On topological data analysis for SHM; an introduction to

persistent homology

T. Gowdridge, N. Dervilis, K. Worden

Dynamics Research Group, Department of Mechanical Engineering, University of Sheffield
Mappin Street, Sheffield S1 3JD, UK
arXiv:2209.06155v1 [math.AT] 12 Sep 2022

Abstract
This paper aims to discuss a method of quantifying the ’shape’ of data, via a methodology called topological data
analysis. The main tool within topological data analysis is persistent homology; this is a means of measuring the shape
of data, from the homology of a simplicial complex, calculated over a range of values. The required background theory
and a method of computing persistent homology is presented here, with applications specific to structural health
monitoring. These results allow for topological inference and the ability to deduce features in higher-dimensional
data, that might otherwise be overlooked.
A simplicial complex is constructed for data for a given distance parameter. This complex encodes information
about the local proximity of data points. A singular homology value can be calculated from this simplicial complex.
Extending this idea, the distance parameter is given for a range of values, and the homology is calculated over this
range. The persistent homology is a representation of how the homological features of the data persist over this
interval. The result is characteristic to the data. A method that allows for the comparison of the persistent homology
for different data sets is also discussed.
Key words: Topological data analysis; Persistent homology; Simplicial complex.

1 Introduction
Topological methods are very rarely used in structural health monitoring (SHM), or indeed in structural dynamics
generally, especially when considering the structure and topology of observed data. Topological methods can provide
a way of proposing new metrics and methods of scrutinising data, the most rudimentary and most powerful of which,
persistence homology, will be discussed in this paper.
SHM has been dominated by the insurgence of machine learning since its introduction to the field [1, 2]. Topological
Data Analysis (TDA) aims to work alongside the vast work already established to provide a new light into ways
that data can be analysed. Previous machine learning operations in SHM have never distinctly considered the data
topology. This paper aims to bring to light the potential importance of TDA in SHM. TDA generalises well into
higher dimensions, so an insight into higher-dimensional data structures can be obtained.
TDA has previously proven useful in other research areas, such as medicine, where virus evolution has been tracked.
TDA has also been used to sequence viruses such as Influenza A and HIV, along with many other uses such as
clustering diabetes types. In economics, TDA has been used to identify topological patterns in multi-dimensional
time series data. It is thought that topological data analysis might be able to give early warning signs of imminent
market crashes.
The contents of this paper aim to walk through the process of performing TDA. A brief outline of the methodology is
given here, to construct a geometric object called a simplicial complex from data. A simplicial complex can be thought
of as a higher-dimensional analogue of a graph. The vertices of the simplicial complexes are the data points, and
the connections between the vertices are less than a prescribed threshold distance. Therein, the simplicial complex
encodes information about connection between the vertices. This geometric object has now attributed shape to
the data. Following from the simplicial complex, this information can be manipulated in order to output algebraic
groups that will capture information about the shape and structure of the connection in the simplicial complex, most
specifically about the number of k−dimensional holes that will be found within the assumed manifold underlying the
data; these groups are called the homology groups. A generalisation of the homology will be discussed with respect to
pioneering work by Edelsbrunner [3, 4] where the homology can be considered over a range of simplicial complexes;
this is called the persistent homology, aptly named as this uncovers how the homology persists over a range of values.
The layout of the paper is as follows: Section 2 give the necessary definitions, listed at the beginning to provide the
required mathematics that is not usually common to engineering. Section 3 will be devoted to persistent homology,
its significance, provide an intuitive understanding and show two common forms of how it can be displayed. Section 4
will show the topology of some known shapes and how the persistent homology links to these. Section 5 will introduce
an interesting engineering specific example which will show how this theory can be used. Following this, the paper
concludes.

2 Background Theory and Definitions

At the expense of labouring multiple required definitions, the theory pertinent to the paper topic will quickly follow.
For more details, the following works can be consulted [5–9].

2.1 Algebra
Definition 2.1. An equivalence relation, denoted by ∼, is a binary operation that is reflexive (a ∼ a), symmetric (if
a ∼ b then b ∼ a) and transitive (if a ∼ b and b ∼ c then a ∼ c). The equivalence relation provides a partition of a set
into elements that share a common property.
Definition 2.2. Groups are an extension of the concept of set, to include a binary operator. This set-operator pair
is written G = {S, ◦}. There are four necessary axioms associated with a group, these are: closure, associativity,
existence of a unique identity, and the existence of an inverse for every element in the group.
Definition 2.3. A subgroup, H, of a group, G, is a subset such that H ⊆ G, and H satisfies all of the group axioms.
The operation and identity of the subgroup is inherited from the parent group.
Definition 2.4. A group is called an abelian group, if the result of the operation is independent of the ordering of
elements. This is the case when the operation is commutative. x ◦ y = y ◦ x, ∀x, y ∈ G.
Definition 2.5. Let G be a group and H a subgroup of G. A left coset of H in G is a subset of G of the form
gH = {gh | h ∈ H} for some g ∈ G, the set of left cosets of H in G is written G/H. Similarly, a right coset of H in
G is a subset of G of the form Hg = {hg | h ∈ H} for some g ∈ G, and the set of right cosets is written H\G [5].
Definition 2.6. A subgroup H of G is said to be normal if gH = Hg, ∀g ∈ G, this is written H / G. In this case,
the set of cosets G/H and H\G are the same by definition. The set of cosets for normal subgroups is called the
quotient group.
Definition 2.7. A group homomorphism is a mapping between two groups, h : G1 → G2 . The mapping is considered
a homomorphism if the identity element in G1 is mapped to the identity element in G2 , and the group operation
distributes over the homomorphism, h(u ∗ v) = h(u) ◦ h(v), ∀u, v ∈ G1 .
Definition 2.8. Given a group homomorphism h : G1 → G2 , the kernel of h, ker(h) ⊂ G1 , is the set of elements x
such that h(x) = e, where e is the identity element. The image of h, im(h) ⊂ G2 , is the set of elements y such that
y = h(x) for some x.
Definition 2.9. A metric space is defined by a pair (X, ∂X ). X refers to the set where the elements of the metric
space live and ∂X is the associated metric or distance function between two points in X. For a space to qualify as a
metric space, the following criteria must be true: ∂(x, y) ≥ 0, ∂(x, y) = ∂(y, x), and ∂(x, z) ≤ ∂(x, y) + ∂(y, z).
Definition 2.10. An open ball is defined as B (x) = {y ∈ X | ∂X (y, x) < }. This encloses a space around the point
x, where all points enclosed are less than the distance  from the point x. This space is often referred to as the
−neighbourhood of x.
Definition 2.11. A topological space is represented by a pair, (X, T ), where X is the set of all the elements and
T is a collection of subsets, referred to as the topology. The open sets in T must satisfy the following axioms: The
set of elements and the empty set are elements of the topology, any union of sets in T is also an element of T , the
intersection of a finite collection of elements of T is an element of T .
Definition 2.12. Given two topological spaces, (X, TX ) and (Y, TY ). These spaces are said to be homeomorphic if
there exists a bijective continuous map between them.
Now that a topological space has been discussed, when taking homeomorphisms between spaces there are quantities
characteristic of that structure called topological invariants; these do not change under homeomorphisms. It will
be the main purpose of this review to develop a method of calculating a topological invariant called the persistent
homology.

2.2 Manifolds
Manifolds are continuous surfaces from which the data are assumed to be sampled. Data that are observed are
assumed to lie on the surface of a manifold. By understanding the topology of the sampled data points, it is the aim
of TDA to extract topological information about the underlying manifold from the sampled data. The manifold is
unknown prior to analysis and persistent homology will identify features within the manifold over a range of length
scales. Thereby, understanding the shape of the sampled data, it is the conjecture of TDA that the shape of the
manifold is also understood. Formally, a manifold is a space that is locally homeomorphic to some n−dimensional
Euclidean space, Rn [6].

2.3 Simplicial Complex

Simplicial complexes are a way of representing data sampled from a manifold; they can be thought of as higher-
dimensional analogues of graphs, giving a way of encoding connections between vertices. The dimension that a
simplicial complex can capture is restricted by the number of points that are fully connected. Simplicial complexes
can be manipulated to output the homology for the data, and following this, the persistent homology.
A simplicial complex is a structure made up of fundamental building blocks called simplices. The first four simplices
can be seen in Figure 1. Each vertex in the simplex is fully connected to all the other vertices and the space enclosed
by the vertices is part of that simplex.

∆0 ∆1 ∆2 ∆3

Figure 1: The first four simplices.

There are many ways to construct a simplicial complex from point data. For simplification, only one method will be
discussed within this paper, the Vietoris-Rips complex. The Vietoris-Rips (VR) complex can be constructed for point
data to output a corresponding complex according to the rules V R (X, ∂X ): let (X, ∂X ) be a finite metric space and
 > 0 be a fixed value. The abstract simplicial complex is determined by the rules [10]:
1. The vertices, v ∈ X, form the vertices in V R (X, ∂X ).
2. A k−simplex is formed when ∂X (vi , vj ) ≤ 2, ∀i, j ≤ k for some ε > 0.
 Figure 2: The process of constructing a VR complex from some data.

This process of constructing a simplicial complex for some data attributes shape to the data, by defining connections
between arbitrarily close points. Before moving on to homology, which is an algebraic construct, a method of
converting the geometric simplicial complex into an algebraic simplicial complex must first be considered. An abstract
simplicial complex represents a method of defining the geometric connections, so that algebra can then be applied to
the shape.
Definition 2.13. An abstract simplicial complex, K, consists of a set of vertices, vert(K) and a set of abstract
simplices, simp(K), such that [11]:
1. Every simplex ∆k ∈ simp(K) is a non-empty subset of vert(K), or the simplices are a union of the vertices.
2. For all vertices, v ∈ vert(K), there is also an abstract simplex {v} ∈ simp(K).
3. For every non-empty abstract simplex, ∆k , a non-empty subset of ∆k is also an abstract simplex. This is
referred to as a face of ∆k .
4. For an abstract simplex ∆k ∈ vert(K), the dimension of ∆k is dim(∆k ) = |∆k | − 1, where |.| denotes the number
of elements in the set.
This notion of an abstract simplicial complex is pivotal to TDA. Abstract simplicial complexes allow for the mapping
between categories: a mapping from a geometric realisation of a simplicial complex which is embedded in Euclidean
space, to an abstract simplicial complex. The abstract simplicial complex form can be algorithmically analysed with
computer packages to output results for that simplicial complex.

2.4 Homology
The homology groups, Hk (X) are invariants for the data set, X, where k refers to the dimension of the feature of the
homology group. The homology groups encode information about the number of k−dimensional holes in the data.
Definition 2.14. The orientation of the vertices of a simplex, ∆k , is an equivalence class of orderings of the vertices
under the equivalence relation that two orderings are the same if they differ by an even permutation. An even
permutation is one that can be expressed as a composition of even permutations [7]. There are only two possible
orientations, a positive and a negative.
Definition 2.15. To each standard simplex ∆ki of a simplicial complex K, an abelian group, Ck (K), called a chain
group is associated to it. The k th chain group is the set of all k−simplices in the simplicial complex K [12].
Now, consider a K containing lk of any standard simplex ∆k for all values of k. The k−chain group of K, Ck (K)
is the free abelian group generated by the oriented k−simplices of K. This means any element ck ∈ Ck (K) can be
thought of in abstract as,
Xlk
ck = fi ∆ki , fi ∈ Z
i=1
where the following criteria are satisfied [8]:
1. A negation of simplices. ∆ki + (−∆ki ) = 0, ∀i, k.
lk
X lk
X lk
X
2. A linearity over the elements. fi ∆ki + gi ∆ki = (fi + gi )∆ki where fi , gi ∈ Z.
i=1 i=1 i=1

Following the previous two definitions, a boundary operator can be formulated. The boundary operator provides
an algebraic formulation for the exterior region of a simplicial complex. Now that simplicial complexes are being
represented by chain groups, the problem has become entirely algebraic.
Definition 2.16. The boundary operator, ∂k , maps between chain groups,

∂k : Ck (K) → Ck−1 (K)

Given an oriented simplex ∆k = [v0 , ..., vk ], a positive sign is assigned to every member of the even permutation class
of ∆k and a negative sign to every member of the odd permutation class. The boundary operator must now obey the
rules [8]:
1. For an oriented simplex,
k
X
∂∆k = (−1)i [v0 , ..., v̂i , ..., vk ]
i=0

where [v0 , ..., v̂i , ..., vk ] represents the face of the simplex with the ith vertex omitted. Note that every successive
omission changes the orientation of the face.
2. Thinking
P of a simplicial complex K, as the abstract sum of all the standard simplices required to construct it,
K = i,k ∆ki , ∆ki ⊂ K. The boundary operator is a linear function over all the simplices in the complex.
X
X
∂(K) = ∂ ∆ki = ∂(∆ki ), ∆ki ⊆ K
i,k i,k

Definition 2.17. A mapping between successive chain groups with the boundary operator,
k ∂ ∂k−1
2 1 0 ∂ ∂ ∂
... → Ck −→ Ck−1 −−−→ ... −→ C1 −→ C0 −→ 0

is called a chain complex [12].

For the purposes of calculations, it is said that Ci (K) = 0 for i > dim(K). It is also the case that Ci (K) = 0 for
i < 0. In both these cases, 0 represents the zero group.
Definition 2.18. For a simplicial complex K, elements of the chain group, zk ∈ Ck (K), are called k−cycles if
∂zk = 0. The group of k−cycles, Zk (K), is given by the kernel of the boundary map,

Zk (K) = ker(∂k : Ck (K) → Ck−1 (K)) = {zk ∈ Ck : ∂k zk = 0}

and Zk (K) is a subgroup of Ck (K) [13].

Definition 2.19. For a simplicial complex K, elements of the chain group, bk ∈ Ck (K), are called k−boundaries
if there exists a (k + 1)−chain group, Ck+1 (K), such that ∂Ck+1 (K) = bk . The group of k−boundaries, Bk (K), is
given by,
Bk (K) = im(∂k+1 : Ck+1 (K) → Ck (K)) = {bk ∈ Ck : ∃b0k ∈ Ck+1 , bk = ∂b0k }
and Bk (K) is a subgroup of Ck (K) [13].
The interesting part of this analysis, which fundamentally results in the homology groups, is that when the composition
of two boundary operations is analysed, applying it twice gives a result of 0. That is the boundary of a boundary is
empty.
Lemma 2.20.
∂k−1 ◦ ∂k = 0
This is a very important result, a proof is supplied in many texts [6, 8]
Corollary 2.21. A very important result follows on from Lemma 2.20. For a simplicial complex, K, any element of
the boundary group bk ∈ Bk (K) has the property ∂k bk = 0. Therefore, Bk (K) ⊆ Zk (K) where Zk (K) is the group of
k−cycles. Since both Zk (K) and Bk (K) are abelian, a property inherited by being subgroups of Ck (K). Bk (K) is a
normal subgroup of Zk (K), and therefore it can divide the parent group.

Cp+1 Cp Cp−1

Zp+1 Zp Zp−1

Bp+1 Bp Bp−1
∂p+2 ∂p+1 ∂p ∂p−1

0 0 0

Figure 3: A chain complex with boundary maps between the chain groups.

Corollary 2.21 brings together a lot of the background theory that has been covered up until this point. It is beneficial
to break this down into bite-size chunks. The first stage is understanding that ∂k bk = 0, as this is essentially the
application of two boundary operators, as the element bk is formed from the boundary operation of a previous chain
element. This can be rewritten as (∂k ◦ ∂k+1 )ck+1 , then using Lemma 2.20 it follows that this result is equal to zero.
If all the elements in Bk (K) are mapped to zero after the application of ∂k then it must be a subset of the group that
is mapped to zero, by Definition 2.18 this is Zk (K). From Definitions 2.18 and 2.19, it is known that Zk (K) ⊂ Ck (K)
and Bk (K) ⊂ Ck (K). Since the chain groups Ck (K) are defined to be abelian, then so are their subgroups, Bk (K)
and Zk (K) by Definition 2.3. It has also been proven that Bk (K) ⊂ Zk (K), and since Bk (K) is abelian it is known
that Bk (K) / Zk (K) and therefore the set of cosets Zk (K)/Bk (K) is the quotient group.
Definition 2.22. The homology groups, Hk (K), are the quotient groups,

Hk = Zk (K)/Bk (K)

Setting aside all this mathematical rigour, the k th homology groups can simply be thought of as the cycles in Ck that
are not boundaries of the elements within Ck+1 . That an element of Ck is a cycle, means it encloses a k−dimensional
region. The fact this is not a boundary means the interior is not part of the underlying space. This is where the idea of
counting the k−dimensional holes springs from. A generalisation of the rule is that Hk (K) counts the k−dimensional
holes in the simplicial complex, K. H0 (K) is the only real exception to this rule, as this encodes information about
the number of path-connected components in K. H1 (K) encodes information about 1D holes; these can be visualised
as circular holes. H2 (K) encodes information about 2D holes; these can be visualised as cavities. Hk (K) encodes
information about kD holes.
Definition 2.23. Hk (K) is a vector space and the elements are the homology classes of K. The homology class of a
cycle zk ∈ Zk (K) is the coset ck + Bk (K) = {ck + bk : bk ∈ Bk (K)}. Cycles are said to be homologous if they are in
the same homology class [13].
Definition 2.24. The Betti number, βk , is the dimension of the k th homology group of a simplicial complex,

βk = dim(Hk (K))

The Betti number represents the number of k−dimensional holes in a simplicial complex [9].
Betti numbers will be the primary topological invariants used throughout this paper. Betti numbers will be used
in persistent homology and they are vital in visualising spaces. Examples of Betti numbers and how they can be
visualised are provided in Section 4. Following this, Betti numbers will be used in the engineering examples in Section
5.
3 Persistent Homology
3.1 Understanding
When the distance, , is smaller than some feature scale, the properties of that feature can be captured. This means
topological invariants that are described at a length scale greater than  can be captured. A problem arises here, as
usually the feature scale is not known prior to analysis.
Obtaining the homology for a single value of  provides very limited information, this notion is almost redundant, due
to potential varying feature length scales in the manifold. For this reason, it is vital to consider what homological
features persist as  is varied. The goal of persistent homology is to track the homology classes as  is varied. This
process of varying  does not bias any disk size, as all are being considered. This process will give an initial value,
min , where a specific homological feature comes to life and max , where the feature is no longer considered for that
simplicial complex. This range of values [min , max ] is called the persistence interval for that homological feature.
Each persistence interval is attributed a Betti number. Following this, the set of all persistence intervals is descriptive
for that manifold, giving information about in which dimension a hole exists in the data and over what range of
values it persists for. Regardless of triangulation of a simplicial complex, or construction method, the information
obtained from persistence is the same; i.e. the persistence of a space is considered a topological invariant.
When varying , the simplicial complex goes from k disconnected vertices to a fully connected (k − 1)−dimensional
simplex (the filtration is likely to not be taken this far, as computational time would be unnecessarily large), where
there are k points sampled from a manifold. The features that persist over the longest relative change are more likely
to represent features of the space. This procedure will then give a strong indication of the likelihood of the topological
invariants associated to that space.
When calculating persistence, a minimum persistence interval length is specified. Given a persistence interval [i , j ],
the length of the interval can be calculated by |j − i | > l where l represents the interval length threshold. Interval
lengths shorter than l will not be considered in the analysis; this is useful as it could be argued that for very small
persistence intervals, whether they even ’persist’ or not.
The persistence intervals obtained can can be represented in two ways: barcodes or persistence diagrams, both having
their merits.
3.2 Diagrams

Figure 4: Persistence barcode with realisations showing which simplicial complexes are present at some values of ε.

In a barcode representation, the x−axis refers to the value of . As  increases, the barcode shows which features
persist. The set of intervals are plotted with each interval beginning at min and ending at max . The colour of
the interval on the barcode refers to the Betti number, βk [14]. The value of the y−axis can simply thought as an
indexing of the intervals in the barcode. An example of a barcode can be seen in Figure 4, with vertical dotted lines
showing the intersections with the intervals, showing which features are present and the corresponding simplicial
complex is found at the end of the dotted line. A few notable tricks to reading the barcodes are: the length of the
interval represents how long the feature persists for. The length of the interval can be thought of as having a higher
probability that this feature is characteristic of the manifold, as the longer the interval, the more prominent that
feature is in the data.
 Figure 5: Birth-death diagram.

The other method of visualising the set of persistence intervals is the persistence diagram, or birth-death diagram. In
this representation, min is plotted on the x−axis and max is plotted on the y−axis, with each interval represented by
the point (min , min ). Intuitively, there is a line defined by y = x below which the points will not be plotted; this line
has the interpretation that the feature must first exist before it can die. Reading these diagrams is as intuitive as
reading the barcodes; the vertical height of the point from the line y = x is analogous to the length of the interval,
that is the further a point is from the line y = x the more the feature persists. An example of a birth-death diagram
can be seen in Figure 5. The Birth-Death diagram shows a more structured method of comparing persistence intervals
as the y−axis is not so arbitrary when compared to the barcodes.
On both the barcode and birth-death diagram, it can be seen that features persist to ∞. This has two meanings. The
first being that there will always be a fully-connected simplex that persists to infinity. There will be a value of fc
that results in a fully-connected simplex where every vertex is connected to every other vertex. For values  > fc the
simplex will remain fully connected, and therefore this will continue to infinity. The second occurrence of this is when
a feature persists past the value of  used in the construction of the simplicial complex, as only smaller simplicial
complexes of up to that value of  are considered.
Each representation has their merits. For barcodes, it is easy to see which features a simplicial complex will have by
drawing a vertical line at x = , as can be seen in Figure 4. The barcode representation shows repeated intervals, as
these are new entries. This is not the case with the birth-death diagram, these intervals will displayed as the same
point and overlap, therefore, information is lost in the birth-death diagram in this situation. Despite this, birth-death
diagrams are less arbitrary and are displayed more compactly.
The space of barcodes forms a metric space; the distance between the barcodes is a measure of similarity of two
barcodes.
As manifolds can be represented by their barcodes, this notion of a metric space allows one to compare the similarity
of manifolds. Metrics between barcodes are well established and the one used in this report is the p−Wasserstein
distance.
Definition 3.1. Given two barcodes B1 and B2 . For p > 0, the p−Wasserstein distance is given by,
! p1
X
p
∂Wp (B1 , B2 ) = inf d∞ (Z, φ(Z))
Z∈B1

where φ is a matching between B1 and B2 and Z is a persistence interval in B1 [7].

4 Understanding TDA
To help elaborate the point and highlight some features of TDA, two examples will be given in this section, where the
points will be sampled from the manifold S 2 (the two-dimensional sphere embedded in R3 ) in two different ways. The
first method used is by taking concentric circles in the plane perpendicular to the z−axis. The second method was to
construct a Fibonacci spiral around the sphere. The formula used for the first method to generate the point cloud
(the set of samples from the manifold) for the embedding of the sphere was,


x = cos u sin v

f (u, v) = y = sin u cos v u ∈ [0, 2π], v ∈ [0, π]

z = cos v


Figure 6: Sphere simplicial complex

When constructing a VR complex, simplices are formed at all points within an open ball B ; this means that
connections will be formed to points past other points in the same direction. In this example, this is problematic as it
results in the shape being over-connected at the poles if one wants the equatorial region to be filled with 2-simplices.
This negatively impacts the computation time.

Figure 7: Sphere persistence representations.

The data and values for the given value of ε, for this sphere and the Fibonacci sphere is shown in Table 1. It should
be noted that the values of the Betti numbers in the table are specific to that value of epsilon. The diagrams contain
a greater depth of information as the value of epsilon can be seen to be varied.
The Betti numbers in Table 1 are representative of the simplicial complex shown in Figure 7. These are β0 = 1, which
says that the simplicial complex is fully connected with no disjoint parts. β1 = 0, tells us that there are no holes in
the simplicial complex. β2 = 1 says that the simplicial complex enclose a volume, which is the inside of the sphere. It
is important to note that the inside of the sphere is not part of the simplicial complex, but the space in which the
simplicial complex is embedded.
The second case of sampling from the sphere, was done with Fibonacci spirals around the sphere; this gives a much
more uniform distribution of points. The formula used to generate the point cloud for the embedding of the sphere
was,


x = cos θ sin φ

√
f (φ, θ) = y = sin θ sin φ φ ∈ [0, π], θ ∈ [0, np π(1+ 5)]

z = cos φ


Figure 8: Fibonacci sphere simplicial complex.

where np is the number of points being sampled, in this case np = 500.

Figure 9: Fibonacci sphere persistence representations.

From Table 1 it can be seen that the first scenario has fewer vertices, but still results in more simplices after
constructing the simplicial complex. This is due to the poor distribution of points and the larger epsilon. The example
goes to show that data cleansing steps and the correct model should be used to output meaningful and efficient results
from TDA.
 Description Sphere Fibonacci sphere
 0.5 0.25
Number of vertices 200 500
Number of simplices 112094 4202
Betti Numbers [1, 0, 1] [1, 0, 1]
Dimension 3 3

Table 1: Topological properties for each sphere.

When using the Wasserstein metric between the two data samples, a relatively small value of ∂Wp = 2.673 is obtained;
this shows that the manifolds are relatively similar. This result is to be expected as the two spheres are sampled from
the same manifold. When qualitatively analysing the persistence data, this backs up the case as the barcodes and
birth-death diagrams, whilst not the same, do exhibit very similar features.

5 Application
In this case, a four-dimensional surface plot will be considered. This example is defined by linear algebraic equations,
where the variables form the axes of the higher-dimensional plots.
The system under analysis is a simple 3DOF system, with springs between masses and ground. A representation of a
general case can be seen in Figure 10. In this modelling scenario, it is assumed that the stiffness of the second spring,
k2 is a function of the coefficient of thermal expansion, α2 , the temperature of the spring, T2 , and the damage present
in the spring, D2 . It is the aim of this analysis to understand the topology of the manifold constructed when the
natural frequency ωi is calculated as a function of the variable parameters α2 , T2 , D2 and all other values are constant.

k1 k2 (α, T, D) k3 k4
m1 m2 m3

Figure 10: General mass-spring-damper system with variable k2 .

A simple relationship is assumed, where an increase in the temperature of a spring causes a reduction of the stiffness,
by the relation k2 (1 − α2 T2 ). It is also assumed that the damage coefficient also reduces the stiffness of the spring by
a factor k2 (1 − D2 ).
After undertaking the appropriate free body analysis, with the assumptions that there is no damping and the system
has no external forcing. The following equation is derived,

MẌ + KX = 0

where,
   
m1 0 0 k1 + k2 (1 − α2 T )(1 − D) −k2 (1 − α2 T )(1 − D) 0
M= 0 m2 0 , K =  −k2 (1 − α2 T )(1 − D) k2 (1 − α2 T )(1 − D) + k3 −k3 
0 0 m3 0 −k3 k3 + k4

Assuming a harmonic response, a solution can be obtained by determining the eigenvalues of the equation |−Mωi2 +K| =
0 The result is a matrix equation that determines the response of the system, where ωi is a function of the parameters
ωi (α2 , D2 , T ). This is a 3DOF system, therefore, it is possible to obtain three eigenvalues (ωi2 ) and eigenvectors
(mode shapes), Xi , where the eigenvalues and eigenvectors satisfy the equation,

M−1 KX = ωi2 X, i ∈ {1, 2, 3}

5.1 TDA of Mass-Spring-Damper Model
This example is embedded in R4 ; for this reason the topological features shown in the persistence barcodes will
have to be studied to determine the topology of the data, rather than visualisation. In this case, analysis is being
undertaken to determine the topology of the manifold from which the data are sampled. This was not the case with
the sphere, as the topology was known prior to analysis. Despite this, it is still possible to determine some features
of the topology of the data. It is known that the manifold is connected, as it is described by continuous functions.
Therefore, the condition that β0 = 1 must be satisfied.
It is also expected that there is a self-intersection in the manifold. This can be inferred from the model, where an
increase in temperature would reduce the spring to become less stiff. This is also the case for the damage parameter.
This must mean that there is a value of αT that will give the same reduction in k, that a value of D will. Since all
the variables are contained within K, this means that the natural frequency must be the same value at these two
points, and thus there is a self intersection in the manifold. It is actually the case that this is a family of values, and
not just a single point. The family of values is parameterised by,

(1 − α2 T )(1 − D2 ) = (1 − α20 T 0 )(1 − D20 )

When creating the data, the inputs were T ∈ [250, 500] with seven divisions, α2 ∈ [0, 0.005] with six divisions and
D ∈ [0, 1] with six divisions; meaning that there are 252 data points. The values for T, α and D can be thought of as
coordinates of the form (T, α, D). Each point is then input into the equations above to give a value ωi . Each point can
then be embedded in 4D in the form (T, α, D, ωi ). These points will then form the vertices of the simplicial complex,
which can be analysed using TDA. When generating data this way, the number of points generated is proportional to
the power of the dimension, so calculations can get cumbersome for high dimensions or high sample sizes. After the
points had been generated, each point was scaled down by the largest value in that dimension, meaning that all the
data are inside [0, 1]4 .
The values used for the calculations in this first example were:

m1 = m2 = m3 = 10kg, k1 = k2 = k3 = k4 = 10000N m−1

Description ω1 ω2 ω3
Epsilon 1 = 0.77 2 = 0.33 3 = 0.31
Num of vert 252 252 252
Num of Simp 188087202 160639 93917
Betti Numbers [1, 0, 0, 0] [1, 0, 0, 0] [1, 0, 0, 0]
Dimension 4 4 4

Table 2: ω1,2,3 simplicial complexes data.

The Betti numbers in Table 2 are relatively uninteresting, but these Betti numbers are for only for the values of
epsilon listed. The persistence diagrams in Figures 11,12 and 13 give more insight into the features for values of
epsilon less than the ones listed in Table 2.
 Figure 11: ω1 persistence representations.

In the case of ω1 , the manifold appears to have a sudden and sharp change with the parameters; this can be seen in
the birth-death diagram in Figure 11, as the second to last red point ≈ (0, 0.69) is the instance at which the simplicial
complex is fully connected. It appears that the data are roughly divided into two clusters either side of this sudden
change. A high value of epsilon is required to span this sudden change. As a result of a high value of epsilon, the
clusters become highly connected within. This can be seen in the relatively large number of simplices in Table 2 when
compared to the other natural frequencies.
Another method of spanning the gap between the sudden change, instead of increasing 1 , is to increase the resolution
of the data by taking more divisions for each parameter. This method also has an issue. More points are added to
the sudden change, but there will also be more points in each cluster. Meaning there are more points to be connected;
this massively compromises the computation time.
It can also be seen on the barcode in Figure 11 that there are numerous instances of 1D holes that persist for about
0.1 units in epsilon. As features only persist for small ranges of epsilon, this is likely caused by the lack of resolution
in the data. In the case of ω1 , there are also instances of cavities forming. These features also do not persist for long,
and are likely artefacts due to the lack of resolution in the data.

Figure 12: ω2 persistence representations.

The persistence representations shown in Figures 12 and 13 are fairly similar, both having nearly the same features
present.

Figure 13: ω3 persistence representations.

Both ω2 , ω3 achieved the criteria of β0 = 1 at a much lower value of  when compared with ω1 , which indicates there
is no sudden abrupt change in the manifold. There are also 1D holes present, again this is likely to be an artefact of
the lack of resolution in the data.
There is also a general trend of the persistence intervals becoming more square in the β0 section of the barcode, going
from ω1 to ω3 . This says that the distance between the connected components is becoming more uniform.

5.1.1 Values at different stiffness

Following the previously-outlined procedure, a new data set was collected for the values,

m1 = m2 = m3 = 10kg, k2 = 5000N m−1 , k1 = k3 = k4 = 10000N m−1

For continuity, the same number of divisions was used along each of the inputs, and the same process of selecting the
lowest value of epsilon that gave a fully-connected simplicial complex. Following these conditions, the results in Table
3 were obtained.

Description ω10 ω20 ω30

Epsilon 01 = 0.38 02 = 0.32 03 = 0.30
Num of vert 252 252 252
Num of Simp 339447 87571 97341
Betti Numbers [1, 0, 0, 0] [1, 0, 0, 0] [1, 0, 0, 0]
Dimension 4 4 4
0
Table 3: ω1,2,3 , k2 = 5000N m−1 simplicial complexes data.

Immediately it is clear that the 01 is much smaller that 1 ; this means that the computation time was much less when
calculating the simplicial complex for k2 = 5000N m−1 than k2 = 10000N m−1 .
Figure 14: ω10 persistence representations.

Figure 15: ω20 persistence representations.

Figure 16: ω30 persistence representations.

The barcodes still exhibit a similar structure when compared with their counterparts in Figures 11, 12 and 13. When
comparing the barcodes using the Wasserstein distance, the following values were obtained,

∂Wp (Bω1 , Bω10 ) = 0.69029, ∂Wp (Bω2 , Bω20 ) = 0.077248, ∂Wp (Bω3 , Bω30 ) = 0.029308

The values may appear relatively small, but the data were scaled by the largest value in that dimension, therefore, the
Wasserstein distance is also scaled. There appears to be a general trend, that the higher the natural frequency, the
more similar the manifolds are. From this, it can be inferred for this example, that as successive natural frequencies
are computed there is a lesser reliance on the stiffness term, k2 .
The idea of comparing two persistence data sets is useful in engineering. If the manifold structure of data from a
machine or structure operating at optimal conditions is known, the shape of the manifold could be used to fine tune
or identify differences in the shape of a manifold of similar machine or structure in an unknown state. This could
have interesting applications in SHM.

6 Conclusions
The results of this paper show that topological inference is a viable analysis strategy for engineering applications.
By opting to use TDA, features in the data can be identified that would not necessarily be found in other analysis
techniques. Features of the data such as holes or cavities in the underlying manifold can help to identify trends or
anomalies. This inference can then be extended to determine engineering examples such as clustering data by their
topological features, this can then be used to identify changes in the operating state of a machine or structure.
An appropriate method for conducting topological data analysis has been outlined, with a special consideration for
conducting the analysis in an effective manner. Examples were given that highlighted where TDA can be inefficient
or return misleading results.
The engineering-specific example given here is a very fundamental start, but not without its issues. Calculating data
at higher dimensions is computationally expensive, and therefore the resolution in the manifold was too low. However,
this is not the case with real world data that has been collected experimentally. Another issue arises when data are
not evenly distributed, as this results in overly-connected areas.
In summary, a novel data analysis strategy has been discussed, previously unfamiliar to engineering, that can provide
new insights into the fundamental structure and shape of the data, even in higher-dimensional analysis where an
intuitive understanding of shape is no longer upheld.

Acknowledgements
The authors would like to thank the UK EPSRC for funding via the Established Career Fellowship EP/R003645/1
and the Programme Grant EP/R006768/1.

References
[1] C R Farrar and K Worden. An introduction to structural health monitoring. Philosophical Transactions of the
Royal Society A: Mathematical, Physical and Engineering Sciences.
[2] C R Farrar and K Worden. Structural health monitoring: a machine learning perspective. John Wiley & Sons,
2012.
[3] H Edelsbrunner and J Harer. Computational Topology: An Introduction. American Mathematical Soc., 2010.
[4] H Edelsbrunner, D Letscher, and A Zomorodian. Topological persistence and simplification. In Proceedings 41st
Annual Symposium on Foundations of Computer Science. IEEE, 2000.
[5] J J Rotman. An Introduction to the Theory of Groups, volume 148. Springer Science & Business Media, 2012.
[6] B F Schutz and Director Bernard F Schutz. Geometrical Methods of Mathematical Physics. Cambridge university
press, 1980.
 [7] R Rabadán and A J Blumberg. Topological Data Analysis for Genomics and Evolution: Topology in Biology.
Cambridge University Press, 2019.
[8] C Nash and S Sen. Topology and Geometry for Physicists. Elsevier, 1988.
[9] R W Ghrist. Elementary Applied Topology, volume 1. Createspace Seattle, 2014.

[10] E W Chambers, V De Silva, J Erickson, and R Ghrist. Vietoris-rips complexes of planar point sets. Discrete &
Computational Geometry, 44(1):75–90, 2010.
[11] F Chazal and B Michel. An introduction to topological data analysis: fundamental and practical aspects for
data scientists. arXiv preprint arXiv:1710.04019, 2017.

[12] R Ghrist. Homological algebra and data. The Mathematics of Data, 25:273, 2018.
[13] J D Boissonnat, Frédéric Chazal, and Mariette Yvinec. Geometric and Topological Inference, volume 57.
Cambridge University Press, 2018.
[14] R Ghrist. Barcodes: the persistent topology of data. Bulletin of the American Mathematical Society, 45(1):61–75,
2008.