Accurate vector velocity measurement requires the joint efforts of various signal processing modules, as illustrated in
Figure 4. After the MU radar data is received by 25 channels, the calibration of each channel is necessary to offset any possible phase and amplitude errors. The MUSIC algorithm is employed to estimate the DOA for angle estimation, fully utilizing the advantage of an array’s large aperture to achieve higher angular resolution, which is crucial for determining the direction of the final vector velocity. In addition, considering the high Doppler shift characteristics of meteor targets, a delay-Doppler matched filter is constructed to obtain the radial velocity and line-of-sight distance of the meteor. The phase of the matched filter output is then extracted, and multiple pairs of channels are interfered to obtain the phase difference change curve. Due to the presence of phase wrapping, phase unwrapping is required prior to obtaining accurate phase differences. Afterwards, the lateral velocity was calculated by using the proposed lateral velocity calculation method in this study. Finally, the results obtained from each signal processing module are jointly used to reconstruct the meteor’s trajectory and infer its vector velocity.
4.1. DOA Determination and Channel Calibration
Meteoroids and hard targets produce coherent radar echoes, which can be used to determine the location of the target by finding the DOA of the incident radar echoes. In this study, the multiple signal classification (MUSIC) algorithm developed by Schmidt [
24] was used to estimate the DOA from the head echo of the MU radar. Considering the MU radar planar array shown in
Figure 1, the received signal model is represented as follows:
where
X(t) is the
M × 1 snapshot data vector of the array,
ξ is a complex Gaussian white additive noise data vector of size
N × 1, and S(t) is the complex envelope
N × 1 data vector of the signal.
is an
M ×
N steering vector matrix, where (
θ,
φ) is the DOA of the ith source, corresponding to its azimuth and elevation angles.
is the standard steering vector, which is given by the following equation:
Here,
r = [
rx,
ry,
rz]
T (a 3 ×
M matrix) represents the center positions of the antenna subarray with respect to the geometric center of the entire array, given in terms of the radar wavelength.
k = 2π[cos
φsin
θ, cos
φcos
θ, sin
φ] is the wavenumber vector (a
K × 1 complex vector), and ⊙ denotes the Hadamard product;
g is the gain patterns for the transmitting and receiving antennas [
19]. In the MU radar, the entire system uses the same type of Yagi antenna, which allows for a common function to be used for all the antennas [
18]. This means that the dependency of the subarray gain on the input DOA is the same, and thus can be omitted (normalized to 1). In addition, for the horizontally oriented MU radar antenna, the value of
rz is zero, which allows Equation (14) to be simplified as follows:
The spectral estimation formula for the MUSIC algorithm is given by
where
H is the Hermitian transpose and
UN is the noise subspace obtained by performing eigendecomposition [
U,
Σ] = eig(
R) on the covariance matrix of the received signals, and then selecting the eigenvectors corresponding to the smaller eigenvalues. The number of larger eigenvalues can be obtained through methods for estimating the number of signal sources. Therefore, the remaining eigenvalues correspond to the smaller eigenvalues. The received data covariance matrix can be expressed as
. When evaluating Equation (16) for different DOAs, the denominator approaches zero near the DOA of the signal, and this results in a narrow peak in the spectrum
P(
θ,
φ). The DOA estimation can then be achieved through angle searching.
In practice, the parameters of the electronic components in antenna systems may vary due to factors such as temperature, time drift, or aging effects, resulting in errors in the phase and amplitude of the array elements. Therefore, it is necessary to perform regular array calibration during the operation of these antennas to maintain their original performance. In the presence of amplitude and phase errors, the array steering vector requires correction as follows:
where
Γ is the phase error matrix as
with
and
as the gain and phase errors of the
m-th antenna channel, respectively, diag(·) returns a diagonal matrix whose diagonal equals the input vector. As the first channel serves as the reference channel, we have
and
.
Using an inaccurate steering vector
for spectrum peak searching may result in peak shifting or indistinguishable peaks. To compensate for array phase errors, the calibration of each receiving channel of the MU radar is required. We should estimate the amplitude and phase errors
Γ by a certain method, and then substitute
for
in Equation (16). Amplitude error estimation can be achieved by computing the signal power of each channel and then normalizing it using a reference channel. The key lies in the estimation of the phase errors. In this study, a statistical method utilizing a large number of strong meteor head echoes was chosen for phase calibration. This method selects multiple strong and well-defined meteor echoes, uses the DOA estimated for each event to generate the best phase estimate through one iteration, and then takes the statistical average of the estimation results for all the strong meteor head echoes [
19,
25,
26]. The 86 strongest meteor head echoes detected by the MU radar from June 28, 2018, 8:00 to June 29, 2018, 8:00 were selected for calibration using this method. The estimation results of the phase errors are shown in
Figure 5, revealing a significant error in channel 7.
The effectiveness of the method is demonstrated through the DOA estimation results for each head echo event using the calibration values estimated in
Figure 5 and substituted into Equations (17) and (18). An example is shown in
Figure 6, where
Figure 6a is a range-time-intensity plot of the raw data with 512 IPPs and 85 sampling points. Due to truncation, the distance at the distance element of 1 is about 73 km and should be added in the final distance calculation. The meteor echo event occurred at 08:19:00 on 26 June 2018 (UT + 8/JST), and strong meteor head echoes spanning multiple IPPs were observed during the meteor event, along with multiple distance element migrations, equivalent to a distance of approximately 15 km being crossed in about 0.2 to 0.3 s.
Figure 6b is the MUSIC spectrum of the DOA estimate after calibration, viewed from the top angle, showing that the meteor is roughly located directly above, with distinct peaks in the spectrum. To clearly demonstrate the effect of calibration,
Figure 7 shows the DOA estimate results before and after calibration for two angular dimensions, indicating a significant improvement in the DOA estimation performance after calibration, as evidenced by the sharper and narrower MUSIC spectral peaks. Remarkably, the presence of five asymmetric array channels in the outermost layer of the MU array results in the absence of a grid lobe near the peak of the MUSIC spectrum. Instead, it appears in the form of sidelobe, as shown in
Figure 6b, which can be verified by theoretical calculations.
4.2. Delay-Doppler Matched Filter
Matched filtering is widely used in various fields of radar signal processing and can be summarized as a technique for filtering signals that match an implemented model. The technique can maximize the signal-to-noise ratio (SNR) of the output signal by computing the cross-correlation function between the received and transmitted signals of the radar.
The meteor head echo comes from the dense region of the plasma that is close to and propagates together with the meteor, resulting in highly transient and Doppler-shifted signals. Because of the large Doppler shift, simple correlation with the transmitted pulse waveform is not sufficient, and correct matching requires Doppler compensation by the radial velocity of the meteor before cross-correlating with the phase-coded pulse transmitted. Next, the matched filtering process of the MU radar received signal is described.
In meteor head mode, the MU radar transmits a pulse waveform encoded by a 13-bit Barker code, which is oversampled by a factor of two. The Barker code with a Baud length of 12 μs is transmitted using a sampling period of
Ts = 6 μs. Therefore, the Barker code modulation part can be expressed in the form of
where
n represents the nth code element used. Therefore, the transmitted signal can be expressed as
where
f0 is the carrier frequency of the transmit signal,
N is the number of sampling points calculated under the pulse length, and
represents the phase modulation part of the transmit signal encoded by the Barker code. The horizontal line in
denotes that the signal contains a carrier and has not yet been demodulated.
Considering a target moving towards the radar with a radial velocity of
vr, the received waveform can be expressed as follows:
where
R0 denotes the initial distance between the target and the radar, and
Tr represents the time between the transmission of the radar pulse and its collision with the target. By removing the carrier frequency from the received signal and normalizing it, we can represent it in the following form:
Equation (24) indicates that the received signal incurs an additional Doppler frequency shift compared to the transmitted signal, in addition to introducing a time delay. The Doppler frequency shift,
fd, is determined by Equation (1). Considering that meteoroids themselves have high velocities, traditional matched filtering can result in significant mismatch in the output. Therefore, compensating for the Doppler shift during matched filtering is necessary. Introducing the compensating frequency term, exp(
−j2
πfnt), the output of delay-Doppler matched filtering (cross-correlation) can be expressed as follows:
where the symbol * denotes the complex conjugate operation, and the following equation is satisfied:
Equation (25) represents the output of the matched filtering, which is a bivariate function that peaks at , corresponding to the actual target delay relative to the radar and the Doppler frequency shift of the received echo. However, the actual radial velocity of the meteoroid is unknown and needs to be determined. The final distance and radial velocity can be obtained by searching the frequency and selecting the delay and Doppler frequency corresponding to the highest peak of the two-dimensional function. Ideally, after matching, the energy dispersed over the Barker code symbols becomes highly concentrated at the first symbol of the Barker code waveform, thus improving the signal-to-noise ratio of the received signal.
Figure 8 presents an example of the matched filtering results obtained from the head echo data using the selected meteor event from
Figure 6a. The matched filtering can provide direct distance and velocity information of the target from each IPP, corresponding to the peak coordinates in the spectrum. As shown in
Figure 8a, the maximum output of the matched filter for the 235th IPP is located at (47.95, 28). By performing matched filtering on all the IPPs, the matched filtering RTI plot is obtained, as illustrated in
Figure 8b. Next, the variation of distance and velocity across consecutive IPPs is obtained, as illustrated in
Figure 9. As shown in the figure, the meteor approaches the Earth at a very high speed, and its radial velocity gradually reduces with time, decreasing from 50 km/s initially to 42 km/s. Additionally, employing the pulse-to-pulse phase matching technique can effectively improve the velocity accuracy after obtaining the Doppler shift of a single IPP echo [
25,
27]. The technique measures the Doppler shift by utilizing the phase difference between the current IPP and the previous IPP, and the phase unwrapping process for adjacent IPPs relies on the velocity results obtained from the matched filter. The additional lines in
Figure 9b show the results obtained through this technique, which are more continuous and stable compared to the results obtained directly.
4.3. Determination of Lateral Velocity
The determination of the vector velocity is achieved by computing the radial and lateral velocities separately. The radial velocity is calculated using Equation (1) based on the Doppler shift, while the lateral velocity is estimated by utilizing the phase difference of the channel, as shown in Equation (7). The MU radar detects the 3D velocity of the meteor head; therefore, multiple channel pairs need to be selected when calculating the lateral velocity, and subsequently synthesized using Equations (8)–(12).
The selected channel pairs follow the principle of symmetry and minimal distance between each other in order to reduce errors and uncertainty. Based on this principle, we selected two groups of channel pairs, which form the basis of comparison for the final calculation results. The first group comprises channels 3, 7, and 11, and channels 15, 19, and 23, wherein each pair consists of a front and a rear channel; and there is a total of 3 pairs, denoted as Ch3–15, Ch7–19, and Ch11–23. The second group includes channels 2, 6, and 10, and channels 14, 18, and 22, denoted as Ch2–14, Ch6–18, and Ch10–22.
Figure 10 shows the selection of these two groups of channel pairs, with each pair connected by a line. Regardless of the choice made, each channel pair can be used to compute a lateral velocity component. The total lateral velocity can be determined by utilizing the lateral velocity components obtained from the remaining channel pairs in the same group.
Next, the relative phase difference between the antennas needs to be calculated. The matching filter output is directly used for this purpose, as it provides a certain signal gain. Using Equation (5), the phase difference can be calculated from the differential signal. Ignoring the change of speed in a single IPP time, the relationship between the phase difference and lateral velocity is given by
where Δ
ϕ is the phase difference between adjacent IPPs and Ts is the IPP interval. However, the phase difference needs to be unwrapped to obtain clear results. By substituting the maximum possible lateral velocity into Equation (27), we obtain a Δ
ϕ value much smaller than 2π, indicating that the phase difference between adjacent PPIs will be less than 2π. Utilizing this constraint, we can successfully unwrap the phase difference.
Figure 11 and
Figure 12 show the changes in the continuous PPI phase differences, with the solid lines representing the uncalibrated phase differences and the dashed lines displaying the unwrapped phase difference version. The phase unwrapping is performed by using the sequential point scanning method (SPSM), which sequentially scans the signal from high SNR sampling points toward each end of the signal, and then segments each section and conducts linear fitting within the segment. The resulting phase curve is then adjusted by adding or subtracting a multiple of 2π from each Δ
ϕ to obtain a smooth phase curve. This is accurate for almost all head echoes from MU due to the slow variation of Δ
ϕ.
The lateral velocity is determined by calculating the phase difference rate
kp using a simple window fitting method to extract the slope, as described in Equation (7). By scanning the phase difference with a sliding time window, as shown in
Figure 11 and
Figure 12, the lateral velocities for each channel pair are then calculated using Equation (7) to obtain the values of
(
i = 3, 7, 11;
j = 15, 19, 23). Taking the first group as an example,
is calculated using the phase difference rate of channels 3 and 15, while
and
are calculated similarly. Finally, the lateral velocity components are combined and substituted into Equations (8)–(12) to calculate the total lateral velocity. The same approach is applied to the second group.
Figure 13 shows the computed lateral velocity results from the two groups of channel pairs.
Figure 13a shows the total lateral velocity magnitude, while
Figure 13b–d display the magnitude of the individual lateral velocity components in a Cartesian coordinate system to indicate the direction of the lateral velocity. Choosing multiple groups of comparison results is crucial because it allows us to verify whether this method is self-consistent. The lateral velocity values obtained from the two groups of channel pairs are very close to each other, especially in the middle section where the SNR is the highest, and a consistent trend in the lateral velocity variation can be seen. Combining the signs of
and
, and considering the sign reversal of
halfway through, we can infer that the meteoroid passed through the zenith direction and that its trajectory was oriented towards the southwest.
Table 2 presents a comparison between Group 1 and Group 2 regarding the total lateral velocity and its three individual components along the x, y, and z axes. It can be observed that the root mean square error (RMSE) of the velocity generally does not exceed 3 km/s, while the mean absolute error (MAE) is around 1 km. Relative to the velocity of the meteor itself, these differences in velocity can be considered small. Therefore, we can calculate the vector velocity of the meteoroid using either group of the lateral velocity results. It is necessary to note that in the practical use of this method, the selection of antenna pairs can be flexible, and researchers should aim to choose as many channel pairs as possible while ensuring that the channel pair signals are enough strong. This ensures the accuracy and stability of the calculations.
4.4. Result of Trajectory and Vector Velocity
Figure 14 shows the plotted trajectory and height variations of the meteoroid, respectively, while
Figure 15 shows the final vector velocity measurement results. For the head echo event shown in
Figure 14, our method estimated the instantaneous velocity, which was found to be consistent with the average velocity computed from the position measurements.
Table 3 displays the range of calculated trajectory azimuth, elevation, height, and velocity variations. It can be observed that the meteoroid traversed the zenith and moved in the southwest direction at an altitude between 90 km and 105 km, which is consistent with the primary distribution range of meteors. Moreover, the velocity was negative in both the x and y directions, with a magnitude nearly twice the relationship, which corresponds to the motion of the meteoroid trajectory inferred from
Figure 14a.
Figure 15a indicates that the velocity of the meteoroid is so great that there is almost no decrease in velocity within the range of IPPs, which is noteworthy. Assuming that the meteoroid did not originate from outside the solar system or encounter third-body perturbation, the total velocity cannot exceed 72.8 km/s [
25]. As this event is already very close to the upper limit, we expect further analysis of more meteor samples. If the velocity of the meteoroid is concentrated at a very high level, the radial component of the velocity vector may not be apparent, and important velocity information may be neglected in conventional radar detection. These are often crucial to research in meteor dynamics. In this article, the method proposed by us enables us to calculate the deceleration in all directions, not just in the radial direction. This provides a more detailed observing capability than previously available, and the motion characteristics of the meteoroid can be obtained using the method proposed in this study.