Eli Lilly - Predictive Analytics for High- Frequency Data with Seeq_.pdf
Predictive Analytics for High- Frequency Data with Seeq Marco Vicentini
Eli Lilly
Eli Lilly – Manufacturing & Quality
Manufacturing & Quality are global functions supporting all Lilly medicines
and business units.
SINCE 2020,
INVESTED $55 BILLION TO BUILD, UPGRADE AND
ACQUIRE FACILITIES IN THE
U.S. AND EUROPE,
WE’VE COMMITTED TO
CREATING MORE THAN 3,000 NEW JOBS GLOBALLY
RELENTLESS FOCUS ON
INNOVATION IN M&Q.
~14,000 employees
20+ Internal
sites
50 CMOs
6000 end items
Our Mission: Safety First Quality Always
Manufacturing Processes Anomalies Detection & Diagnosis In a Nutshell
3
WHAT HOW WHY
Progress from Descriptive to Predictive Machines Performance Monitoring
Embedding Advanced Analytics in Machine Performance Data Analysis
To Support Reliable Supply
CAPACITY INCREASE
PROCESS STABILITY
DESCRIPTIVE ANALYTICS
What happened?
Why did it happen?
DIAGNOSTIC ANALYTICS
PREDICTIVE ANALYTICS
What will happen? PRESCRIPTIVE
ANALYTICS
What should we do?
Running Batch n
Changeover (Planned Downtime)
time
P ro
d u
ct io
n O
u tp
u t
0%
100%
Healthy Machine
Running Batch n
Changeover (Planned Downtime)
Changeover (Planned Downtime)
P ro
d u
ct io
n O
u tp
u t
0%
100%
Running Batch n+1
Component AnomalyHealthy Machine
time
Running Batch n
Changeover (Planned Downtime)
Changeover (Planned Downtime)
P ro
d u
ct io
n O
u tp
u t
0%
100%
Running Batch n+1 Running Batch n+2
Component Anomaly Component FaultHealthy Machine
Corrective Maintenance
(Unplanned Downtime)
time
Running Batch n
Changeover (Planned Downtime)
Changeover (Planned Downtime)
P ro
d u
ct io
n O
u tp
u t
0%
100%
Running Batch n+1 Running Batch n+2
Component Anomaly Component FaultHealthy Machine
Corrective Maintenance
(Unplanned Downtime)
time
Running Batch n
Changeover (Planned Downtime)
Changeover (Planned Downtime)
P ro
d u
ct io
n O
u tp
u t
0%
100%
Running Batch n+1 Running Batch n+2
Component Anomaly Component FaultHealthy Machine
Corrective Maintenance
(Unplanned Downtime)
Availability Loss
Capacity Loss
time
Running Batch n
Changeover (Planned Downtime)
Changeover (Planned Downtime)
P ro
d u
ct io
n O
u tp
u t
0%
100%
Running Batch n+1 Running Batch n+2
Component Anomaly Component FaultHealthy Machine
Corrective Maintenance
(Unplanned Downtime)
Availability Loss
Capacity Loss
EARLY DETECTION
LE VE
RA GI
NG
D IG
ITA L T
O. ..
time
Running Batch n
Changeover (Planned Downtime)
Changeover (Planned Downtime)
P ro
d u
ct io
n O
u tp
u t
0%
100%
Running Batch n+1 Running Batch n+2
Component Anomaly Component FaultHealthy Machine
Corrective Maintenance
(Unplanned Downtime)
Availability Loss
Capacity Loss
EARLY DETECTION
Proactive Maintenance
LE VE
RA GI
NG
D IG
ITA L T
O. ..
time
Running Batch n
Changeover (Planned Downtime)
Changeover (Planned Downtime)
P ro
d u
ct io
n O
u tp
u t
0%
100%
Running Batch n+1 Running Batch n+2
Component AnomalyHealthy Machine
Availability Loss
Capacity Loss
EARLY DETECTION
Proactive Maintenance
Healthy Machine
LE VE
RA GI
NG
D IG
ITA L T
O. ..
time
Running Batch n
Changeover (Planned Downtime)
Changeover (Planned Downtime)
P ro
d u
ct io
n O
u tp
u t
0%
100%
Running Batch n+1 Running Batch n+2
Component AnomalyHealthy Machine
EARLY DETECTION
Proactive Maintenance
Healthy Machine
LE VE
RA GI
NG
DI
GI TA
L T O…
GET STABLE PRODUCTION
OUTPUT time
time
…
Cycle #1 Cycle #2 Cycle #3 Cycle #4 Cycle #n
≈0.7 s
• COLLECT HIGH FREQUENCY COMPONENTS STATUS TIME SERIES (i.e. binary position signals, velocity, torque, force, absorbed current…)
• COMPUTE RELEVANT SIGNAL FEATURES EXTRACTION FROM EACH MACHINE CYCLE (i.e Stroke Duration (Cmd to Feedback): Time taken by the component to respond to the issued command, measured in milliseconds.…)
• TRAIN AN ANOMALY DETECTION MODEL FOR MACHINE COMPONENTS.
The discrete tags are monitored by measuring the time between a given command the feedback response.
Extended Period Overview: Stroke Duration over a broader time frame
Single Actuation Zoom: A focused view on a single actuation.
Daily Distribution: This solution aggregates the actuation of one day (tunable) into a distribution.
Seeq Approach
Actuation Time Distribution Distance Wasserstein Distance is able to quantify the “difference” between two distributions.
If the actuator does not change over time, the distance between the reference distribution and the daily aggregated distribution should remain within acceptable ranges (thresholds).
These thresholds are statistically calculated over a training period.
Day # 1 2 3 4 5
D is
tr ib
u ti
o n
D is
ta n
ce
Statistical Upper Control Limits Anomaly Detected
R 1 2 3 4 5
Python Code Power BI
For each combination of equipment, probe, process,
phase generate:
Condition Capsule Adjustment Metrics computing
Asset definition URL of predefined plot
Get Seeq data using python API.
Integrate Seeq data with data coming from
other sources
Store calculated features into a
database (relational in this
case)
Connect Power BI to the database
and show equipment status
Leverage Seeq Assets and Seeq
plotting/calculation capabilities to allow
quick deepdiving
How to scale up with Seeq
The approach of measuring stroke durations between high-frequency signals faces significant scalability and computational
challenges. Considering the number of actuators and machine components to monitor across multiple machines in a production line, along with the high number of daily actuations, the complexity of the problem increases exponentially.
This high number of distances to be monitored poses a scalability problem that has been solved with the following architecture:
Bottom-Up Approach In Action: From Reactive To Predictive Journey Discrete Manufacturing Components Anomalies Detection & Diagnosis
Maintenance Work Order & Alarms Correlation
Distance Trend
MACHINE X1>>STATION Y1>>ACTUATOR Z1
Key Results
Reduced Unexpected Failures Reduction of unplanned maintenance activities
Early Detection of Faulty Components:
• Scheduled maintenance spread over time.
• Optimal timing for component replacements.
Capacity Increase
Thank you
Marco Vicentini
Eli Lilly
Slide 1: Predictive Analytics for High-Frequency Data with Seeq Slide 2: Eli Lilly – Manufacturing & Quality Slide 3: Manufacturing Processes Anomalies Detection & Diagnosis In a Nutshell Slide 4 Slide 5 Slide 6 Slide 7 Slide 8 Slide 9 Slide 10 Slide 11 Slide 12 Slide 13 Slide 14: Seeq Approach Slide 15: Actuation Time Distribution Distance Slide 16: How to scale up with Seeq Slide 17 Slide 18: Key Results Slide 19: Thank you