Channel estimation scheme for high-mobility communication
The deep-learning assisted channel estimation scheme addresses high computational complexity and inaccurate channel estimation in OTFS systems by predicting significant elements and using a novel pilot structure, achieving reduced complexity and improved bit error rates.
Patent Information
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- HANGZHOU INNOVATION RES INST OF BEIJING UNIV OF AERONAUTICS & ASTRONAUTICS
- Filing Date
- 2024-11-12
- Publication Date
- 2026-05-21
AI Technical Summary
Existing channel estimation schemes for OTFS systems in high-mobility environments face challenges with high computational complexity and inaccurate channel information acquisition due to power leakage and non-static channel characteristics, especially in scenarios with high resolution and non-sparse channels.
A deep-learning assisted channel estimation scheme that includes two stages: channel prediction and significant element identification, using a novel pilot structure to cancel interference and a threshold-based selection method, combined with a recurrent neural network and auto-encoder to reduce complexity and improve accuracy.
The proposed scheme achieves reduced computational complexity and improved bit error rate performance by accurately estimating significant channel elements, outperforming prior art with acceptable performance loss.
Smart Images

Figure CN2024131472_21052026_PF_FP_ABST
Abstract
Description
Channel Estimation Scheme for High-Mobility CommunicationTECHNICAL FIELD
[0001] The present invention relates to a channel estimation scheme for wireless communication system in high-mobility environment.BACKGROUND
[0002] The current cellular communication network is mainly developed based on Orthogonal Frequency Division Multiplexing (OFDM) , which has poor performance in high-mobility environment. The Doppler effect caused by the movement of the receiver / transmitter will destroy the orthogonality between the subcarriers. To solve this problem, the service provider needs to adopt a frequency band which is much lower than the ordinary scenarios to mitigate the Doppler effect. The band with lower center frequency can only support limited number of users and lower data rate. Recently, Orthogonal Time Frequency Space (OTFS) modulation is considered as a promising solution for wireless communication in high-mobility environment. OTFS modulates each information symbol onto one set of two-dimensional orthogonal basis functions that span the frequency and time of the transmission data block. The time-varying channel of traditional OFDM system is transformed into a delay-Doppler channel in a sparse manner. Due to this reason, OTFS can be implemented as a pre-and post-processing block of the OFDM system. After combined with the Multiple-input Multiple-output (MIMO) structure, OTFS system will be extended to space domain. The fully excavate the potential of OTFS, the channel information should be accurately acquired. Most of the studied OTFS system assumes the system is operated with a narrow bandwidth and the channel is quasi-static. However, the future cellular network is supposed to be developed with higher frequency, larger antenna array size and longer transmission time duration. Hence, the resolutions in delay, angle and Doppler, will be extremely high. The channel cannot be treated to be static anymore. Meanwhile, due to the power leakage phenomenon of the OTFS system, the channel is far from sparse. The existing channel estimation scheme is insufficient to accurately acquire channel information due to path change and high computational complexity. A Deep-Learning (DL) based channel estimation method is proposed to estimate the significant elements in the channel while ignoring the insignificant elements.SUMMARY
[0003] The present invention provides a DL assisted channel estimation scheme for OTFS communication system with high resolution.
[0004] The proposed scheme can be roughly divided into 2 stages. In the first stage, the channel is roughly predicted. Then the indices of the significant elements are acquired. In the second stage, the values of the elements are accurately estimated.
[0005] A novel pilot structure is proposed to cancel the interference from the insignificant elements in the value estimation of the significant elements.
[0006] The prediction accuracy of the path existence and indices is critical for the system performance. A threshold coefficient is designed in the significant element selection, which need to be well allocated for an optimal result.
[0007] When Message Passing (MP) method is employed at the receiver side, the proposed scheme in this invention is able to provide acceptable bit error rate (BER) and outperforms the prior art. Compared with the ideal channel situation, the performance loss is acceptable.BRIEF DESCRIPTION OF THE DRAWINGS
[0008] In the following detailed portion of the present description, the framework of the present application will be explained. The following figures are used for the introduction of the invented method:
[0009] Fig. 1 illustrates the system model of the application scenario.
[0010] Fig. 2 shows the phases in the proposed estimation framework.
[0011] Fig. 3 illustrates the developed novel pilot structure.
[0012] Fig. 4 presents the developed RNN network.
[0013] Fig. 5 presents the structure of the developed Auto-Encoder (AE) network.DETAILED DESCRIPTION
[0014] The details of the invention are introduced in this section. The communication is first briefly analyzed in theory. Then the developed channel estimation scheme is illustrated. The mentioned components are marked in appended figures. This description is based on a MIMO-OTFS system with a multi-path propagation environment as shown in Fig. 1. The base station is supposed to support multiple high-mobility users simultaneously. And the users’ trajectories are quasi-static, e.g., high speed trains, automobiles, UAVs, etc. The system deploys millimeter wave band with wide bandwidth. Hence the delay resolution is high.
[0015] The base station equips a Uniform Linear Array (ULA) which consists of M elements. When M is large, the angle resolution of the system can be high. It is worth noting that the proposed scheme can also be used in other type of system, e.g., multiple-input single-output (MISO) OTFS system, single-input single-output (SISO) OTFS system, etc.
[0016] The transmitted data block is denoted as a 3-D matrix in delay-Doppler-space domain: XDDS∈CL×D×M, where C denotes the complex number filed; L denotes the length in the delay domain; D denotes the length in the Doppler domain; M denotes the length in the space domain. The delay resolution is denoted as where B =LΔf is the total bandwidth; Δf is the bandwidth of the subcarriers. The Doppler resolution is denoted as where Ts is the symbol duration. The symbol allocated at the resource element is denoted as: XDDS[l, d, m], where l, d and m are the indices in delay domain, Doppler domain and space domain, respectively. For m-th antenna, the corresponding data block XDDS[:, :, m]is transformed into a matrix in frequency-time (FT) domain. For the brevity of the notations, the data block is denoted as in the following discussion. The transformation is denoted as:
[0017] where FL and FD are the Fourier transform matrix. To transmit the data block by the OFDM system, the data block needs to be transformed into a time series signal as:
[0018] where {·} H denotes the conjugate transpose. The column vectors of S, s.t, S= [s1, …, sD]are the OFDM symbols. The time series signal is denoted as:
[0019] where {·} T denotes the transpose operation. The overall transmission from XDD to s can be denoted as:
[0020] where is the Kronecker product; is the vector form of After stack all the signal of different antennas, the overall transmitted signal of the base station can be denoted as: S∈CM×LD. To avoid inter-user interference, the signal needs to be precoded before transmission. The precoding process in denoted as: Z=PS,
[0021] where P is a hybrid precoder which consists of a digital precoder and an analog precoder. Then the signal is transmitted by a time-varying channel:
[0022] where q is the delay index; is the steering vector; x is the distance between the antenna elements; λ is the half-wavelength of the center carrier; θ is the angle of departure (AoD) ; P is the total number of the paths. It is worth noting that every path has a unique delay τp, Doppler νp and angle θp. The delay, Doppler and angel of p-th path can be arranged to the following indices: lp, dp and mp, respectively. The received signal of k-th user is denoted as:
[0023] where is the [q-lp]LD-th column of Z; [·] LD is the modulation operation. To avoid the inter-symbol interference, the cyclic prefix (CP) needs to be added. Since the process of adding CP and removing CP do not affect the data transmission, the process is eliminated in this description. The received time series signal is denoted as:
[0024] where Πk, p∈CLD×LD is the permutation matrix; Δk, p∈CLD×LDis a diagonal matrix who’s non-zero elements are corresponding phase shifts caused by delay and Doppler; z is the vector form of Z. Finally, the time series signal is transformed back to delay-Doppler domain:
[0025] After combine all the above equations, the equivalent data transmission in delay-Doppler domain can be derived as:
[0026] Where The equivalent channel in delay-Doppler domain is denoted as:
[0027] For the ideal OTFS channel, the delay, Doppler and angle equal to the integer multiples of the resolutions. However, in the practical situation, this assumption can be hardly realized. Large number of non-zero elements might lead to high computational complexity. Luckly, the OTFS system is able to reach high symbol detection accuracy with only significant elements.
[0028] Channel prediction is an efficient way to reduce the complexity of the channel estimation by providing some prior knowledge. The proposed scheme in this invention is shown in Fig. 2. Multiple compressed MIMO-OTFS historical channel 201 forms the input of the Recurrent Neural Network (RNN) 210. And the output of RNN 210 is the predicted channel in a compressed form 202. The prediction results 202 will be extended into a 3-D matrix 203 by a decoder of Auto Encoder (AE) network 221. The size of extended matrix 203 is lmax× (2dmax+1) ×M, where lmax and dmax are the maximum delay index and maximum absolute Doppler index, respectively. To identify the significant elements, a constant false alarm rate (CFAR) -like algorithm 230 is developed. The elements whose magnitude are much larger than the mean value of all the elements will be determined as the significant elements. A threshold parameter∈is created for element filtration. According to the locations of the significant elements, a novel pilot sequence is generated. Compared with the traditional pilot structure of OTFS system, although the proposed pilot consumes more spectrum resource, the absolute value is still low. The detail of the pilot structure is shown in Fig. 3. The resource block is divided into 3 areas, pilot region 3100, data region 3200 and guard region 3300. The pilot region in m-th angle is divided into A×B blocks 3110. The block size is lmax× (2dmax+1) . In every small block, only the corresponding locations of the significant elements are filled with non-zero symbols 3111 i.e., U [l, d, m]≠0, d=a (2dmax+1) -dp, where b= {1, …, B} , a= {1, …, A} . After receiving the pilot signal, the user will perform Least Square method 250 to acquire the values of the significant elements. Different form the conventional methods, the observation region are not arbitrary selected. Only one unique symbol 3112 can be selected as observation region in every block. Hence, elements of the observation are denoted as: The proposed pilot structure guarantees that only the significant elements can contribute to the observation region. The interference from the insignificant elements is cancelled by the zero symbols 3113. The estimation is denoted as:
[0029] where is the estimation result. The elements in this vector are the values of significant elements; is the matrix formed by the corresponding non-zero pilot symbols. Then the 3-D channel matrix will be reconstructed. Since the value estimation is performed at the user end, the estimation results need to be fed back to the base station. To reduce the overhead in this procedure, the encoder part of the AE network 222 is deployed. The encoder network 222 will compress the estimation results into the 1-D feature, which is much smaller than the original size. The feedback will be part of the historical channel 201 of the estimation procedure in next time slot. The details of the AE and RNN 210 are introduce in the following part.
[0030] Since this scheme is supposed to be deployed in the high-mobility scenarios with quasi-static trajectories, it is possible to learn the geometry information of the signal propagation by a data-driven method. Inspired by the powerful information mining capability of DL, the RNN network 210 is developed as shown in Fig. 4. The input of RNN is LI historical channel data 201 and every historical data is processed by a dedicated RNN unit 211. c (t) is the hidden state 212 at time t, which is a function shows the relation between the adjacent time slots. The output of the RNN units will be processed by a fully connect layer 213 to generate the prediction result 202.
[0031] However, directly predicting the 3-D channel will lead to extremely high computational complexity. To solve this problem, an AE network is used to compress the large channel matrix into a smaller feature vector and extend the prediction feature vector back to 3-D matrix. Because the coefficients in the channel matrix 204 are complex numbers, the real part imaginary part will be separated into two matrices 2041 and 2042. In the encoder, these two matrices will be processed by multiple 3-D convolution layers 2221. The output of the last convolution layer will be processed by a dense layer 2222. This compressed channel 205 (201) will be the input of the RNN 210. And the output of RNN 202 will be extended by the decoder 221 which can be treated as the inverse of the encoder 222. The compressed channel feature is first extended by a dense layer 2211 then processed by convolutional layers 2212. The number of the kernels used in each layer need to be adaptive to the channel scenario. The output of the last convolutional layer can be combined to form the final prediction result 203.
Claims
1.A multi-stage channel estimation scheme for communication system with Orthogonal Time Frequency Space modulation in high-mobility environment.2.A scheme according to claim 1, wherein a Deep-Learning network first provides a rough prediction, then the values of the desired coefficients are estimated by using a novel pilot signal.3.A scheme according to claim 1, only estimates small number of channel coefficients that have most of the power share.4.A scheme according to claim 2, wherein the Deep-Learning network has a cascade structure. The input is first compressed by the encoder of an Auto Encoder network, the output is fed into a Recurrent Neural Network. Finally, the output of Recurrent Neural Network is extended by decoder of the Auto Encoder network.5.A scheme according to claim 2, wherein the pilot signal has a sparse structure and the non-zero elements are allocated according to the locations of the significant elements.6.A scheme according to claim2, wherein the significant elements are selected by using constant false alarm rate method.7.A scheme according to claim 4, the Auto Encoder network is implemented based on 3-D Convolutional Neural Network.8.A scheme according to claim 2, the optimal threshold of the deployed constant false alarm rate can be determined by setting a bit error rate requirement.9.A scheme according to claim 5, wherein the guard symbols are added between the pilot symbols and data symbols to avoid interference.10.A scheme according to claim5, wherein the lengths of the guard symbols in delay and Doppler domain are same as the size of the channel in delay and Doppler domain, respectively.11.A scheme according to claim 7, wherein the Convolutional Neural Network has the proposed kernel number and size.12.A scheme according to claim 4, wherein the Recurrent Neural Network has the proposed structure.13.A scheme according to claim 9, the coefficient values are estimated by a Least Squares method with the proposed sparse pilot structure.