Measurement method of metro rail corrugation based on the SEQ2SEQ model and the fusion data of vibration and noise
A metro rail corrugation measurement method based the Seq2Seq model and vibration and noise data fusion comprises the following steps: Constructing a time series data set corresponding to floor longitudinal acceleration and train speed based on historical data of metro trains; Constructing the metro train mileage matching prediction model based on the Seq2Seq model, training the metro train mileage matching prediction model by time series data sets, and inputting the inside floor longitudinal acceleration of the metro train to be detected into the trained metro train mileage matching prediction model to obtain the running speed of the metro train to be detected. According to the vibration and noise data of the metro train to be detected, the vibration and noise composite index of the rail corrugation is calculated to determine the wavelength and amplitude information corresponding to the rail corrugation.
1 . A metro rail corrugation measurement method based on a Seq2Seq model and vibration and noise data fusion, comprising a non-transitory computer readable medium operable on a computer with memory for the metro rail corrugation measurement method, and comprising program instructions for executing the following steps of:
constructing a time series data set corresponding to floor longitudinal acceleration and train speed based on historical data of a metro train, the time series data set comprising a longitudinal acceleration time series and a train speed time series;
wherein constructing the time series data set corresponding to the floor longitudinal acceleration and the metro train based on the historical data of the metro train comprises:
acquiring a vertical and longitudinal vibration acceleration signals from an acceleration sensor arranged on a floor surface above a bogie of the metro train based on the historical data of the metro train, and obtaining vertical and longitudinal vibration time domain information of the floor in the metro train based on the vertical and longitudinal vibration acceleration signals;
acquiring a measured sound pressure signal from a sound pressure sensor arranged in a carriage above the bogie of the metro train, and obtaining a sound pressure time domain information in the metro train accordingly;
acquiring a measured speed signal from a speed measuring wheelset sensor arranged on an axle of the metro train, and obtaining a train speed signal accordingly; and
dividing the longitudinal vibration acceleration signal and the train speed signal into segmented longitudinal acceleration signals based on a preset window length, to generate the time series data set corresponding to the longitudinal vibration acceleration signal and the train speed, wherein the dividing comprises:
the longitudinal vibration acceleration signal and the train speed signal are resampled to ensure that a length of the longitudinal acceleration time series is equal to a length of the train speed time series, setting the length as L s ;
determining a sample number N of the time series data set according to formula (1), wherein the longitudinal acceleration time series is denoted as x i (i=0, 1, . . . , x n ), and the train speed time series is denoted as y j (j=0, 1, . . . , y n ), and a time step number is set to K:
N
=
L
s
-
K
+
1
;
(
1
)
and
generating the time series data set corresponding to the longitudinal vibration acceleration signal and the train speed according to formula (2);
{
[
[
x
0
,
x
1
,
…
,
x
K
-
1
]
,
[
y
0
,
y
1
,
…
,
y
K
-
1
]
]
,
(
2
)
[
[
x
K
,
x
K
+
1
,
…
,
x
2
K
-
1
]
,
[
y
K
,
y
K
+
1
,
…
,
y
2
K
-
1
]
]
,
…
[
[
x
NK
,
x
NK
+
1
,
…
,
x
NK
-
1
]
,
[
y
NK
,
y
NK
+
1
,
…
,
y
NK
-
1
]
]
}
;
constructing a metro train mileage matching prediction model based on the Seq2Seq model, which is trained by the time series data set corresponding to the longitudinal acceleration;
inputting a longitudinal acceleration signal of a floor inside of the metro train to be detected into the trained metro train mileage matching prediction model based on the Seq2Seq model to obtain a running speed of the metro train to be detected;
calculating a vibration and noise composite index of the metro rail corrugation to be detected based on a preset wavelength distribution range of the metro rail corrugation, the running speed of the metro train to be detected, a vertical vibration acceleration of the floor, and a sound pressure signal in the metro train;
determining a specific position of the metro rail corrugation, a corresponding characteristic wavelength, and an amplitude based on the rail corrugation vibration and noise composite index; and
outputting the determined specific position of the metro rail corrugation on a train rail, the characteristic wavelength, and the amplitude determined by the metro rail corrugation measurement method, to mitigate the metro rail corrugation and to reduce vibration and noise.
2 . The method according to claim 1 , wherein constructing and training the metro train mileage matching prediction model based on the Seq2Seq model comprises:
constructing the metro train mileage matching prediction model as a Many to Many model of a recurrent neural network comprising an encode and a decoder, wherein the encoder utilizes a recurrent neural network model to encode a longitudinal acceleration time series and outputs a hidden state of a last step, and the decoder utilizes a recurrent neural network structure and applies the hidden state of the last step of the encoder with a train speed time series as input;
dividing the time series data set corresponding to the longitudinal acceleration and the train speed into a training set, a verification set and a test set;
inputting the training set into the metro train mileage matching prediction model and searching for optimal parameters of the model through K-fold cross verification; and
inputting the verification set into the metro train mileage matching prediction model using optimal parameters to verify the metro train mileage matching prediction model to obtain the trained metro train mileage matching prediction model, wherein input data of the trained metro train mileage matching prediction model is the longitudinal acceleration time series of the floor of the metro train, and the output data is a time series corresponding to a running speed of the metro train.
3 . The method according to claim 1 , wherein calculating the rail vibration and noise composite index comprises:
performing segmented band-pass filtering on the vertical vibration acceleration of the floor and the sound pressure signal in the metro train according to the preset wavelength distribution range and the running speed;
calculating a vibration index and a noise index corresponding to the vertical vibration signal and the sound pressure signal after the band-pass filtering; and
retaining segments of the vertical vibration acceleration and the sound pressure signal having a vibration index and a noise index greater than a preset threshold, and performing further calculation of the rail corrugation vibration and noise composite index using the retained segments.
4 . The method according to claim 3 , wherein performing the band-pass filtering on the vertical vibration acceleration of the floor and the sound pressure signal comprises:
assuming that a floor vertical vibration acceleration signal of the metro train to be detected is equal to the sound pressure signal in the car and is set as L v , setting a segment length as W and overlap ratio as ¾; wherein if a remaining signal length is greater than or equal to W, adding 0 until the remaining signal length is equal to W; otherwise, abandon a remaining signal;
calculating corresponding upper and lower cutoff frequencies [f L , f H ] based on the preset wavelength distribution range [λ L , λ H ] of the metro rail corrugation according to formula (3):
f
L
=
V
_
w
λ
H
,
f
H
=
V
_
W
λ
L
,
(
3
)
wherein V W represents an average speed of segmented trains determined by the metro train mileage matching prediction model; and
designing a band-pass filter based on a window function method with a filter order n, and filtering the segmented vertical vibration acceleration and the segmented sound pressure signal to obtain a band-pass filtered segmented vertical vibration acceleration signal (sig vb ) and a band-pass filtered sound pressure signal (sig no ).
5 . The method according to claim 3 , wherein calculating the vibration index and noise index comprises:
calculating a self-power spectral density of the segmented floor vertical vibration acceleration signal sig vb , identifying a peak frequency f vmax in the [f L , f H ] frequency band, performing a second band-pass filtering on the signal sig vb according to a frequency range [f vmax −D, f vmax +D], and determining the vibration index according to formula (4):
VI
(
i
)
=
S
Vi
(
filter
(
f
vmax
-
D
,
f
vmax
+
D
)
[
sig
vb
]
)
S
Vtotal
,
(
4
)
wherein VI(i) is the vibration index of an i th frame of the floor vertical vibration acceleration signal, sig vb is the vibration signal after the first filtering, “filter” represents the band-pass filter, D is half of a pass-band width, f vmax is the peak frequency of the i th frame vibration signal in the self-power spectral density, S Vtotal is an effective value of the vibration signal after the first filtering, and S Vi is an effective value of the vibration signal after the second filtering determined by formula (5):
S
Vi
=
∑
n
=
i
i
+
W
x
V
n
2
W
V
,
(
5
)
in the formula (5), W V is a length of each frame of the vibration signal; and
calculating a self-power spectral density of the segmented inside sound pressure signal sig no , and identifying a peak frequency f nmax in the frequency band [f L , f H ], performing a second band-pass filtering on the signal sig no according to a frequency range [f nmax −D, f nmax +D], and determining the noise index according to formula (6):
NI
(
i
)
=
S
Ni
(
filter
(
f
nmax
-
D
,
f
nmax
+
D
)
[
sig
no
]
)
S
Ntotal
,
(
6
)
wherein NI(i) is the noise index of an i th frame of the noise signal, sig no is the noise signal after the first filtering, f nmax is the peak frequency of the i th frame noise signal in the self-power spectral density, S Ntotal is an effective value of the noise signal after the first filtering, and S Ni is an effective value of the noise signal after the second filtering determined by formula (7):
S
Ni
=
∑
n
=
i
i
+
W
x
N
n
2
W
N
,
(
7
)
wherein W N is a length of each frame of the noise signal.
6 . The method according to claim 3 , wherein performing the further calculation of the rail corrugation vibration and noise composite index comprises:
determining a vibration index threshold VI m and a noise index threshold Nlm;
calculating a cross-power spectrum between the vertical vibration acceleration of the floor and the sound pressure signal if the vibration index VI(i) is greater than VI m and the noise index NI(i) is greater than NI m ,
identifying a peak frequency f vnmax in the frequency band [f L , f H ], and band-pass filtering an acoustic-vibration composite signal according to a frequency range [f vnmax −D, f vnmax +D] to determine the rail corrugation vibration and noise composite index according to formula (8):
CVNI
(
i
)
=
S
VNi
(
filter
(
f
vnmax
-
D
,
f
vnmax
+
D
)
[
sig
vn
]
)
S
VNtotal
,
(
8
)
wherein CVNI(i) is the vibration and noise composite index of an i th frame, sig vn is the acoustic-vibration composite signal, f vnmax is the peak frequency in the cross-power spectrum of the i th frame acoustic-vibration signal, S VNtotal is an effective value of the acoustic-vibration composite signal after the first filtering, and S VNi is an effective value of the acoustic-vibration composite signal after the second filtering determined by formula (9):
S
VNi
=
∑
n
=
i
i
+
W
x
V
N
n
2
W
V
N
.
(
9
)
7 . The method according to claim 1 , wherein determining a specific location of the metro rail corrugation, the corresponding characteristic wavelength and the amplitude comprises:
determining a train position x(t) based on the speed v(t) obtained by the metro train mileage matching prediction model according to formula (10);
x
(
t
)
=
∫
0
t
v
(
t
)
dt
;
(
10
)
identifying the characteristic wavelength of the metro rail corrugation based on the peak frequency f vnmax and an average vehicle speed V corresponding to the segmented acoustic-vibration composite signal according to formula (11):
λ
i
,
detect
=
{
0
,
no
corrugation
in
frame
i
V
f
vnmax
,
corrugation
in
frame
i
;
(
11
)
and
identifying a severity of the metro rail corrugation representing the amplitude based on the rail corrugation vibration and noise composite index (CVNI) according to formula (12):
{
CVNI
(
i
)
<
0
.
4
→
general
corrugation
0.3
<
CVNI
(
i
)
<
0
.
6
→
severe
corrugation
CVNI
(
i
)
>
0
.
6
→
v
e
r
y
severe
corrugation
.
(
12
)