Machine learning in mineral processing

Lidia Auret
Process Engineering

23 March 2018 – Maties Machine Learning

Continuous, connected, controlled, circulating, complex, changing

2
 Industrial data
• Online physical property sensor data
– E.g. mass flow rate, density, temperature, pressure
– ~ seconds
• Online image data
– E.g. rocks on conveyor belts, flotation froth (mud and bubbles)
– ~ minutes
• Offline laboratory data
– E.g. metal content, particle size distribution
– ~ hours
• Offline image data
– E.g. microscopic grain shape and colour
– ~ days

3
 Abnormal event detection
• Many faults and failures can occur in complex processes
• Large variation in normal operating conditions due to
range of allowable disturbances

Disturbance
D
processes

4
 Abnormal event detection
• Missed detections can lead to suboptimal performance,
equipment failure, safety and environmental violations
• False alarms can lead to unnecessary downtime and loss
of trust in alarm systems

5
 Abnormal event detection
• Unsupervised learning problem
– Abundance of one class of data: Normal operating conditions
• Fault detection
– Feature extraction
– Data description / support estimation
• Fault identification
– Topology extraction
– Supervised learning model inspection:
• Variable importance
• Partial dependence

6
 Abnormal event detection
• Feature extraction
– Sensor data correlated
(through mass and energy balances, control instructions)
– Sensor data noisy
– Feature space represents lower dimensional, noise-free
information
– Residual space represents feature extraction model validity

7
 Abnormal event detection
• Feature extraction
– Principal component analysis
෡ = 𝑻∗ 𝑷∗
• 𝑻∗ = 𝑿𝑷∗ ; 𝑿 𝑻

– Kernel principal component analysis

8
 Abnormal event detection
• Feature extraction
– Autoassociative neural networks (NLPCA)

9
 Abnormal event detection
• Feature extraction
– Autoassociative neural networks (NLPCA)

10
 Abnormal event detection
• Data description / support estimation
– Kernel density estimation
– One-class support vector machines

11
 Abnormal event detection
• Data description / support estimation
– Kernel density estimation
– One-class support vector machines

12
 Abnormal event detection
• Data description / support estimation
– Kernel density estimation
– One-class support vector machines

13
 Abnormal event detection
• Topology extraction
– Identification of propagation path of fault
– Transfer entropy / lagged cross-correlation used to determine
direction and strength of connections between variables

14
 Abnormal event detection
• Topology extraction
– Identification of propagation path of fault
– Transfer entropy / lagged cross-correlation used to determine
direction and strength of connections between variables

15
 Abnormal event detection
• Topology extraction
– Identification of propagation path of fault
– Transfer entropy / lagged cross-correlation used to determine
direction and strength of connections between variables

16
 Abnormal event detection
• Research approach
– Scarcity of industrial data with faults detected and identified
– Simulation of complex, dynamic processes with known faults
– Repository with dynamic models and simulated data

github.com/ProcessMonitoringStellenboschUniversity

17
 Soft sensors
• Ore characteristics
– Metal content, particle size  correlated to process performance
– Captured by image data
• Soft sensor
– Trained model for prediction of process performance from
measured process data

Ore grade
Particle size
Process state
GLCM k-NN
Wavelet LDA
Textons
Steerable pyramids
Etc.
18
 Soft sensors
• Flotation grade prediction with convolutional neural
networks texture features and classification

19
 Soft sensors
• Flotation grade prediction with convolutional neural
networks texture features and classification

Class labels
used for
training
Variants
generated to
supplement
data set Bottleneck
introduced to
create lower
dimensional
feature space

20
 Soft sensors
• Flotation grade prediction with convolutional neural
networks texture features and classification

21
 Soft sensors
• Flotation grade prediction with convolutional neural
networks texture features and classification

Interpretability important for industrial adoption

22
 Challenges and opportunities
• Data size, quality and fusion
– Potentially massive data sets
– Shifting process conditions
– Online process data + offline process data + maintenance records
+ mine plan + purchase orders + etc.
• Exploiting process knowledge
– Dynamic Bayesian networks
– Hybrid modelling
• Process recovery
– Actionable insights
– Reinforcement learning

23
 Questions?
[email protected]

24