Lesson 13.

DATA SCIENCE AND

AUTOMATION COURSE

MASTER DEGREE SMART

TECHNOLOGY ENGINEERING

Fault diagnosis I
TEACHER

Introduction Mirko Mazzoleni

PLACE
University of Bergamo
Outline

1. Introduction and motivation

2. Terminology

3. Taxonomy

4. Defining a predictive maintenance problem

2 /19
Outline

1. Introduction and motivation

2. Terminology

3. Taxonomy

4. Defining a predictive maintenance problem

3 /19
Introduction and motivation
Intelligent fault diagnosis is one of the cornerstone of the industry 4.0 paradigm
• Maintenance costs represent an important part of a production industry total costs

• Optimizing the interventions, reducing the dead times, can drastically improve the
production and reduce the costs

Common maintenance strategies are:

• Reactive maintenance (run-to-failure): use the component until it breaks
• Preventive maintenance (time-based): do maintenance at regular time instants
• Predictive maintenance (condition-based): do maintenance on the basis of
measured data and algorithms

4 /19
Data sources
In order to develop an algorithmic solution, it is mandatory to have data on the problem
• Measurements made during the machine functioning
• Information about mean time to failure of the component and working conditions

The usefulness of a certain measure and technique depends on the fault type
• Vibration analysis: mechanical faults, rotating components
• Thermography: emitted infrared energy, proportional to the surface temperature

• Tribology analysis: of lubricant oil, wear particle analysis

• Ultrasound analysis: high frequency vibrations (≥ 30 kHz), leak detection

• Other: electro-magnetic tests, visual inspection,…

5 /19
Data sources
Public datasets for predictive maintenance include:

• Gearbox Fault Detection Dataset: Dataset for fault detection in a transmission

system using accelerometer data and information regarding bearing geometry.
https://c3.nasa.gov/dashlink/resources/997/

• Robot Execution Failures Dataset: Force measurements and twisting moment to

detect faults in robots; it is composed of 463 samples and 30 features.
https://archive.ics.uci.edu/ml/datasets/Robot+Execution+Failures

• IDA2016Challenge Data Set: Dataset for the detection of faults in the air pressure
systems of a truck: it is composed of 76000 samples and 171 features.
https://archive.ics.uci.edu/ml/datasets/IDA2016Challenge

6 /19
Outline

1. Introduction and motivation

2. Terminology

3. Taxonomy

4. Defining a predictive maintenance problem

7 /19
 Terminology
• The aim of fault diagnosis is to discover anomalous behaviours in physical plants

1. Fault detection: detect if there is an anomaly

2. Fault isolation: detect the faulty component

3. Fault identification: Fault Diagnosis

Fault Estimation
characterize the type,
Fault Detection Fault Isolation
size, nature of the fault (FD) (FI)

4. Fault estimation: Fault Identification

Fault Detection and Isolation (FDI)
estimate the fault signal

8 /19
Terminology
Condition monitoring (CM) refers to the continued oversight of the progression of the
degradation of a system or a component

The differences with fault diagnosis (FD) is not so sharp, but, in general:
• FD produces a discrete output, CM produces a continuous output
• FD is usually a more structured approach which is applicable component-wise

• CM deals with a more system-wise evaluation, for which simple procedures can be
applied

Fault estimation methods, when applicable, lead to effective condition monitoring

9 /19
Outline

1. Introduction and motivation

2. Terminology

3. Taxonomy

4. Defining a predictive maintenance problem

10 /19
Taxonomy
Fault diagnosis

Hardware
redundancy Analytical redundancy schemes
schemes

Knowledge
Plausibility Model- Signal-
based Hybrid Active
tests based based
or Data-Driven

Data-driven
design

11 /19
Model-based fault diagnosis
A model of the system is developed to design a residual generator 𝑸 𝑧

Faults 𝒖(𝑡)
𝒓(𝑡)
𝒚(𝑡) 𝑸 𝑧
𝒖(𝑡) 𝒚(𝑡)
Plant

+ Residuals 𝜽(𝑡) Decision Diagnostic

Residual
processing logic decision
− 𝒓(𝑡)
Process 𝜄(𝑡)
model

Residual generation Residual evaluation

Model based fault diagnosis system

12 /19
Signal-based fault diagnosis
Certain process signals carry information about the faults to be detected
• From them, fault symptoms are computed…
• …and compared with prior knowledge about the faults
Faults

Examples are: Knowledge of

𝒖(𝑡) 𝒚(𝑡) symptoms
• Diagnosis of rotating Plant

components defects
• Stator\rotor current faults Diagnostic
Symptoms Symptoms decision
Symptoms
generation analysis

13 /19
Knowledge-based fault diagnosis
Great amount of historical data available:
• There is not a prior knowledge about the faults, and it has to be extracted from data
• Training vs testing phases Faults
𝒖(𝑡) 𝒚(𝑡)
Plant
Common approaches are:
Diagnostic
• Machine learning classifiers Consistency checking & decision
classifier
• Statistical Process monitoring methods
Knowledge
(control charts, PCA indicators,…)
Historical data base
of the plant Training &
learning

14 /19
Outline

1. Introduction and motivation

2. Terminology

3. Taxonomy

4. Defining a predictive maintenance problem

15 /19
Defining a predictive maintence problem
Suppose you want to develop a predictive maintenance algorithm on a machine, made
up of several components.
• AIM: to understand when the machine is degrading

What do we have to consider? Machine

1. Problem definition
Component 1 … Component N
2. Data acquisition

3. Development of the solution Sub- Sub-

Component 1.1 Component 1.2

16 /19
Defining a predictive maintence problem
1. PROBLEM DEFINITION (PURPOSE OF THE ALGORITHM)

• What answer do I want to get from the algorithm?

 Fault detection & isolation  fault \ not fault, location of the fault
 Condition monitoring  monotonic indicator of machine health status

• What are the critical components?

 FMECA / FTA – criticality, frequency of failure
 Exchange of information with the maintenance department

• Do I have all the information that I need?

 Faults history, physical quantities related to machine operation

17 /19
Defining a predictive maintence problem
2. DATA ACQUISITION

• Test plan
 Typical operation of the machine or in its working range
 Tests in Healthy state - Test in faulty state (fault injection or endurance?)

• Measurements acquisition
 Additional sensors? They must be acquired with specific hardware: which one?
 Syncronize data from multiple data sources
 Amount of data to be stored
 Always check at the end of the day (even during) what has been acquired 

18 /19
Defining a predictive maintence problem
3. DEVELOPMENT OF THE SOLUTION

• Chosen technique

 Model-based: based on a mathematical model of the machine or component

 Signal-based: based on a symptom of failure visible within a signal

 Data-driven: based only on the observed data (machine learning). I suppose that
there is symptom of faults inside the data, but I have to discover it

 Supervised model: estimate the relation from data (A) and condition (B), A→B

 Unsupervised model: anomaly detection methods

19 /19