5 Machine Learning for IoT
255
Classificatio
n
Normal
Abnormal
What is my health status?
Predictive maintenance: when will the wind turbine fail?
Regression
In next 36 hours
Clustering
Anomaly detection: cluster motors based on sensor readouts
Which cluster ID
Fig. 5.7 A few use cases of machine learning in IoT
•
Faults are reported by end-user
•
Afterwards, inventory and the team
should be scheduled and dispatched
Traditional Corrective Maintenance
•
Faults are detected by connected
sensors in near real-time
•
Afterwards, inventory and the team
should be scheduled and dispatched
Real-time Monitoring (IoT)
•
Faults are predicted before they
really occur
•
There is enough time to schedule the
team and inventory in advance
Machine Learning
(Predictive Maintenance)
Time:
Time:
v
Time:
v
Fig. 5.8 A simple example of predictive maintenance in the energy industry
• Regression (Supervised): Predictive maintenance (PdM) is a cutting-edge maintenance strategy, which has been adopted in several domains (e.g., manufacturing,
energy and supply chain). The key idea of PdM is to identify which equipment
needs maintenance and which component will fail in the future and to predict
the remaining useful life (RUL) of machine parts (see Fig. 5.8). In this context,
regression techniques can be utilized to predict RULs accurately.
• Clustering (Unsupervised): An example of unsupervised learning in IoT is data
processing in a factory producing car engines. Suppose that we want to design a
machine to detect engines that require further adjustments. It is almost impossible
to build a system to detect the defects visually, but this can be achieved by
collecting several key parameters from each engine and then using a clustering
algorithm to find groups/clusters. For example, if the parameters are temperature
and the produced sound, the clustering algorithm (e.g., K-means clustering) will
group the engines into different categories based on their similarity in producing
sound at a specific range of temperature (Fig. 5.9). This will help engineers of
that factory to detect the engines that belong to the problematic group quickly.
255
Classificatio
n
Normal
Abnormal
What is my health status?
Predictive maintenance: when will the wind turbine fail?
Regression
In next 36 hours
Clustering
Anomaly detection: cluster motors based on sensor readouts
Which cluster ID
Fig. 5.7 A few use cases of machine learning in IoT
•
Faults are reported by end-user
•
Afterwards, inventory and the team
should be scheduled and dispatched
Traditional Corrective Maintenance
•
Faults are detected by connected
sensors in near real-time
•
Afterwards, inventory and the team
should be scheduled and dispatched
Real-time Monitoring (IoT)
•
Faults are predicted before they
really occur
•
There is enough time to schedule the
team and inventory in advance
Machine Learning
(Predictive Maintenance)
Time:
Time:
v
Time:
v
Fig. 5.8 A simple example of predictive maintenance in the energy industry
• Regression (Supervised): Predictive maintenance (PdM) is a cutting-edge maintenance strategy, which has been adopted in several domains (e.g., manufacturing,
energy and supply chain). The key idea of PdM is to identify which equipment
needs maintenance and which component will fail in the future and to predict
the remaining useful life (RUL) of machine parts (see Fig. 5.8). In this context,
regression techniques can be utilized to predict RULs accurately.
• Clustering (Unsupervised): An example of unsupervised learning in IoT is data
processing in a factory producing car engines. Suppose that we want to design a
machine to detect engines that require further adjustments. It is almost impossible
to build a system to detect the defects visually, but this can be achieved by
collecting several key parameters from each engine and then using a clustering
algorithm to find groups/clusters. For example, if the parameters are temperature
and the produced sound, the clustering algorithm (e.g., K-means clustering) will
group the engines into different categories based on their similarity in producing
sound at a specific range of temperature (Fig. 5.9). This will help engineers of
that factory to detect the engines that belong to the problematic group quickly.