Analytics at Industrial Internet of Things
.. let me recap what is Industrial Internet of Things (IIoT). Industrial IoT is a branch of IoT including Asset Management, Remote Monitoring, Condition Based Maintenance, Service Optimisation andManufacturing Operations applied to Industry.
We list some typical uses cases in the Industrial IoT:
●Highly customized products; the goal is to have the single consumer customize the product according to his needs
●Asset Performance Management: Industrial equipment monitored and their performance controlled
●Remote or unmanned operation
●Condition Based Maintenance: Equipment maintenance not according a fixed schedule but according to their real condition; less cost and downtime with more productivity
●Personalized services and new business model
IIoT differs from consumers IoT (Smart, Home, Smart City, Smart Transportation, Smart Grid) due to the already in place system for industry (legacy), the long life cycle, the numbers of signals and the deterministic approach.
Recently we held a live session (organized by packt) in which we provided an introduction to IIoT. You can find more details on our book:
https://www.amazon.com/Hands-Industrial-Internet-Things-infrastructure-ebook/dp/B07HH9Y5PY
Flow of IIoT can be summarised by the following slide:
1)The fist pillar is the Factory where the signals and production data are generated and locally stored
2)The second one is the Edge; the Edged is connected from one side to the industrial data sources for gathering the data and on the other side to the cloud for transferring them through a secure channel over the internet
3)The Cloud is the third one; The Cloud provides storage, computing resources, and API for administrative purposes
4)The fourth pillar are the orchestration services and collaborative microservices available in the cloud for processing and analysing the data
5)The last one are applications living in the cloud and providing dashboards and drill-drawn capabilities to visualize the data from different perspective according the need of the different type of users.
The IIoT is based on five pillars: collecting, sending, storing, processing (thru analytics) and visualising data.
Analytics are the most advanced and interesting part of IIoT. On IIoT we can identify 6 classes of analytics
●3 are general purpose and can be purely data driven (ML): Anomaly Detection, RUL and Production Prediction: these
●3 are industry specific. It means that change from industry to industry: Predictive and Preventive Maiantenance, KPI Performance indicator, Product Optimization
Anomaly Detection
Anomaly Detection algorithms can be implemented supervised or unsupervised:
- Supervised: you know the anomaly and you try to recognise (also called Known Anomaly)
- Unsupervised: you classify as “anomaly” an unexpected behaviour (also called Novelty Detection)
You can implement Anomaly Detection using various unsupervised algorithms. The most common is the IsolationForest :
from sklearn.ensemble import IsolationForest# training the modelclf = IsolationForest(max_samples=100, random_state=rng)clf.fit(X_train)# predictionsy_pred_outliers = clf.predict(X_outliers)
…but remember that:
90% of what you want can be reached using threshold plus heuristic formulas
Production Prediction
Production Prediction algorithm uses regression algorithms or more complex algorithms to predict the performance accordingly to weather business etc.
The most common algorithm is the ARIMA (Autoregressive integrated moving average)
In this figure you ca see a production of a chemical dept during the last 60 days, the prediction vs the real production 20days before
Can be purely Data Driven but for better performance we need a model of the process
Residual Useful Life
Let me recap maintenance in industry …. this is a slide from Mathworks: there are 3 type od strategies: reactive, preventive predictive
●Reactive: failure then we repair, it’s our son’s bicycle maintenance: no so smart.
●Preventive: is the maintenance of our car, periodically we perform maintenance also if we do not need it, its more or less adopted by every industry. It is the standard maintenance policy
●Predictive: in Industry 4.0 we can implement predictive maintenance using IOT and data analytics: in other words we repair only if we need
On RUL, normally we estimate the degradation during time, given the last samples and a model of the equipment, then we change parameters of the model to estimate the uncertainty; when the degradation exceeds the threshold (point F) we calculate the so called time to failure or time to trip (point P).
A RUL is the time between time P and the expected time to failure plus an uncertainty quantification, such as a probability distribution
Optimisation
Optimisation is not properly an IOT algorithm, but can be improved thanks to IOT DATA. You build the model from data, you define the cost function, finally you use a simple optimisation algorithm to minimize the cost function
Some example of optimisation can be
- Multi Components optimisation: for instance on a chemical plant
- Emission/energy consumption reduction
KPIs
KPI is good for a panel operator who controls continuously the plant. KPI can be also a business measure for instance the number of open cases, the level of production and so on, but in the case of IIOT, we build the model from data or we calculate the performance from data then we show or we can decide to compare with a similar asset. For instance we want to see our pump in china how it is performing compared to the pump in Russia.
Sometime KPI is a simple threshold or an heuristic rule applied to data.
Predictive or Preventive Maintenance
Predictive or Preventive Maintenance algorithm is based in RUL or similar, from RUL you compare the expected maintenance with the programmed maintenance and you try to optimize it.
The best predictive maintenance is the so called CBM (conditioning based maintenance) where we extended the maintenance based on predictive maintenance.
CBM is the most interesting and profitable algorithm on Industry…
Implementation
The following picture summarises, the algorithms you can use to implement the six classes of analytics:
MIMOSA
This picture depicts the flow of our “data processing”
This picture is an adaptation from the MIMOSA OSA-CBM standard. From left to right:
●data is acquired then the signal is filtered and preprocessed.
●then we can apply simple rule for condition monitoring and advanced rules such as anomaly detection to check health of the system or RUL.
● finally we can apply Prognostic and Decision Making algorithms to suggest insights to the operators.
At the end you need to visualize the data.
I hope you enjoyed this reading, please reach me at
