JID:AESCTE AID:105747 /FLA [m5G; v1.261; Prn:28/01/2020; 16:18] P.

1 (1-8)
Aerospace Science and Technology ••• (••••) ••••••

1 67
Contents lists available at ScienceDirect
2 68
3 69
4 Aerospace Science and Technology 70
5 71
6 72
7
www.elsevier.com/locate/aescte 73
8 74
9 75
10 76
11
12
Information-reusing alignment technology for rotating inertial 77
78
13 navigation system 79
14 80
15 a,∗ a a b 81
Qiangwen Fu , Sihai Li , Yang Liu , Feng Wu
16 82
a 83
17 School of Automation, Northwestern Polytechnical University, 710072 Xi’an, China
b
18 Shanghai Aerospace Control Technology Institute, Shanghai, China 84
19 85
20 86
21 a r t i c l e i n f o a b s t r a c t 87
22 88
Article history: This paper proposes a real-time information-reusing alignment strategy for rotating inertial navigation
23 89
Received 30 January 2019 system (INS), which is devoted to improving alignment performance in a limited time by making full use
24 Received in revised form 11 December 2019 90
of all sensor data. In the forward coarse alignment stage, all sensor data was recorded for subsequent
25 Accepted 24 January 2020 91
fine alignment while the rough attitude was determined, and the well-designed data extraction strategy
26 Available online xxxx 92
Communicated by Roberto Sabatini
significantly reduced the memory space and the computation load without losing solution accuracy.
27 In the fine alignment stage, the stored data was used for backtracking navigation update and Kalman 93
28 filtering in the local level geographic frame rather than the inertial frame, simplifying the fine alignment 94
Keywords:
29 Initial alignment algorithm. The lever arm between the center of the inertial measurement unit (IMU) and the rotating 95
30 Backtracking navigation center of the turntable was automatically estimated by the fine alignment Kalman filter, which helped to 96
31 Rotation modulation suppress motion interference and improve alignment performance. Finally, the simulation tests and the 97
Inertial navigation system (INS) practical experiments justified the proposed information-reusing alignment method.
32 98
33 © 2020 Published by Elsevier Masson SAS. 99
34 100
35 101
36 102
37 103
1. Introduction with 180◦ azimuth rotation has proven to be optimal, and this
38 104
scheme is commonly used to improve the heading estimate accu-
39 105
40 The inertial navigation system (INS), as a dead-reckoning nav- racy [5–7]. However, the multi-position alignment technique can 106
41 igation method, has been widely used in aerospace, aviation, ma- do nothing for the variable sensor errors. Instead, the rotation 107
42 rine and other fields due to its existing advantages of autonomy, modulation technology is presented as a systematic error auto- 108
43 concealment and continuity [1,2]. The initial alignment process is compensation method, which can modulate not only the constant 109
44 of vital importance for an INS because this process largely deter- biases but also the low-frequency variable errors of the inertial 110
45 mines the subsequent navigation accuracies [3]. On a stationary sensors into zero-mean periodical values by continuously rotat- 111
46 base, the INS has such poor observability that the initial self- ing the inertial measurement unit (IMU) around the azimuth axis 112
47 alignment accuracy is restricted by the performance of the inertial [8–10]. Therefore, the rotation modulation technology can achieve 113
48 sensors. Specifically, the level attitude alignment accuracy is re- more accurate alignment results than the multi-position technique, 114
49 stricted by the accelerometer biases, while the eastern gyro bias in principle. However, the continuous rotation introduces extra lin- 115
50 mainly determines the azimuth alignment accuracy, and the con- ear and angular movement interferences. Thus, the alignment con- 116
51 vergence time of the azimuth misalignment is seriously affected dition is further contaminated and need to be taken seriously. 117
52 by the gyro random walk [4,5]. Due to the expensive cost and low 118
Traditional initial alignment methods are usually composed of
53 yield for ultra-high accuracy gyros, an accurate azimuth determi- 119
coarse alignment and fine alignment stages [11,12]. Typically, the
54 nation within a given time is left as a daunting task in the initial 120
coarse alignment stage is required to estimate the vehicle heading
55 alignment stage. 121
with an accuracy of a few degrees and to estimate the pitch/roll
56 Given the same inertial sensors, a multi-position alignment 122
with an accuracy of a few tenths of a degree to allow the fine
57 technique can improve the observability of the errors and achieve 123
58 higher attitude accuracy by separating the estimation results from alignment filter to operate within its linear region [13]. Under 124
59 the constant sensor biases. The two-position alignment strategy swaying conditions, a longer time is necessary in the coarse align- 125
60 ment stage to suppress the linear and angular movement inter- 126
61 ferences. However, the coarse alignment time is only used to de- 127
62 * Corresponding author. termine a rough initial attitude matrix, without directly contribut- 128
63 E-mail address: [email protected] (Q. Fu). ing to the final alignment accuracy. Therefore, a certain period of 129
64 130
https://doi.org/10.1016/j.ast.2020.105747
65 131
1270-9638/© 2020 Published by Elsevier Masson SAS.
66 132
 JID:AESCTE AID:105747 /FLA [m5G; v1.261; Prn:28/01/2020; 16:18] P.2 (1-8)
2 Q. Fu et al. / Aerospace Science and Technology ••• (••••) ••••••

εE ∇E
1 the coarse alignment time (typically 30–120 s [14]) is inevitably φU = − + tan L (1) 67
2 wasted. ωie cos L g 68
3 To make full use of all of the alignment time, some tradi- where φU denotes the azimuth misalignment error in the n-frame,
69
4 tional methods record the inertial sensor data in each sampling ε E and ∇ E denote the equivalent eastern gyro drift and the eastern
70
5 period during the forward process for a reciprocating calculation 71
accelerometer bias, L is the local latitude, ωie is the earth rotation
6 [15–17]. This kind of reciprocating navigation strategy relies on 72
rate, and g is the gravitational acceleration.
7 a powerful navigation computer to complete the forward calcu- 73
Considering the amplitudes of ωie and g in Equation (1), as well
8 lation, storage, and backward calculation of massive amounts of 74
as the current accuracy level of the inertial sensors, the equivalent
9 data at the same time. However, no matter how powerful the nav- 75
eastern gyro drift ε E is the dominant error source of the azimuth
10 igation computer is, vast storage space and a strong computing 76
11 capability result in an increase of cost and power, and the sud- alignment result. Three gyro errors of the INS are denoted by ε b = 77
12 den computational burden inevitably complicates the scheduling [ εxb εby εbz ]T , and the corresponding error model can be described 78
13 of online navigation software. A data extraction strategy is a good as 79
14 solution to this problem, in which the needed sensor data are com- 80
15 pressed in the inertial frame before being recorded, so the memory εib = ε0i
b
+ εsib + εbwi (2) 81
16 space and the computation load are significantly reduced without b b
82
where ε represents repeatability (constant) biases; ε represents s
17 the loss of the navigation accuracy. To avoid the overshoot caused 0 83
the slow-varying drifts, which are usually modeled as the first-
18 by an error model switching, the fine alignment and subsequent 84
order Markov process ε̇sb = −1/τG εsb + w s [21], where τG is a
19 navigation are arranged in the inertial frame rather than the or- 85
correlated time and w r is a white noise with initial standard de-
20 dinary local level frame in [18,19]. Clearly, the existing approach 86
viation R s (0); εbw represents the random walk noises (fast-varying
21 conflicts with the common usage conventions. Thus, the additional 87
22 error model mechanization and corresponding software design are errors); and the subscript i denotes the x, y, and z gyros. 88
23 needed in the fine alignment stage and the following navigation In the classical multi-position alignment strategy, the constant 89
24 stage. error ε0b of the horizontal gyroscopes can be separated by the fine 90
25 In view of this, we propose an improved information-reusing alignment Kalman filter, but the variable errors εsb and εbw cannot 91
26 alignment strategy for the rotation modulation INS. In the coarse be effectively estimated. 92
27 alignment stage, the dual-inertial vectors attitude determination There are some differences in the situation of a rotation mod- 93
28 (DIVAD) method is applied to estimate the rough attitude matrix ulation INS. Because the IMU rotates with a constant rotation rate 94
29 at the beginning of the alignment. Meanwhile, the attitude quater- ωc while the carrier body frame stays stationary, the projection of 95
30 nions and velocity increments are recorded with a lower frequency. εb on the n-frame can be expressed as follows [8,22]: 96
31 In the subsequent fine alignment stage, the stored data are used  T 97
32 to accomplish the backtracking navigation and the Kalman filter εn = ε E ε N εU = C nbc C bbc εb 98
33 calculation in the local geographic level frame. Thus, the classical ⎡ ⎤ 99
34 strapdown INS (SINS) error model can be directly followed. Con- cos ωc t − sin ωc t 0 100
= C nbc ⎣ sin ωc t cos ωc t 0 ⎦ε b
35 sidering the installing location relationship, the lever arm between 101
36 the sensitive center of the IMU and the rotating center of the 0 0 1 102
turntable is able to deteriorate the velocity measurement noises ⎡ ⎤
37
ε cos ωc t − ε
b
x
b
y sin ωc t 103
by coupling with the rotational angular rates. Since the lever arm ⎣ ε sin ωc t − ε ωc t ⎦
38 104
= C nbc b
x
b
y cos (3)
39 is difficult to measure accurately and to provide compensation for 105
40 in advance, the proposed method treats it as an error state to be εbz 106
41 estimated by the fine alignment Kalman filter in real time. 107
where C nb
is the transition matrix from the bc -frame to the
42 The remainder of this paper is organized as follows. The ro- c 108
n-frame, and t is the rotation time.
43 tation alignment principle is given in Section 2. In Section 3, the 109
During the rotation process, the matrix C nb is almost a constant
44 well-designed information-reusing alignment algorithm is derived c 110
in detail, and the backtracking navigation and the Kalman filter value, and the errors of the two horizontal gyros are modulated to
45 111
method are provided. Section 4 validates the proposed method sine functions [8]. From Equations (2) and (3), the accumulation
46 b 112
through simulation tests and practical experiments. Finally, the results of the horizontal constant biases, ε0x and ε0b y , are zeros in
47 113
48 conclusions are drawn in Section 5. one rotation period. Thus, they have no effect on the equivalent 114
49
eastern gyro drift ε E . In addition, as long as the rotation frequency 115
50 2. Rotation alignment principle is substantially faster than the time scale of the error variation, the 116
51
eastern gyro drift ε E is nearly independent of the slowly-varying 117
52 The coordinate systems involved in this paper are defined as drifts εsb [23]. Similarly, the eastern accelerometer bias ∇ E has the 118
53 follows: The local level east–north–up (ENU) geographic coordinate same modulation effects. 119
54 was chosen as the navigation frame, denoted by n-frame. The IMU Considering the error model Equation (2), the only residual er- 120
55 body frame, denoted by b-frame, was implicitly predefined by the ror in Equation (1) is the gyro random walk noise εbw , which affects 121
56 sensitive axes of the inertial sensors, with axes pointing right, for- the azimuth alignment accuracy as follows: 122
57 ward, and upward. The carrier body frame, denoted by bc -frame, 123
σE
58 was fixed to the carrier with axes pointing right, forward, and up- φU = (4) 124
59 ward. The bc -frame coincided with the b-frame only if the rotary ωie cos L 125
60 encoder output of turntable was zero. where σ E is the standard deviation of the equivalent eastern gyro 126
61 The final aim of the INS alignment process is to determine the random walk. Assuming that the level gyros have independent ran- 127
62 transition matrix from the b-frame to the n-frame, denoted by C nb , dom walk noises with the same standard deviation σ g , such as 128
63 with satisfactory accuracy within a limited time. For the stationary σ gx = σ g and σ g y = σ g , the relationship σ E = σ g can be derived 129
64 base, the autonomous alignment accuracy is determined by the in- from Equation (3). 130
65 ertial sensor errors, and the ultimate azimuth accuracy is limited Denoting the gyro random walk coefficient as N g , the limit az- 131
66 as [4,20] imuth alignment accuracy for a rotary INS is 132
 JID:AESCTE AID:105747 /FLA [m5G; v1.261; Prn:28/01/2020; 16:18] P.3 (1-8)
Q. Fu et al. / Aerospace Science and Technology ••• (••••) •••••• 3

Ng
1
φU = √ (5) 67
2 t ωie cos L 68
3 69
where t denotes the alignment time.
4 70
Equation (5) shows that the azimuth misalignment error is
5 71
mainly caused by the gyro random walk noises in a rotary INS,
6 72
and the only way to reduce its influence is to increase the mea-
7 73
surement time. Since the alignment time is usually limited and
8 74
preset, the most effective way to improve alignment accuracy is to
9 75
make the full use of all of the alignment time.
10 76
11 3. Proposed information-reusing alignment strategy 77
12 78
13 In a traditional alignment method, the entire time is divided 79
14 into coarse and fine alignment stages to accomplish the initial 80
15 alignment. As shown in Fig. 1(a), the coarse and fine alignment 81
16 stages always occupy separate periods of time. The turntable re- 82
17 mains stationary for the coarse alignment in a time interval of Fig. 1. (a) Traditional alignment strategy. (b) Proposed information-reusing align-
83
18 [0, t c ], and it continuously rotates from time t c for the fine align- ment strategy.
84
19 ment. Under swaying conditions, the coarse alignment stage takes 85
20 a longer time to determine a rough attitude matrix that should n b
86
21 meet the linearization requirements of the fine alignment. Since matrices C nb (t ), C nn0 (t ), C b0 , and C b0 (t ) with certain transformation 87
0
22 the coarse alignment has no direct contribution to the final align- relationships [4]. 88
23 ment accuracy, the time segment [0, t c ] is not being used effec- According to the attitude quaternion differential equation, 89
b b
24 tively. Q b0 (t ) can be updated according to the gyro outputs ω̃ ib , as fol- 90
25 The proposed information-reusing alignment strategy is also lows: 91
26 composed of the coarse alignment stage and the fine alignment 92
b0 1 b0
27 stage, as shown in Fig. 1(b). The turntable rotates at the beginning Q̇ b (t ) = Q b
(t ) ⊗ ω̃bib (8) 93
28 of the alignment. The time range [0, t c ] is used for real-time coarse 2 94
n
29 alignment to determine the initial attitude matrix C b0 at the start b0 b0 95
0 where the initial value of quaternion Q b
(t ) is Q b
(0) =
30 time 0. Meanwhile, the attitude quaternions and the velocity incre- 96
31
[1 0 0 0 ]T . 97
ments are extracted in the inertial frame and recorded with a low
32
Similarly, Q nn0 (t ) can be obtained as 98
frequency (1 Hz in this paper). After the coarse alignment is com-
33 pleted at time t c , the fine alignment stage uses the initial attitude 99
n n0 1 n
34 matrix C b0 and recorded data to complete the backtracking navi- Q̇ n (t ) = Q n0 (t ) ⊗ ωnin (9) 100
35
0
2 101
gation and the optimal linear Kalman filtering. The time range [t c ,
n n
36 t f ] is left to execute the software program of the backtracking fine where the initial value of quaternion Q n0 (t ) is Q n0 (0) = 102
37 alignment. Because the stored data amount has been reduced sig- [1 0 0 0 ]T , and ωnin = ωnie + ωnen , in which ωnie denotes the earth 103
38 104
nificantly, the left time interval [t c , t f ] can be shortened to several rotation rate vector in the n-frame, and ωnen denotes the angular
39 seconds for a normal navigation computer. Therefore, the coarse 105
rate of the n-frame with respect to the e-frame. Given no change
40 106
and fine alignment stages consume nearly the same time period. in position, ωnen can be regarded as zero, and ωnie is only deter-
41 107
mined by the initial latitude.
42 3.1. Coarse alignment process b 108
Since C nn0 (t ) and C b0 (t ) can be solved as shown in Equations
43 109
(8) and (9), the determination of time-varying matrix C nb (t ) is con-
44 To suppress the linear and angular movement interferences n 110
verted into the calculation of constant attitude matrix C b0 . If the
45 caused by continuous rotation, the DIVAD method is used to ac- 0 111
46
carrier is not moving, the specific force equation in the n-frame is 112
complish the coarse alignment. Based on the inertial freezing the-
47 simplified as 113
ory [24,25], the b0 -frame and n0 -frame are established by fixing
n0  
48 b 114
C b0 (t ) f˜ (t )
the b-frame and the n-frame relative to the inertial space at the
49 −C nn0 (t ) g n = C b0 b
(10) 115
beginning of the initial alignment. According to the chain rule of
50 116
the attitude matrix, the direction cosine matrix C nb at any time t b
51 where f˜ denotes the specific force measured by accelerometers, 117
can be decomposed into three parts, as follows [16]:
52 and g n is the gravity vector in the n-frame. 118
n b
53
C nb (t ) = C nn0 (t )C b0 C b0 (t ) (6) Using Equation (10), many methods can be used to solve 119
0 n
54 the desired constant matrix C b0 , such as the three-axis attitude 120
55 where C nn0 (t ) represents the time-varying transition matrix from 0
determination (TRIAD) method [26], the q-method [27], or the 121
b 122
56 the n0 -frame to the n-frame, C b0 (t ) represents the time-varying optimization-based method [28]. Since the coarse alignment is
57 n 123
matrix from the b-frame to the b0 -frame, and C b0 denotes the con- only used to determine an approximate attitude matrix, the align-
58 0 124
stant matrix between the b0 -frame and the n0 -frame. ment accuracy, computational complexity, and anti-interference
59 125
Since the attitude matrix can be parameterized in terms of ability of the alignment algorithm should be comprehensively con-
60 126
quaternion to reduce the computational amount and storage space, sidered. Thus, the TRIAD method with two inertial apparent veloc-
61 127
62
Equation (6) can be re-expressed as ity vectors is selected to solve the initial matrix, as follows: 128
63 n
=Q n n0 b0 ⎡ ⎤−1 ⎡ ⎤ 129
n0 (t ) ⊗ ⊗Q
Q b
64 b (t ) Q b0 b
(t ) (7) ( v nt10 )T ( ṽ t10 )T 130
65 where ⊗ denotes the quaternion multiplication, and the quater-
n0
Ĉ b0 =⎣ ( v t10 × v nt20 )T
n ⎦ ⎢ ⎣
b0 b
( ṽ t1 × ṽ t20 )T
⎥
⎦ (11) 131
n0 n n
66 n b
nions Q nb (t ), Q nn0 (t ), Q b0 , and Q b0 (t ) correspond to the attitude ( v t1 × v t20 × v t10 )T b0 b0 b0 T
( ṽ t1 × ṽ t2 × ṽ t1 ) 132
0
 JID:AESCTE AID:105747 /FLA [m5G; v1.261; Prn:28/01/2020; 16:18] P.4 (1-8)
4 Q. Fu et al. / Aerospace Science and Technology ••• (••••) ••••••

n n
1 where the vectors v t10 and v t20 are the integral results of the left- During the backtracking navigation update process, a fine align- 67
2 hand side of Equation (10) at different moments t 1 and t 2 , while ment Kalman filter is used to estimate the misalignment angles 68
b b
3 the vectors ṽ t10 and ṽ t20 are the integral results of the right-hand and the sensor errors. The Kalman filter calculation includes the 69
4 side of Equation (10). This is shown in the equations below: state update and measurement update stages [4,29], and both can 70
5 be done per second. 71
t1 t2
6 Because the proposed fine alignment algorithm is arranged in 72
n n n n
7 v t10 = − C n0 (t ) g n dt , v t20 = − C n0 (t ) g n dt (12) the ordinary local level frame but not in the inertial frame, the 73
8 74
0 0 classical INS error model can be followed. Under non-motion con-
9 75
t1 t2 ditions, the attitude and velocity error models are simplified as
10 b b 76
f˜ (t )dt , f˜ (t )dt
b b b b follows:
11 ṽ t10 = C b0 (t ) ṽ t20 = C b0 (t ) (13) 77
12 0 0 n 78
13 φ̇ = −ωnie × φ n − C nb εb (17) 79
where t 1 and t 2 are selected as t 1 = t 2 /2; and t 2 is the current
14 n 80
15
coarse alignment time, which is equal to t c of Fig. 1(b) at the end δ v̇ = f n × φ n + C nb ∇b (18) 81
of coarse alignment.
16 82
During the real-time coarse alignment process, some necessary where φ n and δ v n denote the misalignment angles and the veloc-
17 83
information must be extracted and recorded for later use. To re- ity errors in the n-frame, while εb and ∇b denote the gyro biases
18 84
duce the storage space and the computational load, the storing fre- and the accelerometer biases, respectively.
19 85
quency is selected as 1 Hz, which determines the subsequent back- Because the sensitive center of the IMU is usually not consis-
20 86
tracking navigation update frequency and the Kalman filter update tent with the rotating center of the turntable, the lever arm is
21 87
frequency in the fine alignment stage. There are ten floating-point augmented to the state vector and treated as a random constant
22 88
data points to be recorded per second, including to be estimated. Thus, a 15-dimensional error state vector is de-
23 b 89
a) Q b0 , four-dimensional attitude quaternion at time tk up- signed as
24 k 90
dated by Equation (7);
25
b
b)  ṽ k 0 , three-dimensional velocity increments in the kth sec-
 T 91
26 X = (φ n )T (δ v n )T (εb )T (∇b )T ( L b )T (19) 92
27 ond as follows: 93
28 tk where L b is the lever arm vector projected in the b-frame. 94
b
 ṽ kb0 = C b0 (t ) f˜ (t )dt ; The corresponding state equation is represented as
29 b 95
(14)
30 96
⎡ ⎤
31 t k −1
−C nb εbw 97
FI 012×3
32
c) ω̃
b Ẋ = X + ⎣ C nb ∇bw ⎦ (20) 98
33 ib,k ,
three-dimensional instantaneous angular rate outputs 03×12 03×3 99
by gyros at time tk , which are ready for the lever arm estimation 09×1
34 100
in the subsequent fine alignment Kalman filter.
35
where ε and b
w ∇bw
are white noises measured by the gyros and 101
36 102
3.2. Fine alignment process accelerometers, and F I denotes a 12 × 12 state transition matrix:
37 103
n0
⎡ ⎤
38
Since the rough initial attitude matrix and the stored data Ĉ b0 −ω n
ie
× 03×3 −Cnb 03×3 104
39 ⎢ −g × n
03×3 03×3 Cnb ⎥
105
are available after the coarse alignment stage, the fine alignment ⎢
FI =⎣ ⎥ (21)
03×3 ⎦
40 106
can be updated by the second, including the backtracking naviga- 03×3 03×3 03×3
41 107
tion update and the Kalman filter update. To follow common usage, 03×3 03×3 03×3 03×3
42 108
simplify the algorithms, and take advantage of the classical error
43 n 109
model, the fine alignment is well designed in the n-frame. The measurement vector is the navigation velocity v updated
44 110
Like the normal inertial navigation update, the backtracking by Equation (16). Considering the lever arm L b between the IMU
45 111
navigation update also includes the attitude and velocity updat- center and the turntable center, the velocity at time tk is
46 112
47 ing equation. The backtracking attitude updating equation at time   113
n b n b
48 tk is v nk = C bk ω̃ eb,k ×L b
≈ C bk ω̃ ib,k
b
×L , (22) 114
k k
49 n n n0 b 115
50
C bk = C nk0 Ĉ b0 C b0 (15) b b b b
where ω̃eb = ω̃ib − ω̃ ie ≈ ω̃ ib . Thus, the measurement equation can 116
k k
51 n0 be formulated as 117
where Ĉ b0 denotes the initial attitude matrix obtained in the
52 118
n
coarse alignment, C nk0 is acquired by the solution of differential  
53 n b 119
b0 Zk = δ v nk = v nk = 03×3 I3×3 03×6 C bk (
ω̃ ib,k ×) X + wv
54 equation (9), and C b can be obtained by parameterizing the stored k 120
k
55 b (23) 121
quaternion Q b0 in terms of the direction cosine matrix.
56 k 122
Because the theoretical velocity is zero, the backtracking veloc- b
57 where ω̃ ib,k denotes the instantaneous angular rates at time tk 123
ity updating equation at time tk is simplified as
58 recoded in coarse alignment, and w v is the measurement noise 124
nk−1/2 n0 b
59 v nk = v nk−1 + C n0 Ĉ b0  ṽ k 0 + g n T s (16) vector representing the carrier disturbance and the rotation veloc- 125
60 ity interference. 126
61 where the initial velocity v n0 = 0, the update cycle T s = 1 s, and Therefore, all of the needed parameters in the state equation 127
b
62  ṽ k 0 represents the velocity increments recorded in the coarse and the measurement equation can be obtained from the stored 128
63 alignment. In order to compensate for the rotation of the navi- data or the backtracking navigation update process. The only dif- 129
64 gation frame during one integration period in high-accuracy appli- ference between the backtracking fine alignment and the normal 130
nk−1/2
65 cations, C n0 is used to represent the transformation matrix of forward alignment is the update cycle, so it can be predicted that 131
66 the n-frame from time 0 to time (tk−1 − T s /2). the resulting errors will be small enough to be ignored. 132
 JID:AESCTE AID:105747 /FLA [m5G; v1.261; Prn:28/01/2020; 16:18] P.5 (1-8)
Q. Fu et al. / Aerospace Science and Technology ••• (••••) •••••• 5

1 67
2 68
3 69
4 70
5 71
6 72
7 73
8 74
9 75
10 76
11 77
12 78
13 79
14 80
15 81
16 82
17 83
18 84
19 85
20 86
Fig. 2. Three different rotation schemes. (For interpretation of the colors in the fig- Fig. 3. The alignment errors (3 RMSE) of three rotation schemes.
21 ure(s), the reader is referred to the web version of this article.) 87
22 88
23 89
24 Table 1 90
Specifications of the inertial sensors.
25 91
26 Gyros Bias 0.002◦ /h 92
√
27 Random Walk 0.0002◦ / h 93
28
Markov process τ = 600 s, R (0) = 0.002◦ /h 94
Scale factor error 1 ppm
29 95
Accelerometers Bias 20 μg
30 √ 96
31 Random Walk 1 μg/ Hz 97
Markov process τ = 600s, R (0) = 1 μg
32 98
33 99
34 100
4. Simulation and experimental analysis
35 101
36 102
37 4.1. Simulation analysis 103
38 104
39 Since the azimuth alignment accuracy of a rotary INS is ex- 105
40 tremely high, we verified the proposed alignment method by sim- 106
41 ulation tests. First, to verify the advantages of the continuous ro- 107
42 tation scheme, we compared it with two other different rotation 108
Fig. 4. The misalignment errors (RMSE) estimated in the coarse alignment.
43 schemes. Three different rotation schemes are shown in Fig. 2. 109
44 In Fig. 2, the total time is 300 s in one alignment process, and 110
45 ψ1 , ψ2 and ψ3 represent three different rotation azimuths of the In Fig. 3, the two-position rotation (Scheme 1) has the worst ac- 111
46 turntable. Scheme 1 is a classic two-position rotation strategy with curacy. One reason is that the slow-change error is not effectively 112
47 an azimuthal rotation of 180◦ . Scheme 2 is a three-position rota- modulated, and the other is that this scheme needs longer time 113
48 tion strategy, in which the turntable stays at 0◦ and 180◦ for equal to identify the gyro biases. The alignment accuracy of Scheme 2 114
49 time. Scheme 3 is the proposed continuous rotation scheme. To de- is better, because it can partially modulate the slow drift, and the 115
50 crease the influence of the scale factor error of the azimuth gyro, continuous rotation (Scheme 3) has the undisputed highest accu- 116
51 the continuous rotation scheme of the turntable was suitably ar- racy. It can be seen that the continuous rotation scheme can well 117
52 ranged with positive–negative rotations. The rotation angular rate modulate the constant errors and slowly varying drifts, and obtain 118
53 was designed to 30◦ /s, so ten positive–negative rotating cycles higher alignment accuracy. 119
54 were included in Scheme 3. Next, based on the continuous rotation scheme, we focused 120
55 The lever arm between the IMU center and the turntable center on the simulation analysis of the proposed real-time information- 121
56 was set as [−0.04; 0.05; 0.06] m. All of the inertial sensors were reusing alignment (RA) method. With the same conditions, 30 122
57 sampled at 200 Hz, and the specifications are listed in Table 1. Monte-Carlo simulation tests were carried out, and the coarse 123
58 For three different rotation schemes, we use the same post- alignment results of the RA method are shown in Fig. 4. Without 124
59 processing alignment (PA) approach for simulations. In the PA estimation and compensation for the lever arm, the linear motion 125
60 method, all of the sensor data were recorded in each cycle for disturbances that were caused by the coupling of the lever arm 126
61 non-real-time processing. The initial attitude at the start time was and rotation rate would deteriorate the coarse alignment results. 127
62 obtained by the first run of the coarse alignment, while the second Therefore, the curve of the azimuth misalignment angle φU had an 128
63 run used the Kalman filter to accomplish the fine alignment. With obvious periodic fluctuation characteristic. The final attitude errors 129
64 the same sensor errors, 30 Monte-Carlo simulation tests were car- of the coarse alignment were about [0.001; 0.001; 0.1]◦ (RMSE). 130
65 ried out for each rotation scheme, and the alignment errors (3 root Although perfect alignment accuracy cannot be achieved at the 131
66 mean square error, 3 RMSE) are shown in Fig. 3. end of coarse alignment stage, after 100 s the misalignment an- 132
 JID:AESCTE AID:105747 /FLA [m5G; v1.261; Prn:28/01/2020; 16:18] P.6 (1-8)
6 Q. Fu et al. / Aerospace Science and Technology ••• (••••) ••••••

1 67
2 68
3 69
4 70
5 71
6 72
7 73
8 74
9 75
10 76
11 77
12 78
13 79
14 80
15 81
16 82
17 83
18 84
19 85
20 86
21 Fig. 6. The differences in attitude estimation results between the RA and PA meth- 87
Fig. 5. The attitude and velocity differences between the RA and PA methods for
22 pure inertial navigation. ods. 88
23 89
24 gles were accurate enough to meet the linear requirements of the 90
25 subsequent fine alignment. 91
26 In the fine alignment stage, the attitude estimation accuracy of 92
27 the proposed RA method depended on the results of the backtrack- 93
28 ing pure inertial navigation update. We could evaluate the back- 94
29 tracking pure inertial navigation results of the RA method by com- 95
30 paring it with the PA method, using the same initial attitude and 96
31 sensor data. With 30 Monte-Carlo simulation tests, Fig. 5 shows 97
32 the attitude and velocity differences of the pure inertial navigation 98
33 between the two approaches. It can be seen that the attitude dif- 99
34 ferences caused by the low update cycle in the RA method were at 100
35 the 10−13 ◦ level, and the velocity differences were approximately 101
36 10−7 m/s. These differences were mainly caused by numerical in- 102
37 tegration calculation errors, and they were small enough to be 103
38 negligible. This means that the backtracking navigation strategy 104
39 can also be applied to compress the inertial sensor data in some 105
40 situations where the storage space or communication speed is lim- 106
41 ited. 107
42 The differences in attitude estimation results between the RA 108
43 Fig. 7. The misalignment errors (3 RMSE) estimated by the four alignment method. 109
and PA methods are illustrated in Fig. 6, which shows that the final
44 level and azimuth estimation differences were all less than 10−3 110
45 arc-seconds. Clearly, the differences were very insignificant. There- method with the other three methods. The statistical results (3 111
46 fore, by adding a small amount of storage space, without changing RMSE) are shown in Fig. 7. 112
47 the error model, the proposed RA method can achieve the same For the optimization-based methods ADIA1 and ADIA2, the sen- 113
48 accuracy as that of post-processing. sor data recorded in each cycle was used to construct different 114
49 To further evaluate the effectiveness of the proposed RA integration vectors to suppress the rotation interference. However, 115
50 method, we introduced three other alignment approaches for com- the accelerometer’s biases and the lever arm cannot be estimated 116
51 parison. The first approach was the NA method. As shown in by these two methods, which will cause certain level misalignment 117
52 Fig. 1(a), a period of time [0, t c ] was used for the coarse alignment, errors. As shown in Fig. 7, the eastern and northern misalign- 118
53 and the residual time period [t c , t f ] was left for the real-time fine ment errors of the ADIA1 and ADIA2 methods are about 2 seconds, 119
54 alignment. In the NA approach, we assumed that at least 100 which are larger than that of the Kalman filtering-based methods. 120
55 s would be required to determine satisfactory attitude informa- For the azimuth alignment accuracy, the proposed RA method out- 121
56 tion under a specific swaying interference environment. Thus, in performs the NA method because more time was used for the fine 122
57 the simulation tests, 100 s was used for the coarse alignment Kalman filtering. At the end of the alignment, the azimuth align- 123
58 stage, and the remaining 200 s was used for the fine alignment ment accuracy of the proposed RA method is basically the same 124
59 stage. The two other approaches were the backtracking attitude as that of ADIA2 method, and higher than the ADIA1 and NA 125
60 determination-based initial alignment (ADIA1) method described methods. Moreover, Fig. 7 shows that the azimuth misalignment 126
61 in [15] and ADIA2 method described in [16]. The two methods estimated by the proposed RA method converges faster than other 127
62 were optimization-based alignment method [28], and the main methods. Therefore, the well-designed information-reusing align- 128
63 difference between the two methods was the special integration ment strategy could shorten the alignment time by making full 129
64 process for backtracking vector observation construction. use of the sensor data. 130
65 With the same sensor data, 30 sets of simulation tests were For the proposed RA method, the average gyro and accelerom- 131
66 performed to compare the alignment accuracy of the proposed RA eter biases estimated in the fine alignment are shown in Fig. 8. 132
 JID:AESCTE AID:105747 /FLA [m5G; v1.261; Prn:28/01/2020; 16:18] P.7 (1-8)
Q. Fu et al. / Aerospace Science and Technology ••• (••••) •••••• 7

1 67
2 68
3 69
4 70
5 71
6 72
7 73
8 74
9 75
10 76
11 77
12 78
13 79
14 80
15 81
16 82
17 83
18 84
19 Fig. 10. The homemade RLG INS for testing. 85
20 86
21 Fig. 8. The gyro and accelerometer biases estimated by the backtracking fine align- signal processor (DSP) produced by Texas Instruments Incorpo- 87
22 ment. rated [30], which had 256 KB of internal memory space and a 200 88
23 MHz operating frequency. Under these configurations, the needed 89
24 backtracking calculation time [t c , t f ] was less than 2 s. Therefore, 90
25 the storage space and the computational capability required by 91
26 the proposed RA strategy were easy to satisfy with this kind of 92
27 general-purpose processor. 93
28 Azimuth alignment accuracy is a key indicator of a rotary INS. 94
29 Since it was very difficult to obtain a more accurate continuous 95
30 attitude reference in an actual rotary INS, we mainly compared 96
31 the azimuth alignment accuracy of different methods in practical 97
32 experiments. We took the following steps to obtain a continuous 98
33 azimuth reference. The rotary INS was fixed on the platform, and 99
34 the initial pitch and roll angles are obtained by the PA method 100
35 with continuous rotation. The initial heading angle was obtained 101
36 from the light aiming results of the prism. Based on the known 102
37 initial pitch, roll and heading angles, a pure attitude update could 103
38 be performed according to the IMU data, and the heading angle 104
39 obtained per cycle was used as a continuous azimuth reference. 105
40 Considering the accuracy of the optical sight, the zero repeatability 106
41 of the indexing mechanism, and the gyroscope accuracy, the ac- 107
42 curacy of the continuous azimuth reference can be guaranteed for 108
43 about 6 arc-seconds within 300 s. 109
44 Fig. 9. The lever arm estimated by the backtracking fine alignment. 110
During the alignment tests, all IMU data was record for the NA,
45 ADIA1 and ADIA2 methods, and the alignment results were com- 111
46 Due to the rotation modulation effects, the gyro biases, which had pared with the proposed real-time RA method. We carried out 10 112
47 a slow influence on the accumulation of the calculated velocity alignment tests at a fixed position, and the statistical results (3 113
48 errors, could not be effectively estimated by the Kalman filter. RMSE) of azimuth alignment errors are shown in Fig. 11. 114
49 However, the accelerometer biases could still be estimated with The alignment results shown in Fig. 11 are basically consistent 115
50 certain accuracy due to their direct effect on the velocity errors. with the simulation results. The final azimuth alignment accuracies 116
51 The average lever arm estimated by the backtracking fine align- (3 RMSE) of the ADIA1, ADIA2, NA, and proposed RA methods are 117
52 ment is shown in Fig. 9. Referring to the set value, the level arm 82.70 arc-seconds, 48.78 arc-seconds, 65.94 arc-seconds, and 47.25 118
53 components L bx and L by had a clear estimation effect, while the ver- arc-seconds, respectively. Therefore, the proposed RA method could 119
54 tical component L bz did not. The main reason for this was that the be successfully applied to the practical system, and the satisfactory 120
55 single-axis turntable rotated around the z-axis of the IMU, and the azimuth alignment accuracy were achieved within 300 s. In addi- 121
56 vertical arm L bz was unobservable. tion, compared with the ADIA1 and ADIA2 methods, the memory 122
57 space and the computation load of the proposed RA method are 123
58 4.2. Experimental analysis significantly reduced. 124
59 125
60 The performance of the proposed RA approach was also eval- 5. Conclusion 126
61 uated through practical experiments. The experimental data and 127
62 the results were collected from a homemade ring laser gyroscope In the proposed information-reusing alignment strategy, all of 128
63 (RLG) rotary INS, as shown in Fig. 10. The sensor specifications of the sensor data were effectively used in both the coarse align- 129
64 the INS were similar to those shown in Table 1, the sampling pe- ment and fine alignment phases, so the alignment time required 130
65 riod was set as 0.005 s, and the total alignment time was still 300 for achieving a specific accuracy was reduced. In the coarse align- 131
66 s. The master computer of the INS was a TMS320C6713B digital ment stage, the inertial sensor data were extracted and recorded 132
 JID:AESCTE AID:105747 /FLA [m5G; v1.261; Prn:28/01/2020; 16:18] P.8 (1-8)
8 Q. Fu et al. / Aerospace Science and Technology ••• (••••) ••••••

1 [6] D. Chung, J.G. Lee, C.G. Park, H.W. Park, Strapdown INS error model for multi- 67
2 position alignment, IEEE Trans. Aerosp. Electron. Syst. 32 (4) (1996) 1362–1366. 68
3 [7] X.L. Wang, X.J. Guan, J.C. Fang, et al., A high accuracy multiplex two-position 69
alignment method based on SINS with the aid of star sensor, Aerosp. Sci. Tech-
4 70
nol. 42 (2015) 66–73.
5 [8] P. Lv, J.Z. Lai, J.Y. Liu, M.X. Nie, The compensation effects of gyros’ stochas- 71
6 tic errors in a rotational inertial navigation system, J. Navig. 67 (6) (2014) 72
7 1069–1088. 73
[9] F.J. Pei, L. Zhu, J. Zhao, Initial self-alignment for marine rotary SINS using novel
8 74
adaptive Kalman filter, Math. Probl. Eng. (2015) 320536.
9 75
[10] Y. Qin, J. Lai, Q. Wu, J. Liu, MEMS rotary strapdown INS with low-resolution
10 rotary encoder, Gyro. Navig. 7 (4) (2016) 311–317. 76
11 [11] P.G. Savage, Strapdown Analytics, 2nd ed., Strapdown Associates, Inc., 2000. 77
12 [12] S.Y. Cao, L. Guo, W.H. Chen, Anti-disturbance fault tolerant initial alignment 78
for inertial navigation system subjected to multiple disturbances, Aerosp. Sci.
13 79
Technol. 72 (2018) 95–103.
14 [13] P.M.G. Silson, Coarse alignment of a ship’s strapdown inertial attitude reference 80
15 system using velocity loci, IEEE Trans. Instrum. Meas. 60 (6) (2011) 1930–1941. 81
16 [14] J. Hu, X.H. Cheng, A new in-motion initial alignment for land-vehicle SINS/OD 82
17 integrated system, in: Proceedings of IEEE Position, Location and Navigation 83
Symposium, Monterey, CA, 2014, pp. 407–412.
18 84
[15] L.B. Chang, F.J. Qin, A. Li, A novel backtracking scheme for attitude
19 determination-based initial alignment, IEEE Trans. Autom. Sci. Eng. 12 (1) 85
20 Fig. 11. Azimuth alignment errors (3 RMSE) of the rotary alignment experiments. (2015) 384–390. 86
21 [16] L.B. Chang, B.Q. Hu, Y. Li, Backtracking integration for fast attitude 87
determination-based initial alignment, IEEE Trans. Instrum. Meas. 64 (3) (2015) 88
22 at 1 s cycles, which reduced the storage space and computational 795–803.
23 load and greatly promoted the engineering practicability. In the 89
[17] W. Gao, Y.Y. Ben, X. Zhang, Q. Li, F. Yu, Rapid fine strapdown INS alignment
24 fine alignment, the backtracking navigation and the corresponding method under marine mooring condition, IEEE Trans. Aerosp. Electron. Syst. 90
25 Kalman filter were arranged in the local level geographic frame 47 (4) (2011) 2887–2896. 91
26 [18] W.L. Li, W.Q. Wu, J.L. Wang, M.P. Wu, A novel backtracking navigation scheme 92
instead of the inertial frame. Thus, the classical error model of
for autonomous underwater vehicles, Measurement 47 (2014) 496–504.
27 the INS could continue to be used without modification. The pro- 93
[19] W.L. Li, W.Q. Wu, J.L. Wang, L.Q. Lu, A fast SINS initial alignment scheme for
28 posed method was justified in the simulation tests and practi- underwater vehicle applications, J. Navig. 66 (2) (2013) 181–198. 94
29 cal experiments, and an azimuth alignment accuracy better than [20] F.O. Silva, E.M. Hemerly, W.C.L. Filho, Error analysis of analytical coarse align- 95
30 50 arc-seconds was accomplished in the physical tests. Based on ment formulations for stationary SINS, IEEE Trans. Aerosp. Electron. Syst. 52 (4) 96
(2016) 1777–1796. 97
31 the well-designed backtracking navigation scheme, the proposed [21] R.M. Rogers, Applied Mathematics in Integrated Navigation Systems, 2nd ed.,
32 information-reusing alignment strategy can also be used as a good 98
American Institute of Aeronautics and Astronautics, Inc., Reston, Virginia, 2003.
33 solution for a swaying alignment or in-motion alignment after a [22] L.D. Zhang, J.X. Lian, M.P. Wu, X.P. Hu, An improved computation scheme of 99
34 slight change. strapdown inertial navigation system using rotation technique, J. Cent. South 100
35 Univ. Technol. 19 (5) (2012) 1258–1266. 101
[23] J.S. Lee, S.W. Jang, J.G. Choi, T.G. Lee, North-finding system using multi-position
36 Declaration of competing interest method with a two-axis rotary table for a mortar, IEEE Sens. J. 16 (16) (2016)
102
37 6161–6166. 103
38 The authors declared that they have no conflicts of interest to [24] Y.Y. Qin, G.M. Yan, D.Q. Gu, J.B. Zheng, A clever way of SINS coarse alignment 104
39 this work. despite rocking ship, J. Northwest. Polytech. Univ. 23 (2005) 681–684. 105
40 [25] D.Q. Gu, N. El-Sheimy, T. Hassan, Z. Syed, Coarse alignment for marine SINS us- 106
ing gravity in the inertial frame as a reference, in: Proceedings of IEEE Position,
41 References Location and Navigation Symposium, Monterey, CA, 2008, pp. 961–965.
107
42 [26] J.R. Wertz, Spacecraft Attitude Determination and Control, Springer, Nether- 108
43 [1] G.G. Hu, W. Wang, Y.M. Zhong, et al., A new direct filtering approach to lands, 1978. 109
INS/GNSS integration, Aerosp. Sci. Technol. 77 (2018) 755–764. [27] J.E. Keat, Analysis of Least-Squares Atitude Determination Routine DOAOP,
44 110
[2] M. Liu, G.C. Li, Y.B. Gao, et al., Improved polar inertial navigation algorithm Computer Sciences Corp., 1977.
45 based on pseudo INS mechanization, Aerosp. Sci. Technol. 77 (2018) 105–116. 111
[28] M.P. Wu, Y.X. Wu, X.P.D.W. Hu, Optimization-based alignment for inertial navi-
46 [3] Q.W. Fu, Y. Liu, Z.B. Liu, S.H. Li, B.F. Guan, Autonomous in-motion alignment for 112
gation systems: theory and algorithm, Aerosp. Sci. Technol. 15 (2011) 1–17.
47 land vehicle strapdown inertial navigation system without the aid of external [29] Z.S. Zhu, B. Zhao, Y.Y. Guo, et al., Research on gravity vertical deflection on 113
48 sensors, J. Navig. 71 (6) (2018) 1312–1328. attitude of position and orientation system and compensation method, Aerosp. 114
[4] Y.Y. Qin, Inertial Navigation, 2nd ed., China Science Press, 2014. Sci. Technol. 85 (2019) 495–504.
49 115
[5] Q.W. Fu, Y.Y. Qin, S.H. Li, H.M. Wang, Optimal two-position alignment method [30] TMS320C6713/TMS320C6713B Floating-point Digital Signal Processors, Texas
50 based on parameter identification, J. Chin. Inert. Technol. 21 (4) (2013) 116
Instruments Incorporated, 2004.
51 430–434. 117
52 118
53 119
54 120
55 121
56 122
57 123
58 124
59 125
60 126
61 127
62 128
63 129
64 130
65 131
66 132