The blockchain part serves three main functions:
1. immutable time-chain of transactions between users and customers for analytics results
2. aggregate of authenticated data sources
3. digital market place for data analytics between trusted parties
(details on the right of diagram)
GCP services like DataFlow and Pub/Sub encrypt both in transit and at rest. Customer managed encryption plus pinning certs with the ATECC608a on the DPU are additional. Proof of concepts can be seen here for BigQuery and here for DataFlow at Google GCP blog.
We use Cloud ML engine for machine learning and GCP IoT Core for provisioning. “Other Cloud Services” hosts the blockchain part; having hosted ERC20 and a Nano node, we will probably go with Guardtime or Google Cloud Spanner.
The Data Processing Unit (DPU) fits into a standard circuit breaker or can be integrated into OEM products for fittings in walls or ceilings:
Q. What was the motivation for making the prototype?
A. After working with data scientists from different companies for a few years, they taught us what clean data means. Here a summary of their 7 pain points:
1 IIoT data is generally not time stamped at source
“Even basic time series models are not really possible”
2 IIoT clusters are not synced
“Time sync deltas between sensor readings at different locations, even just 70m², means significant correlation errors - garbage in, garbage out!”
3 IIoT is not contextual
“Data context from the field cannot be instantly understood or classified”
4 IIoT data can’t be sequenced for ML / AI
(consequence of the above 3 issues!)
5 IIoT data formats are not standardized
“Little compatibility between industrial hardware vendors, data types or clouds”
6 IIoT systems have high latency
“Takes too long to determine anomalous patterns (hours instead of milliseconds)”
7 IIoT data is not clean
“Data needs 3 days prep for every 2 days analysis per week”
“Industrial hardware data sources and real time data processing at scale still costs too much for just getting to know the value of the data for customer solutions”
The first customer to consider in any data driven product (hardware or software) is the data scientist.
“How are we supposed to turn raw data into gold for “customers” if the raw data is dirty?” a pain DAX data scientists often express for IIoT.
The DPU is a first attempt to produce high octane real time data for the scientists, machine learning systems or AI.
- Synchronizes data points in real time at the same time from 14 vectors
- Material cost less than 90 EUR per DPU (bill of material)
- Each DPU generates 2.5 Mio “data science ready” points/day
Small batches of DPU reference designs are produced at a secure automated facility with Future Lab. Derivative products can then manufactured at scale by pre-certifying the reference design.
GCP offers superb AI and ML plugins provided the data sources are precise.
Example: Real time energy analysis at low cost. RMS current of each phase, Power Factors, Apparent Power and Voltage:
Energy usage is analysed with 9 other vectors: both DPU and GCP instances are synced to atomic clocks
Energy events can then be correlated
Note change in energy graphs at 10:15 with CO² spike from cooking. The two DPUs are 10 metres distant from each other.
Sand-boxed customer labs will scale to 100Mio IIoT data points per day in the next six months (approx 20% of world Twitter traffic)
DPUs run at high resolution to statistically catch every correlation
Once analytics are done, they are re-programmed in the ARM M4 edge chip (with FPU)
Final products made from the reference design only then need 1/100th of the data rate once a good balance between edge and cloud analytics is determined depending on customer needs
Google this code:
93E3 BEBC C164 D766
(with spaces) to contact
auditable identity here
I don’t use LinkedIn or Facebook