Systems and methods for photovoltaic fault detection using a feedback-enhanced positive unlabeled learning
View Patent ↗Various embodiments of a system and associated method for identifying and classifying faults in a photovoltaic array using relatively little labeled data are described herein. In particular, the system builds on existing PU classification techniques by addition of a feedback loop that enables classification of limited operational data of a photovoltaic array by expanding a plurality of features within the operational data based on a learned importance of each feature.
1 . A system, comprising:
a processor in communication with a memory and one or more sensors in operative communication with a photovoltaic array wherein the one or more sensors are operable to capture operational data for the photovoltaic array and communicate the operational data to the processor, the memory including instructions, which, when executed, cause the processor to:
(a) access a training dataset comprising a plurality of data samples, each data sample having a plurality of features descriptive of sensor-derived operational data of a photovoltaic panel of the photovoltaic array, wherein a first subset of the plurality of data samples is labeled as positive and a second subset is unlabeled;
(b) access a plurality of weights of a non-traditional classification model, the non-traditional classification model being trained to estimate, for a data sample and using the training dataset, a labeling probability that the data sample is labeled as positive, wherein the non-traditional classification model is trained using the first subset labeled as positive and the second subset unlabeled to estimate the labeling probability, and determine an estimated labeling frequency value corresponding to an upper bound on the labeling probability that a data sample corresponding to a positive class is labeled;
(c) generate an augmented feature set based on the training dataset by application of one or more non-linear transformations to a plurality of most influential features of the plurality of features based on a magnitude of each weight of the plurality of weights of the non-traditional classification model, wherein the plurality of most influential features comprises a top k highest-weighted features selected based on the magnitude of the plurality of weights, and wherein the one or more non-linear transformations comprise adding p-level polynomial combinations of the top k highest-weighted features to obtain an expanded feature space; and
(d) apply the non-traditional classification model to the training dataset as represented in the expanded feature space to estimate, for each data sample of the training dataset, the labeling probability that the data sample is labeled as positive;
wherein the processor is operable to: (i) determine, for a new data sample of operational data corresponding to the photovoltaic panel as obtained by the one or more sensors and sing the non-traditional classification model, a labeling probability that the new data sample is labeled as positive, and (ij) classify the photovoltaic panel into a first class or a second class using a fault classification model configured to classify based on the labeling probability and the estimated labeling frequency value corresponding to an upper bound on the labeling probability that a data sample corresponding to a positive class is labeled, wherein the first class is indicative of a fault and wherein the second class is indicative of no fault.
2 . The system of claim 1 , wherein the memory further includes instructions which, when executed, cause the processor to:
receive, by the processor, the operational data from the one or more sensors, wherein the operational data includes photovoltaic data including the plurality of features for each photovoltaic panel of the photovoltaic array.
3 . The system of claim 1 , wherein the non-traditional classification model is a Modified Logistic Regression classifier.
4 . The system of claim 3 , wherein the Modified Logistic Regression classifier defines a weighted combination of the plurality of features, wherein each feature of the plurality of features is associated with a respective weight of the plurality of weights of the non-traditional classification model.
5 . The system of claim 1 , wherein the processor as configured increases a dimensionality of each feature by adding p-level polynomial combinations of the top k highest-weighted features to the training dataset.
6 . The system of claim 1 , wherein the processor as configured applies the non-traditional classification model to the training dataset as represented in the expanded feature space thereby introducing a non-linear decision boundary into a feature space descriptive of the plurality of features of the training dataset.
7 . The system of claim 1 , wherein the memory further includes instructions which, when executed, cause the processor to:
receive an amount k of highest-weighted features to be selected.
8 . The system of claim 1 , wherein the processor as configured minimizes one or more features of the expanded feature space that do not substantially contribute to classification using a dimensionality reduction algorithm.
9 . The system of claim 8 , wherein the dimensionality reduction algorithm is a Principal Component Analysis algorithm.
10 . The system of claim 1 , wherein the processor as configured minimizes one or more features of the expanded feature space that do not substantially contribute to classification using a regularization algorithm.
11 . The system of claim 1 , further comprising: receiving a value indicative of a level of polynomial enhancements (p).
12 . A system, comprising:
a photovoltaic array including a plurality of photovoltaic panels and one or more sensors in operative communication with the plurality of photovoltaic panels;
a processor in communication with a memory and the photovoltaic array, wherein the one or more sensors are operable to capture operational data for the photovoltaic array and communicate the operational data to a processor, the memory including instructions, which, when executed, cause the processor to:
access a plurality of weights of anon-traditional classification model, the non-traditional classification model trained to estimate, for a data sample and using a training dataset, a labeling probability that the data sample is labeled as positive, wherein the non-traditional classification model is trained using a first subset labeled as positive and a second subset unlabeled to estimate the labeling probability, and determine an estimated labeling frequency value corresponding to an upper bound on the labeling probability that a data sample corresponding to a positive class is labeled;
generate, for a new data sample of operational data corresponding to a photovoltaic panel as obtained by the one or more sensors, an augmented feature set by application of one or more non-linear transformations to a plurality of most influential features based on a magnitude of each weight of a plurality of weights of the non-traditional classification model, wherein the plurality of most influential features comprises atop k highest-weighted features selected based on the magnitude of the plurality of weights, and wherein the one or more non-linear transformations comprise adding p-level polynomial combinations of the top k highest-weighted features to obtain an expanded feature space;
apply the non-traditional classification model to the new data sample as represented in the expanded feature space to determine a labeling probability that the new data sample is labeled as positive; and
classify the photovoltaic panel into a first class or a second class using a fault classification model configured to classify based on the labeling probability and the estimated labeling frequency value of the non-traditional classification model corresponding to an upper bound on the labeling probability that a data sample corresponding to a positive class is labeled, wherein the first class is indicative of a fault and wherein the second class is indicative of no fault.
13 . The system of claim 12 , wherein the non-traditional classification model is a Modified Logistic Regression classifier.
14 . The system of claim 13 , wherein the Modified Logistic Regression classifier defines a weighted combination of a plurality of features, wherein each feature is associated with a respective weight of the plurality of weights of the non-traditional classification model.
15 . The system of claim 12 , wherein the processor as configured increases a dimensionality of each feature by adding p-level polynomial combinations of the top k highest-weighted features to the new data sample.
16 . The system of claim 12 , wherein the processor as configured minimizes one or more features of the expanded feature space that do not substantially contribute to classification using a dimensionality reduction algorithm.
17 . A method, comprising:
(a) accessing, at a processor in communication with a memory and one or more sensors in operative communication with a photovoltaic array wherein the one or more sensors are operable to capture operational data for the photovoltaic array and communicate the operational data to the processor, a training dataset comprising a plurality of data samples, each data sample having a plurality of features descriptive of sensor-derived operational data of a photovoltaic panel of the photovoltaic array, wherein a first subset of the plurality of data samples is labeled as positive and a second subset is unlabeled;
(b) accessing a plurality of weights of a non-traditional classification model, the non-traditional classification model being trained to estimate, for a data sample and using the training dataset, a labeling probability that the data sample is labeled as positive, wherein the non-traditional classification model is trained using the first subset labeled as positive and the second subset unlabeled to estimate the labeling probability, and determining an estimated labeling frequency value corresponding to an upper bound on the labeling probability that a data sample corresponding to a positive class is labeled;
(c) generating an augmented feature set based on the training dataset by application of one or more non-linear transformations to a plurality of most influential features of the plurality of features based on a magnitude of each weight of the plurality of weights of the non-traditional classification model, wherein the plurality of most influential features comprises atop k highest-weighted features selected based on the magnitude of the plurality of weights, and wherein the one or more non-linear transformations comprise adding p-level polynomial combinations of the top k highest-weighted features to obtain an expanded feature space;
(d) applying the non-traditional classification model to the training dataset as represented in the expanded feature space to estimate, for each data sample of the training dataset, the labeling probability that the data sample is labeled as positive;
(e) determining, for a new data sample of operational data corresponding to the photovoltaic panel as obtained by the one or more sensors and using the non-traditional classification model, a labeling probability that the new data sample is labeled as positive; and
(f) classifying the photovoltaic panel into a first class or a second class using a fault classification model configured to classify based on the labeling probability and the estimated labeling frequency value corresponding to an upper bound on the labeling probability that a data sample corresponding to a positive class is labeled, wherein the first class is indicative of a fault and wherein the second class is indicative of no fault.
18 . The method of claim 17 , wherein the non-traditional classification model is a Modified Logistic Regression classifier, wherein the Modified Logistic Regression classifier defines a weighted combination of the plurality of features, wherein each feature is associated with a respective weight of the plurality of weights.
19 . The method of claim 17 , wherein the processor as configured increases a dimensionality of each feature by adding p-level polynomial combinations of the top k highest-weighted features to the training dataset.
20 . The method of claim 17 , wherein the processor as configured minimizes one or more features of the expanded feature space that do not substantially contribute to classification using a dimensionality reduction algorithm.
21 . The method of claim 17 , wherein the processor as configured minimizes one or more features of the expanded feature space that do not substantially contribute to classification using a regularization algorithm.