Method of analysing flight data
View Patent ↗The invention relates to a method of analyzing flight data recorded during N flights of at least one aircraft, by means of a flight data recorder of an aircraft, the data being grouped together by flight i in a signature vector of the flight X i of size d, the components of which correspond to data recorded during said flight i of the aircraft, a flight i being thus defined by the signature vector X i , the method comprising the following steps: Gaussian kernel entropy component analysis of the flight signatures X i to obtain a zone of normal flights and classifying the flight signatures X i with respect to their distance to said zone; determining, for each flight i, an abnormality score z i defined by the distance of a flight signature X i with respect to the zone of normal flights; detecting, as a function of the abnormality score z i , at least one abnormal flight.
1. Method of analysing flight data comprising the following steps:
obtaining flight data recorded during N flights of at least one aircraft, said flight data have been recorded with a flight data recorder of an aircraft;
grouping in a signature vector X i of a flight i, the data corresponding to the flight i said signature vector X i being of size d, the components of the signature vector being the data recorded during said flight i;
obtaining a zone of normal flights and classifying the flight signatures X i with respect to their distance to said zone, using a Gaussian kernel entropy component analysis of the flight signatures X i ;
determining, for each flight i, an abnormality score z i defined by the distance of a flight signature X i with respect to the zone of normal flights;
detecting, as a function of the abnormality score z i , at least one abnormal flight;
determining for each abnormal flight, a phantom flight the nearest to the abnormal flight while being in the zone of normal flights,
wherein the Gaussian kernel entropy component analysis comprises the following sub-steps:
determination of a matrix of similarity K of size N×N, the components of which quantify the proximity between two flight signatures X i ;
breakdown into eigen vectors of the matrix of similarity K to obtain N eigen vectors a 1 , . . . , a N and N eigen values λ 1 , . . . , λ N such that ∀i=1, . . . , N K·a i =λ i ·a i ;
determination for each eigen vector of its entropy coefficient γ i ;
selection of a sub-assembly of eigen vectors {a m } mε{1, . . . , N} such that the sum of the entropies γ m is greater than the percentage of the sum of the N entropies γ i .
2. The analysis method according to claim 1 , in which the parameters of the abnormal flight detected are compared with those of the phantom flight determined in order to detect at least one parameter of the abnormal flight which has rendered abnormal said abnormal flight detected.
3. The analysis method according to claim 1 , in which the entropy coefficient is defined by:
∀
i
=
1
,
…
,
N
γ
i
=
λ
i
·
(
∑
j
=
1
N
a
i
,
j
)
N
2
.
4. The analysis method according to claim 1 in which the abnormality score of a flight i is defined by:
z
i
=
1
-
∑
k
∈
m
(
K
i
T
a
k
)
2
λ
k
.
5. The analysis method according to claim 1 , in which the components of the matrix of similarity K are defined by:
∀
i
=
1
,
…
,
N
K
i
,
j
=
exp
(
-
X
i
-
X
j
2
2
σ
2
)
,
where σ 2 is a predetermined bandwidth parameter.
6. The analysis method according to claim 5 , in which the bandwidth parameter σ 2 is determined in the following manner:
a matrix D of size N×N corresponding to the distance between two flights is determined and is defined in the following manner ∀i=1, . . . , N D i,j =∥X i −X j ∥ 2 ;
each component of each column is arranged in ascending order to obtain a matrix D′;
k first lines of the matrix D′ thereby obtained are selected and the others are eliminated to obtain a matrix D′ of size k×N;
the average of each column of the matrix D′ to obtain N values y 1 , . . . , y N is determined;
the median absolute deviation of the values y 1 , . . . , y N defined by mad=mediane{|y i −med|} with med=mediane{y1, . . . , y N } is determined;
the bandwidth parameter σ 2 is determined from the median absolute deviation of the values y 1 , . . . , y N by the following functional
σ
2
=
max
y
i
<
med
+
20
·
mad
{
y
i
}
.
7. The analysis method according to claim 1 , in which the data of a flight i are grouped together in a matrix F i of dimension T×P with T the number of data recorded during the flight i and P the number of parameters recorded, the vector X i having for components the columns of the matrix F i end to end, the vector X i being then of dimension d=T×P and is defined by
X i =[F i ( t= 1, p= 1), . . . , F i ( t= 1, p=P ), . . . , F i ( t=T,p=P ).
8. The analysis method according to claim 1 , in which the data of a flight i are grouped together in a matrix F i of dimension T×P with T the number of data recorded during the flight i and P the number of parameters recorded, the vector X i having for components the columns of the matrix F i sampled to select n<T recordings t 1 i , . . . , t n i of parameters, the vector X i being then of dimension d=n×P and is defined by
X i =[F i ( t=t 1 i ,p= 1), . . . , F i ( t=t n i ,p= 1), . . . , F i ( t=t 1 i ,p=P ), . . . , F i ( t=t n i ,p=P ).
9. The analysis method according to claim 1 , in which the data of a flight i are grouped together in a matrix F i of dimension T×P with T the number of data recorded during the flight i and P the number of parameters recorded, the vector X i having for components the average, the variance, the minimal value, the maximal value of a parameter among all the recordings of this parameter, the vector X i being then defined by:
X
i
=
[
mean
(
F
i
(
p
=
1
)
)
,
var
(
F
i
(
p
=
1
)
)
,
max
(
F
i
(
p
=
1
)
)
,
max
(
F
i
(
p
=
1
)
)
,
⋮
mean
(
F
i
(
p
=
P
)
)
,
var
(
F
i
(
p
=
P
)
)
,
max
(
F
i
(
p
=
P
)
)
,
max
(
F
i
(
p
=
P
)
)
]
.