Develop Predictive Maintenance

Algorithms using MATLAB

Dr. Sarah Drewes, MathWorks Consulting Services

© 2015 The MathWorks, Inc.

1
Different Types of Learning

Machine Learning

Supervised Unsupervised • Discover a good internal representation

Learning Learning • Learn a low dimensional representation

• Output is a choice between classes

Classification • (True, False) (Red, Blue, Green)

Regression • Output is a real number (temperature,

stock prices)

2
Classification in Predictive Maintenance

 Parameters/Predictors: Sensor data, control settings

 Classes/States: Failure states, time horizon until failure/ material fatigue

Goal: Predict failure from sensor data

Prerequisites:

- Machine-readable data format

- Sufficient historical data containing meaningful information

3
Classification model generation
@MONDI Gronau
Sensor Data Parameters/Predictors
(10-100 /plant)

update ~ 60-90
min.
Quality State
Classes/States:
1: ok
2: failure

Which sensor measurements indicate machine failure?

4
Classification model generation

Basic Workflow

Preprocess Data

Choose Algorithm

Fit Model

Evaluate Model

Choose Model

Make Predictions
5
Classification model generation-
Prepare data

• Preprocess sensor data: clean invalid data, disregard constant values,

identify data types

• Aggregate per time stamp

Sensor Data
(10-100 /plant)

update ~ 60-90
min.
Quality State

6
Classification model generation

Basic Workflow

Preprocess Data

Choose Algorithm

Fit Model

Evaluate Model

Choose Model

Make Predictions
7
Classification model generation
Choose algorithms
Possible Classification Methods

Discriminant Analysis

8
Classification model generation
Choose an algorithm
 Distinguish 'categorical' (= discrete) and other (= continuous) predictors

 A priori analysis of data, e.g., test for normal distribution

 Reduce dimension of predictor variables, e.g., principal component analysis

(PCA)

 Use ensemble learning to reduce sensitivity of learning algorithms, e.g.

TreeBagger for classification trees

9
Classification model generation
Choose an algorithm

Y N

Y ?
Y

10
Classification model generation

Basic Workflow

Preprocess Data

Choose Algorithm

Fit Model

Evaluate Model

Choose Model

Make Predictions
11
Classification model generation
Fit model

Fit model based on historic data

PredictionModel = fitctree(PARAMETER,
fitcxxx
fitcknn (PARAMETER, STATE)
fitcnb
myfitnn

Training Data,
e.g. 70% of
historic data

12
Classification model generation
Fit model

13
Classification model generation

Basic Workflow

Preprocess Data

Choose Algorithm

Fit Model

Evaluate Model

Choose Model

Make Predictions
14
 Classification model generation
Evaluate model

predictedState = PredictionModel(Parameter)
PredictionModel

predictedState

1
Validation
1
Data, e.g. 1
30% of 1

historic data 2
2
1

Misclassification rate 1 of 7: 14.28 %

Accuracy: 85.72 %
15
Classification model generation
Evaluate model - using Classification Learner App

16
Classification model generation

Basic Workflow

Preprocess Data

Choose Algorithm
For each
classification
Fit Model method

Evaluate Model

Choose Model

Make Predictions
17
Classification model generation
Choose model
Choose Model with best misclassification rate

18
Classification model generation
Choose model

19
Classification model generation

Basic Workflow

Preprocess Data

Choose Algorithm

Fit Model

Evaluate Model

Choose Model

Make Predictions
20
Predictive monitoring at MONDI Gronau -
Use the predictive model
Predict current machine states during operation.

Sensor Data
(10-100 /plant)

update ~ 60-90
min.
Quality State
ok
State is: not ok
Sensor data
(now)

Prediction Update
Train Prediction
Prediction
Model
Model
(historic data)
Model
Predicted State (now)
21
Process monitoring at MONDI Gronau –
Domain knowledge and tools

Tools:
 MATLAB
 Database Toolbox
 Statistics and Machine Learning Toolbox
 Neural Network Toolbox
 MATLAB Compiler

22