used to analyze time series data for monitoring purposes. In recent years, Recurrent Neural Networks (RNNs) have become increasingly popular. RNNs allow for training Machine Learning (ML) models to detect specific events or situations in a classification training setting, using a training set consisting of an input time series at each time step and a desired output.

[0005] "Intermediate fused network with multiple timescale for anomaly detection" by W. Wang et al., Neurocomputing 433(2021)37-49, https: / / doi.org / 10.1016 / j.neucom.2020.12.025, ISSN0925-2312, proposes an intermediate fusion network with multiple timescales to predict future video segments for video anomaly detection. Video prediction techniques for anomaly detection need to derive anomalously identifiable future frames from a normal distribution provided by training data. Then, by measuring the difference between the generated frames and the reference frames, the model can know which frames represent abnormal events in a particular video sequence. To synthesize more characteristic future frames, the authors propose to enhance image prediction by integrating several input video segments at multiple timescales. Specifically, the diversity of timescales means the diversity of sampling frequencies of video frames. Furthermore, the authors introduce a novel intermediate fusion strategy by concatenating feature maps from intermediate layers to better preserve different characteristics from different timescales. The authors evaluate the proposed method on several public video surveillance datasets and achieve competitive results with respect to state-of-the-art methods.

[0006] "WaveletAE: A Wavelet-enhanced Autoencoder for Wind Turbine Blade Icing Detection" by B. Yuan et al., arXiv:1902.05625, teaches that wind power is abundant and inexhaustible as an alternative to burning fossil fuels. To fully utilize wind power, wind farms are usually located in high altitude regions and face severe icing conditions, which can lead to serious consequences. Rapid detection of blade icing is important for the maintenance of wind farms. Unlike the traditional method of installing expensive physical detectors on wind blades, data-driven techniques are increasingly popular to inspect wind turbines for faults. The authors propose a wavelet enhanced autoencoder model (WaveletAE) to identify wind turbine malfunctions by analyzing multivariate time series monitored by SCADA systems. WaveletAE is enhanced with wavelet detail coefficients to allow the autoencoder to incorporate information from multiple scales, and a CNN-LSTM architecture is applied to learn channel-wise and time-wise relationships. Experimental studies show that the proposed model outperforms other state-of-the-art time series anomaly detection methods for real-world blade icing detection.

[0007] The use of computer-aided techniques to generate features that can be used as a basis for control, monitoring, and / or analytical purposes can suffer from a variety of challenges. For example, it is difficult to provide computer-aided techniques that generate features that relate to and are useful for describing different temporal, spatial, or logical aspects of a physical asset, its operation, health, and / or deterioration. Summary of the Invention [Problem to be solved by the invention]

[0008] overview There is a need in the art for enhanced methods and systems for automatically or semi-automatically generating signal processing logic that can be executed by an apparatus to control, monitor, and / or analyze a physical asset. There is also a need in the art for improved apparatus, methods, and systems for controlling, monitoring, and / or analyzing a physical asset. For example, there is a need for methods and systems operable to generate signal processing logic that processes a set of input signals into one or several features that reflect characteristics of the physical asset on different temporal and / or spatial and / or logical (e.g., hierarchical) scales. [Means for solving the problem]

[0009] According to exemplary embodiments, there are provided inventions, methods, devices and systems as set forth in the independent claims. The dependent claims define preferred embodiments.

[0010] Methods, apparatus, and systems according to embodiments operate to generate or update signal processing logic by training one or more ML models comprising one or more encoders and one or more decoders. At least the encoder and / or decoder of the trained ML model may be implemented using at least one integrated circuit for processing signals including measurements related to a physical asset to perform control, monitoring, and / or analysis functions. Optionally, both the encoder and the decoder of the trained ML model may be implemented using at least one integrated circuit for processing signals including measurements related to a physical asset to perform control, monitoring, and / or analysis functions.

[0011] The physical asset may be an asset of an electrical power system. The measurements associated with the physical asset may be or may include measurements of electrical properties associated with the physical asset.

[0012] The signals may be processed to control, monitor, and / or analyze the physical asset during field use of the asset in the power system, for example during operation of the asset in the power system.

[0013] In particular, signals including measurements of electrical properties associated with physical assets may be processed to control the operation of the physical assets in a power system during field use. For example, the signals may be processed to perform preventive or corrective control actions on the physical assets by operating at least one circuit breaker, disconnector, and / or converter.

[0014] In this way, automated and improved control over the operation of an asset can be provided, for example, based on measurements of physical input parameters.

[0015] Training of the ML model can be performed to ensure separation of latent features of the code generated by the encoder of the autoencoder according to one or several characteristics. The separation can be by different time domains (e.g., into one or several first latent features that vary over a slower first time scale representing a sub-sampling of the second time scale over which the one or several second latent features vary). The separation can be by different spatial domains (e.g., into one or several first latent features that vary over a first spatial scale that is larger than the second spatial scale over which the one or several second latent features vary). The separation can also be by different logical domains.

[0016] The architecture of the ML model being trained may be configured to provide the desired feature separation. The autoencoder may operate such that some latent features of the code generated by the encoder are updated more frequently than other latent features. For example, a first set of latent features may be updated each time a new input is provided to the encoder (e.g., at a time interval corresponding to a sampling interval of the measurements), and a second set of latent features is updated only at integer multiples of that time. Several different encoders of the ML model may comprise different machine learning (ML) architectures, with a first ML portion providing a first set of latent features that are updated more frequently than a second set of latent features provided by a second ML portion.

[0017] Separation of latent features from different time, space, and / or logical domains can be achieved by performing ML model training to ensure separation. Training can be performed such that latent features included in a second set of latent features (which are updated less frequently) depend on the input in a way that they are automatically learned while providing adequate information about characteristics that change slowly over time and / or space scales.

[0018] In some implementations, adversarial decoder training may be performed in which a decoder of an ML model (e.g., an autoencoder) and an adversarial decoder of the ML model (e.g., an autoencoder) are trained in parallel. This may be done with iterative techniques. Only a subset of the latent features included in the code generated by the encoder of the autoencoder may be provided as decoder inputs to the adversarial decoder. For example, only latent features that vary more slowly (in time and / or spatial scales) than other latent features may be provided to the adversarial decoder.

[0019] Alternatively or additionally, signals that affect the operation of a physical asset, but are independent of the operation of the physical asset, may be processed differently than signals that are dependent on the operation of the physical asset, for example, signals that affect the operation of a physical asset may be fed to both the encoder and decoder portions of an autoencoder during autoencoder training.

[0020] ML model training may be performed such that an objective function is minimized or maximized, respectively. The objective function may be or include a loss function.

[0021] The objective function may be or may include the reconstruction loss or prediction loss.

[0022] The objective function may be or may include a loss quantification (e.g., using a metric or loss function) of the difference between the signal input to the encoder of the autoencoder and the reconstruction of the signal output by the decoder of the ML model.

[0023] The methods and systems may be used in connection with physical asset control within a power system and / or asset health monitoring of power system assets. The methods and systems may be used in connection with health monitoring or control of power transformers. The methods and systems may be used more broadly in connection with asset monitoring and control.

[0024] According to one embodiment, there is provided a method for generating or updating signal processing logic for processing signals, particularly time series signals, that include measurements related to a physical asset. The method includes training a machine learning (ML) model using at least one integrated circuit. The ML model comprises: a first encoder having a first input layer for receiving an input signal, in particular a time series input signal, the first encoder being operative to generate a first code comprising first latent features; at least one second encoder having at least one second input layer for receiving an input signal, the at least one second encoder operative to generate at least one second code including a second latent feature; a first decoder having an output layer, the first decoder operative to process a first decoder input to generate a first decoder output; and at least one second decoder having at least one second output layer, the at least one second decoder operative to process the second decoder inputs to generate a second decoder output.

[0025] Training the ML model may include providing training signals as input signals to a first input layer and a second input layer, generating first decoder inputs from at least some first latent features included in the first code, and generating second decoder inputs from at least some second latent features included in at least one second code.

[0026] The method may further include providing at least one of the first encoder or the second encoder and / or at least one of the first decoder or the second decoder of the trained ML model as at least part of the signal processing logic to a device executing the signal processing logic to control, monitor, and / or analyze the physical asset.

[0027] The physical asset may be an asset of an electrical power system. The measurements associated with the physical asset may be or may include measurements of electrical properties associated with the physical asset.

[0028] The device may execute signal processing logic to control, monitor, and / or analyze the physical asset during field use of the asset in the power system, for example during operation of the asset in the power system.

[0029] In particular, by executing the signal processing logic, the device may process signals including measurements of electrical properties associated with a physical asset to control operation of the physical asset within the power system during field use. For example, the device may execute the signal processing logic to operate at least one circuit breaker, disconnector, and / or converter to perform preventive or corrective control action of the physical asset during field use of the asset within the power system.

[0030] In this manner, for example, an automated and improved method for controlling the operation of a physical asset based on measurements of physical input parameters may be provided.

[0031] Generating the first decoder input may include generating the first decoder input from the first code and from an additional signal that may include at least some second potential features included in at least one second code.

[0032] The first latent feature and the second latent feature may change on different time scales. The first latent feature may vary over a slower time scale than at least some of the second latent features.

[0033] Training the ML model may include performing iterative training including a number of iterations, each of the number of iterations respectively including training a first encoder, a first decoder, and at least one second encoder with the objective of minimizing a loss function (such as a reconstruction loss or a prediction loss) of the first decoder and maximizing a loss function (such as a reconstruction loss or a prediction loss) of the at least one second decoder.

[0034] Training the ML model may include training at least one second decoder with the objective of minimizing a loss function (such as a reconstruction loss or a prediction loss) of the at least one second decoder.

[0035] The loss functions used in the training process (e.g., the loss function of the first encoder, the first decoder, and the at least one second decoder used to train the at least one second encoder, and the loss function of the at least one second decoder used to train the at least one second decoder) may be the same or different.

[0036] The loss function may be the reconstruction loss or the prediction loss. The ML model may include or be at least one autoencoder.

[0037] The first decoder output may include a first reconstruction of at least some of the input signal.

[0038] The second decoder output may include a second reconstruction of at least some of the input signal.

[0039] The additional signals may include variables that affect the operation of a physical asset, but exist independent of the presence or absence of the physical asset.

[0040] The additional signal may include a variable that is measured at a first sampling time that is longer than a second sampling time at which the measurement is sampled.

[0041] The additional signals may include ambient condition variables of the physical asset. The ambient condition variables may include weather data that affects at least one power system characteristic associated with the physical asset.

[0042] The first decoder input may include replicas of at least some of the second latent features included in the at least one second code.

[0043] One or more of the first latent variables may represent a temporal characteristic regarding the physical asset.

[0044] The one or several first latent variables of the trained first encoder may represent temporal properties that may vary over a first time scale that is longer than the sampling time interval of one, some, or all of the measurements.

[0045] The first time scale may be a sub-sampling of the sampling time interval. The apparatus may process one or more of the first latent variables to perform an asset health analysis of the physical asset.

[0046] According to an aspect of the invention, which may be used in combination with or independently of the aspects disclosed above, a method is provided for generating or updating signal processing logic for processing signals including measurements related to a physical asset. The method includes training an ML model. The ML model comprises an encoder having an input layer for receiving an input signal, in particular a time series input signal. The encoder is operative to generate a code including latent features. The ML model comprises a decoder having an output layer. The decoder is operative to process a decoder input including a code to generate a decoder output. Training the autoencoder may include providing a training signal as an input signal to the input layer.

[0047] Training the ML model may include generating decoder inputs from the code and, optionally, from an additional signal that is distinct from the code.

[0048] The decoder output may include a reconstruction of at least some of the input signal. The input signal may be or may include an offline signal that may be obtained from a data repository.

[0049] The input signal may be or may include an online signal received from a field device (eg, when updating signal processing logic during field use).

[0050] The method may include providing (e.g., by storing or transporting) at least an encoder and / or a decoder of the trained ML model as signal processing logic to a device that executes the signal processing logic to control, monitor, and / or analyze the physical asset.

[0051] The method may include providing (e.g., by storing or transporting) at least a decoder of the trained ML model as signal processing logic to a device that executes the signal processing logic to control, monitor, and / or analyze the physical asset.

[0052] The physical asset may be an asset of an electrical power system. The measurements associated with the physical asset may be or may include measurements of electrical properties associated with the physical asset.

[0053] The device may execute signal processing logic to control, monitor, and / or analyze the physical asset during field use of the asset in the power system, for example during operation of the asset in the power system.

[0054] In particular, by executing the signal processing logic, the device may process signals including measurements of electrical properties associated with a physical asset to control operation of the physical asset within the power system during field use. For example, the device may execute the signal processing logic to operate at least one circuit breaker, disconnector, and / or converter to perform preventive or corrective control action of the physical asset during field use of the asset within the power system.

[0055] Thus, an automated and improved method for controlling the operation of an asset based on measurements of physical input parameters can be provided.

[0056] The method can include separating the latent features into at least two groups.

[0057] The method may include adversarial decoder training. In adversarial decoder training, only a portion of the code may be provided to the adversarial decoder input layer of the adversarial decoder.

[0058] Separating the latent features into at least two groups can include processing the signal differently depending on whether the signal varies more rapidly or more slowly.

[0059] Separating the latent features into at least two groups can include feeding signals that are sampled at a lower rate than other signals (e.g., weather data) to both an input layer of the encoder and a decoder input layer of the decoder during ML model training.

[0060] The method may be performed by or using at least one Integrated Circuit (IC).

[0061] The method may be a computer-implemented method. The signal may be a time-series signal.

[0062] The additional signal may include a second latent feature contained in a second code generated using a second encoder.

[0063] The ML model can include a second encoder. Training the ML model may include training a second encoder.

[0064] Training the ML model may include performing adversarial training of the decoder and the second decoder.

[0065] The second decoder is operable to process the second decoder input to generate a second decoder output.

[0066] The second decoder output may include a second reconstruction of at least some of the signal.

[0067] Performing adversarial training can include providing only a subset of the latent features included in the code to the second decoder as second decoder inputs.

[0068] The subset of latent features can be selected based on the time scale over which the latent features change.

[0069] The subset may include latent features that are updated by the encoder less frequently than other latent features.

[0070] Alternatively or additionally, the subset of latent features may be selected based on the spatial scale over which the latent features vary.

[0071] Alternatively or additionally, the subset of potential features may be selected based on the logic stage to which the potential features are associated.

[0072] Training the ML model can include performing iterative training that includes several iterations.

[0073] Each of the several iterations may include training the encoder, the decoder, and the second encoder with the objective of minimizing a loss function (e.g., reconstruction loss or prediction loss) of the decoder and maximizing a loss function (e.g., reconstruction loss or prediction loss) of the second decoder, and training the second decoder with the objective of minimizing a loss function (e.g., reconstruction loss or prediction loss) of the second decoder, respectively.

[0074] The encoder can include an encoder output layer operative to output a first set of latent features at a first rate.

[0075] The second encoder can include a second encoder output layer operative to output a second set of latent features at a second rate different from the first rate.

[0076] The second rate may be less than the first rate such that the latent features in the second set are updated less frequently than the latent features in the first set.

[0077] The subset of latent features may include only latent features from the second set.

[0078] ML model training can include adjusting parameters of at least one Artificial Neural Network (ANN), such as weights and / or propagation functions.

[0079] ML model training may include tuning parameters of at least one artificial neural network (ANN) to minimize an objective function and / or maximize another objective function.

[0080] ML model training can include adjusting parameters of at least one recurrent neural network (RNN), such as weights and / or a propagation function.

[0081] ML model training may include tuning parameters of at least one recurrent neural network (RNN) to minimize an objective function and / or maximize another objective function.

[0082] ML model training can include adjusting parameters of at least one Gated Recurrent Unit (GRU) or Long-Short-Term Memory (LSTM) cell, such as weights and / or biases.

[0083] ML model training may include adjusting parameters of at least one gated recurrent unit (GRU) or long short-term memory (LSTM) cell, such as weights and / or biases, to minimize an objective function and / or maximize another objective function.

[0084] The objective function may depend on a loss function (such as reconstruction loss or prediction loss). The objective function may depend, for example, on a metric of the difference between the encoder input and the decoder output or on an entropy function.

[0085] Adversarial training can maximize or minimize a loss function that depends on the training signal, the decoder output, and the second decoder output.

[0086] The loss function may include a metric, an entropy measure, or a distance determined according to a distance measure of a probability distribution.

[0087] The adversarial training may be iterative training involving several iterations in which the decoder and the second decoder are trained sequentially in each of several iterations.

[0088] The additional signals may include variables that affect the operation of a physical asset, but exist independent of the presence or absence of the physical asset.

[0089] The additional signal may vary over a time scale that is slower than the operating time scale of the physical asset.

[0090] The additional signal may include a variable that is measured at a first sampling time that is longer than a second sampling time at which the measurement is sampled.

[0091] The first sampling time may be a sub-sampling of the second sampling time.

[0092] The additional signals may include ambient condition variables of the physical asset. The ambient condition variables may include weather data that affects at least one power system characteristic associated with the physical asset.

[0093] The decoder input may include replicas of at least some of the latent features of the second code produced by the second encoder.

[0094] The decoder input may include replicas of latent features of the second code that are produced by a second encoder and that vary over slower time scales than other latent features.

[0095] The code can contain N-dimensional tensors. N may be 3 or may be greater than 3.

[0096] The N-dimensional tensor may be defined in a space having a first time axis and a second time axis, the first time axis representing changes related to the physical asset over a first time scale and the second time axis representing changes related to the physical asset over a second time scale, the first time scale being different from the second time scale.

[0097] The final stage of the encoder can map the sequence into a vector. The final stage of the encoder may operate to update some vector elements of the vector more frequently than other vector elements of the vector.

[0098] The encoder can have an encoder output layer operative to output a first set of latent features at a first rate and to output a second set of latent features at a second rate different from the first rate.

[0099] The first rate may be an integer multiple of the second rate. Training the ML model can include replicating all of the latent features in the second set along a time axis.

[0100] Training the ML model can include generating decoder inputs during training that include combining (e.g., concatenating) replicas of latent features in the second set with features in the first set.

[0101] One or more of the latent variables may represent temporal characteristics related to the physical asset.

[0102] One or several latent variables of the trained ML model can represent temporal properties that vary over a first time scale that is longer than the sampling time interval of one, some, or all of the measurements.

[0103] The first time scale may be a regular or irregular sub-sampling of the sampling time interval.

[0104] The device can process one or several latent variables to perform asset health analysis of a physical asset.

[0105] One or several latent variables of a trained ML model can represent spatial and / or logical properties.

[0106] The method may further include training at least one additional encoder and / or decoder to design features for three or more different temporal, spatial, and / or logical domains.

[0107] The method may further include performing adversarial training of the adversarial decoder. The encoder of the ML model may comprise a first ANN. The first ANN may be operable to process a sequence of signals (such as a time series) into at least one vector. The first ANN may comprise a recurrent neural network (RNN), in particular a gated recurrent unit (GRU) and / or a long short-term memory (LSTM) cell. The encoder of the autoencoder may comprise the first ANN.

[0108] The decoder of the ML model may comprise a second ANN. If an adversarial (ie, second) decoder is used, it may comprise a third ANN.

[0109] The topologies of the first ANN, the second ANN, and, if present, the third ANN (e.g., RNN) may remain unchanged during ML model training.

[0110] Alternatively, the ML model training may include varying one or more hyperparameters that define the topology of at least one or some of the first ANN, the second ANN, and, if present, the third ANN.

[0111] The signals used to train the method may include past measurements. The signals used to train the method may include labeled historical measurement data.

[0112] Any one of the methods of the embodiments or aspects disclosed herein may be performed to commission an automation system for a utility or industrial system, for example, for an electrical power generation, transmission, and / or distribution system.

[0113] Any one of the methods of the embodiments or aspects disclosed herein may be performed to configure an automation system for a utility or industrial system, for example, for an electrical power generation, transmission, and / or distribution system.

[0114] The method of any one of the embodiments or aspects disclosed herein may be performed to update signal processing logic of equipment of an automation system for a utility or industrial system during field operation.

[0115] The method of any one of the embodiments or aspects disclosed herein may include monitoring an apparatus that controls, monitors, and / or analyzes a physical asset during field operation of the apparatus.

[0116] The method of any one of the embodiments or aspects disclosed herein may include selectively replacing logic executed by the device depending on the performance of the device during field operations.

[0117] The physical asset may be or may include at least one power transformer, disconnect switch, tap changer, converter (eg, AC / DC converter), and the like.

[0118] According to another aspect of the invention, a method for generating or updating signal processing logic for processing signals including measurements related to a physical asset is provided. The method includes training an ML model. The ML model comprises an encoder having an input layer for receiving an input signal, in particular a time series input signal. The encoder is operative to generate a code including latent features. The ML model comprises a decoder having an output layer. The decoder is operative to process a decoder input including the code to generate a decoder output that may include a reconstruction of at least some of the input signal. Training the ML model includes providing a training signal as an input signal to the input layer. Training the ML model includes separating the latent features into at least two groups, where one or some of the latent features included in the first group of latent features are associated with a first time scale and / or spatial scale and one or some of the latent features included in the second group of latent features are associated with a second time scale and / or spatial scale different from the first time scale and / or spatial scale.

[0119] The encoder can have an encoder output layer operative to output a first set of latent features at a first rate and to output a second set of latent features at a second rate different from the first rate.

[0120] The first rate may be an integer multiple of the second rate. Training the ML model can include replicating all of the latent features in the second set along a time axis.

[0121] Training the ML model can include generating decoder inputs during training that include combining (e.g., concatenating) replicas of latent features in the second set with features in the first set.

[0122] Separating the latent features into at least two groups can include adversarial decoder training.

[0123] In adversarial decoder training, only a portion of the code (e.g., only the latent features included in the second set of latent features) may be provided to an adversarial decoder input layer of the adversarial decoder.

[0124] Separating the latent features into at least two groups can include processing the signal differently depending on whether the signal varies more rapidly or more slowly.

[0125] Separating the latent features into at least two groups can include feeding signals that are sampled at a lower rate than other signals (e.g., weather data) to both an input layer of the encoder and a decoder input layer of the decoder during ML model training.

[0126] The method may include providing (e.g., by storing or transporting) at least an encoder of the trained ML model as signal processing logic to a device that executes the signal processing logic to control, monitor, and / or analyze the physical asset.

[0127] The physical asset may be an asset of an electrical power system. The measurements associated with the physical asset may be or may include measurements of electrical properties associated with the physical asset.

[0128] The device may execute signal processing logic to control, monitor, and / or analyze the physical asset during field use of the asset in the power system, for example during operation of the asset in the power system.

[0129] In particular, by executing the signal processing logic, the device may process signals including measurements of electrical properties associated with a physical asset to control operation of the physical asset within the power system during field use. For example, the device may execute the signal processing logic to operate at least one circuit breaker, disconnector, and / or converter to perform preventive or corrective control action of the physical asset during field use of the asset within the power system.

[0130] In this way, for example, an automated and improved method for controlling the operation of an asset based on measurements of physical input parameters can be provided.

[0131] The method may be performed by or using at least one integrated circuit (IC).

[0132] The method may be a computer-implemented method. The signal may be a time-series signal.

[0133] The additional signal may include a second code generated using a second encoder.

[0134] Training the ML model may include performing adversarial training of the decoder and the second decoder.

[0135] The second decoder is operable to process the second decoder input to generate a second decoder output.

[0136] The second autoencoder output may include a second reconstruction of at least some of the signal.

[0137] Performing adversarial training can include providing only a subset of the latent features included in the code to the second decoder as second decoder inputs.

[0138] The subset of latent features can be selected based on the time scale over which the latent features change.

[0139] Alternatively or additionally, the subset of latent features may be selected based on the spatial scale over which the latent features vary.

[0140] Alternatively or additionally, the subset of potential features may be selected based on the logic stage to which the potential features are associated.

[0141] Training the ML model can include performing iterative training that includes several iterations.

[0142] Each of the several iterations may include training the encoder, the decoder, and the second encoder with the objective of minimizing a loss function (e.g., reconstruction loss or prediction loss) of the decoder and maximizing a loss function (e.g., reconstruction loss or prediction loss) of the second decoder, and training the second decoder with the objective of minimizing a loss function (e.g., reconstruction loss or prediction loss) of the second decoder, respectively.

[0143] The encoder can include an encoder output layer operative to output a first set of latent features at a first rate.

[0144] The second encoding can include a second output layer operative to output a second set of latent features at a second rate different from the first rate.

[0145] The second rate may be less than the first rate such that the latent features in the second set are updated less frequently than the latent features in the first set, and the subset of the latent features includes only latent features from the second set.

[0146] ML model training may include adjusting parameters of at least one artificial neural network (ANN), such as weights and / or propagation functions.

[0147] ML model training may include tuning parameters of at least one artificial neural network (ANN) to minimize an objective function and / or maximize another objective function.

[0148] ML model training can include adjusting parameters of at least one recurrent neural network (RNN), such as weights and / or a propagation function.

[0149] ML model training may include tuning parameters of at least one recurrent neural network (RNN) to minimize an objective function and / or maximize another objective function.

[0150] ML model training may include adjusting parameters of at least one gated recurrent unit (GRU) or long short-term memory (LSTM) cell, such as weights and / or biases.

[0151] ML model training may include adjusting parameters of at least one gated recurrent unit (GRU) or long short-term memory (LSTM) cell, such as weights and / or biases, to minimize an objective function and / or maximize another objective function.

[0152] The objective function may depend on a loss function (such as reconstruction loss or prediction loss). The objective function may depend on the reconstruction loss, e.g., a metric of the difference between the autoencoder input and the autoencoder output, or an entropy function.

[0153] Adversarial training can maximize or minimize a loss function that depends on the training signal, the autoencoder output, and the second autoencoder output.

[0154] The loss function may include a metric, an entropy measure, or a distance determined according to a distance measure of a probability distribution.

[0155] The adversarial training may be iterative training involving several iterations in which the decoder and the second decoder are trained sequentially in each of several iterations.

[0156] The additional signals may include variables that affect the operation of a physical asset, but exist independent of the presence or absence of the physical asset.

[0157] The additional signal may vary over a time scale that is slower than the operating time scale of the physical asset.

[0158] The additional signal may include a variable that is measured at a first sampling time that is longer than a second sampling time at which the measurement is sampled.

[0159] The first sampling time may be a sub-sampling of the second sampling time.

[0160] The additional signals may include ambient condition variables of the physical asset. The ambient condition variables may include weather data that affects at least one power system characteristic associated with the physical asset.

[0161] The decoder input may contain replicas of at least some of the latent features of the code.

[0162] The decoder input may contain replicas of latent features that change over slower time scales than other latent features.

[0163] The code can contain N-dimensional tensors. N may be 3 or may be greater than 3.

[0164] The N-dimensional tensor may be defined in a space having a first time axis and a second time axis, the first time axis representing changes related to the physical asset over a first time scale and the second time axis representing changes related to the physical asset over a second time scale, the first time scale being different from the second time scale.

[0165] The final stage of the encoder can map the sequence into a vector. One or more of the latent variables may represent temporal characteristics related to the physical asset.

[0166] One or several latent variables of the trained ML model can represent temporal properties that vary over a first time scale that is longer than the sampling time interval of one, some, or all of the measurements.

[0167] The first time scale may be a sub-sampling of the sampling time interval. The device can process one or several latent variables to perform asset health analysis of a physical asset.

[0168] One or several latent variables of a trained ML model can represent spatial and / or logical properties.

[0169] The method may further include training at least one additional encoder and at least one additional decoder to design features for three or more different temporal, spatial, and / or logical domains.

[0170] The method may further include performing adversarial training of the adversarial decoder stage. The encoder of the ML model may comprise a first ANN. The first ANN may be operable to process a series of signals (such as a time series) into at least one vector. The first ANN may comprise a recurrent neural network (RNN), in particular a gated recurrent unit (GRU) and / or a long short-term memory (LSTM) cell. The encoder of the ML model may comprise a first ANN.

[0171] The decoder of the ML model may comprise a second ANN. If an adversarial decoder is used, it may comprise a third ANN.

[0172] The topologies of the first ANN, the second ANN, and, if present, the third ANN, may remain unchanged during ML model training.

[0173] Alternatively, the ML model training may include varying one or more hyperparameters that define the topology of at least one or some of the first ANN, the second ANN, and, if present, the third ANN.

[0174] The signals used to train the method may include past measurements. The signals used to train the method may include labeled historical measurement data.

[0175] The method may be performed for commissioning an automation system for a utility or industrial system, for example, for an electrical power generation, transmission, and / or distribution system.

[0176] The method may be performed to configure automation systems for utilities or industrial systems, for example, for power generation, transmission, and / or distribution systems.

[0177] The method may be performed to update signal processing logic of equipment of an automation system for a utility or industrial system during field operation.

[0178] The method may include monitoring an apparatus that controls, monitors, and / or analyzes a physical asset during field operation of the apparatus.

[0179] The method may include selectively replacing logic executed by the device depending on the performance of the device during field operations.

[0180] The physical asset may be or may include at least one power transformer, disconnect switch, tap changer, converter (eg, AC / DC converter), and the like.

[0181] A method for controlling, analyzing, and / or monitoring a physical asset includes executing, by the control, analysis, and / or monitoring device, an encoder of the trained ML model (e.g., an autoencoder) as at least part of signal processing logic for processing signals, in particular time series signals including measurements, and / or executing, by the control, analysis, and / or monitoring device, a decoder of the trained ML model (e.g., an autoencoder) as at least part of the signal processing logic.

[0182] The physical asset may be an asset of an electrical power system. The measurements may be or include measurements of electrical properties associated with the physical asset.

[0183] By executing the signal processing logic, signals including measurements may be processed to perform control, monitoring, and / or analytical functions on the physical asset during field use of the asset within the power system, e.g., during operation of the asset within the power system.

[0184] In particular, executing the signal processing logic may process signals including measurements of electrical properties associated with the physical assets to perform control over operation of the physical assets in the power system during field use. For example, the signals may be processed to perform preventive or corrective control actions of the physical assets during field use of the assets in the power system by operating at least one circuit breaker, disconnector, and / or converter.

[0185] In this manner, for example, an automated and improved method for exercising control over the operation of an asset based on measurements of physical input parameters can be provided.

[0186] The signal processing logic can process the signal into latent features, which can include latent features associated with at least two different time, space, and / or logical domains.

[0187] The method may include performing at least one control, analysis, and / or monitoring action based on at least a subset of the latent characteristics.

[0188] The method may include performing at least one control, analysis, and / or monitoring action based on the subset of the potential features, independent of the remaining features generated by the signal processing logic.

[0189] The signal processing logic may be signal processing logic generated using a method according to an embodiment.

[0190] The method may include performing, by a control, analysis, and / or monitoring device, at least one control action based on one or more of the latent characteristics.

[0191] The at least one control action may include one or more of: triggering a protection function, controlling an interface to perform an output action, controlling at least one operating parameter of a physical asset, and controlling at least one operating parameter of a power system asset other than the physical asset.

[0192] The physical asset may be a power system asset. The method may be a method of monitoring asset health of at least one power system asset, such as at least one power transformer, disconnect switch, tap changer, converter (eg, AC / DC converter), etc.

[0193] The method may be a method of controlling at least one power system asset, such as at least one power transformer, disconnect switch, tap changer, converter (eg, AC / DC converter), or the like.

[0194] A method for automatically designing features for controlling, analyzing, and / or monitoring a physical asset is provided, the method including training an ML model comprising an encoder and a decoder, and processing signals using at least the encoder of the trained ML model into potential features for controlling, analyzing, and / or monitoring the physical asset.

[0195] The encoder may be generated by a method for generating signal processing logic according to an embodiment.

[0196] According to one aspect of the present invention, there is provided a computing system for generating or updating signal processing logic for processing signals including measurements related to a physical asset, the computing system comprising at least one integrated circuit operative to train an ML model using any one of the techniques disclosed herein.

[0197] The physical asset control, analysis and / or monitoring device comprises an interface for receiving signals, in particular time series signals, comprising measurements related to the physical asset. The device may comprise at least one integrated circuit, the at least one integrated circuit being operative to process the signals using an encoder (e.g., an autoencoder), the encoder having an input layer for receiving the signals, and operable to generate a code comprising latent features, and to perform at least one control action based on one or several latent features. Alternatively or additionally, the device may comprise at least one integrated circuit being operative to process the signals using a decoder (e.g., an autoencoder), and to perform at least one control action based on the signal processing performed using the decoder.

[0198] The signal processing logic may be generated by a method according to any embodiment disclosed herein.

[0199] The physical asset may be an asset of an electrical power system. The measurements may be or include measurements of electrical properties associated with the physical asset.

[0200] Performing the at least one control action may include performing the at least one control action during field use of a physical asset in the power system, for example during operation of the physical asset in the power system.

[0201] In particular, the measurements may be or may include measurements of electrical properties associated with the physical asset, and performing the at least one control action may include performing at least one control action to control operation of the physical asset in the power system during field use. Thus, for example, at least a preventive or corrective control action of the physical asset may be performed during field use of the asset in the power system by operating at least one circuit breaker, disconnector, and / or converter.

[0202] In this way, for example, an automated and improved apparatus can be provided for exercising control over the operation of an asset based on measurements of physical input parameters.

[0203] The encoder may be or may include a first and / or second encoder of an ML model trained in a manner that generates signal processing logic.

[0204] The decoder may be or may include a first and / or second decoder of an ML model trained in the method for generating signal processing logic.

[0205] The asset may be a power system asset. The power system comprises a physical asset, at least one measurement device operative to capture signals associated with the physical asset, and control, analysis, and / or monitoring device for processing the captured signals.

[0206] The captured signals may include measurements of electrical properties associated with the physical asset. The captured signals may be processed to control, monitor, and / or analyze the physical asset during field use of the asset in the power system, for example during operation of the asset in the power system.

[0207] In particular, signals including measurements of electrical properties associated with a physical asset may be processed to control operation of the physical asset in an electric power system during field use. For example, the captured signals may perform preventive or corrective control actions of the physical asset during field use of the asset in an electric power system by operating at least one circuit breaker, disconnector, and / or converter.

[0208] In this way, for example, a system can be provided that allows automated and improved control over the operation of an asset based on measurements of physical input parameters.

[0209] The control, analysis, and / or monitoring devices may be Intelligent Electronic Devices (IEDs).

[0210] The physical asset may be or may include at least one power transformer, disconnect switch, tap changer, converter (eg, AC / DC converter), and the like.

[0211] Various effects and advantages are achieved by the methods, systems, and devices according to the embodiments. As already mentioned several times, the embodiments may provide methods, devices, programs, and systems that enable automated and improved control over the operation of a physical asset based on measurements of physical input parameters. The signal processing logic may be automatically generated or updated. The signal processing logic may be executed by an intelligent electronic device (IED) to automatically make decisions that may include triggering control measures, outputting monitoring reports, and / or outputting analysis results (such as alarms). The signal processing logic may provide potential features that are different from other potential features in terms of time and / or logical domains (e.g., vary more slowly than the sampling time of the measurements processed by the signal processing logic).

[0212] BRIEF DESCRIPTION OF THE DRAWINGS The subject matter of the present invention will now be explained in more detail with reference to preferred exemplary embodiments illustrated in the accompanying drawings. [Brief description of the drawings]

[0213] [Figure 1] FIG. 1 is a schematic diagram of a system comprising an apparatus operative to control, monitor, and / or analyze power system assets and a computing system for generating signal processing logic. [Diagram 2] FIG. 1 is a schematic diagram of a system comprising an apparatus operative to control, monitor, and / or analyze power system assets and a computing system for generating signal processing logic. [Diagram 3] 1 is a flow chart of a method. [Figure 4] 1 is a flow chart of a method. [Diagram 5] FIG. 1 is a block diagram of an autoencoder. [Figure 6] 1 is a flow chart of a method. [Figure 7] FIG. 1 is a block diagram of an autoencoder system with an adversarial decoder. [Figure 8] FIG. 1 is a block diagram of an autoencoder. [Figure 9] FIG. 1 is a block diagram of an autoencoder system with an adversarial decoder. [Figure 10] 1 illustrates the operation of a method, apparatus, and system. [Figure 11] FIG. 2 illustrates a portion of the operation of an autoencoder. [Figure 12] FIG. 2 illustrates a portion of the operation of an autoencoder. [Figure 13] FIG. 2 illustrates a portion of the operation of an autoencoder. [Figure 14] FIG. 2 illustrates a portion of the operation of an autoencoder. [Figure 15] FIG. 2 illustrates a portion of the operation of an autoencoder. [Figure 16]FIG. 1 is a block diagram of an autoencoder to explain autoencoder training. [Figure 17] FIG. 1 is a block diagram of an autoencoder to explain autoencoder training. [Figure 18] FIG. 1 is a block diagram of an autoencoder system to illustrate autoencoder training. [Figure 19] FIG. 1 is a block diagram of an autoencoder system to illustrate autoencoder training. [Figure 20] FIG. 13 illustrates the time evolution of latent features determined by signal processing logic generated according to one embodiment. [Figure 21] FIG. 1 illustrates the time evolution of latent features determined by untrained conventional signal processing logic, according to one embodiment. [Figure 22] FIG. 1 is a block diagram of an autoencoder system to illustrate autoencoder training. [Figure 23] FIG. 1 is a block diagram of an autoencoder system to illustrate autoencoder training. [Figure 24] FIG. 2 is a schematic diagram of the signal processing logic. [Figure 25] FIG. 1 is a block diagram of a system comprising an apparatus operative to control, monitor, and / or analyze power system assets and a computing system for updating signal processing logic. [Figure 26] FIG. 2 is a schematic diagram of an encoder. [Figure 27] FIG. 2 is a schematic diagram of an encoder. [Figure 28] 1 is a flow chart of a method. [Figure 29] 1 is a flow chart of a method. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0214] Detailed Description of the Preferred Embodiments Exemplary embodiments of the invention will be described with reference to the drawings, in which like or similar reference numerals indicate like or similar elements. Although some embodiments are described in the context of distance or time domain protection of an electrical distribution or transmission system, the methods and apparatus described in detail below can be used in a wide variety of systems.

[0215] Features of the embodiments can be combined with each other unless otherwise stated. Although some embodiments are described in the context of signal processing logic executed by a monitoring device for asset health monitoring, the embodiments are not so limited. The embodiments of the invention can be used in general for systems requiring the features for which they are designed. For example, as mentioned several times in the previous section, the signal processing logic can be used to process signals including, for example, measurements of electrical properties associated with physical assets of a power system to control, monitor, and / or analyze the assets during their field use in the power system. Thus, in particular, control over the operation of an asset during field use can be based on signals including, for example, measurements of physical input parameters. Thus, as one example, preventive or corrective control actions can be performed on a physical asset, for example, by operating at least one circuit breaker, disconnector, and / or converter.

[0216] The systems, methods, and apparatus address challenges related to processing signals (such as time series signals) to generate features on which decisions about a physical asset can be based. The decisions can be, but are not limited to, triggering corrective or protective actions related to the physical asset and / or determining the health of the asset.

[0217] As used herein, the term "signal" is not limited to signals that are continuously received during field use. A "signal" may be or include online signals that are processed during live operation of the asset being controlled, monitored, and / or analyzed, and / or signals that may have been previously recorded and stored in a storage system. This applies to both signals that are used in generating signal processing logic and signals that are used to perform actions in field use.

[0218] As used herein, an ML model may also be referred to as, without limitation, an autoencoder.

[0219] Recurrent neural networks (RNNs) can be trained, for example, to classify samples in one or several data series. Identifying (or designing) features from input signals that include one or several measurements related to physical assets remains a significant challenge.

[0220] Embodiments of the present invention provide techniques that allow for automatically or semi-automatically generating or updating signal processing logic executed by at least one integrated circuit for operating, monitoring, and / or analyzing a physical asset in response to input signals that include measurements of electrical or other characteristics associated with the physical asset.

[0221] An embodiment of the invention utilizes ML model (e.g., autoencoder) training for feature engineering. At least the encoder of the trained ML model can be used as signal processing logic during field operation of a physical asset. The ML model training can be performed in a manner that ensures that latent features included in the code output by the encoder of the autoencoder include a first latent feature associated with a first (spatial, temporal, spatiotemporal, logical or other) characteristic scale and a second feature associated with a second (spatial, temporal, spatiotemporal, logical or other) characteristic scale that is different from the first characteristic scale.

[0222] In some implementations, the ML model training is performed such that the latent features included in the code output by the encoder of the autoencoder include one or several first latent features that vary over a first time scale and one or several second latent features that vary over a second time scale different from the first time scale. The second time scale may correspond to a sampling interval in which one or several measurements included in the signal input to the autoencoder are samples. The first time scale may correspond to a regular or irregular sub-sampling of the second time scale. That is, the first time scale may be associated with slower changes than the second time scale.

[0223] In some implementations, the ML model training is performed such that the latent features included in the code output by the encoder of the autoencoder include one or more first latent features associated with a first spatial domain and one or more second latent features associated with a second spatial domain different from the first spatial domain.

[0224] In the techniques disclosed herein, a particular ML model architecture may be used. The ML model may include a first encoder that may include a first encoder output layer that operates to output a first set of latent features that are updated at a first rate. The ML model may include a second encoder that may include a second encoder output layer that may operate to output a second set of latent features that are updated at a second rate different from the first rate. The second rate may be less than the first rate such that the latent features included in the second set are updated less frequently than the latent features included in the first set.

[0225] During autoencoder training, information can be "forced" into the latent features contained in the second set. This can be done in a variety of ways, as disclosed herein.

[0226] Ensuring the desired separation of latent features according to time scale, spatial domain, and / or logical grouping can be implemented in various ways. In some implementations, adversarial decoder training can be performed. The adversarial decoder and the decoder of the ML model can be trained in an iterative procedure in which only a subset of the (e.g., more slowly varying) latent features in the encoder's code are provided to the adversarial decoder. Alternatively or additionally, additional signals known to vary more slowly than the characteristic operating time scale of the physical asset may be provided to the decoder of the ML model during ML model training.

[0227] Historical data obtained by measurement of a physical asset for which the signal processing logic is generated can be used as part of the training data when generating the signal processing logic. Alternatively or additionally, historical data obtained by measurement of assets similar to the physical asset for which the signal processing logic is generated (e.g., assets with the same or similar specifications and the same or similar deployment scenarios) can be used as part of the training data when generating the signal processing logic.

[0228] The signal processing logic may be used to perform control, monitoring, and / or analysis functions during field use of the physical asset. Asset health monitoring and outputting the results of the asset health monitoring is one example of such an operation that may be performed by the signal processing logic. Executing the signal processing logic to perform a preventive or corrective control action is another example. The preventive or corrective control action may include operating at least one circuit breaker, disconnector, transducer, etc.

[0229] 1 and 2 show a power system 10 comprising a device 40, 80 having one or several integrated circuits (ICs) 41, 81 operative to execute a signal processing logic 42, 82. The signal processing logic 42, 82 receives signals as inputs and processes the signals into one or several characteristics. The signal processing logic 42, 82 performs control, monitoring, and / or analysis operations in response to signals, including signals from one or several measurement devices 31, 32, 33, collectively referred to as a measurement system 30. The signal processing logic 42, 82 may perform control, monitoring, and / or analysis functions during field use of physical assets. The signal processing logic 42, 82 may perform asset health monitoring and output results of the asset health monitoring. Alternatively or additionally, the signal processing logic 42, 82 may perform preventive or corrective control actions. The preventive or corrective control actions may include operating at least one breaker, disconnector, converter, or the like.

[0230] The ICs 41, 81 may be or comprise any one or any combination of a processor, microprocessor, controller, microcontroller, Application Specific Integrated Circuit (ASIC), Field-Programmable Gate Array (FPGA), or other integrated circuits.

[0231] Measurement system 30 may include current and / or voltage transformers, temperature sensors, circuit breaker position sensors, other sensors installed on or associated with circuit breakers, sensors installed on or associated with power transformers (e.g., insulating fluid temperature sensors, Dissolved Gas Analysis (DGA) sensors, etc.), sensors installed on or associated with load tap changers, and / or sensors installed on or associated with converters (such as AC / DC converters).

[0232] The system 10 comprises one or several physical assets. The physical assets may be or may include power system assets. The power system may include, but is not limited to, microgrids, distributed energy resources (DERs), renewable energy resources, or other assets. For example, the system 10 may comprise power generation assets 11-13 that may be or may include renewable energy generation assets. The system 10 may comprise controllable loads 14, which may include one or several controllable loads 15, 16. The system 10 may comprise one or several controllable generators 18, 19 that consume fossil fuels to generate power. The signal processing logic 42 may be operative to execute signal processing logic that may include at least an encoder of a trained ML model as described in more detail herein to perform one or several of asset monitoring, asset control, and / or asset analysis for one or several of the energy generation systems 11-13, the generators 18, 19, the loads 15, 16, and / or other devices deployed in the system 10.

[0233] The techniques disclosed herein are not limited to microgrids or DERs, but can be used for a wide variety of different power systems or other systems involving system automation. For example, as shown in FIG. 2, a system 70 can include a power system having a line 71 connected to a bus 72 via a power transformer 74. A circuit breaker (CB) 75 may be provided. A protection device 80 (which may be an intelligent electronic device (IED)) can process measurements detected during field operation by one or several detection devices 31, 32 using a signal processing logic 84, which can include at least an encoder of a trained ML model, trained as described in more detail herein. The signal processing logic 82 can generate an output that can be used to control a control device (such as the CB 75) of the power system 70 to perform preventive and / or corrective control actions. A control command output via an interface 83 can be triggered based on the output of the signal processing logic 82. Alternatively or additionally, the signal processing logic 82 may generate outputs that can be used to display alarms, warnings, asset health information, or other information about the power system determined by the signal processing logic 82 via a Human Machine Interface (HMI) 93. The HMI 93 may be provided in a control center 90 operative to receive the output of the signal processing logic 82. The control center 90 may be a national or regional control center, or a control center dedicated to the microgrid. The control center 90 may include one or several ICs 91 operative to process the output of the signal processing logic 82 and control the HMI based thereon.

[0234] The systems 10, 70 include a signal processing logic generation computing system 50. The computing system 50 generally operates to perform ML model training to generate the signal processing logic 42, 82. The computing system 50 may perform the ML model training such that the latent features included in the code output by the encoder of the trained ML model include different sets associated with different time, space, spatiotemporal, and / or logical domains. An architecture of the ML model may be used in which some latent features are updated less frequently than other features (e.g., due to the architecture used in the ML model).

[0235] The computing system 50 can perform ML model (e.g., autoencoder) training in a manner that forces the latent features to include one or several first latent features associated with a first time scale and one or several latent features associated with a second time scale different from the first time scale. Alternatively or additionally, the computing system 50 can perform ML model training in a manner that forces the latent features to include one or several first latent features associated with a first spatial domain size and one or several latent features associated with a second spatial domain size different from the first spatial domain size. Alternatively or additionally, the computing system 50 can perform ML model training in a manner that forces the latent features to include one or several first latent features associated with a first spatio-temporal domain and one or several latent features associated with a second spatio-temporal domain different from the first spatio-temporal domain. Alternatively or additionally, computing system 50 may perform ML model training in a manner that forces the latent features to include one or more first latent features associated with a first logical domain and one or more latent features associated with a second logical domain that is different from the first logical domain. Various techniques that may be used by computing system 50 are described in more detail with reference to Figures 3-29 below.

[0236] The computing system 50 operates to perform signal processing logic generation to generate and / or update the signal processing logic 42, 82. The signal processing logic generation includes performing ML model training. The signal processing logic deployed to the IEDs 40, 80 for execution may be, or may be based at least in part on, at least the encoder and decoder of the trained ML model, and may optionally be based on both the encoder and decoder of the trained ML model.

[0237] The computing system 50 comprises one or several ICs 51 for performing ML model training and optional further post-tuning steps that can be used to generate signal processing logic from the trained ML model. The ICs 51 can be or comprise any one or any combination of a processor, microprocessor, controller, microcontroller, application specific integrated circuit (ASIC), field programmable gate array (FPGA), or other integrated circuits.

[0238] The computing system 50 may include storage 52, which may include volatile and / or non-volatile storage components. The storage 52 may store data 54 that defines possible ML model architectures. The data 54 may define different candidate neural networks or other ML architectures, one or several of which may be trained in the ML model training.

[0239] The computing system 50 can use training and validation data sets 53, 60 that are used to train the ML model and perform validation. Some or all of the training and validation data sets 53, 60 may be stored locally in the storage device 52. Alternatively or additionally, some or all of the training and validation data sets may be retrieved from an external data repository. The training and / or validation data may include measurements performed on the particular system for which the signal processing logic is to be generated (which may be particularly suitable for improving the signal processing logic during field use) and / or physical assets of similar or the same structure and / or function as used in the system 10, 70.

[0240] Computing system 50 may be or comprise one or several computers, one or several servers, or other computing devices.

[0241] The computing system 50 operates to generate or update the signal processing logic 42, 82 for processing the signals. The signals may be time-series signals. The signals include measurements related to the physical assets of the system 10, 70.

[0242] The computing system 50 is operative to train an ML model comprising an encoder having an input layer for receiving an input signal, in particular a time series input signal, the encoder being operative to generate a code including latent features.

[0243] The ML model includes a decoder having an output layer, the decoder operative to process a decoder input that includes code for generating an autoencoder output that includes a reconstruction of at least some of the input signal.

[0244] The computing system 50 is operative such that training the ML model includes providing a training signal as an input signal to the input layer.

[0245] The computing system 50 may be operated such that during training, the decoder input is generated from the code and an additional signal that is different from the code.

[0246] The computing system 50 may be operable to provide at least an encoder and / or a decoder of the trained ML model as at least part of the signal processing logic to a device 40, 80 that executes the signal processing logic to control, monitor and / or analyze the physical asset. The computing system 50 may have an interface 55 for outputting at least the encoder of the trained ML model to the device 40, 80 for execution.

[0247] 3 is a flow chart of the method 100. The method 100 may be performed automatically by the computing system 50 to generate the signal processing logic.

[0248] In step 101, ML model training is performed. The ML model training may be performed by the computing system 50. The ML model training may be performed using measurement data including measurements obtained from a system including a physical asset for which the signal processing logic is determined. The ML model training may be performed in a manner that implements code generated by an encoder of an autoencoder to include one or several first latent features associated with a first temporal, spatial, logical, or spatiotemporal scale and one or several second latent features associated with a second temporal, spatial, logical, or spatiotemporal scale different from the first temporal, spatial, logical, or spatiotemporal scale.

[0249] The ML model can include an encoder having an output layer that outputs or updates first latent features at a first rate and second latent features at a second rate different from the first rate.

[0250] Step 101 may include using adversarial decoder training to separate the latent features into at least two groups (e.g., by ensuring that relevant information is forced to latent features that are updated less frequently). In adversarial decoder training, only a portion of the code may be provided to an adversarial decoder input layer of the adversarial decoder.

[0251] Alternatively or additionally, step 101 may include separating the latent features into at least two groups (e.g., by ensuring that relevant information is forced towards latent features that are updated less frequently) by processing the signal differently depending on whether the signal changes more rapidly or more slowly.

[0252] Alternatively or additionally, step 101 may include separating the latent features into at least two groups (e.g., by ensuring that relevant information is forced towards latent features that are updated less frequently) during ML model training by feeding signals that are sampled at a lower rate than other signals (e.g., weather data) to both the input layer of the encoder and the decoder input layer of the decoder.

[0253] At step 102, validation is performed. Validation may include verifying that the distance between the encoder input (e.g., an input time series) and the decoder output (e.g., a reconstruction of the input generated by the decoder from the code provided by the encoder) meets a quality criterion. The quality criterion may be a threshold criterion. Validation may be performed using a portion of the training and validation data that was not used for training.

[0254] In step 103, at least a portion (e.g., the encoder and / or decoder) of the trained ML model may be stored in a device (e.g., an IED). The device may execute the encoder to identify characteristics of the asset associated with a first temporal, spatial, logical, or spatio-temporal scale that is different from a second temporal, spatial, logical, or spatio-temporal scale. The features grouped to correspond to different time scales for the ML model training may be used, for example, to perform at least one control, monitoring, and / or analytical function during field use of the physical asset.

[0255] The device may execute a decoder for anomaly detection and / or further analysis. For example, the decoder may operate on signals (such as asset health indicators) captured during field use of the asset to estimate a loss function (such as reconstruction loss or prediction loss). This information may be used, optionally in combination with loss function values ​​(e.g., reconstruction loss) determined during ML model training, for asset analysis.

[0256] Thus, ML model training can be used for feature design, where at least some of the features are used during field use to perform at least one control, monitoring, and / or analytical function during field use of the physical asset.

[0257] 4 is a flow chart of method 110. Method 110 may be performed automatically by apparatus 40, 80 by executing signal processing logic based on the trained ML model. The signal processing logic may be or may include at least an encoder of the trained ML model.

[0258] In step 111, a time series signal is received. The time series signal may be received from a field device (such as an IED, a gateway, and / or a merging unit). The time series signal may be obtained from a storage system.

[0259] The time series signal may include or depend on measurements obtained from the physical asset upon which the control, monitoring, and / or analysis operations are performed, or from components coupled (e.g., electrically coupled) thereto.

[0260] The time series signal may include voltage and / or current measurements, which may be sampled at sampling intervals.

[0261] In step 112, the time series signal may be processed using signal processing logic. The time series signal may be input to an input layer of an encoder of an ML model trained as described in relation to step 101. The output layer of the encoder may provide one or more first latent features associated with a first temporal, spatial, logical, or spatio-temporal scale, as determined from the input signal, and one or more second latent features associated with a second temporal, spatial, logical, or spatio-temporal scale different from the first temporal, spatial, logical, or spatio-temporal scale.

[0262] In step 113, a control, monitoring, and / or analysis action is performed. The control, monitoring, and / or analysis action may be based on at least one of the latent features in the code. Alternatively or additionally, the action may be determined by executing a decoder of the trained ML model. The analysis action may be performed based on one or more first latent features associated with a first time scale that is longer than the sampling interval and / or longer than a second time scale to which at least a different one of the latent features is associated. The at least one control action may include one or more of triggering a protection function, controlling an interface to perform an output action, controlling at least one operating parameter of a physical asset, controlling at least one operating parameter of a power system asset different from the physical asset.

[0263] FIG. 5 is a schematic diagram of an ML model 120 that may be used in the methods, apparatus, and systems disclosed herein. The ML model 120 includes an encoder 121 and a decoder 122. The encoder 121 may include one or several machine learning (ML) models, such as, but not limited to, an artificial neural network (ANN), in particular a recurrent neural network (RNN). The decoder 122 may include one or several other machine learning (ML) models, such as, but not limited to, an artificial neural network (ANN), in particular a recurrent neural network (RNN). The ANN of the decoder 122 may be distinguished from the ANN of the encoder 121 in some input nodes. The ANN of the decoder 122 may be distinguished from the ANN of the encoder 121 in some output nodes. The ANN of the decoder 122 may be distinguished from the ANN of the encoder 121 in some hidden layers. The ANN of the decoder 122 may be distinguished from the ANN of the encoder 121 in the activation functions used in the ANN. If an RNN including gated recurrent units (GRUs) and / or long short-term memory (LSTM) cells is used, the weights and / or activation functions and / or biases may be different in the decoder 122 and the encoder 121.

[0264] The encoder 121 is operable to process an input signal 131 into a code that includes a number of latent features 132. The number of latent features is less than the number of input nodes of the encoder 121. That is, the encoder 121 is capable of "condensing" the information contained in the input signal (which may be an N-dimensional vector) into K latent features, where K <Nである。

[0265] The decoder 122 is operable to process a decoder input into a decoder output, which may be a reconstruction 133 of the input signal. The decoder input may include some or all of the latent features 132 contained in the code generated by the encoder 121.

[0266] The architecture (e.g., topology) of the ML model may be such that several encoders operate to output different sets of latent features, each including one or several latent features. FIG. 26 shows an example of an ML model 250 comprising a first encoder (which may be trained using the methods disclosed herein to generate signal processing logic and / or may be executed as signal processing logic after training) and a second encoder (which may be trained simultaneously with the first encoder). The ML model 250 has an architecture 252 comprising a first encoder that processes an input 251 into a first set of one or several first latent features 253, and a second encoder that processes the input 251 into a second set of one or several second latent features 255. The first and second encoders are configured such that the first set 253 is output or updated more frequently than the second set 255.

[0267] For purposes of training, and the techniques described herein, the latent features included in second set 255 are guaranteed to depend on input 251 to provide relevant information about features, e.g., features that vary slowly over time scales (such as may be relevant for asset health assessment).

[0268] FIG. 27 illustrates an example of GRU cells 256, 257 included in an encoder (which may be trained using methods disclosed herein to generate signal processing logic and / or may be executed as signal processing logic after training). The GRU cells 256, 257 are configured such that an input 251 is processed into a first set 253 of one or several latent features and into a second set 255 of one or several latent features. The ML model 250 is configured such that the first set 253 is output or updated more frequently than the second set 255. Due to the training, and techniques described herein, the latent features included in the second set 255 are guaranteed to depend on the input 251 to provide relevant information about features, e.g., features that vary slowly over time scales (such as may be relevant for asset health assessment).

[0269] It will be appreciated that all parameters of the encoder (i.e., both those associated with the part of the architecture providing the first set of latent features 253 and those associated with another part of the architecture providing the second set of latent features 255) are updated in a training procedure, which may be an iterative training procedure.

[0270] During ML model training, latent features in the second set (i.e., latent features that are updated less frequently) may be replicated along the time axis and combined (e.g., concatenated) with latent features in the first set (i.e., latent features that are updated more frequently).

[0271] In ML model training, the ML model 120 is trained with the goal of minimizing a loss function. The loss function may quantify the difference between the signal 131 received at the input layer of the encoder 121 and the output 133 provided by the output layer of the decoder 122.

[0272] As described in more detail below, the number of input nodes of decoder 122 may differ from the number of output nodes of encoder 121 when autoencoder 122 is trained. An additional signal may be provided to decoder 122 during training. The additional signal may be selected, for example, such that one or some of latent features 132 are associated with a different temporal, spatial, spatiotemporal, and / or logical scale than at least another one or some others of latent features 132, as described in more detail herein. The additional signal may be provided by a second encoder of the ML model.

[0273] The autoencoder 120 is operable to receive and process records of operational data from one or more assets. The operational data constitutes time series data.

[0274]

number

[0275] where x t x is an N-dimensional vector and can have real- or complex-valued vector elements. T denotes the number of time points. t is an index that specifies the time points. The time points may be regularly or irregularly spaced. t can include N real-valued or complex-valued signals. The input signals may also be of a different type, e.g., categorical. It will be appreciated that for use with an ANN, such data may be converted to real-valued data before further processing (e.g., categorical data may be represented by the well-known one-hot encoding). Thus, the discussion may be limited to real-valued signals (it being understood that an N-dimensional complex-valued vector may be mapped to a 2N-dimensional real-valued vector).

[0276] The operational data may contain missing values ​​that can be handled by known techniques. In some cases, any vector x containing missing values ​​may be tcan be removed. If present, the missing values ​​can be replaced using other techniques such as interpolation between adjacent vectors, hard value imputation, etc. Missing data replacement can be performed using a machine learning (ML) model trained to replace missing data, for example as disclosed in WO 2019 / 141856.

[0277] The operational data is typically recorded on a time scale referred to herein as the second time scale. This time scale may also be referred to as the operational time scale. The second time scale may relate to a sampling interval that may be smaller (meaning that slower trends occur than the first time scale described below). For example, the operational data may be recorded on a time scale (e.g., defined by an average sampling interval) of 1 hour or less, 2 hours or less, 3 hours or less, 5 hours or less, 10 hours or less, 1 minute or less, 2 minutes or less, 3 minutes or less, 5 minutes or less, 10 minutes or less, 1 second or less, 2 seconds or less, 3 seconds or less, 5 seconds or less, 10 seconds or less, or sub-second, e.g., 1 second or less, 800 milliseconds or less, 500 milliseconds or less, 300 milliseconds or less, 200 milliseconds or less, 100 milliseconds or less.

[0278] The methods, apparatus, and systems disclosed herein are useful for determining the evolution of an asset (e.g., its health state) over a time scale (called the first time scale) that is larger relative to the operational time scale, such as 1 day or more, 2 days or more, 3 days or more, 5 days or more, 10 days or more, 1 week or more, 2 weeks or more, 3 weeks or more, 5 weeks or more, 10 weeks or more, 1 month or more, 2 months or more, 3 months or more, 5 months or more, 10 months or more, 1 year or more, 2 years or more, 3 years or more, 5 years or more, 10 years or more, etc. For example, gradual evolutions such as aging, gradual deterioration, wear, leakage, etc. can be observed in the slow first time scale. However, considering only raw data, this is a challenging problem, since effects in the operational time scale are superimposed on the slow time scale and their changes are usually much larger than those in the slow time scale.

[0279] To facilitate analysis of the evolution of an asset's operating characteristics at slow time scales, methods, apparatus, and systems provide automatic feature design designed to extract features at a slow first time scale. This is done in a manner that enhances (e.g., maximizes) the importance of the engineered features. For example, with reference to Figures 26 and 27, while the encoder architecture and the number of potential features included in the first and second sets 253, 255 may be fixed, the techniques disclosed herein define how the potential features included in the second set 255 (which may be particularly useful for identifying slowly changing trends) depend on the input 251.

[0280] Methods The methods, apparatus and systems are based on an ML model, which may be or include an ANN-implemented autoencoder.

[0281] Alternatively, or in addition to, feature design for time scale separation, the methods, apparatus, and systems may also operate to perform automated design of multiple sets of features corresponding to different logical portions of the system, thus achieving, for example, spatial separation or separation by different subsystems, and / or to use such features in field use of power system assets or other physical assets.

[0282] 6 is a flowchart of a method 115 that may be performed by the system 10, 70. The method may be performed automatically by a computing system 50 that performs the ML model training and at least one device 40, 80 that executes signal processing logic based on the trained ML model.

[0283] Training begins at step 116. Initiating training can include selecting a type of physical asset, obtaining information about the system in which the physical asset is included, selecting training and / or validation cases based on the asset type and system topology, etc. Information about the system can include information about the system topology, which can be obtained from a Substation Configuration Description (SCD) of the power system. For example, a process section of a Substation Configuration Language (SCL) file can be analyzed to identify suitable test and / or validation cases for ML model training.

[0284] In step 117, ML model training is performed. The latent features may be processed differently depending on whether they relate to a first time, space, or spatiotemporal scale, or to a different second time, space, or spatiotemporal scale. This can be done in a variety of ways. For example, as described in more detail herein, Parameters that are exogenous to the physical assets and / or power system while influencing its operation may be fed to the decoder 122 during ML model training in addition to the latent features 132 in the code generated by the encoder 121; weather conditions or changing loads are examples of such conditions; and / or - Adversarial machine learning (ML), in particular adversarial decoder training, can be performed, where features generated by the encoder of the adversarial system are fed to the decoder 122 during ML model training, in addition to the latent features 132 in the code generated by the encoder 121.

[0285] Other techniques, such as regularization techniques, may additionally or alternatively be used to effect separation of latent features in the code generated by the encoder 131 according to the time, space, or spatio-temporal scales to which the features are associated.

[0286] Techniques that cause encoder 121 to generate one or several first latent features associated with a first time scale or a first spatial scale and one or several second latent features associated with a different second time scale or a different second spatial scale may be used only during generation of the signal processing logic, and more specifically, during ML model training, but not during subsequent execution of the signal processing logic by a field device.

[0287] In step 118, the signal processing logic is executed to control, monitor and / or analyze operation of the physical asset based on one or more latent features in the code. In field use, the device 40, 80 may take control actions based at least on features associated with a faster first time scale, and the control actions may be dependent or independent of latent features associated with a second, slower time scale. In field use, the device 40, 80 may take analytical actions based at least on features associated with a second, slower time scale, and the analytical actions may be dependent or independent of latent features associated with the first, faster time scale.

[0288] 7 is a schematic diagram showing ML model training performed to achieve a desired grouping of latent features according to time and / or spatial scale. The ML model 120 being trained to evolve in the signal processing logic generates code that includes latent features 132.

[0289] The decoder 122 has a number of decoder input nodes I D is the number of output nodes of the encoder 121. E For example, the number of decoder input nodes of the decoder 122, I D is the number of output nodes of the encoder 121. E The number of decoder input nodes I of the decoder 122 may be at least 10%, at least 20%, at least 30%, or at least 50% greater than I. D is the number of output nodes of the encoder 121. EThe number of output nodes of the decoder 122 may be twice or more. D is the number of input nodes of the encoder 121, I E may be equal to

[0290] The input of the decoder 122 is coupled to a source of an additional signal 139 during ML model training. The additional signal 139 may include a code generated by another encoder. The additional signal 139 may include measurements sampled at a second sampling interval or a second average sampling interval that is greater than the first sampling interval or the first average sampling interval at which at least one, some, most, or all of the encoder inputs are sampled.

[0291] ML model training may be performed with the goal of minimizing or maximizing an objective function. The objective function may include a loss function, such as a reconstruction loss or a prediction loss. The objective function may include one reconstruction loss or prediction loss (of the "main" encoder 121 and decoder 122) that is minimized and another reconstruction loss or prediction loss (of the "auxiliary" signal processing path through a second adversarial decoder) that is maximized. The loss functions (or reconstruction loss or prediction loss) may be the same, but do not have to be the same.

[0292] The objective function may depend on a distance between the encoder input 131 and the decoder output 133, determined according to a metric. For example, the objective function that may be optimized (e.g., minimized) may be or may include the reconstruction loss.

[0293] The objective function may be equal to or may include:

[0294]

number

[0295] The objective function may be equal to or may include:

[0296]

number

[0297] Here, t j where is the sampling time and J is the sampling count. The autoencoder inputs may be available at times spaced apart in the time domain by a constant or varying sampling interval.

[0298] The ML model training may include adjusting parameter values ​​of a first ANN configured by the encoder 121 and parameter values ​​of a second ANN configured by the decoder 122. The parameter values ​​adjusted during training may include weights between nodes of the first and second ANNs. The parameter values ​​adjusted during ML model training may include parameters of a recurrent neural network, such as parameters of a GRU or LSTM cell. The parameter values ​​adjusted during ML model training may include biases and / or weights of a GRU or LSTM cell, or other parameter values ​​defining an ANN.

[0299] ML model training may be performed using known techniques, such as gradient descent or other known techniques. In contrast to traditional ML model training techniques, ML model training may be performed to separate latent features in the code into two or more groups, where the separation represents different time scales, spatial scales, spatiotemporal scales, or logical domains.

[0300] The additional signal 139 may include measurements, optionally decoupled, different from the signal 131 provided to the encoder 121 as an encoder input during ML model training. The additional signal 139 may include measurements, optionally decoupled, different from the signal 131 provided to the encoder 121 as an encoder input during ML model training. The additional signal 139 may include additional measurements that vary over a slower time scale than one, some, or all of the measurements included in the signal 131 provided to the encoder 121 as an encoder input during ML model training. Alternatively or additionally, the additional signal 139 may include additional measurements that vary over a larger spatial scale than one, some, or all of the measurements included in the signal 131 provided to the encoder 121 as an encoder input during ML model training. Alternatively or additionally, the additional signal 139 may include additional measurements associated with a hierarchical level of the logical hierarchy that is different from one, some, or all of the measurements included in the signal 131 provided to the encoder 121 as an encoder input during ML model training.

[0301] One, some, or all of the additional signals 139 may optionally be included in the signal 131 provided 131 to the encoder 121 as an encoder input during ML model training.

[0302] In field use of the signal processing logic generated from the trained ML model, it is possible that the signal processing logic is deployed as being or including only the encoder 121 of the trained ML model, i.e. the question does not arise as to which additional signal 139 should be provided to the decoder 122 of the trained ML model during field use. Optionally, the additional signal 139 may also be used during field use to operate the decoder 122.

[0303] As described in more detail below, the additional signal 139 may also be obtained in other ways. For example, the additional signal 139 may include at least some signals obtained from an adversarial decoder or encoder.

[0304] A desired separation of latent features 132 into two or more groups associated with different time, space, spatiotemporal, or logical scales can be achieved by feeding additional signals 139 to the decoder stage 122 that vary on a different time, space, spatiotemporal, or logical scale than the signals 131 during ML model training.

[0305] The inventive technique provides automatic feature design with separation of time scales, but different forms of separation (spatial, spatiotemporal, logical) can also be achieved.

[0306] 8 illustrates a method and operation of a system in which an additional signal provided to the decoder 122 of the autoencoder 120 during training is based on latent features in the code 137 of the second ML model portion 125. The autoencoder 121 and the second ML model portion 125 may be trained in parallel. Below, adversarial training of the autoencoder (or the decoder stage of the autoencoder) is described, although a different ML model may be used as an adversary to the autoencoder 120 during ML model training.

[0307] By implementing adversarial training, the desired separation of features according to their time scales, spatial scales, etc. in the code generated by the encoder 121 can be enforced. Adversarial techniques exploit the fact that much of the relevant information from the signal 131 tends to be represented in features that vary only on faster time scales (hereinafter also referred to as dynamic features for brevity), while features that vary on slower time scales (hereinafter also referred to as static features for brevity) may contain little meaningful information. Adversarial techniques enable the successful construction of meaningful static features. Thus, automatic feature design is achieved that provides slowly varying features indicative of slowly changing asset properties.

[0308] The technique of Figure 8 is based on adversarial training. The overall ANN / computation graph is augmented with a second parallel decoder stage 127 (the "adversary"). The decoder 127 has access to dynamic latent features (i.e., latent features that change more rapidly than other features). Training then alternates between the two steps.

[0309] (i) Training the adversary, i.e., the decoder 127. Only the weights corresponding to the dashed connections in FIG. 8 can be trained to minimize a loss function (e.g., reconstruction loss or prediction loss) between the data to be reconstructed 131 and the adversarial output 138.

[0310] (ii) Training the Encoder 121 and the Decoder 122. Only the weights corresponding to the solid black lines in the autoencoder 120 (shown on the right side of FIG. 8) are trained to minimize the overall objective function.

[0311] During training, one, some, or all of the latent features 137 generated by the encoder 126 may be provided to the decoder 122 during training of the autoencoder 120.

[0312] The objective function may include a term related to a loss function (e.g., reconstruction loss or prediction loss) of the decoder 122. The objective function may include a second term, the adversarial loss. The objective function may be the sum of the reconstruction loss and the adversarial loss.

[0313]

number

[0314] In equation (3), fmain denotes the output of the decoder 122 (i.e., the output of the autoencoder 120). The output of the decoder 122 is the reconstruction r i (t j ) (as a function of time, if applicable). recmay be in any of the forms described in connection with Equation (1) and Equation (2), including but not limited to, any metric that quantifies how closely the autoencoder output 133 represents the encoder input 131 may be used to define the reconstruction loss.

[0315] In formula (3), f aux denotes the output of the adversarial decoder 127. The output of the autoencoder 120 is a reconstruction r aux,i (t j ) (as a function of time, if applicable). As noted above, only a portion of signal 131 may be fed to second portion 125 during adversarial training.

[0316] The adversarial loss may be a reconstruction loss, which may be the same as or include:

[0317]

number

[0318] The adversarial loss may be a reconstruction loss, which may be the same as or include:

[0319]

number

[0320] Here, t j where is the sampling time and J is the sampling count. The autoencoder inputs may be available at times spaced apart in the time domain by a constant or varying sampling interval.

[0321] In equation (3), α>0 is a constant that defines the weight of the second adversarial training criterion. Due to the negative sign in equation (3), increasing the adversarial loss reduces the overall objective function (which is the purpose of adversarial training). As mentioned above, the reconstruction loss can also be used for the adversarial loss. In this case, the training criterion for the main autoencoder is to simultaneously achieve a low reconstruction loss for the main decoder and a high reconstruction loss for the adversarial decoder. The adversarial loss is not limited to the reconstruction loss as shown in relation to equations (4) and (5).

[0322] Thus, adversarial training can be an iterative process. Each iteration training the encoder 121, the decoder 122 and the second encoder 126 with the objective of minimizing the reconstruction loss of the autoencoder 120 and maximizing the reconstruction loss (adversarial loss) of the second autoencoder 125; training the second decoder 127 with the objective of minimizing the reconstruction loss of the second decoder 127.

[0323] Although we refer to reconstruction loss above, other loss functions (such as prediction loss) or other objective functions can be used.

[0324] The techniques disclosed in relation to FIGS. 7 and 8 may be combined as shown in FIG.

[0325] FIG. 9 illustrates operation of a method and computing system during ML model training in which the decoder 122 of an autoencoder 120 (at least whose encoder 121 is used to generate signal processing logic) receives a decoder input including code generated by the encoder 121, one, some, or all of the latent features generated by a second machine learning model (which may be an adversary of the decoder 122 or the encoder 121), and an additional signal 139 including measurements different from (and optionally decoupled from) those included in the signal 131.

[0326] An additional signal 139 may also be fed to the second ML model portion 125 during adversarial training. The additional signal 139 may include measurements sampled at a second sampling interval or a second average sampling interval that is greater than the first sampling interval or the first average sampling interval at which at least one, some, most, or all of the encoder inputs are sampled.

[0327] 7 and 9, if the additional signal 139 may include measurements sampled at a second sampling interval or a second average sampling interval that is greater than the first sampling interval or the first average sampling interval at which at least one, some, most, or all of the encoder inputs 131 are sampled, the additional signal 139 may be replicated along the time axis for ML model training. For example, to accommodate different sampling times, multiple different samples of the signal included in the signal 131 may be input to the autoencoder architectures 120, 125, and the same sample values ​​of the signal included in the additional signal 139 may be input to the decoder stages 122, 127.

[0328] The adversarial autoencoder technique described in relation to Figures 7-9 provides a discriminative model, which is different from the adversarial autoencoder technique for generative models.

[0329] In the techniques disclosed herein, the separation of the two time scales can be achieved by the corresponding autoencoder architecture and the alignment of the input / output data. The two time scales are a faster time scale (e.g., an operating time scale) and a slower time scale. The faster scale typically depends on the original time sampling of the data (i.e., the original time point at which the data was specified). In general, it may be any time scale faster than the second time scale specified below, and thus the original data may be pre-processed by any resampling, interpolation, filtering, etc. method before further processing. On the other hand, the slower time scale features depend on a (regular or irregular) sub-sampling of the faster time scale (typically the original time point). For example, if the original sampling rate is at regular intervals of 1 minute and the slower time scale is at regular intervals of 1 day, there will be only one slow feature vector per day, which is 1440 original samples.

[0330] Technically, the encoder 121 converts data (i.e., the input signal 131) into latent features. The latent features (also called codes) may be three-dimensional tensors with (1) a slow time scale, (2) an action time scale, and (3) the dimensions of the signal.

[0331] FIG. 10 shows the encoder input 141. The encoder processes the encoder input 121 into latent features. The latent features include one or several first groups of latent features 160 associated with a first time scale. The latent features include one or several second groups of latent features 170 that change at a different second time scale. Different points on the slow time scale are shown by different filling patterns 161-164 in FIG. 10. Features associated with the faster second time scale are shown as regions 171-174. Since at a constant time of the slow scale there is only a single feature vector of the slow features, these features are also called static. Features that depend on the operating time scale are called dynamic. The latent features collectively form the code 150.

[0332] The output of the autoencoder is a reconstruction 143 of the encoder input 141. During training, to generate (part of) the decoder input, the less frequently updated latent features 161-164 may be replicated along the time axis shown in FIG.

[0333] The generation of decoder inputs during ML model training can be performed using a method 260 as shown diagrammatically in FIG. 28. The method 260 can be used to generate the decoder inputs. In step 261, a first and a second set of latent features are obtained from the encoder output layers of the first and second encoders, respectively. In step 262, the latent features included in the second set are replicated along the time axis (as shown in FIG. 10 by the extension regions 161-164) until a new second set of latent features is available. The replicas are combined with the latent features included in the first set, which are updated more frequently. That is, when K new first sets of latent features are output by the encoder while the same second set of latent features are being output, K-1 replicas of the latent features are generated and combined (e.g., by concatenation) with the first set of latent features, respectively, to generate (a part of) the decoder inputs in step 263.

[0334] Any combination of layers in the encoder can be used to map the input tensors to static features. The last layer of the encoder can map sequences to vectors.

[0335] Some different options of encoder and decoder architectures that may be used in the autoencoder 120 and, if present, the second autoencoder 125, are illustrated in FIGS.

[0336] 11 illustrates a feature-wise fully connected architecture 183. This can be achieved by applying the same ANN (e.g., with the same and / or shared weights) over time.

[0337] 12 illustrates an RNN architecture 184. The RNN architecture may be an LSTM or GRU with a time series output, in which case a time series output is provided, i.e., the architecture 185 has an output tensor with a time dimension.

[0338] 13 shows an architecture 185 that can be used as the last layer of the encoder 121 (and optionally the encoder 126). First, all features are processed along the motion time dimension, and all features are processed in the same way (weight sharing). This information is then processed into a static feature vector.

[0339] 14 shows an architecture 186 that can be used as the last layer of the encoder 121 (and optionally the encoder 126). All features are first aligned along the motion time dimension, and all features are treated differently (no weight sharing). This information is then further processed into a static feature vector.

[0340] Figure 15 shows an architecture 187 in which a sequence-to-vector RNN is used. For input to the subsequent decoder, the static features are replicated over the operating time so that the decoder network operates on the operating time scale and can therefore use, for example, elements from Figures 11 and 12.

[0341] As described with reference to Figures 7 and 9 and further described below, feature design, in which latent features are separated according to their associated time scales, may be achieved by processing signals differently depending on the typical scale at which the signals vary. For example, for a physical asset that is a power system asset, voltage and / or current may vary on one time scale, while parameters such as ambient conditions (which may depend on weather or time of day, such as time-dependent load changes) vary on a slower time scale. The latter parameters are sometimes referred to as "extrinsic" parameters, since they exist independently of the presence of the physical asset. This is in contrast to signals that emanate from or are directly related to the presence of the physical asset. These two types of signals may be processed differently, as described with reference to Figures 7 and 9 and further described with reference to Figures 16-19. One, some, or all of the signals that exist independently of the presence of the physical asset and that are available for ML model training may be used as additional signals 139, at least some of which are used as decoder inputs for decoder 122 during training. One, some, or all of the signals that exist independent of the presence of a physical asset may be included in the encoder input of encoder 121.

[0342] Thus, the additional signal 139 may be or include an external signal that exists independently of the presence of the asset: in the wind turbine example, for example, the wind speed and direction are external signals, while the active power generated by the wind turbine is a signal that is directly dependent on the presence of the physical asset.

[0343] To facilitate the selection of a low-dimensional latent representation, external signals and measurements that depend on the presence of a physical asset may be treated differently, as may external signals to the decoder 122. Thus, the autoencoder 120 can focus on representing information indicative of the asset's behavior in the latent representation, and does not need to include information in the code 132 that cannot be indicative of the asset's behavior (because it belongs to the external signals).

[0344] Additionally or alternatively, it is possible to focus on a single or limited set of failure modes for a given ML model. Domain knowledge can be used to determine which of the signals 131 that depend on the physical assets can indicate the failure mode of interest. By having the decoder 122 decode only these relevant signals, the focus of the autoencoder can be further improved and / or the latent dimensionality can be reduced. That is, the number of output nodes of the decoder 122 can be selected depending on which of the signals 131 that depend on the presence of physical assets are considered particularly relevant. This architecture is shown in FIG. 16.

[0345] FIG. 17 shows a further architecture in which several ANNs 121a, 121b are included as encoders. Some ANNs may be or include RNNs. Some ANNs 121a, 121b may be trained simultaneously, but the weights of some ANNs 121a, 121b may be adjusted independently of each other in training. Several groups of latent features 132a, 132b are provided by the ANNs 121a, 121b. The groups 132a, 132b may be used in combination with a signal 139 that exists independent of the presence of a physical asset to provide a decoder input for the decoder 122 during ML model training.

[0346] 18 shows a further architecture in which an adversarial decoder 127 is provided. The ML model having one or several encoders 121a, 121b and an adversarial decoder 127 is trained to minimize a composite objective function that may have the form as described in relation to equation (3).

[0347] The adversarial decoder 127 and the encoders 121a, 121b and decoder 122 may each be trained in an iterative process with the goal of minimizing a composite objective function. Training may include adjusting weights and / or other parameter values ​​(such as biases for GRUs or LSTMs). The objective function used in the training process may be implemented as described above.

[0348] The training step can use techniques such as, but not limited to, gradient descent.

[0349] 18, the adversarial decoder 127 may receive as a decoder input only a subset of the latent features generated by the encoders 121a, 121b (e.g., latent features 132b). For example, only latent features 132b associated with a slower time scale may be selectively provided to the adversarial decoder 127. The decoder 122 may further receive latent features 132a associated with a different time scale as a decoder input.

[0350] FIG. 19 shows an architecture that combines the differential processing of external signals that exist independent of the presence of physical assets (such as ambient conditions of a power system) and signals that exist only because of physical assets (e.g., current, voltage, power, etc. measured on or in relation to a physical asset), as well as adversarial training techniques.

[0351] During adversarial training, the decoder 122 can receive a first latent feature 132a associated with one time scale, a second latent feature 132b associated with a different time scale, and an external signal 139 that exists independent of the presence of a physical asset.

[0352] During adversarial training, the adversarial decoder 126 may select only a subset of the latent features (e.g., latent features 132b), depending on the characteristic time scales over which the latent features vary, and on external signals 139 that exist independent of the presence of physical assets.

[0353] The adversarial training may be performed as described in more detail above. For example, a composite objective function may be minimized in the adversarial training, which may involve adjusting the weights of the encoders 121a, 121b to reduce or minimize the composite objective function described with reference to equations (3)-(5).

[0354] Figures 20 and 21 show how features are automatically designed using adversarial ML model training (Figure 20) compared to the conventional technique (Figure 21). ML model training is performed with a dataset where the external signal is sampled as a 3-dimensional standard normally distributed random variable over T = 1000 time steps. A 1-dimensional true latent signal is generated by superimposing a linear trend (the slow evolving part that should be obtained), a high-frequency sinusoid, and Gaussian noise. From the input and the linear trend (representing the latent signal to be identified), a 5-dimensional output signal is generated by random linear mapping. The input and output constitute the input of an autoencoder with one dynamic latent feature and one static latent feature. The signal to be reconstructed is the autoencoder output.

[0355] Results with adversarial training (Figure 20) show that linear trends are captured by static latent features, but without adversarial training (Figure 21) the trend is much lower, with the full information instead being represented by dynamic latent features.

[0356] In this simple example used to demonstrate the concept, linear trends were visible to the naked eye in both the dynamic latent features shown in Figure 20 and the 5D autoencoder input (not shown in Figures 20 and 21).

[0357] In practical applications, similar slow trends are often not clearly evident in the input signal as faster dynamics dominate. Moreover, the slow trends do not have to be linearly increasing functions but may be more complex, non-monotonic, multivariate, and / or dynamic in nature (i.e., visible only when multiple time steps are considered together). In this case, the proposed invention is well suited to extract these more complex, slowly evolving trends.

[0358] In all of the methods, computing systems, and systems disclosed herein that use an adversarial loss, the adversarial loss does not have to be a reconstruction loss (as illustratively described with reference to equations (4) and (5)). Various variations of the adversarial loss can be used instead of the reconstruction loss. This list is not exhaustive. It should be noted that any combination of the listed variations (linear combination, weighted average, etc.) can also be used.

[0359] (i) The adversarial loss does not necessarily have to be the reconstruction loss for the signal to be reconstructed, but may quantify the deviation between the output of the second (adversarial) decoder 127 and the output of the decoder 122.

[0360] For example, the adversarial loss may be the reconstruction loss, which may be the same as or include:

[0361]

number

[0362] The adversarial loss may be a reconstruction loss, which may be the same as or include:

[0363]

number

[0364] Here, t jwhere is the sampling time and J is the sampling count. The autoencoder inputs may be available at times spaced apart in the time domain by a constant or varying sampling interval.

[0365] (ii) To determine the reconstruction loss (whether determined for the decoder 122 or the second decoder 127), a distance / divergence measure of the probability distribution may be used, e.g., Kullback-Leibler divergence, maximum mean divergence, Wasserstein distance, or others.

[0366] (iii) Instead of formulating an adversarial training criterion in the second decoder 127, a criterion on the probability distribution of the dynamic latent features may be used. The invariance of the dynamic features can be promoted by implementing some statistical test for invariance, such as Dickey-Fuller. Alternatively or additionally, the dynamic features may be regularized towards a probability distribution that is given a priori or to be determined during training.

[0367] The autoencoder architecture is not limited to the architecture described in relation to Figures 1 to 21. Modifications and variations may be implemented, some of which are described in relation to Figures 22 and 23.

[0368] 22 illustrates an autoencoder architecture having an encoder that operates to generate three or more groups of latent features 211-213. The autoencoder may have an encoder with three or more ANNs 191-193 to generate the groups of latent features 211-213. More than two groups of latent features may be generated.

[0369] All of the latent features may be processed by the decoder 201 of the autoencoder to be trained. Only latent features of a subset of the groups 211-213 may be processed by the adversarial decoders 202, 203.

[0370] External signals that exist independent of physical assets (such as signals indicative of ambient conditions or load) may be fed to both the various encoders 191-193, the decoder 201, and the adversarial decoders 202, 203. The number of adversarial decoders 202, 203 may be equal to the number of groups 211-213 minus one. Each of the adversarial decoders 202, 203 may be associated with a different time scale, spatial scale, spatiotemporal scale, or logical hierarchical level.

[0371] The number of main decoders (e.g., decoders that receive more latent features than the adversarial decoders) may be greater than 1. The number of adversarial decoders may be greater than 1.

[0372] 23 shows an architecture with two or more main decoders 231, 232 and two or more adversarial decoders 233, 234. The main decoders 231, 232 and the adversarial decoders 233, 234 can all receive common latent features 241 generated by the encoder 221 as decoder input.

[0373] At least the main decoder stages 231, 232 may receive as decoder inputs the latent features 242, 243 generated by the additional encoders 222, 223.

[0374] The adversarial decoders 233, 234 may receive as input at least the external signal 139 and common latent features 241. Each adversarial decoder 233, 234 may selectively receive latent features included in one group 242, 243 of latent features or a subset of all available groups of latent features.

[0375] Each decoder stage 231-234 (main or adversarial) can be configured to reconstruct all the signals or only a selected subset of the signals.

[0376] While ML models can be trained to perform separation of latent features according to time scales (suitable for detecting slowly evolving trends), alternative or additional goals of automatic feature design can be achieved. Below is a non-exhaustive list of some possibilities:

[0377] (i) Separation of more than two time scales can be achieved by setting up two or more adversarial decoder stages (as shown in Figure 22).

[0378] (ii) Different separations, e.g., separation in space, or subsystem / logical separation, can be achieved, for example, by setting up multiple main and adversarial decoder stages and associating subsets of signals to the decoder stages according to their spatial location / subsystem affiliation. Figure 23 shows an example of latent features in two spatial / logical domains a, constructed such that a first constructed feature set is trained to represent information in the first set of signals but not in the second set of signals. Conversely, a second constructed feature set is trained to represent information in the second set of signals but not in the first set of signals. This formulation does not require a time setting, i.e., time does not need to be a dimension of data here.

[0379] At least one encoder of the trained ML model can be used as (part of) the signal processing logic. If the autoencoder architecture has two or more encoders (as shown in Figures 17-19, 22, and 23), a selected one of the encoders that provides latent features associated with a desired time scale (e.g., latent features evolving over time scales slow compared to grid frequency), a desired geographic region (e.g., a given substation), or a desired hierarchical level (e.g., voltage or bay level) can be deployed for use in the signal processing logic.

[0380] The signal processing logic may include additional pre-processing and / or post-processing components. The pre-processing logic may condition the signal before it is input to the input layer of the encoder of the trained ML model. The pre-processing logic may comprise a filter (e.g., a high-pass, low-pass, or band-pass filter), may operate to process some measurements into the encoder's input signal (e.g., calculate conductance, impedance, or admittance from some measurements), or may perform other counters. In general, the pre-processing logic may operate to perform signal conditioning during field use of the encoder.

[0381] The post-processing logic may operate to process trends represented by one or several latent features of the encoder. The post-processing logic may perform feature adjustment operations (such as filtering). The post-processing logic may operate to process the latent features output by the encoder during field use and to identify asset health degradation, general trends in asset health, etc. The post-processing logic may include a prediction component that operates to predict whether and / or when a safety margin of asset operation can be reached based on the past time evolution of at least one latent feature over a future time horizon. Proactive warnings and / or mitigation actions may be taken based thereon. The post-processing logic may operate to process the latent features by comparing them to safety margins that define safe, critical, and optionally failure states of the asset.

[0382] 24 is a diagram of a machine learning model that is the encoder portion of an autoencoder trained using the techniques disclosed herein, specifically, signal processing logic that comprises an ANN 252. The ANN 252 may be an RNN. The ANN 252 may comprise, but is not limited to, an LSTM or a GRU.

[0383] The ANN 252 has an input layer 251. The input layer 251 operates to receive measurements. The measurements may include measurements of electrical or other properties dependent on the physical asset. The measurements may include voltage, current, and / or power measured on the physical asset or conductors connected to it. The measurements may include non-electrical measurements such as gas concentrations (e.g., for dissolved gas analysis) and / or temperature measurements.

[0384] The ANN 252 has an output layer 253. The output layer 253 can provide one or several latent features. The one or several latent features can indicate trends over time of the physical asset condition. The one or several latent features can indicate trends as a function of location and / or logical association.

[0385] Control decisions may be made automatically, such as automatic operation of power system components and / or user interface control, depending on the values ​​output by output layer 253. Thus, output layer 253 may provide control signals that operate to operate power system components, user interfaces, trigger physical actions, or trigger other actions related to physical assets in the power system.

[0386] The techniques disclosed herein may be used to monitor and update the signal processing logic during field use. The data used to train the autoencoder may include measurements captured during field use of the physical asset. Once the signal processing logic including at least one encoder is deployed, the training of the autoencoder may continue using measurements captured during field use. The signal processing logic may be selectively updated using updated versions of the at least one encoder obtained by continued training in parallel with field operation of the physical asset and the signal processing logic.

[0387] 25 shows a diagram of a system 30 comprising an IED 40 and a computing system 50. The IED 40 comprises one or several ICs 41 executing signal processing logic 42. The signal processing logic 41 may comprise at least one encoder of a trained ML model that may be generated as disclosed herein.

[0388] Measurements captured during field use of a physical asset using sensors 31, 32 may be processed by the IED 40 to perform control, monitoring and / or analytical functions.

[0389] Measurements captured during field use of the physical asset using the sensors 31, 32 may also be used by the computing system 50 to continue ML model training and / or for performance monitoring 56. Based on results of the performance monitoring 56 that identify whether the signal processing logic 42 should be updated, new signal processing logic generated by signal processing logic generation 55 may be deployed to the IED 40. In this manner, the signal processing logic of the IED 40 may be updated during field use.

[0390] 26 and 27 show further examples of signal processing logic that may include an encoder for a trained ML model.

[0391] In addition to or instead of using an encoder of the trained ML model as (part of) the signal processing logic, a decoder of the trained ML model can also be used.

[0392] 29 is a flow chart of a method 270 in which a signal (which may be an online or offline signal) is processed using signal processing logic including or otherwise based on a decoder in step 271. The signal processing in step 271 may include performing anomaly detection. The asset health indicator (or other latent variable) may be artificially altered, a loss function (e.g., reconstruction loss or prediction loss) may be analyzed, or the reconstruction may be analyzed in step 272. In step 273, a control action is performed based on the results of the analysis in step 272.

[0393] The method for generating signal processing logic is embodied in a computer-implemented system. The (adversarial) training process can be used for automatic feature design of data recorded from one or more assets.

[0394] The trained ML model (or at least one of its encoders) can be used to analyze data offline or online, where "online" means that the trained ML model is evaluated against new data recorded from the asset, either continuously or at regular or irregular time intervals.

[0395] The trained ML model can be used to analyze data and monitor assets by analyzing and / or evaluating the diagnostic output of the autoencoder, which may consist of latent features, reconstruction loss, or any other signal derived from the reconstruction, or any subset of features further derived from the autoencoder's output by subsequent computation.

[0396] Methods, apparatus, computing systems, and systems according to embodiments may be used to generate diagnostic output that may be output via a UI or automatically analyzed by another computer-implemented system.

[0397] Methods, apparatus, computing systems, and systems according to embodiments can be used for offline generation of features for monitoring (pre-training), online application for monitoring, diagnostics, fault prediction, etc. The evolution of diagnostic output may be monitored online and output via a UI, or may be automatically analyzed by another computer-implemented system to track asset performance, make operational and / or maintenance decisions, generate alarms, etc.

[0398] Methods, apparatus, computing systems, and systems according to embodiments can be used for online training or retraining of signal processing logic, keeping the signal processing logic up to date with changes in asset behavior, and for online use of diagnostic outputs as described above.

[0399] Methods, apparatus, computing systems, and systems according to embodiments can be used to implement a digital twin of a system (e.g., a power system) that includes IEDs that execute signal processing logic. When the signal processing logic that includes or is based on at least one encoder of a trained ML model is deployed in the IEDs, the digital twin can be modified accordingly in an automated manner. When the signal processing logic that includes or is based on at least one encoder of a trained ML model is updated in the IEDs, the digital twin can be updated accordingly in an automated manner.

[0400] The methods, systems, and apparatus disclosed herein provide techniques for enhanced monitoring, control, and / or analysis of physical assets. Features used to perform monitoring, control, and / or analysis operations can be automatically designed using ML model training, which can optionally include adversarial training.

[0401] No manual or specialized knowledge is required to manipulate features from a potentially large number of different measurements that may be time-dependent. A difficult task with the potential for suboptimal feature design likely to be encountered with ergonomic features is improved using the techniques disclosed herein.

[0402] Although embodiments may be used for power system protection, embodiments are not limited thereto.

[0403] Further examples are provided below as part of this disclosure. 1. A method for generating or updating signal processing logic for processing signals, particularly time series signals, containing measurements related to a physical asset, comprising the steps of: Training a machine learning (ML) model, the ML model being a first encoder having a first input layer for receiving an input signal, in particular a time series input signal, the first encoder being operative to generate a first code comprising first latent features; at least one second encoder having at least one second input layer for receiving an input signal, the at least one second encoder operative to generate at least one second code including a second latent feature; a first decoder having an output layer, the first decoder operative to process a first decoder input to generate a first decoder output; at least one second decoder having at least one second output layer, the at least one second decoder operative to process the second decoder inputs to generate a second decoder output; Training the ML model includes providing training signals as input signals to a first input layer and a second input layer, generating first decoder inputs from at least some first latent features included in the first code, and generating second decoder inputs from at least some second latent features included in at least one second code; providing at least one of the first encoder or the second encoder and / or at least one of the first decoder or the second decoder of the trained ML model as at least a part of a signal processing logic to a device that executes the signal processing logic to control, monitor, and / or analyze a physical asset; A method comprising:

[0404] 2. The method of Example 1, wherein generating the first decoder input includes generating the first decoder input from the first code and from an additional signal that includes at least some second latent features included in at least one second code.

[0405] 3. The method of Example 1 or Example 2, wherein the first latent feature and the second latent feature vary on different time scales and / or the first latent feature varies on a slower time scale than at least some of the second latent features.

[0406] 4. Training the ML model includes performing an iterative training including a number of iterations, each of the number of iterations training a first encoder, a first decoder and at least one second encoder with the objective of minimizing a loss function, in particular a reconstruction loss, of the first decoder and maximizing a loss function, in particular a reconstruction loss, of the at least one second decoder; The method according to any one of the preceding Examples 1 to 3, each comprising training at least one second decoder with the objective of minimizing a loss function, in particular a reconstruction loss, of the at least one second decoder.

[0407] 5. The method of any one of the preceding examples 1-4, wherein the ML model comprises at least one autoencoder, and / or the first decoder output includes a first reconstruction of at least some of the input signals, and / or the second decoder output includes a second reconstruction of at least some of the input signals.

[0408] 6. The method of any one of the preceding Examples 1-5 when dependent on Example 2, wherein the additional signal includes a variable that affects the operation of the physical asset but exists regardless of the presence or absence of the physical asset, and / or the additional signal includes a variable that is measured at a first sampling time that is longer than a second sampling time at which the measurement is sampled.

[0409] 7. The method of any one of the preceding Examples 1-6 when dependent on Example 2, wherein the additional signal includes an ambient condition variable of the physical asset, and optionally, the ambient condition variable includes weather data affecting at least one power system characteristic associated with the physical asset.

[0410] 8. The method of any one of the preceding Examples 1-7, wherein the first decoder input includes replicas of at least some of the second latent features included in the at least one second code.

[0411] 9. One or more of the first latent variables represent temporal characteristics related to the physical asset, and optionally one or more of the first latent variables of the trained first encoder represent temporal characteristics that vary over a first time scale that is longer than a sampling time interval of one, some, or all of the measurements; The method according to any one of the preceding Examples 1-8.

[0412] 10. The method of example 9, wherein the first time scale is a sub-sampling of the sampling time interval.

[0413] 11. The method of any one of the preceding Examples 1-10, wherein the apparatus processes one or more of the first latent variables to perform an asset health analysis of the physical asset.

[0414] 12. A method for controlling, analyzing and / or monitoring a physical asset, in particular a power system asset, comprising: - executing, by a control, analysis and / or monitoring device, a signal processing logic generated by the method of any one of the preceding claims, the signal processing logic processing signals comprising measurements related to a physical asset, in particular time series signals; performing, by a control, analysis, and / or monitoring device, at least one control action based on one or more of the latent characteristics; In some cases, the method, wherein the at least one control action includes one or more of: triggering a protective function, controlling an interface to perform an output action, controlling at least one operating parameter of a physical asset, or controlling at least one operating parameter of a power system asset other than the physical asset.

[0415] 13. A method for automatically designing features for controlling, analyzing, and / or monitoring a physical asset, comprising: A method comprising using the method of any one of Examples 1 to 11 to design features whose measurements are processed to control, analyze, and / or monitor a physical asset.

[0416] 14. Devices for controlling, analyzing, and / or monitoring physical assets, particularly power system assets, comprising: an interface for receiving signals, in particular time series signals, comprising measurements related to a physical asset; At least one integrated circuit, Processing the signal using a trained Machine Learning (ML) model, in particular an encoder of a trained autoencoder, the encoder having an input layer for receiving the signal, the encoder generating a code including latent features, and / or processing the signal using a decoder, in particular a decoder of a trained autoencoder; an integrated circuit operative to perform at least one control action based on one or more of the latent features or based on a decoder output of the decoder; 1. A control, analysis and / or monitoring device for a physical asset, in particular a power system asset, comprising:

[0417] 15. An electric power system comprising: Physical assets; at least one measurement device operative to capture a signal associated with the physical asset; A control, analysis and / or monitoring device according to example 14 for processing the captured signal, optionally the control, analysis and / or monitoring device being an intelligent electronic device (IED); A power system comprising:

Claims

1. A method for generating or updating signal processing logic for processing signals, in particular time series signals, containing measurements of electrical properties associated with physical assets of a power system, the method comprising the steps of: Training a Machine Learning (ML) model, the ML model comprising: a first encoder having a first input layer for receiving an input signal, in particular a time series input signal, the first encoder being operative to generate a first code comprising first latent features; at least one second encoder having at least one second input layer for receiving the input signal, the at least one second encoder operative to generate at least one second code including second latent features; a first decoder having an output layer, the first decoder operative to process a first decoder input to generate a first decoder output; at least one second decoder having at least one second output layer, the at least one second decoder operative to process the second decoder inputs to generate a second decoder output; training the ML model includes providing a training signal as the input signal to the first input layer and the at least one second input layer; generating the first decoder inputs from at least some first latent features included in the first code; and generating the second decoder inputs from at least some second latent features included in the at least one second code; providing at least one of the first encoder or the second encoder and / or at least one of the first decoder or the second decoder of the trained ML model as at least a part of signal processing logic to a device executing the signal processing logic to control, monitor, and / or analyze the physical asset during field use of the physical asset in the power system; A method comprising:

2. 2. The method of claim 1, wherein the device executes the signal processing logic to operate at least one circuit breaker, disconnector, and / or converter to perform preventive or corrective control actions on the physical asset during field use of the physical asset in the power system.

3. 3. The method of claim 1, wherein generating the first decoder input comprises generating the first decoder input from the first code and from an additional signal that includes at least some second latent features included in the at least one second code.

4. 4. The method of claim 1, wherein the first latent feature and the second latent feature vary on different time scales and / or the first latent feature varies on a slower time scale than at least some of the second latent features.

5. Training the ML model includes performing an iterative training including a number of iterations, each of the number of iterations: training the first encoder, the first decoder and the at least one second encoder with the objective of minimizing a loss function, in particular a reconstruction loss, of the first decoder and maximizing a loss function, in particular a reconstruction loss, of the at least one second decoder; and training the at least one second decoder with the aim of minimizing a loss function of the at least one second decoder, in particular a reconstruction loss.

6. 10. The method of any one of the preceding claims, wherein the ML model comprises at least one autoencoder, and / or the first decoder output comprises a first reconstruction of at least some of the input signals, and / or the second decoder output comprises a second reconstruction of at least some of the input signals.

7. 11. A method according to any one of the preceding claims when dependent on claim 3, wherein the additional signals include variables that affect the operation of the physical assets of the power system but exist independent of the presence or absence of the physical assets, and / or the additional signals include variables measured at a first sampling time that is longer than a second sampling time at which the measurements are sampled.

8. 4. The method of claim 3, wherein the additional signals include ambient condition variables of the physical assets of the power system, and optionally the ambient condition variables include weather data affecting at least one power system characteristic associated with the physical assets.

9. 2. A method according to any one of the preceding claims, wherein the first decoder input comprises replicas of at least some of the second latent features contained in the at least one second code.

10. one or more of the first latent variables represent temporal characteristics related to the physical assets of the power system, and optionally the one or more first latent variables of the trained first encoder represent temporal characteristics that vary over a first time scale that is longer than a sampling time interval of one, some, or all of the measurements of the electrical properties associated with the physical assets; 10. A method according to any one of the preceding claims.

11. The method of claim 10 , wherein the first time scale is a sub-sampling of the sampling time interval.

12. 2. The method of any one of the preceding claims, wherein the apparatus processes one or more of the first latent variables to perform an asset health analysis of the physical assets of the power system.

13. 1. A method for controlling, analyzing, and / or monitoring physical assets of an electric power system, comprising: - executing, by a control, analysis and / or monitoring device, a signal processing logic generated by the method of any one of the preceding claims, said signal processing logic processing signals, in particular time series signals, comprising measurements of electrical properties related to said physical assets; performing, by the control, analysis, and / or monitoring device, at least one control action of the physical asset during field use of the physical asset within the power system based on one or more of the latent characteristics; A method comprising:

14. 14. The method of claim 13, wherein performing the at least one control action of the physical asset comprises performing at least one preventive or corrective control action of the physical asset during field use of the physical asset in the power system by operating at least one breaker, disconnector, and / or converter.

15. 1. A method for automatically designing features for controlling, analyzing, and / or monitoring physical assets of an electric power system, comprising:

13. A method comprising using the method of any one of claims 1 to 12 to design features in which measurements of electrical properties associated with a physical asset are processed to control, analyze and / or monitor the physical asset during field use of the physical asset in the power system.

16. 16. The method of claim 15, wherein controlling, analyzing, and / or monitoring the physical asset includes performing at least one preventive or corrective control action on the physical asset during field use of the physical asset in the power system by operating at least one breaker, disconnect switch, and / or converter.

17. 1. A control, analysis, and / or monitoring device for a physical asset of an electric power system, comprising: an interface for receiving a signal, in particular a time series signal, comprising a measurement of an electrical property related to said physical asset; At least one integrated circuit, Processing the signal using a trained Machine Learning (ML) model, in particular an encoder of a trained autoencoder, the encoder having an input layer for receiving the signal, the encoder generating a code including latent features, and / or processing the signal using a decoder, in particular a decoder of a trained autoencoder; and an integrated circuit operative to perform at least one control action of the physical asset during field use of the physical asset in the power system based on one or more of the latent characteristics or based on a decoder output of the decoder; 1. A control, analysis, and / or monitoring device for a physical asset of a power system comprising:

18. 20. The apparatus of claim 17, wherein performing the at least one control action of the physical asset comprises performing at least one preventive or corrective control action of the physical asset during field use of the physical asset in the power system by operating at least one breaker, disconnect switch, and / or converter.

19. 1. An electric power system comprising: Physical assets; at least one measurement device operative to capture a signal related to an electrical characteristic of the physical asset; - a control, analysis and / or monitoring device according to claim 17 or claim 18 for processing the captured signals, optionally being an Intelligent Electronic Device (IED); A power system comprising:

Images