Towards Additive Manufacturing Process Control using Semi- - - PowerPoint PPT Presentation

▶
towards additive manufacturing process control using semi
SMART_READER_LITE
LIVE PREVIEW

Towards Additive Manufacturing Process Control using Semi- - - PowerPoint PPT Presentation

Sep 21, 2023 457 likes •862 views

Towards Additive Manufacturing Process Control using Semi- Supervised Learning Dr Ikenna A Okaro Miss Sarini Jayasinghe Dr Kate Black Prof Chris Sutcliffe Dr Paolo Paoletti Dr Peter L Green 26-06-18 People Ikenna A Okaro Project PDRA

slide-1
SLIDE 1

Towards Additive Manufacturing Process Control using Semi- Supervised Learning

Dr Ikenna A Okaro Miss Sarini Jayasinghe Dr Kate Black Prof Chris Sutcliffe Dr Paolo Paoletti Dr Peter L Green 26-06-18

slide-2
SLIDE 2

People

Ikenna A Okaro Project PDRA Sarini Jayasinghe PhD Student Kate Black Lecturer Chris Sutcliffe Professor and Renishaw Contact Paolo Paoletti Senior Lecturer

slide-3
SLIDE 3

Motivation

  • Additive Manufacturing is revolutionising UK industry.
  • Potential in more risk-averse sectors (aerospace, healthcare etc.)
  • We must de-risk AM technology to maximise its impact.
  • Current issues stem from a lack of process control.

Can machine learning help us to pioneer robust process control for Additive Manufacturing ?

slide-4
SLIDE 4

Certification ? Uncertain part quality hinders the adoption of AM in aerospace and medical sectors

slide-5
SLIDE 5

Certification ? Uncertain part quality hinders the adoption of AM in aerospace and medical sectors Process measurements Machine Learning? Train an algorithm to identify faulty components from AM process measurements.

slide-6
SLIDE 6

Certification ? Uncertain part quality hinders the adoption of AM in aerospace and medical sectors Process measurements Machine Learning? Train an algorithm to identify faulty components from AM process measurements. 100s of parts need to be manually certified before the algorithm can be

  • trained. This is far too expensive for AM.
slide-7
SLIDE 7

Certification ? Uncertain part quality hinders the adoption of AM in aerospace and medical sectors Process measurements Machine Learning? Train an algorithm to identify faulty components from AM process measurements. 100s of parts need to be manually certified before the algorithm can be

  • trained. This is far too expensive for AM.

Process measurements Semi-supervised learning (SSL): large amounts of ‘unlabelled data’ and small amounts of ‘labelled data’. Applied to new data sets, from Renishaw AM machines.

slide-8
SLIDE 8

Certification ? Uncertain part quality hinders the adoption of AM in aerospace and medical sectors Process measurements Machine Learning? Train an algorithm to identify faulty components from AM process measurements. 100s of parts need to be manually certified before the algorithm can be

  • trained. This is far too expensive for AM.

Process measurements Semi-supervised learning (SSL): large amounts of ‘unlabelled data’ and small amounts of ‘labelled data’. Applied to new data sets, from Renishaw AM machines.

DID IT WORK?

slide-9
SLIDE 9

Methodology

  • 2 SLM builds, each of 25 tensile test bars.
  • During each build we measure:
  • Back reflected light (2 photodiodes, infrared and visible)
  • Laser position
  • 400GB of data per build (!)
  • Conducted tensile tests of each specimen.
slide-10
SLIDE 10

Methodology

  • 1. Label each bar as ‘good’ or ‘bad’ depending on tensile test results.
  • 2. Extract the measurements that relate to each test bar.
  • 3. Extract statistically significant indicators of build quality from photodiode

measurements (called features).

  • 4. Semi-supervised learning applied to features.
  • Can we identify faulty components?
  • Does semi-supervised learning reduce the number of certification

experiments needed?

slide-11
SLIDE 11

Methodology

  • 1. Label each bar as ‘good’ or ‘bad’ depending on tensile test results.
  • 2. Extract the measurements that relate to each test bar.
  • 3. Extract statistically significant indicators of build quality from photodiode

measurements (called features).

  • 4. Semi-supervised learning applied to features.
  • Can we identify faulty components?
  • Does semi-supervised learning reduce the number of certification

experiments needed?

Feature extraction - high risk part of the project!

slide-12
SLIDE 12
  • 1. Labelling each specimen
200 400 600 800 1000 1200 1400 1600 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19

Stress [ MPa] St rain [ %]

y

Specim en # 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
  • Ultimate Tensile Strength > 1400MPa labelled as ‘good’.
  • A little arbitrary but sufficient for a feasibility study.
  • Fatigue tests and/or CT scans will be used in the future.
slide-13
SLIDE 13
  • 2. Extract measurements relating to each bar
  • Figure shows X-Y coordinates of a single layer.
  • We identify the photodiode measurements that are obtained when the laser

is in a red region.

  • We also omit data obtained when the laser is not running.
slide-14
SLIDE 14
  • 3. Feature Extraction (per Photodiode)
slide-15
SLIDE 15
  • 3. Feature Extraction (per Photodiode)

Bar 1 Bar 2 Layer 1 Layer 2 Layer 3 …

  • …
  • Singular Value Decomposition (SVD)
  • Each vector can be written as a linear

combination of basis vectors.

  • …
slide-16
SLIDE 16
  • 3. Feature Extraction (per Photodiode)

Bar 1 Bar 2 … Layer 1 Layer 2 Layer 3 …

  • …
  • Singular Value Decomposition (SVD)
  • Each vector can be written as a linear

combination of basis vectors.

  • Basis vectors

Constants

slide-17
SLIDE 17
  • 3. Feature Extraction (per Photodiode)

Bar 1 Bar 2 Layer 1 Layer 2 Layer 3 …

  • …
  • Singular Value Decomposition (SVD)
  • Each vector can be written as a linear

combination of basis vectors.

  • ∗
  • ∗ 25
  • …
slide-18
SLIDE 18
  • 3. Feature Extraction (per Photodiode)

Bar 1 Bar 2 Layer 1 Layer 2 Layer 3 …

  • …
  • Singular Value Decomposition (SVD)
  • Each vector can be written as a linear

combination of basis vectors.

  • ∗
  • ∗ 25

These constants become our features

  • …
slide-19
SLIDE 19
  • 3. Feature Extraction (per Photodiode)
  • Including both builds, we have a 50 by 3 10 data matrix.
  • Computational cost prevents a standard SVD being applied here.
  • We (Sarini!) circumvented this issue using methods form Big Data

analytics: Probabilistic Singular Value Decomposition.

  • For the feasibility study, we kept just 1 basis vector per photodiode.
  • 2 photodiodes => 2 dimensional feature space.
slide-20
SLIDE 20
  • 3. Feature Extraction (per Photodiode)

These figures give an impression of the ‘information lost’ by only projecting onto 1 basis vector. Projecting onto more basis vectors increases the dimensionality of the feature space. This trade-off can be investigated in the future.

slide-21
SLIDE 21
  • 4. Semi-Supervised Machine Learning

Our 2D feature space. Green and red represent ‘good’ and ‘bad’ specimens respectively.

slide-22
SLIDE 22

To investigate the semi-supervised approach, we delete half of our labels. These are chosen at random.

  • 4. Semi-Supervised Machine Learning
slide-23
SLIDE 23
  • 4. Semi-Supervised Machine Learning

We fit our Gaussian Mixture Model. In this case, ‘bad’ specimens were identified with a 77% success rate. Above 75% was a key project objective.

slide-24
SLIDE 24
  • 4. Semi-Supervised Machine Learning

We fit our Gaussian Mixture Model. In this case, ‘bad’ specimens were identified with a 77% success rate. Above 75% was a key project objective.

slide-25
SLIDE 25
  • 4. Semi-Supervised Machine Learning

We fit our Gaussian Mixture Model. In this case, ‘bad’ specimens were identified with a 77% success rate. Above 75% was a key project objective. Triangles are coloured in to represent the probability that a component is ‘faulty’. UQ must be included in machine learning!

slide-26
SLIDE 26
  • 4. Semi-Supervised Machine Learning

We repeat this many times, where data is randomly unlabelled in each iteration.

slide-27
SLIDE 27
  • 4. Semi-Supervised Machine Learning

For the case where half of our specimens are unlabelled, this is the histogram of the resulting success rates:

slide-28
SLIDE 28
  • 4. Semi-Supervised Machine Learning

Finally, we conducted this for different numbers of labelled and unlabelled data: It is encouraging that there is no sudden ‘drop off’ in performance.

slide-29
SLIDE 29

New Results

  • So far we have classified specimens after the builds have finished.
  • Moving towards machine-learnt control we would like to identify faults

layer-by-layer.

  • This will allow us to take corrective actions.
  • It will also give us a much ‘richer’ dataset (more data points,

essentially).

  • We now have builds where faults are deliberately introduced at certain

layers.

  • We are trying to detect these flaws automatically using data-based

techniques.

slide-30
SLIDE 30

New Results

Here, each data point corresponds to a layer. We are automatically detecting a layer where 10% less powder was deployed.

slide-31
SLIDE 31

Final Outputs

slide-32
SLIDE 32

Final Outputs

  • Machine-Learnt process control for SLM is feasible.
slide-33
SLIDE 33

Final Outputs

  • Machine-Learnt process control for SLM is feasible.
  • Ikenna has gone on to secure a permanent position at the

Manufacturing Technology Centre.

slide-34
SLIDE 34

Final Outputs

  • Machine-Learnt process control for SLM is feasible.
  • Ikenna has gone on to secure a permanent position at the

Manufacturing Technology Centre.

  • Awarded EPSRC Impact Acceleration Account to deploy our algorithm

into a user-friendly GUI.

slide-35
SLIDE 35

Final Outputs

  • Machine-Learnt process control for SLM is feasible.
  • Ikenna has gone on to secure a permanent position at the

Manufacturing Technology Centre.

  • Awarded EPSRC Impact Acceleration Account to deploy our algorithm

into a user-friendly GUI.

  • Journal paper submitted to Additive Manufacturing (Automatic Fault

Detection for Selective Laser Melting using Semi-Supervised Machine Learning).

slide-36
SLIDE 36

Final Outputs

  • Machine-Learnt process control for SLM is feasible.
  • Ikenna has gone on to secure a permanent position at the

Manufacturing Technology Centre.

  • Awarded EPSRC Impact Acceleration Account to deploy our algorithm

into a user-friendly GUI.

  • Journal paper submitted to Additive Manufacturing (Automatic Fault

Detection for Selective Laser Melting using Semi-Supervised Machine Learning).

  • Machine-Learnt process control => EPSRC Responsive Mode proposal

submitted.

slide-37
SLIDE 37

Final Outputs

  • Machine-Learnt process control for SLM is feasible.
  • Ikenna has gone on to secure a permanent position at the

Manufacturing Technology Centre.

  • Awarded EPSRC Impact Acceleration Account to deploy our algorithm

into a user-friendly GUI.

  • Journal paper submitted to Additive Manufacturing (Automatic Fault

Detection for Selective Laser Melting using Semi-Supervised Machine Learning).

  • Machine-Learnt process control => EPSRC Responsive Mode proposal

submitted.

  • Currently extending to Inkjet Printing (Kate Black).
slide-38
SLIDE 38

Final Outputs

  • Machine-Learnt process control for SLM is feasible.
  • Ikenna has gone on to secure a permanent position at the

Manufacturing Technology Centre.

  • Awarded EPSRC Impact Acceleration Account to deploy our algorithm

into a user-friendly GUI.

  • Journal paper submitted to Additive Manufacturing (Automatic Fault

Detection for Selective Laser Melting using Semi-Supervised Machine Learning).

  • Machine-Learnt process control => EPSRC Responsive Mode proposal

submitted.

  • Currently extending to Inkjet Printing (Kate Black).
  • Has led to 3 PhD projects (including 1 Risk CDT and 1 Risk+AM CDT).
slide-39
SLIDE 39

Thank you for your attention. ANY QUESTIONS?

p.l.green@liverpool.ac.uk liverpool.ac.uk/engineering/staff/peter-green/ Peter_Green17 @plgreen4