Develop Predictive Maintenance Algorithms using · PDF file2 Different Types of Learning...

22
1 © 2015 The MathWorks, Inc. Develop Predictive Maintenance Algorithms using MATLAB Dr. Sarah Drewes, MathWorks Consulting Services

Transcript of Develop Predictive Maintenance Algorithms using · PDF file2 Different Types of Learning...

Page 1: Develop Predictive Maintenance Algorithms using · PDF file2 Different Types of Learning Machine Learning Supervised Learning Classification Regression Unsupervised Learning • Discover

1© 2015 The MathWorks, Inc.

Develop Predictive Maintenance

Algorithms using MATLAB

Dr. Sarah Drewes, MathWorks Consulting Services

Page 2: Develop Predictive Maintenance Algorithms using · PDF file2 Different Types of Learning Machine Learning Supervised Learning Classification Regression Unsupervised Learning • Discover

2

Different Types of Learning

Machine Learning

Supervised Learning

Classification

Regression

Unsupervised Learning

• Discover a good internal representation

• Learn a low dimensional representation

• Output is a real number (temperature,

stock prices)

• Output is a choice between classes

• (True, False) (Red, Blue, Green)

Page 3: Develop Predictive Maintenance Algorithms using · PDF file2 Different Types of Learning Machine Learning Supervised Learning Classification Regression Unsupervised Learning • Discover

3

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

Page 4: Develop Predictive Maintenance Algorithms using · PDF file2 Different Types of Learning Machine Learning Supervised Learning Classification Regression Unsupervised Learning • Discover

4

Classification model generation

@MONDI Gronau

Sensor Data

(10-100 /plant)

Quality State

update ~ 60-90

min.

Parameters/Predictors

Classes/States:

1: ok

2: failure

Which sensor measurements indicate machine failure?

Page 5: Develop Predictive Maintenance Algorithms using · PDF file2 Different Types of Learning Machine Learning Supervised Learning Classification Regression Unsupervised Learning • Discover

5

Classification model generation

Basic Workflow

Evaluate Model

Fit Model

Choose Algorithm

Preprocess Data

Choose Model

Make Predictions

Page 6: Develop Predictive Maintenance Algorithms using · PDF file2 Different Types of Learning Machine Learning Supervised Learning Classification Regression Unsupervised Learning • Discover

6

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)

Quality State

update ~ 60-90

min.

Page 7: Develop Predictive Maintenance Algorithms using · PDF file2 Different Types of Learning Machine Learning Supervised Learning Classification Regression Unsupervised Learning • Discover

7

Classification model generation

Basic Workflow

Evaluate Model

Fit Model

Choose Algorithm

Preprocess Data

Choose Model

Make Predictions

Page 8: Develop Predictive Maintenance Algorithms using · PDF file2 Different Types of Learning Machine Learning Supervised Learning Classification Regression Unsupervised Learning • Discover

8

Classification model generation

Choose algorithms

Possible Classification Methods

Discriminant Analysis

Page 9: Develop Predictive Maintenance Algorithms using · PDF file2 Different Types of Learning Machine Learning Supervised Learning Classification Regression Unsupervised Learning • Discover

9

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

Page 10: Develop Predictive Maintenance Algorithms using · PDF file2 Different Types of Learning Machine Learning Supervised Learning Classification Regression Unsupervised Learning • Discover

10

Classification model generation

Choose an algorithm

?

Y N

Y

?

Y

Page 11: Develop Predictive Maintenance Algorithms using · PDF file2 Different Types of Learning Machine Learning Supervised Learning Classification Regression Unsupervised Learning • Discover

11

Classification model generation

Basic Workflow

Evaluate Model

Fit Model

Choose Algorithm

Preprocess Data

Choose Model

Make Predictions

Page 12: Develop Predictive Maintenance Algorithms using · PDF file2 Different Types of Learning Machine Learning Supervised Learning Classification Regression Unsupervised Learning • Discover

12

PredictionModel = fitcknn (PARAMETER, STATE)PredictionModel = fitcxxx (PARAMETER, STATE)

Classification model generation

Fit model

Fit model based on historic data

Training Data,

e.g. 70% of

historic data

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

Page 13: Develop Predictive Maintenance Algorithms using · PDF file2 Different Types of Learning Machine Learning Supervised Learning Classification Regression Unsupervised Learning • Discover

13

Classification model generation

Fit model

Page 14: Develop Predictive Maintenance Algorithms using · PDF file2 Different Types of Learning Machine Learning Supervised Learning Classification Regression Unsupervised Learning • Discover

14

Classification model generation

Basic Workflow

Evaluate Model

Fit Model

Choose Algorithm

Preprocess Data

Choose Model

Make Predictions

Page 15: Develop Predictive Maintenance Algorithms using · PDF file2 Different Types of Learning Machine Learning Supervised Learning Classification Regression Unsupervised Learning • Discover

15

Classification model generation

Evaluate model

Validation

Data, e.g.

30% of

historic data

predictedState

1

1

1

1

2

2

1

PredictionModel

Misclassification rate 1 of 7: 14.28 %

Accuracy: 85.72 %

predictedState = PredictionModel(Parameter)

Page 16: Develop Predictive Maintenance Algorithms using · PDF file2 Different Types of Learning Machine Learning Supervised Learning Classification Regression Unsupervised Learning • Discover

16

Classification model generation

Evaluate model - using Classification Learner App

Page 17: Develop Predictive Maintenance Algorithms using · PDF file2 Different Types of Learning Machine Learning Supervised Learning Classification Regression Unsupervised Learning • Discover

17

Classification model generation

Basic Workflow

Evaluate Model

Fit Model

Choose Algorithm

Preprocess Data

Choose Model

Make Predictions

For each

classification

method

Page 18: Develop Predictive Maintenance Algorithms using · PDF file2 Different Types of Learning Machine Learning Supervised Learning Classification Regression Unsupervised Learning • Discover

18

Classification model generation

Choose model

Choose Model with best misclassification rate

Page 19: Develop Predictive Maintenance Algorithms using · PDF file2 Different Types of Learning Machine Learning Supervised Learning Classification Regression Unsupervised Learning • Discover

19

Classification model generation

Choose model

Page 20: Develop Predictive Maintenance Algorithms using · PDF file2 Different Types of Learning Machine Learning Supervised Learning Classification Regression Unsupervised Learning • Discover

20

Classification model generation

Basic Workflow

Evaluate Model

Fit Model

Choose Algorithm

Preprocess Data

Choose Model

Make Predictions

Page 21: Develop Predictive Maintenance Algorithms using · PDF file2 Different Types of Learning Machine Learning Supervised Learning Classification Regression Unsupervised Learning • Discover

21

Predictive monitoring at MONDI Gronau -

Use the predictive model

Sensor Data

(10-100 /plant)

Quality State

Predict current machine states during operation.

Train Prediction Model

(historic data)Prediction

Model

Sensor data

(now)

Predicted State (now)

update ~ 60-90

min.

State is: not okState is: ok

Update Prediction Model

(historic data)Prediction

Model

Page 22: Develop Predictive Maintenance Algorithms using · PDF file2 Different Types of Learning Machine Learning Supervised Learning Classification Regression Unsupervised Learning • Discover

22

Process monitoring at MONDI Gronau –

Domain knowledge and tools

Tools:

MATLAB

Database Toolbox

Statistics and Machine Learning Toolbox

Neural Network Toolbox

MATLAB Compiler