Automated Electric Submersible Pump (ESP) failure analysis
A method, apparatus, and program product may utilize data associated with one or more electric submersible pumps (ESPs) to train a machine learning model and/or use a machine learning model to perform ESP failure analysis. In addition, one or more features from the data may be encoded into a machine-readable format to facilitate ingestion by the machine learning model.
1 . A method for automating root cause identification of electric submersible pump (ESP) failures post-breakdown, the method comprising:
receiving data associated with at least one ESP, the received data including numerical features, categorical features, and textual features;
encoding the numerical features, the categorical features, and the textual features of the received data into a machine-readable format, the encoding including:
encoding the numerical features using a data engineering pipeline to generate standardized numerical values;
encoding the categorical features using one-hot vector encoding or cardinality encoding based on a threshold comparison against a number of categories; and
encoding the textual features by:
matching one or more non-repeated words for a first textual feature of the textual features to a set of pretrained oil and gas specific word embeddings to create a first matrix of vocabulary words and corresponding word embeddings;
classifying failure components using the first matrix to generate a second matrix of trained word embeddings;
matching the second matrix to one or more tokenized sentences; and
reducing dimensionality of the matched second matrix to generate a third compressed embeddings matrix,
wherein the classifying the failure components is performed with a bi-directional Long Short Term Memory (LSTM) recurrent neural network, and
wherein the reducing the dimensionality of the matched second matrix is performed with an autoencoder neural network; and
predicting at least one failure component of the at least one ESP using the encoded numerical features, the encoded categorical features, and the encoded textual features with a trained multi-input multinomial classification model,
wherein the multi-input multinomial classification model has been trained using second numerical features, second categorical features, and second textual features from at least one other ESP.
2 . The method of claim 1 , wherein the received data further includes one or more of: Dismantle Inspect Failure Analysis (DIFA) tracker data, well tracker data, or DIFA reports.
3 . The method of claim 1 , wherein the received data further includes one or more of: a pull reason, findings, installation dates, failure dates, pull dates, sensor information, motor information, protector information, immediate cause information, failure component findings, a general pull reason, a general failed component, a general failed category, a general failure cause, a specific pull reason, a specific failed component, a specific failure category, or a specific failure cause.
4 . The method of claim 1 , further comprising processing the received data with a data engineering pipeline to process, clean, and standardize the received data.
5 . The method of claim 1 , wherein the trained multi-input multinomial classification model comprises an ensemble of multiple machine learning algorithms.
6 . The method of claim 5 , wherein the ensemble of multiple machine learning algorithms includes at least two of: an XGBoost algorithm, an AdaBoost algorithm, a Random Forest algorithm, a Multinomial Logistic Classifier algorithm, or a Neural Network algorithm.
7 . The method of claim 1 , wherein:
the received data is first received data; and
the method further comprises:
receiving second data associated with one or more additional ESPs; and
processing the received second data using the trained multi-input multinomial classification model to predict at least one failure component of the one or more additional ESPs.
8 . An apparatus, comprising:
a computing system including one or more processors; and
a memory storing program code that, upon execution by the one or more processors, causes the apparatus to perform the method of claim 1 .
9 . A non-transitory computer-readable medium having program code stored thereon, the program code configured to, upon execution by a computing system including one or more processors, cause the one or more processors to perform the method of claim 1 .
10 . A method for automating root cause identification of electric submersible pump (ESP) failures post-breakdown, the method comprising:
receiving data associated with at least one ESP, the received data including numerical features, categorical features, and textual features;
encoding the numerical features, the categorical features, and the textual features of the received data into a machine-readable format, the encoding including:
encoding the numerical features using a data engineering pipeline to generate standardized numerical values;
encoding the categorical features using one-hot vector encoding or cardinality encoding based on a threshold comparison against a number of categories; and
encoding the textual features by:
matching one or more non-repeated words for a first textual feature of the textual features to a set of pretrained oil and gas specific word embeddings to create a first matrix of vocabulary words and corresponding word embeddings;
classifying failure components using the first matrix to generate a second matrix of trained word embeddings;
matching the second matrix to one or more tokenized sentences; and
reducing dimensionality of the matched second matrix to generate a third compressed embeddings matrix,
wherein the classifying the failure components is performed with a bi-directional Long Short Term Memory (LSTM) recurrent neural network, and
wherein the reducing the dimensionality of the matched second matrix is performed with an autoencoder neural network; and
predicting at least one failure component class or category of the at least one ESP using the encoded numerical features, the encoded categorical features, and the encoded textual features with a trained multi-input multinomial classification model.
11 . The method of claim 10 , wherein the trained multi-input multinomial classification model includes an ensemble of multiple machine learning algorithms configured to receive at least a portion of the numerical features, the categorical features, and the textual features in the received data.
12 . A method for automating root cause identification of electric submersible pump (ESP) failures post-breakdown, the method comprising:
receiving data associated with at least one ESP, the data including numerical features, categorical features, and textual features;
encoding the numerical features, the categorical features, and the textual features of the received data into a machine-readable format, the encoding including respective pathways for each of the numerical features, the categorical features, and the textual features, the pathways including:
for the numerical features, encoding the numerical features using a data engineering pipeline to generate standardized numerical values;
for the categorical features, encoding the categorical features using one-hot vector encoding or cardinality encoding based on a threshold comparison against a number of categories; and
for the textual features, encoding the textual features by:
matching one or more non-repeated words for a first textual feature of the textual features to a set of pretrained specific word embeddings to create a first matrix of vocabulary words and corresponding word embeddings;
classifying failure components using the first matrix to generate a second matrix of trained word embeddings;
matching the second matrix to one or more tokenized sentences; and
reducing dimensionality of the matched second matrix to generate a third compressed embeddings matrix,
wherein the classifying the failure components is performed with a bi-directional Long Short Term Memory (LSTM) recurrent neural network, and
wherein the reducing the dimensionality of the matched second matrix is performed with an autoencoder neural network; and
predicting at least one failure component of the at least one ESP using the encoded numerical features, the encoded categorical features, and the encoded textual features with a trained multi-input multinomial classification model.
13 . An apparatus for automating root cause identification of electric submersible pump (ESP) failures post-breakdown, comprising:
a computing system including one or more processors; and
a non-transitory computer-readable medium storing program code that, upon execution by the one or more processors, causes the apparatus to:
receive data associated with at least one ESP, the received data including numerical features, categorical features, and textual features;
encode the numerical features, the categorical features, and the textual features of the received data into a machine-readable format, the encoding including:
encoding the numerical features using a data engineering pipeline to generate standardized numerical values;
encoding the categorical features using one-hot vector encoding or cardinality encoding based on a threshold comparison against a number of categories; and
encoding the textual features by:
matching one or more non-repeated words for a first textual feature of the textual features to a set of pretrained oil and gas specific word embeddings to create a first matrix of vocabulary words and corresponding word embeddings;
classifying failure components using the first matrix to generate a second matrix of trained word embeddings;
matching the second matrix to one or more tokenized sentences; and
reducing dimensionality of the matched second matrix to generate a third compressed embeddings matrix,
wherein the classifying the failure components is performed with a bi-directional Long Short Term Memory (LSTM) recurrent neural network, and
wherein the reducing the dimensionality of the matched second matrix is performed with an autoencoder neural network; and
predict at least one failure component of the at least one ESP using the encoded numerical features, the encoded categorical features, and the encoded textual features with a trained multi-input multinomial classification model,
wherein the multi-input multinomial classification model has been trained using second numerical features, second categorical features, and second textual features from at least one other ESP.
14 . The apparatus of claim 13 , wherein the received data further includes one or more of: Dismantle Inspect Failure Analysis (DIFA) tracker data, well tracker data, or DIFA reports.
15 . The apparatus of claim 13 , wherein the received data further includes one or more of: a pull reason, findings, installation dates, failure dates, pull dates, sensor information, motor information, protector information, immediate cause information, failure component findings, a general pull reason, a general failed component, a general failed category, a general failure cause, a specific pull reason, a specific failed component, a specific failure category, or a specific failure cause.
16 . The apparatus of claim 13 , wherein the program code further causes the apparatus to generate a data engineering pipeline configured to process, clean, and standardize the received data.
17 . The apparatus of claim 13 , wherein the trained multi-input multinomial classification model comprises an ensemble of multiple machine learning algorithms.
18 . The apparatus of claim 17 , wherein the ensemble of multiple machine learning algorithms includes at least two of: an XGBoost algorithm, an AdaBoost algorithm, a Random Forest algorithm, a Multinomial Logistic Classifier algorithm, or a Neural Network algorithm.
19 . The apparatus of claim 13 , wherein:
the received data is first received data; and
the program code further causes the apparatus to:
receive second data associated with one or more additional ESPs; and
process the received second data using the trained multi-input multinomial classification model to predict at least one failure component of the one or more additional ESPs.
20 . An apparatus for automating root cause identification of electric submersible pump (ESP) failures post-breakdown, comprising:
a computing system including one or more processors; and
a memory storing program code that, when executed by the one or more processors, causes the one or more processors to:
process, clean, and standardize received data associated with at least one ESP, the received data including numerical features, categorical features, and textual features;
encode the numerical features, the categorical features, and the textual features of the received data into a machine-readable format, the encoding including:
encoding the numerical features using a data engineering pipeline to generate standardized numerical values;
encoding the categorical features using one-hot vector encoding or cardinality encoding based on a threshold comparison against a number of categories; and
encoding the textual features by:
matching one or more non-repeated words for a first textual feature of the textual features to a set of pretrained oil and gas specific word embeddings to create a first matrix of vocabulary words and corresponding word embeddings;
classifying failure components using the first matrix to generate a second matrix of trained word embeddings;
matching the second matrix to one or more tokenized sentences; and
reducing dimensionality of the matched second matrix to generate a third compressed embeddings matrix,
wherein the classifying the failure components is performed with a bi-directional Long Short Term Memory (LSTM) recurrent neural network, and
wherein the reducing the dimensionality of the matched second matrix is performed with an autoencoder neural network; and
predict at least one failure component of the at least one ESP using the encoded numerical features, the encoded categorical features, and the encoded textual features with a trained multi-input multinomial classification model,
wherein the multi-input multinomial classification model has been trained using second numerical features, second categorical features, and second textual features from at least one other ESP.