Part 2

Lecture Notes on
Dynamic Network Reliability Analysis
In using this material and other related material, please read and cite:
Abdullah Alkaff, M. Nur Qomarudin, and Yusuf Bilfaqih, Network reliability
analysis: matrix-exponential approach, Reliability Engineering & System Safety,
Vol. 204, December 2020.
https://doi.org/10.1016/j.ress.2020.107192
Download MATLAB script:
https://github.com/mochamadnurq/reliability-networksystem-
matrixexponential
 This is Part 2 of lecture notes series in Dynamic System Reliability Analysis.
Reading Part 1 is highly recommended before reading Part 2.

Part 1
Lecture Notes on
Functional System Reliability Analysis: An Algorithmic Approach
Available on ResearchGate
based on:
Abdullah Alkaff and Mochamad Nur Qomarudin, Modeling and Analysis of System
Reliability Using Phase-Type Distribution Closure Properties, Applied Stochastic Models
in Business and Industries, 2020; 1– 22
https://doi.org/10.1002/asmb.2509
Contents
1. Concept of dynamic network reliability analysis
2. An approach in network reliability modeling
3. The main feature of networks systems
4. SDP in network reliability analysis
5. SDP representation of network reliability
6. Matrix-exponential distribution
7. SDP algorithm for dynamic network reliability analysis
8. Experiment using simplified ARPA Network
9. Extension to general structure systems
10. Advantages of using ME distribution
1. Concept of dynamic network reliability analysis
Conventional network reliability analysis
𝑅1 , 𝑅2 , … , 𝑅𝑁 Network 𝑅𝑆

- 𝑅1 , 𝑅2 , … , 𝑅𝑁 Structure - 𝑅𝑆 : system reliability value

(components’ reliability values)
Modified conventional static reliability algorithm

𝑅1 (𝑡), 𝑅2 (𝑡), … , 𝑅𝑁 (𝑡) Network 𝑅𝑆 (𝑡)

Structure

- 𝑅1 𝑡 , 𝑅2 , 𝑡 , … , 𝑅𝑁 (𝑡) - 𝑅𝑆 (𝑡): system reliability function

(components’ reliability functions)
Dynamic network reliability analysis
2. An approach in network reliability modeling
• Using combinatorial model instead of state space model
• 𝑋𝑆 and 𝑋𝑖 , 𝑖 = 1, 2, … , 𝑁, are Boolean variables representing states of the system
and the 𝑖𝑡ℎ component: 0 for failed and 1 for good.
• 𝑋𝑆 is a Boolean function of 𝑋𝑖 , 𝑖 = 1, 2, … , 𝑁,
𝑋𝑆 = 𝜙 𝑋𝑖
𝜙 is the network structure function
• Combinatorial model based on combination of components successes (failures) that
lead to the system success (failure)
• Using the minimum paths (or minimum cuts) of the network to represent the
combinations of components successes (failures).
• This lecture notes uses minimum paths, but the method is applicable as well for
minimum cuts with simple modification
2. An approach in network reliability modeling
• Let 𝑀𝑃𝑗 , 𝑗 = 1, 2, … , 𝑀, be the 𝑗𝑡ℎ minimum path of the network.
• The network structure function is given by
𝑀

𝑋𝑆 = ሧ ෑ 𝑋𝑖 (1)
𝑗=1 𝑖∈𝑀𝑃𝑗
• Example: Consider a parallel-series system

✓ The network has minimum paths: {1}, {2,3}, {4,5,6}

✓ Its minimum products are 𝑋1 , 𝑋2 𝑋3 , 𝑋4 𝑋5 𝑋6 .
✓ Its structure function in Boolean operations is
𝑋𝑆 = 𝑋1 ∨ 𝑋2 𝑋3 ∨ 𝑋4 𝑋5 𝑋6
✓ In arithmetic operation is 𝑋𝑆 = 1 − (1 − 𝑋1 )(1 − 𝑋2 𝑋3 )(1 − 𝑋4 𝑋5 𝑋6 )
✓ The system reliability function can be obtained by replacing 𝑋𝑖 with 𝑅𝑖 (𝑡)
𝑅𝑆 = 1 − 1 − 𝑅1 1 − 𝑅2 𝑅3 1 − 𝑅4 𝑅5 𝑅6
3. The main feature of network systems
• The main feature of a network is the presence of bridge components
• Bridge components may appear in several minimum paths
• Example: Consider a simple network of 5 components with bridge structure

✓ Its minimum paths are {1,2}, {1,3,5}, {4,3,2}, and {4,5}

✓ Its structure function in Boolean function is
𝑋𝑆 = 𝑋1 𝑋2 ∨ 𝑋1 𝑋3 𝑋5 ∨ 𝑋2 𝑋3 𝑋4 ∨ 𝑋4 𝑋5
✓ Two minimum paths involve 𝑋3 → dependent
✓ 𝑋𝑆 cannot be transformed directly into arithmetic function
✓ Must be disjointed to become sum of disjoint products (SDP)!
4. SDP in network reliability analysis: definition
• Boolean OR operation
𝐶 =𝐴∨𝐵
• Its equivalent arithmetic operation is obtained by multiplying 𝐵 with 𝐴 invers
𝐶 = 𝐴 + 𝐴𝐵 ҧ
• 𝐴𝐵ҧ and A are disjoint products (DP), and 𝐴 + 𝐴𝐵 ҧ is sum of disjoint products (SDP)
• They are disjoint since their product (intersection) is zero
• Many algorithms are available for obtaining the simplest SDP representation from a
Boolean function
• The most efficient algorithms are based on the multivariable inversion proposed by
Klauss D. Heidtmann (1989), known as KDH88 algorithm
• Originally, SDP is used for constant failure probabilities
• In this lecture note, it will be extended to be used for dynamic failure probabilities
K. D. Heidtmann, "Smaller sums of disjoint products by subproduct inversion," IEEE Transactions on Reliability, vol.
38, no. 3, pp. 305-311, 1989.
4. SDP in network reliability analysis: Properties
• Some properties of Boolean operations:
𝐴 ∨ 𝐴𝐵 = 𝐴
𝐴𝐵 = 𝐴ҧ ∨ 𝐵ത
𝐴 ∨ 𝐵 = 𝐴.ҧ 𝐵ത
𝐴𝐵. 𝐴ҧ = 𝐴ҧ
𝐴𝐵. 𝐴 = 𝐴𝐵ത
𝐴𝐵 ∨ 𝐴𝐶 = 𝐴 𝐵 ∨ 𝐶
• Example: Find SDP of 𝑍 = 𝐴𝐵𝐶 ∨ 𝐵𝐷
𝐴𝐵𝐶 ∨ 𝐵𝐷 ⇒ 𝐴𝐵𝐶 + DISJOINT 𝐵𝐷, 𝐴𝐵𝐶 = 𝐴𝐵𝐶 + 𝐴𝐵𝐶𝐵𝐷
= 𝐴𝐵𝐶 + 𝐴𝐶 ∨ 𝐵ത 𝐵𝐷 = 𝐴𝐵𝐶 + 𝐴𝐶𝐵 ∨ 𝐵𝐵 ത 𝐷 = 𝐴𝐵𝐶 + 𝐴𝐶𝐵𝐷
• This example shows that 𝐵𝐷 needs to be multiplied by variables in 𝐴𝐵𝐶 that do not
appear in 𝐵𝐷
4. SDP in network reliability analysis: Example
• Example: Find the SDP representation of a simple network in Slide 6
𝑋𝑆 = 𝑋1 𝑋2 ∨ 𝑋1 𝑋3 𝑋5 ∨ 𝑋2 𝑋3 𝑋4 ∨ 𝑋4 𝑋5
The process is executed recursively as follows:
1. Make the second term disjointed from the first term
DISJOINT 𝑋1 𝑋3 𝑋5 , 𝑋1 𝑋2 = 𝑋1 𝑋2 𝑋3 𝑋5 .
2. Make the third term disjointed
a. From the first term
DISJOINT 𝑋2 𝑋3 𝑋4 , 𝑋1 𝑋2 = 𝑋1 𝑋2 𝑋3 𝑋4 ,
b. From the second term, but it is already disjointed from the second term (their
product are zero)
DISJOINT 𝑋1 𝑋2 𝑋3 𝑋4 , 𝑋1 𝑋3 𝑋5 = 𝑋1 𝑋2 𝑋3 𝑋4 .
Continue…
4. SDP in network reliability analysis: Example
3. Make the fourth term disjointed
a. From the first term:
DISJOINT 𝑋4 𝑋5 , 𝑋1 𝑋2 = 𝑋1 𝑋2 𝑋4 𝑋5 ,
a. From the second term:
DISJOINT 𝑋1 𝑋2 𝑋4 𝑋5 , 𝑋1 𝑋3 𝑋5 = 𝑋1 𝑋3 . 𝑋1 𝑋2 𝑋4 𝑋5 = 𝑋1 𝑋3 ∨ 𝑋1 𝑋2 𝑋4 𝑋5
= 𝑋1 𝑋2 ∨ 𝑋3 𝑋4 𝑋5 = 𝑋1 ∨ 𝑋2 ∨ 𝑋3 𝑋4 𝑋5 = 𝑋1 ∨ 𝑋ത2 𝑋3 𝑋4 𝑋5
= 𝑋ത1 + 𝑋1 𝑋ത2 𝑋ത3 𝑋4 𝑋5 = 𝑋ത1 𝑋4 𝑋5 + 𝑋1 𝑋ത2 𝑋ത3 𝑋4 𝑋5 ,
b. From the third term, but the second resulted term has been disjointed from the
third term (their product are zero):
DISJOINT 𝑋ത1 𝑋4 𝑋5 + 𝑋1 𝑋ത2 𝑋ത3 𝑋4 𝑋5 , 𝑋2 𝑋3 𝑋4 = 𝑋ത1 𝑋2 𝑋3 𝑋4 𝑋5 + 𝑋1 𝑋ത2 𝑋ത3 𝑋4 𝑋5 .
The resulting SDP representation of the network is
𝑋𝑆 = 𝑋1 𝑋2 + 𝑋1 𝑋ത2 𝑋3 𝑋5 + 𝑋ത1 𝑋2 𝑋3 𝑋4 + 𝑋ത1 𝑋2 𝑋3 𝑋4 𝑋5 + 𝑋1 𝑋ത2 𝑋ത3 𝑋4 𝑋5
4. SDP in network reliability analysis: algorithm
The algorithm proposed in Heidtmann (1989) is summarized as

For all 𝑀𝑃𝑖 , i = 2, 3… m

For all 𝑀𝑃𝑗 , j = 1, 2… (i-1)
If disjoint
Continue;
Else
Obtain the disjoint terms and apply the process
for next j with each term produced in this step.
End
End
End
Heidtmann, K. D. (1989). Smaller sums of disjoint products by subproduct inversion. IEEE Transactions on
Reliability, 38(3), 305-311.
4. SDP in network reliability analysis: algorithm
• Example:
𝑋𝑆 = 𝑋1 𝑋2 ∨ 𝑋1 𝑋3 𝑋5 ∨ 𝑋2 𝑋3 𝑋4 ∨ 𝑋4 𝑋5
• Representation:
• 0 : the variable is present in the term
• - : the variable is absent in the term
• 1, 2, … : the variable is inverted together with other variable having the same
representation in the term
• The original structure function is represented as
𝒌 1 2 3 4 5 Minimum path
1 0 0 - - - 𝑋1 𝑋2
2 0 - 0 - 0 𝑋1 𝑋3 𝑋5
3 - 0 0 0 - 𝑋2 𝑋3 𝑋4
4 - - - 0 0 𝑋4 𝑋5
4. SDP in network reliability analysis: algorithm
The resulting SDP representation of the network
𝑋𝑆 = 𝑋1 𝑋2 + 𝑋1 𝑋ത2 𝑋3 𝑋5 + 𝑋ത1 𝑋2 𝑋3 𝑋4 + 𝑋ത1 𝑋2 𝑋3 𝑋4 𝑋5 + 𝑋1 𝑋ത2 𝑋ത3 𝑋4 𝑋5
is represented as

𝒌 1 2 3 4 5 Disjoint product
1 0 0 - - - 𝑋1 𝑋2
2 0 1 0 - 0 𝑋1 𝑋ത2 𝑋3 𝑋5
3 1 0 0 0 - 𝑋ത1 𝑋2 𝑋3 𝑋4
4 1 3 3 0 0 𝑋ത1 𝑋2 𝑋3 𝑋4 𝑋5
5 0 1 2 0 0 𝑋1 𝑋ത2 𝑋ത3 𝑋4 𝑋5
5. SDP representation of network reliability: formulation
• Using an SDP algorithm, 𝑀𝑃𝑗 in Equation (1) is transformed into disjoint product

• Let 𝐷𝑃𝑘 , where 𝑘 = 1, 2, … , 𝐾, are the resulted disjoint products

• Equation (1) can be written as
𝐾

𝑋𝑆 = ෍ ෑ 𝑍𝑖 (2)
𝑘=1 𝑖∈𝐷𝑃𝑘
𝑍𝑖 can be 𝑋𝑖 , its inverse 𝑋ത𝑖 , or the inverse of multivariable 𝑋𝑖 𝑋𝑗 , depending on
the form of 𝐷𝑃𝑘 .
• Equation (2) is called the SDP representation of the network.
5. SDP representation of network reliability: formulation
• SDP representation can be summarized as
𝐾

𝑋𝑆 = ෍ 𝑌𝑘 , where 𝑌𝑘 = ෑ 𝑍𝑖
𝑘=1 𝑖∈𝐷𝑃𝑘
• Each DP term contains at least one variable without inversion
• The network reliability function 𝑅𝑆 (𝑡) can be obtained by replacing 𝑋𝑖 , 𝑋ത𝑖 and 𝑋𝑖 𝑋𝑗
in 𝑍𝑖 by 𝑅𝑖 𝑡 , 1 − 𝑅𝑖 𝑡 and 1 − 𝑅𝑖 𝑡 𝑅𝑗 𝑡 , respectively

• The result is
𝐾

𝑅𝑆 (𝑡) = ෍ 𝐺𝑌𝑘 (𝑡)

𝑘=1
where 𝐺𝑌𝑘 (𝑡) is a function obtained when 𝑍𝑖 is replaced by its associated reliability
function
5. SDP representation of network reliability: example
• Example: The simple bridge network with five components
• Its SDP representation
𝑋𝑆 = 𝑋1 𝑋2 + 𝑋1 𝑋ത2 𝑋3 𝑋5 + 𝑋ത1 𝑋2 𝑋3 𝑋4 + 𝑋ത1 𝑋2 𝑋3 𝑋4 𝑋5 + 𝑋1 𝑋ത2 𝑋ത3 𝑋4 𝑋5
• Assume each component has reliability function of exponential distribution
𝑅𝑖 𝑡 = 𝑒 −𝜆𝑖 𝑡
• The network reliability function can be obtained by replacing 𝑋𝑖 , 𝑋ത𝑖 , and 𝑋𝑖 𝑋𝑗 with
𝑅𝑖 𝑡 , 1 − 𝑅𝑖 𝑡 and 1 − 𝑅𝑖 𝑡 𝑅𝑗 𝑡 , respectively
𝑅𝑆 𝑡 = 𝑅1 𝑡 𝑅2 𝑡 + 𝑅1 𝑡 𝑅3 (𝑡)𝑅5 𝑡 (1 − 𝑅2 (𝑡)) +
𝑅2 𝑡 𝑅3 (𝑡)𝑅4 𝑡 (1 − 𝑅1 (𝑡)) + 𝑅4 𝑡 𝑅5 (𝑡)(1 − 𝑅1 𝑡 )(1 − 𝑅2 (𝑡)𝑅3 (𝑡)) +
𝑅1 𝑡 𝑅4 (𝑡)𝑅5 𝑡 (1 − 𝑅2 (𝑡))(1 − 𝑅3 (𝑡))
• Thus,
𝑅𝑆 𝑡 = 𝑒 − 𝜆1 +𝜆2 𝑡 + 𝑒 − 𝜆1 +𝜆3 +𝜆5 𝑡 1 − 𝑒 −𝜆2 𝑡 + 𝑒 − 𝜆2 +𝜆3 +𝜆4 𝑡 1 − 𝑒 −𝜆1 𝑡 +
𝑒 − 𝜆4 +𝜆5 𝑡 1 − 𝑒 −𝜆1 𝑡 1 − 𝑒 − 𝜆2 +𝜆3 𝑡 + 𝑒 − 𝜆1 +𝜆4 +𝜆5 𝑡 1 − 𝑒 −𝜆2 𝑡 1 − 𝑒 −𝜆3 𝑡
5. SDP representation of network reliability: operations
• SDP comprises of three basic operations:
• Multiplication of variables without variable inversion: 𝑋𝑖 𝑋𝑗 ,
• Multiplication of variables with variable inversion: 𝑋𝑖 𝑋ഥ𝑗 or 𝑋𝑖 𝑋𝑗 𝑋𝑘 ,
• Addition of disjoint product terms: 𝑌𝑖 + 𝑌𝑗 .
• Each basic operation is extendable to more than two variables recursively
• Operations are applied to reliability functions not Boolean variables
• The reliability functions are represented in their parameters
• Operations are on the parameters not the functions
• Needs a special algebraic technique, known as tensor algebra
• For simplicity, use a distribution that is closed under those three operations
• Matrix-exponential (ME) distribution satisfies this requirement
6. ME distribution: mathematical model
• Matrix-exponential (ME) distribution is a univariate distribution on [0, ∞) having a
rational Laplace transform (Bladt and Nielsen, 2017)
• Probability density function for an ME distribution is
𝑓𝑋 (𝑥) = 𝒔𝑒 𝑨𝑥 𝒄 𝑥≥0
for a 1 × 𝑛 starting vector 𝒔, 𝑛 × 1 closing vector 𝒄, and 𝑛 × 𝑛 matrix 𝑨
• Its cumulative distribution function
𝐹𝑋 𝑥 = 1 + 𝒔𝑒 𝑨𝑥 𝑨−1 𝒄 𝑥≥0
• For reliability analysis, it will be simpler to write as
𝐹𝑋 𝑥 = 1 − 𝒂𝑒 𝑨𝑥 𝒗 𝑥≥0
where 𝒗 = −𝑨−1 𝒄 and 𝒂 = 𝒔

M. Bladt and B. F. Nielsen, Matrix-Exponential Distributions in Applied Probability, New York: Springer, 2017.
6. ME distribution: mathematical model
• 𝑋~𝑀𝐸𝑛(𝒂, 𝑨, 𝒗): 𝑋 has an ME distribution of order 𝑛 with representation (𝒂, 𝑨, 𝒗)
• The 𝑖𝑡ℎ moment of 𝑋 is 𝐸[𝑋 𝑖 ] = (−1)𝑖 𝑖! 𝒂𝑨−𝑖 𝒗 𝑖 = 1, 2, …
• ME distribution is more general and flexible than PH distribution because it is not
restricted to CTMC model
• A PH distribution is an ME distribution, but not vice versa
• Reliability measures in ME distribution is the same as in PH distribution (see Part 1)
• Reliability function 𝑅 𝑡 = 𝒂𝑒 𝑨𝑡 𝒗 𝑡≥0
• ℎ 𝑡 = −(𝒂𝑨𝑒 𝑨𝑡 𝒗)(𝒂𝑒 𝑨𝑡 𝒗)−1
• 𝑀𝑇𝑇𝐹 = −𝒂𝑨−1 𝒗
• 𝑉𝑇𝑇𝐹 = 2𝒂𝑨−2 𝒗 − 𝑀𝑇𝑇𝐹 2

M. Bladt and B. F. Nielsen, Matrix-Exponential Distributions in Applied Probability, New York: Springer, 2017.
6. ME distribution: mathematical model
• The differences of (𝒂, 𝑨, 𝒗) in PH and ME distributions

PH ME
Its elements are all non-negatives and, May have negative values or values greater
𝒂 under no atom assumption, their sum is than one (it is not a probability vector.)
equal to 1.
Must be a state transition matrix among Any nonsingular matrix having eigenvalues
𝑨 the transient states in a CTMC with one with negative real parts that satisfies −𝒂𝑨𝒗 ≥
absorbing state. 0, 𝒂𝑒 𝑨𝑥 𝒗 → 0 as 𝑥 → ∞.
𝒗 Always be a vector of ones 𝒆. Any vector which satisfies 𝒂𝒗 = 1.
• Advantages of using ME distribution:
• ME distribution may have smaller order than PH distribution
• ME distribution may be represented in Jordan canonical form by using matrix
similarity transformation
𝑨𝑥 −1 𝑨𝑥 −1 𝑷−1 𝑨𝑷𝑥 −1 ′ 𝑨′ 𝑥 ′
𝒂𝑒 𝒗 = 𝒂𝑷𝑷 𝑒 𝑷𝑷 𝒗 = 𝒂𝑷𝑒 𝑷 𝒗=𝒂𝑒 𝒗
6. ME distribution: closure properties
Assume 𝑋𝑖 has ME distribution with representation (𝒂𝑖 , 𝑨𝑖 , 𝒗𝑖 )
1. Multiplication of variables: 𝑌 = 𝑋1 𝑋2
Substituting 𝑅𝑖 𝑡 = 𝒂𝑖 𝑒 𝑨𝒊 𝑡 𝒗𝑖 for 𝑋𝑖 yields
𝐺𝑌 𝑡 = 𝑅1 𝑡 𝑅2 𝑡 = (𝒂1 𝑒 𝑨1 𝑡 𝒗1 )(𝒂2 𝑒 𝑨2 𝑡 𝒗2 )
= 𝒂1 𝑒 𝑨1 𝑡 𝒗1 ⊗ 𝒂2 𝑒 𝑨2 𝑡 𝒗2 (use 𝑔 ⊗ ℎ = 𝑔ℎ)
= 𝒂1 ⊗ 𝒂2 𝑒 𝑨1𝑡 ⊗ 𝑒 𝑨2𝑡 𝒗1 ⊗ 𝒗2 (use 𝑨𝑩 ⊗ 𝑪𝑫 = 𝑨 ⊗ 𝑪 𝑩 ⊗ 𝑫 )

= 𝒂1 ⊗ 𝒂2 𝑒 𝑨1 ⊕𝑨2 𝑡 𝒗1 ⊗ 𝒗2 (use 𝑒 𝑨𝑡 ⊗ 𝑒 𝑩𝑡 = 𝑒 𝑨⊕𝑩 𝑡 )

Theorem 1: 𝑋1 𝑋2 has ME representation (𝒂𝑌 , 𝑨𝑌 , 𝒗𝑌 ) where

𝒂𝑌 = 𝒂1 ⊗ 𝒂2 𝑨𝑌 = 𝑨1 ⊕ 𝑨2 𝒗𝑌 = 𝒗1 ⊗ 𝒗2

Note: Read Part 1 for explanation on Kronecker product ⊗ and Kronecker sum ⊕
6. ME distribution: closure properties
2. Multiplication with variable inversion 𝑌 = 𝑋1 𝑋ത2
Since 𝑋ത2 = 1 − 𝑋2 , it follows that
𝑌 = 𝑋1 1 − 𝑋2 = 𝑋1 − 𝑋1 𝑋2
Substituting 𝑅𝑖 𝑡 = 𝒂𝑖 𝑒 𝑨𝒊 𝑡 𝒗𝑖 for 𝑋𝑖 yields
𝐺𝑌 𝑡 = 𝑅1 𝑡 − 𝑅1 𝑡 𝑅2 𝑡
= 𝒂1 𝑒 𝑨1 𝑡 𝒗1 − 𝒂1 𝑒 𝑨1 𝑡 𝒗1 𝒂2 𝑒 𝑨2 𝑡 𝒗2
= 𝒂1 𝑒 𝑨1 𝑡 𝒗1 − (𝒂1 ⊗ 𝒂2 )(𝑒 𝑨1⊕𝑨2 𝒕 )(𝒗1 ⊗ 𝒗2 )
Theorem 2: 𝑋1 𝑋ത2 has ME representation (𝒂𝑌 , 𝑨𝑌 , 𝒗𝑌 ) where
𝑨1 𝟎 𝒗1
𝒂𝑌 = 𝒂1 −𝒂1 ⊗ 𝒂2 𝑨𝑌 = 𝒗𝑌 = 𝒗 ⊗ 𝒗
𝟎 𝑨1 ⊕ 𝑨2 1 2

Multiplication with a multivariable inversion such as 𝑌 = 𝑋1 𝑋2 𝑋3 is conducted by

writing as 𝑌 = 𝑋1 𝑍ҧ with 𝑍 = 𝑋2 𝑋3 and ME representation of 𝑍 is given by Theorem 1.
6. ME distribution: closure properties
• The two types of multiplication operations can be combined into one and can be
generalized for more than two variables by using recursion.
• Let 𝐷𝑘 be the number of variables in disjoint product 𝑌𝑘 and let 𝑖 = 1, 2, … , 𝐷𝑘 is
the sequence of variables in 𝑌𝑘 . The first variable can be assumed without inversion.
Lemma 1: 𝑌𝑘 has ME representation given by the following recursion
for 𝑖 = 2, … , 𝐷𝑘 ,
if the 𝑖𝑡ℎ variable is 𝑋𝑖 , then calculate
𝑖 𝑖−1 𝑖 𝑖−1 𝑖 𝑖−1
𝒂𝑌 = 𝒂𝑌 ⊗ 𝒂𝑖 𝑨𝑌 = 𝑨𝑌 ⊕ 𝑨𝑖 𝒗𝑌 = 𝒗𝑌 ⊗ 𝒗𝑖
else calculate
𝑖−1 𝑖−1
𝑖 𝑖−1 𝑖−1 𝑖 𝑨𝑌 𝟎 𝑖 𝒗𝑌
𝒂𝑌 = 𝒂𝑌 − 𝒂𝑌 ⊗ 𝒂𝑖 𝑨𝑌 = 𝑖−1
𝒗𝑌 = 𝑖−1
𝟎 𝑨𝑌 ⊕ 𝑨𝑖 𝒗𝑌 ⊗ 𝒗𝑖
with initial conditions
(1) (1) (1)
𝒂𝑌 = 𝒂1 𝑨𝑌 = 𝑨1 𝒗𝑌 = 𝒗1
6. ME distribution: closure properties
3. Addition of disjoint product: 𝑋𝑆 = 𝑌1 + 𝑌2
The matrix exponential function for 𝑋𝑆 , which is the network reliability function
𝑅𝑆 𝑡 , is
𝑅𝑆 𝑡 = 𝒂𝑌1 𝑒 𝑨𝑌1 𝑡 𝒗𝑌 1 + 𝒂𝑌2 𝑒 𝑨𝑌2 𝑡 𝒗𝑌 2

𝑒 𝑨𝑌 1 𝑡 𝟎 𝒗𝑌1
= 𝒂𝑌1 𝒂𝑌2 𝒗𝑌2
𝟎 𝑒 𝑨𝑌 2 𝑡
= 𝒂𝑆 𝑒 𝑨𝑆𝑡 𝒗𝑆
Theorem 3. 𝑋𝑆 has an ME representation (𝒂𝑆 , 𝑨𝑆 , 𝒗𝑆 ) where
𝑨𝑌1 𝟎 𝒗𝑌1
𝒂𝑆 = 𝒂𝑌1 𝒂𝑌2 𝑨𝑆 = 𝒗𝑆 = 𝒗
𝟎 𝑨𝑌2 𝑌2

In a few cases, 𝑌1 and 𝑌2 have the same expression. Then, 𝒂𝑌1 𝑒 𝑨𝑌1 𝑡 𝒗𝑌1 +
𝒂𝑌2 𝑒 𝑨𝑌2 𝑡 𝒗𝑌 2 = 2𝒂𝑌1 𝑒 𝑨𝑌1 𝑡 𝒗𝑌1 , i.e., without augmenting the representation.
6. ME distribution: closure properties
• For a general addition operation involving more than two variables, the
representation of 𝑋𝑆 can be obtained recursively.
• Let 𝐾 be the number of disjoint products in SDP and let 𝑘 = 1, 2, … , 𝐾 represent
their sequence.
Lemma 2: 𝑋𝑆 has an ME representation (𝒂𝑆 , 𝑨𝑆 , 𝒗𝑆 ) calculated recursively as follows
For 𝑘 = 2, … , 𝐾, calculate
𝑘 𝑘−1
𝒂𝑆 = [𝒂𝑆 𝒂𝑌𝑘 ]
𝑘−1
𝑘 𝑨𝑆 𝟎
𝑨𝑆 =
𝟎 𝑨𝑌𝑘
𝑘−1
𝑘 𝒗𝑆
𝒗𝑆 =
𝒗𝑌𝑘
The initial conditions are
1 1 1
𝒂𝑆 = 𝒂𝑌1 𝑨𝑆 = 𝑨𝑌1 𝒗𝑆 = 𝒗𝑌1
7. SDP algorithm for dynamic network reliability analysis
Procedures to conduct dynamic network reliability analysis:
1. Define 𝑋𝑆 and 𝑋𝑖 , where 𝑖 = 1, 2, … , 𝑁, as the states of the network and its 𝑖𝑡ℎ
component
2. Find the network structure function from its minimum paths
𝑀

𝑋𝑆 = ሧ ෑ 𝑋𝑖
𝑗=1 𝑖∈𝑀𝑃𝑗
4. Find the network SDP representation
𝐾

𝑋𝑆 = ෍ ෑ 𝑍𝑖
𝑘=1 𝑖∈𝐷𝑃𝑘
5. Find ME representation of 𝑋𝑆 . The result is an ME representation of the
network reliability function.
6. Generate the network hazard function
7. Calculate the network MTTF and VTTF
7. SDP algorithm for dynamic network reliability analysis
The procedure is summarized below.

− SDP representation of the network (output of an SDP generating

Input algorithm)
− 𝒂𝑖 , 𝑨𝑖 , 𝒗𝑖 for 𝑖 = 1, 2, … , 𝑁
− Calculate 𝒂𝑌𝑘 , 𝑨𝑌𝑘 , 𝒗𝑌𝑘 using multiplication operations
− Calculate 𝒂𝑆 , 𝑨𝑆 , 𝒗𝑆 using addition operation
− Plot 𝑅𝑆 (𝑡) using 𝑅𝑆 𝑡 = 𝒂𝑆 𝑒 𝑨𝑆𝑡 𝒗𝑆 for several values of 𝑡
Process
− Generate and plot ℎ𝑆 (𝑡) using ℎ𝑆 𝑡 = − 𝒂𝑆 𝑨𝑆 𝑒 𝑨𝑆𝑡 𝒗𝑆 (𝒂𝑆 𝑒 𝑨𝑆𝑡 𝒗𝑆 )−1
for several values of 𝑡
− Calculate network’s MTTF, VTTF, and CVTTF
Output − 𝑅𝑆 (𝑡), ℎ𝑆 (𝑡), network’s MTTF, network’s CVTTF
8. Experiment: modified ARPA network
• The algorithms is demonstrated to generate the ME representation of the lifetime
distribution of a network whose components’ lifetimes have an ME distribution
(including a PH distribution).
• The SDP representation of the network is obtained using the KDH88 algorithm.
• The resulting ME representation is utilized to generate the network’s reliability and
hazard functions and calculate the network’s MTTF and CVTTF.
• The network for experiment is a modified ARPA network
8. Experiment: component lifetime distributions
Several ME distributions (including PH distributions) with MTTF 100 represent
component lifetime as inputs.
No Distribution Type CVTTF Representation
1 Exponential (𝜆) CFR 1 𝒂= 1 𝑨= −𝜆 𝒗=𝒆
−𝑘𝜆 𝑘𝜆 0 0
1/√𝑘 0 −𝑘𝜆 𝑘𝜆 0
2 𝑘-stage Erlang (𝑘𝜆) IFR 𝒂 = 1, 0, … , 0 𝑨 = 𝒗=𝒆
(< 1) ⋮ ⋮ ⋱ ⋮
0 0 0 −𝑘𝜆
Two-stage Coxian −𝜆1 𝑝𝜆1
3 DFR >1 𝒂 = 1, 0 𝑨= 𝒗=𝒆
(𝜆1 , 𝜆2 , 𝑝) 0 −𝜆2
Hyperexponential (HE)
11 4 −24 3
4 in ME representation DFR 1.1391 𝒂 = −1, 2 𝑨= 𝒗=
with full matrix 𝑨 18000 13 −33 2

Pure ME for 1 0 −2 2
5 IFR 0.5 𝜶 = 1, 0, 0 𝑨= 3 2 −6 𝒗=𝒆
𝑡 ∊ (0,200) 50
2 2 −5
8. Experiment: network reliability & hazard functions
−1
𝑅𝑆 𝑡 = 𝒂𝑆 𝑒 𝑨𝑆 𝑡
𝒗𝑆 ℎ𝑆 𝑡 = −𝒂𝑆 𝑨𝑆 𝑒 𝑨𝑆𝑡 𝒗𝑆 𝑨
𝒂𝑆 𝑒 𝒗𝑆
𝑆 𝑡

Reliability function Hazard function

8. Experiment: network reliability measure statistics

Components and network statistics

Component Network
No
Distribution Order MTTF CVTTF Order MTTF CVTTF
1 Coxian 2 100 1.5 17,304 52.7480 0.7007
3 HE in Full Matrix 2 100 1.1391 17,304 62.8638 0.6356
4 Exponential 1 100 1 141 66.9841 0.6109
5 Genuine ME 3 100 0.5 380,781 88.4575 0.2891
6 Erlang 2 100 0.7071 17,304 79.5767 0.4087
7 Erlang 3 100 0.5774 380,781 84.6241 0.3273
8 Erlang 4 100 0.5 3,748,032 87.4050 0.2806
9 Erlang 5 100 0.4472 23,074,125 89.1917 0.2495
8. Experiment: Weibull distribution
Weibull distribution may be represented as PH distribution
1.5
𝑅 𝑡 = 𝑒 −(0.1𝑡) whose CVTTF 2 = 0.46
−0.270 0.147 0.099
Its approximate PH representation: 𝒂 = 1 0 0 𝑨 = 0 −0.270 0.270 𝒗 = 𝒆
0 0 −0.270

The maximum difference is 0.0071.

The maximum difference is 0.005

Component reliability function Network reliability function

9. Extension to general system: an approach
• General systems usually comprise many components with different structures.
1. Series
2. Parallel
3. Standby
4. K-out-of-N
5. Bridge
• To be more efficient, components with structures other than a bridge should be
consolidated into subsystems.
• The ME representations of subsystems’ lifetime distributions are derived from their
components’ lifetime distributions using closure properties.
9. Extension to general system: simplification
Let the ME representations of component 1 and 2 be (𝒂1 , 𝑨1 , 𝒗1 ) and (𝒂2 , 𝑨2 , 𝒗2 ). The
ME representation 𝒂𝑆𝑆 , 𝑨𝑆𝑆 , 𝒗𝑆𝑆 of a subsystem comprising components 1 and 2.
1. Series structure
Its SDP representation is 𝑋𝑆𝑆 = 𝑋1 𝑋2 . Hence,
𝒂𝑆𝑆 = 𝒂1 ⊗ 𝒂2 𝑨𝑆𝑆 = 𝑨1 ⊕ 𝑨2 𝒗𝑆𝑆 = 𝒗1 ⊗ 𝒗2
2. Parallel structure
Its SDP representation is 𝑋𝑆𝑆 = 𝑋1 + 𝑋ത1 𝑋2 = 𝑋1 + 𝑋2 − 𝑋1 𝑋2 . Hence,
𝑨𝟏 𝟎 𝟎 𝒗1
𝒂𝑆𝑆 = 𝒂1 𝒂2 − 𝒂1 ⊗ 𝒂2 𝑨𝑆𝑆 = 𝟎 𝑨𝟐 𝟎 𝒗𝑆𝑆 = 𝒗2
𝟎 𝟎 𝑨𝟏 ⊕ 𝑨𝟐 𝒗1 ⊗ 𝒗2
3. Standby structure
This structure does not have an SDP representation
𝑨1 −𝑨1 𝒗1 𝒂2 𝒗1
𝒂𝑆𝑆 = 𝒂1 𝟎 𝑨𝑆𝑆 = 𝒗𝑆𝑆 = 𝒗
𝟎 𝑨2 2
9. Extension to general system: example
Consider a general system structure of 14 components

Some components can be consolidated as subsystems

9. Extension to general system: simpilification example
Formulas for obtaining ME representation of each subsystem
Subsystem ME representation
𝑨𝑎 −𝑨𝑎 𝒗𝑎 𝒂𝑏 𝒗𝑎
Standby: 𝒂𝑆𝑆 = 𝒂𝑎 𝟎 𝑨𝑆𝑆 = 𝒗𝑆𝑆 = 𝒗
𝟎 𝑨𝑏 𝑏
𝑆𝑆1 , 𝑆𝑆4
For 𝑆𝑆1 : 𝑎 = 1, 𝑏 = 2; 𝑆𝑆4 : 𝑎 = 7, 𝑏 = 8
Series: 𝒂𝑆𝑆 = 𝒂𝑎 ⊗ 𝒂𝑏 𝑨𝑆𝑆 = 𝑨𝑎 ⊕ 𝑨𝑏 𝒗𝑆𝑆 = 𝒗𝑎 ⊗ 𝒗𝑏
𝑆𝑆2 , 𝑆𝑆3 , For 𝑆𝑆2 : 𝑎 = 3, 𝑏 = 4; 𝑆𝑆3 : 𝑎 = 9, 𝑏 = 10; 𝑆𝑆5 : 𝑎 = 5, 𝑏 = 6;
𝑆𝑆5 , 𝑆𝑆6 𝑆𝑆6 : 𝑎 = 11, 𝑏 = 12.

𝒂𝑆𝑆 = 𝒂13 𝒂14 −𝒂13 ⊗ 𝒂14

Parallel: 𝑨13 𝟎 𝟎 𝒗13
𝑆𝑆7 𝑨𝑆𝑆 = 𝟎 𝑨14 𝟎 𝒗𝑆𝑆 = 𝒗14
𝟎 𝟎 𝑨13 ⊕ 𝑨14 𝒗13 ⊗ 𝒗14
9. Extension to general system: graphical results

Reliability function Hazard function

9. Extension to general system: statistical results

Components Network
Order/Stage Distribution CVTTF Order MTTF CVTTF
2 Coxian 1.5 115,712 29.0879 0.6771
2 2-HE in ME 1.1391 115,712 35.7482 0.6570
1 Exponential 1.0 126 38.6761 0.6455
2 Erlang 0.7071 115,712 56.4948 0.4175
3 Genuine ME 0.5 9,484,290 69.9703 0.3021
10. Advantages of ME distribution: network for experiment
• Comparison with using PH distribution closure properties (see Part 1)
• Three methods are compared
1. Based on PH distribution closure properties (CP)
2. Based on SDP with ME representation (SDP)
3. Based on SDP with ME in Jordan canonical form (SDP Jordan)
• PH distribution closure properties is applicable for ME distribution with 𝒗 = 𝒆
• Network for making comparison must not contain bridge since PH distribution
closure properties are not applicable for network with bridges
10. Advantages of ME distribution: PH distribution
• PH representation of component reliability function
−6 4 2
𝒂 = 1, 0, 0 , 𝑨 = 2 −5 2 , 𝒗=𝒆
1 2 −4
• Comparison of computing time

Component Network Computation time

No lifetime Method lifetime MTTF and 21 points of R(t)
representation representation CVTTF and h(t)
1 CP PH 99.73 sec ≈ 5.9 days
2 PH SDP ME 13.30 sec 17.51 hours
3 SDP Jordan ME 0.0063 sec 0.1321 sec
10. Advantages of ME distribution: minimal representation
−1 1 0 0
275 55 8 2 0 −1 1 0
• PH representation 𝒂 = , , ,
306 714 1071 119
, 𝑨= 0 0 −1 1 , 𝒗 = 𝒆
9
0 0 0 −
2

−1 1 0
275 55 9
has a minimal ME representation 𝒂 = 238
,− ,
238 119
, 𝑨= 0 −1 1 , 𝒗 = 𝒆
0 0 −1
• Comparison of computing time using the original PH representation, minimal ME
representation, and its Jordan canonical form
Component Network Computation time
No lifetime Method lifetime MTTF and 21 points of
representation representation CVTTF R(t) and h(t)
1 CP PH 0.3595 sec 18 hours
PH
2 SDP ME 0.2054 sec 4.1 hours
3 CP ME 0.0277 sec 333.7 sec
Minimal ME
4 SDP ME 0.0205 sec 57.28 sec
Discussion
• Dynamic network reliability analysis is made possible by utilizing some closure
properties of PH and ME distribution
• The method is exact and applicable for systems with general structures that may
contains DFR, IFR, CFR, and even BFR components.
• The result is algorithms readily programmed in MATLAB, available from the article
link or (including all experiments reported in the article):
https://github.com/mochamadnurq/reliability-networksystem-matrixexponential
• Possible extensions of the model are to include:
• Dependent components
• Discrete time analysis (accepted for publication)
• General standby systems (submitted for publication)
• Application for binary decision diagram (BDD) and phased mission system
• The challenge is to reduce computation time by:
• exploiting the resulted matrix structures
• using numerical techniques specifically for large matrices with Kronecker structures.