559
2.3 Supervised and unsupervised learning
Machine Learning (ML) can be described as an application of Artificial Intelligence (AI),
based on the idea that algorithms can autonomously “learn” from data in order to make
predictions or help decision-making without being explicitly programmed. Predictive maintenance algorithms development by Newtrax take advantage of two general categories of
machine learning algorithms; supervised and unsupervised learning. The supervised approach
is based on the idea that by training the algorithms with labeled data, it will “learn” to differentiate between the characteristics of the different labels and acquire the ability to predict the
right label for new entries. A classic example of this is character recognition, where a model
is trained on labeled images of character e.g. a picture of a handwritten four labeled “4”, to
develop the ability to read number and letters from pictures. See (Yalcın 2018) for a simple
and visual example. Applications of supervised learning to predictive maintenance can be
found in the work of (Li et al. 2014) and (Susto et al. 2015).
Unsupervised learning, on the other hand, is based on the idea that by simply looking at
the unlabeled data, the model will be able to split entries in different categories based on the
similarity of their characteristics. A common example of this is in customer segmentation,
where these algorithms are used to regroup customers into a limited number of categories in
order to develop tailored marketing strategies for each of them. (Vivek 2018) display simple
and visual example of the problem. An application to predictive maintenance is illustrated in
the article from (Beghi et al. 2014).
Each of these approaches has advantages and inconveniences and cannot be applied in all
situations. For instance, supervised techniques generally require a lot of labeled data for their
training, which can pose problems in the case where few or no data are available for certain
labels or can require a lot of preliminary work only to label the data. That being said, if unsupervised techniques do not require any label, their results are generally much more uncertain
since the characteristics of each group formed cannot be known in advance.
3 MODELS AND RESULTS
This section will discuss the different techniques used by Newtrax in the context of predictive maintenance and some of its applications. Before they are put into the machine learning
algorithms, the different channels of information undergo a series of transformations specific to each algorithm including but not limited to: rolling window mean/max/min/standard
deviation, categorization, Principal Component Analysis (PCA), standardization and special
events identification.
3.1 Supervised learning
Newtrax uses supervised learning for specific events detection that have already been observed
in historical data e.g. engine or transmission failures. The labeling for the various equipment
failure is derived from maintenance logs supplied by our clients. The algorithms used for
this type of task are classifier, meaning that they classify the real time data coming from the
equipment in different categories. Support Vector Machines (SVM) and Neural Network
(NN) are used for this type of application. Put in really simple terms, the way SVM works
is that it finds the best hyperplane (e.g. a line in two dimensions) that separates the points
from different categories based on the training data. Once trained, the algorithm makes predictions for new data by simply looking at which side of these hyperplanes the new points
are located. As for neural networks, they make use of a network of mathematical “neurons”
and find the best setting for these neuron parameters by optimization with the training data.
Predictions are made by passing new data through this optimized neuron network which
returns an answer.
One of our most prominent result in supervised learning was in the area of truck engine
failure detection. The engine, taken as a whole, is definitely the most critical component of a
haulage truck since a failure will stop the truck for extended periods of time and require costly
2.3 Supervised and unsupervised learning
Machine Learning (ML) can be described as an application of Artificial Intelligence (AI),
based on the idea that algorithms can autonomously “learn” from data in order to make
predictions or help decision-making without being explicitly programmed. Predictive maintenance algorithms development by Newtrax take advantage of two general categories of
machine learning algorithms; supervised and unsupervised learning. The supervised approach
is based on the idea that by training the algorithms with labeled data, it will “learn” to differentiate between the characteristics of the different labels and acquire the ability to predict the
right label for new entries. A classic example of this is character recognition, where a model
is trained on labeled images of character e.g. a picture of a handwritten four labeled “4”, to
develop the ability to read number and letters from pictures. See (Yalcın 2018) for a simple
and visual example. Applications of supervised learning to predictive maintenance can be
found in the work of (Li et al. 2014) and (Susto et al. 2015).
Unsupervised learning, on the other hand, is based on the idea that by simply looking at
the unlabeled data, the model will be able to split entries in different categories based on the
similarity of their characteristics. A common example of this is in customer segmentation,
where these algorithms are used to regroup customers into a limited number of categories in
order to develop tailored marketing strategies for each of them. (Vivek 2018) display simple
and visual example of the problem. An application to predictive maintenance is illustrated in
the article from (Beghi et al. 2014).
Each of these approaches has advantages and inconveniences and cannot be applied in all
situations. For instance, supervised techniques generally require a lot of labeled data for their
training, which can pose problems in the case where few or no data are available for certain
labels or can require a lot of preliminary work only to label the data. That being said, if unsupervised techniques do not require any label, their results are generally much more uncertain
since the characteristics of each group formed cannot be known in advance.
3 MODELS AND RESULTS
This section will discuss the different techniques used by Newtrax in the context of predictive maintenance and some of its applications. Before they are put into the machine learning
algorithms, the different channels of information undergo a series of transformations specific to each algorithm including but not limited to: rolling window mean/max/min/standard
deviation, categorization, Principal Component Analysis (PCA), standardization and special
events identification.
3.1 Supervised learning
Newtrax uses supervised learning for specific events detection that have already been observed
in historical data e.g. engine or transmission failures. The labeling for the various equipment
failure is derived from maintenance logs supplied by our clients. The algorithms used for
this type of task are classifier, meaning that they classify the real time data coming from the
equipment in different categories. Support Vector Machines (SVM) and Neural Network
(NN) are used for this type of application. Put in really simple terms, the way SVM works
is that it finds the best hyperplane (e.g. a line in two dimensions) that separates the points
from different categories based on the training data. Once trained, the algorithm makes predictions for new data by simply looking at which side of these hyperplanes the new points
are located. As for neural networks, they make use of a network of mathematical “neurons”
and find the best setting for these neuron parameters by optimization with the training data.
Predictions are made by passing new data through this optimized neuron network which
returns an answer.
One of our most prominent result in supervised learning was in the area of truck engine
failure detection. The engine, taken as a whole, is definitely the most critical component of a
haulage truck since a failure will stop the truck for extended periods of time and require costly