Frequency offset estimation method, system, device and medium based on clustering neural network

Abstract

The application belongs to the technical field of satellite communication, and discloses a frequency offset estimation method, system, device and medium based on a clustering neural network. Firstly, a random access preamble root index selection principle is proposed, so that the relationship between the power delay spectrum generated by the receiving end preamble sequence and the carrier frequency offset is unique and is not affected by timing advance compensation error and multipath effect. Then, power delay spectrum dataset is generated by all available preambles under different frequency offsets, and for each root index, the sparsity and regularity of the power delay spectrum matrix are fully utilized to construct an efficient and lightweight comprehensive frequency offset estimation model based on the semi-supervised K-means algorithm optimized based on the initial value and the back propagation neural network optimized based on the sparse dimension reduction, which can be respectively used for estimating the integer part and the decimal part of the frequency offset. Through the optimization of the neural network parameters in the training process by using the dataset, the final frequency offset estimation model is obtained, and complete frequency offset estimation under each root index can be realized.

Description

Technical Field

[0001] This invention belongs to the field of satellite communication technology, and in particular relates to a method, system, device and medium for frequency offset estimation of 5G non-terrestrial networks based on clustering neural networks. Background Technology

[0002] Currently, to expand the coverage and service of terrestrial networks, non-terrestrial networks (NTNs) are considered a promising solution due to their ability to provide ubiquitous and continuous network connectivity. The 3rd Generation Partnership Project (3GPP) has been actively promoting the development of new radios for 5G (5th Generation Mobile Communications Technology) to support NTNs and is continuously exploring new application areas, such as direct satellite connectivity. However, unlike terrestrial communication networks, the characteristics of NTNs impact air interface transmission. In particular, the Doppler frequency shift caused by the rapid movement of NTN platforms, such as low-Earth orbit satellite communication systems, leads to a significant increase in carrier frequency offset (CFO), which severely degrades the performance of 5G uplinks based on orthogonal frequency division multiplexing (OFDM) waveforms. Therefore, designing an effective frequency offset estimation scheme is crucial in highly dynamic NTN scenarios.

[0003] For frequency offset estimation, existing methods can compensate for the CFO in advance using the frequency offset estimate from downlink synchronization. However, due to the rapid changes in non-terrestrial network channels and the instability of transceiver oscillators, this may lead to untimely and inaccurate CFO adjustments. Some more reasonable approaches are to estimate the frequency offset directly using the uplink reference signal, such as a random access preamble. Commonly used methods include integer multiple frequency offset estimation based on conjugate symmetric ZC (Zadoff-Chu, ZC) sequences and preambles connected by random access preambles with two different root indices. However, these methods depend on specific channel models and preamble formats, and when fractional multiple frequency offsets exist, they require fractional multiple frequency offset compensation methods or fractional multiple frequency offset estimation schemes based on cyclic prefixes to mitigate the impact of fractional multiple frequency offsets.

[0004] Based on the above analysis, the problems and shortcomings of the existing technology are as follows:

[0005] (1) In high-mobility non-terrestrial communication scenarios, due to the rapid movement of satellites, the frequency offset estimated in downlink synchronization will be outdated and inaccurate if used in uplink synchronization.

[0006] (2) In terms of algorithms, the frequency offset estimation range of existing frequency offset estimation algorithms is generally limited. They can only estimate integer multiples or fractional multiples of frequency offset, and the remaining frequency offset will significantly reduce the performance of the algorithm.

[0007] (3) In terms of applicable scenarios, existing frequency offset estimation algorithms are greatly affected by multipath effects and are not robust to channel models.

[0008] (4) In terms of complexity, existing deep learning-based frequency offset estimation methods are not suitable for satellite scenarios with limited payload due to their high complexity.

[0009] Therefore, it is necessary to consider improving the existing methods to achieve complete frequency offset estimation in high-dynamic NTN scenarios, so as to improve the frequency offset estimation performance under various channel models. Summary of the Invention

[0010] To address the problems existing in the prior art, this invention provides a frequency offset estimation method, system, device, and medium based on clustering neural networks.

[0011] This invention is implemented as follows: a frequency offset estimation method based on a clustering neural network. The method selects root indices that are unaffected by timing advance (TA) compensation errors and multipath effects within a set frequency offset range of the dataset, according to the designed preamble principles. All available preamble sequences are generated from different root indices. Then, under the set frequency offset, a power delay profile (PDP) dataset is generated for all available root indices. Finally, for each dataset generated by a root index, a frequency offset estimation model based on a clustering neural network is constructed. The model is trained offline to achieve efficient frequency offset estimation in the online stage.

[0012] Furthermore, the frequency offset estimation method based on clustering neural networks includes the following steps:

[0013] Step 1: Analyze the impact of different CFOs on the PDP spectrum, and design the leader selection principle based on timing advance TA compensation error, maximum multipath delay spread and PDP spectrum peak position;

[0014] Step 2: Construct the PDP spectral dataset based on the leading design principles;

[0015] Step 3: For each root index, based on the impact of CFO on the PDP spectrum obtained in Step 1, a semi-supervised K-means algorithm is proposed to cluster the generated PDP spectrum data:

[0016] Step 4: Combining the impact of CFO on PDP spectrum obtained in Step 1, perform sparse dimensionality reduction on the clustered data to construct a new dataset;

[0017] Step 5: Train the backpropagation (BP) neural network using the new dataset to obtain the frequency offset estimation model;

[0018] Step 6: Estimate the frequency offset based on the frequency offset estimation model based on clustering neural network.

[0019] Furthermore, in the first step, a 5G random access preamble is generated from the ZC sequence. The ZC sequence is generated in the following manner: Where n is the number of sample points, N ZC Let u be the length of the ZC sequence, and u be the root index where u∈{1,...,N}. ZC -1}; Using different root indices, all available preambles are generated. The user randomly selects one for transmission on the physical random access channel. After passing through the satellite multipath fading channel, the preamble sequence received by the receiver is: Where L is the total number of paths, ω(n) is the noise, l is the l-th path to the destination, τ is the round-trip time, and h l and τ l ε represents the channel gain and relative path delay of the l-th path, and ε is the normalized frequency offset after normalizing the subcarrier spacing, including both integer and fractional parts. The PDP spectrum is generated by correlation operation between the received signal and the local root sequence with root index u as follows:

[0020]

[0021] in(·) * For complex conjugate operations, m is the time index. Assuming the transmitted sequence has undergone TA pre-compensation, the channel model only includes the line-of-sight channel with a channel gain of 1, and noise is ignored (the noise has no effect on the position of the PDP spectrum peak), the received signal is: The generated PDP spectrum is as follows:

[0022]

[0023] Expanding the above equation, we find that when ε is an integer, the value of p(m) can only be 0 or 1, and the position where 1 appears is... When ε is an integer multiple of the frequency offset, the peak position is unique and fixed; when ε is a decimal, At this point, there is more than one non-zero position, and the largest peak value is no longer 1. In other words, when the frequency offset has a fractional part, the energy of the largest peak value will be distributed to other positions, and these distribution positions will only be... When the frequency offset is an integer multiple, i.e. a fractional multiple, several pseudo-peaks will appear at fixed positions; therefore, we can see that the PDP spectral matrix has regularity and sparsity.

[0024] To establish the correspondence between CFO and PDP spectra, a root sequence selection principle is designed as follows:

[0025] The root index u is a prime number;

[0026]

[0027] Where Δτ is the TA compensation error, τ L This is for maximum multipath delay extension. z1 and z2 are two arbitrary integers. Using the root sequence selection principle, select the available root indices and generate all available leading sequences using different root indices.

[0028] Furthermore, in the second step, for all available leader sequences, within a preset frequency offset range, a dataset of PDP spectra generated at different frequency offsets is produced, denoted by D:

[0029] D={(p1,ε1),(p2,ε2),...,(p G ,ε G )}

[0030] Where G represents the size of the dataset, which contains data for multiple channel models, p g The frequency offset value is ε g The generated PDP spectrum data, where g represents the g-th one, and g∈{1,2,...,G};

[0031] The third step involves establishing a frequency offset estimation model based on a clustering neural network for each root index. Based on the analysis in the first step, the peak positions of the PDP spectrum follow a pattern; that is, the non-zero features of the PDP spectrum vector are regular. Specifically, when fractional frequency offsets exist, spurious peaks appear in the PDP spectrum, but their positions remain fixed. Therefore, within a preset frequency offset range for the sample set, we use PDP spectrum sample data generated by integer frequency offsets as supervisory information to guide a semi-supervised K-means algorithm, clustering the entire dataset into several classes. Specifically, in the clustering initialization phase, the number of integer frequency offsets and their corresponding samples are used as the K value and the initial cluster centers, and the labels of these samples remain unchanged during the clustering process. After the iteration termination condition is met, all samples will be clustered into K classes.

[0032] Furthermore, the fourth step, based on the analysis in the first step, involves the N of the PDP spectrum. ZCOf the features, only a very small number are affected by frequency offset, resulting in sparse sample features. By extracting only the frequency offset-affected features from the PDP spectrum, a more accurate low-dimensional feature vector p' is obtained.

[0033] The fifth step uses the k-th category with size G. k The new dataset is fed into the BP neural network for training, where k∈[1,K]. Specifically, the data of this category is divided into training and test sets in a ratio of 8:2. The training set is used to train the network, and the test set is used to evaluate whether the network has good frequency offset estimation performance. The network error is backpropagated using the BP algorithm, and the network parameters are further tuned to obtain the optimal parameters.

[0034] Furthermore, the sixth step specifically includes:

[0035] First, the target receives the preamble signal and the local root sequence, performing correlation operations to obtain the PDP spectrum. Based on its peak positions, the root index used by the user is determined. Then, the obtained PDP spectrum is used as an input vector and fed into a pre-trained frequency offset estimation model based on a clustering neural network under that root index. Specifically, the frequency offset category is estimated by calculating the minimum Euclidean distance between the maximum peak position of the target PDP spectrum and the maximum peak position of each cluster center. Where (x1,n1) and (x k ,n k ) represent the maximum peak position of the target PDP spectrum and the maximum peak position of the k-th cluster center, respectively. Finally, after dimensionality reduction of the PDP spectrum, a complete estimate of the frequency offset is obtained through network regression. Target frequency offset The precise value is estimated as follows: Where f(·) is the mapping function between the PDP spectrum and CFO, p t 'This is the data after dimensionality reduction of the target PDP spectrum.' It is an estimate of the target frequency offset.

[0036] Another object of the present invention is to provide a computer device including a memory and a processor, the memory storing a computer program, which, when executed by the processor, causes the processor to perform the frequency offset estimation method based on a clustering neural network.

[0037] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the frequency offset estimation method based on a clustering neural network.

[0038] Another objective of this invention is to provide an information data processing terminal for implementing the frequency offset estimation method based on clustering neural networks.

[0039] Another object of the present invention is to provide a frequency offset estimation system based on a clustering neural network for implementing the aforementioned frequency offset estimation method based on a clustering neural network, the frequency offset estimation system based on a clustering neural network comprising:

[0040] The principle design module proposes a leader design principle to select root indices that are not affected by TA compensation errors and multipath effects within the frequency offset range set in the dataset, and generates all available leaders.

[0041] The dataset generation module is used to generate PDP spectrum datasets for all available root indices under a set frequency offset.

[0042] The model building module is used to build a frequency offset estimation model based on a clustering neural network using the dataset of each root index. The model is trained offline and used to achieve efficient frequency offset estimation in the online stage.

[0043] Based on the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by this invention are as follows:

[0044] First, regarding the technical problems existing in the aforementioned prior art and the difficulty in solving these problems, we discuss some innovative technical effects that would result from solving these problems. A detailed description follows:

[0045] This invention utilizes a proposed frequency offset estimation method based on clustering neural networks to realize the mapping relationship between PDP spectrum and CFO. It can estimate not only integer multiples of frequency offset but also fractional multiples of frequency offset, i.e., it can estimate the complete frequency offset. Moreover, because it is not limited by the preamble format, it has stable estimation performance for different frequency offsets. The dataset contains data under various channel models, and the proposed method can have similar estimation performance under different channel models, demonstrating robustness to different channel models. Based on the regularity of the sample data, partial labeled sample data is used to guide the semi-supervised K-means clustering process, improving the clustering effect. Based on the sparsity of the sample data, the dimensionality of the samples is reduced, decreasing the impact of redundant features on sample features, reducing algorithm complexity, and improving the estimation accuracy of frequency offset.

[0046] Second, this invention first proposes a random access preamble root index selection principle, ensuring a unique relationship between the PDP spectrum and CFO generated by the receiver preamble sequence, unaffected by TA compensation errors and multipath effects. Next, it generates PDP spectrum datasets using all available preambles at different frequency offsets. For each root index, it fully utilizes the sparsity and regularity of the PDP spectrum matrix to construct an efficient and lightweight comprehensive frequency offset estimation model based on a semi-supervised K-means algorithm optimized by initial value and a BP neural network optimized by sparse dimensionality reduction. This model can be used to estimate the integer and fractional parts of the frequency offset, respectively. By optimizing the neural network parameters during training using the dataset, the final frequency offset estimation model is obtained, enabling complete frequency offset estimation for each root index.

[0047] Third, the expected benefits and commercial value of the technical solution of this invention after transformation are as follows:

[0048] During random access, based on the timing and carrier frequency information of the received signal, the receiver can use a suitable receiving algorithm to demodulate the received signal. During demodulation, the receiver recovers the original data or codewords of the transmitted signal for further processing or extraction of required information. This invention primarily focuses on providing accurate frequency offset to ensure frequency matching between the received and transmitted signals for correct demodulation. If the frequency offset is not accurately estimated and compensated during random access, the receiver cannot correctly identify and decode the random access codes of other devices, leading to decreased access performance, potential access failures or increased access delays, increased probability of signal collisions with other signals, and potential failure to correctly decode the received signal, resulting in an increased transmission error rate. This invention ensures demodulation accuracy, improves communication quality, and enhances anti-interference capabilities, contributing to optimized performance and user experience of wireless communication systems.

[0049] The technical solution of this invention overcomes technical bias:

[0050] Traditional frequency offset estimation algorithms estimate the frequency offset based on the phase difference between received signals through correlation operations. However, due to limitations imposed by the method and specific channel models, the estimated frequency offset range is limited, typically only estimating fractional or integer multiples of the frequency offset. The remaining portion requires further processing, and the extraction of correlation peaks is susceptible to multipath effects and noise, significantly reducing the performance of frequency offset estimation. To overcome these problems, this invention utilizes a data-driven intelligent frequency offset estimation method based on the PDP spectrum of the received signal and a clustering neural network. This method automatically extracts features from the PDP spectrum, establishes the relationship between the CFO and the PDP spectrum, and leverages the denoising capabilities of the neural network to significantly improve the overall frequency offset estimation performance under various channel models. Attached Figure Description

[0051] Figure 1 This is a flowchart of the frequency offset estimation method based on clustering neural networks provided in an embodiment of the present invention;

[0052] Figure 2 This is a schematic diagram of the frequency offset estimation method based on clustering neural network provided in the embodiments of the present invention;

[0053] Figure 3 This is a schematic diagram of the mean square error of integer multiple frequency offset estimation between the conventional method and the proposed method provided in this embodiment of the invention;

[0054] Figure 4 This is a schematic diagram illustrating the mean square error of fractional frequency offset estimation between the conventional method and the proposed method provided in this embodiment of the invention. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0056] like Figure 1 As shown, the frequency offset estimation method based on clustering neural networks provided in this embodiment of the invention includes the following steps:

[0057] S101: A leader design principle is proposed to select the root index that is not affected by TA compensation error and multipath effect within the frequency offset range set in the dataset, and to generate all available leaders.

[0058] S102: Generate a PDP spectrum dataset for all available root indices under the set frequency offset;

[0059] S103: Construct a comprehensive frequency offset estimation model based on a clustering neural network using the dataset of each root index. Train the model offline to achieve efficient full frequency offset estimation in the online stage.

[0060] Example 1:

[0061] The frequency offset estimation method for 5G non-terrestrial networks based on clustering neural networks provided in this invention includes the following steps:

[0062] 1) Based on the influence of frequency offset on PDP spectrum, a leading root index selection principle is designed according to TA compensation error, maximum multipath delay spread and PDP spectrum peak position;

[0063] 2) For each available root index, construct the PDP spectrum dataset under different frequency offsets;

[0064] 3) For any one of the root indices, based on the regularity of the PDP spectrum samples, a semi-supervised K-means algorithm is proposed to cluster the entire dataset;

[0065] 4) Combining the sparsity of PDP spectrum samples, each cluster of data is sparsely processed as a new dataset, and a single-hidden-layer BP neural network is trained to obtain a frequency offset estimation model based on the clustering neural network.

[0066] 5) The signal frequency offset is estimated using the comprehensive frequency offset estimation model based on clustering neural network.

[0067] The specific steps in step 1) are as follows: analyze the impact of frequency offset on the PDP spectrum, and select the available root sequences based on the range of frequency offset in the sample set.

[0068] The specific method for constructing the dataset in step 2) is as follows: within the set frequency offset range, for each available root index, generate the PDP spectrum under different frequency offsets as the dataset.

[0069] Based on the constructed dataset, in step 3), the PDP spectrum exhibits regularity under the influence of frequency offset, and PDP sample data generated by integer multiples of frequency offset is used as label data to guide the clustering process of a semi-supervised K-means algorithm.

[0070] In step 4), the clustered dataset can be used to extract data on potential peak locations based on the root index, resulting in a new low-dimensional dataset. Next, a single-hidden-layer BP neural network is trained using this new dataset, yielding a frequency offset estimation model based on the clustering neural network, which can be used to implement complete frequency offset estimation online.

[0071] Example 2:

[0072] like Figure 2 As shown in the figure, the 5G non-terrestrial network frequency offset estimation method based on clustering neural network provided in this embodiment of the invention includes the following steps:

[0073] Step 1: Analyze the impact of frequency offset on the PDP spectrum and design the prelead selection principle:

[0074] The 5G random access preamble is generated from the ZC sequence. The ZC sequence is generated in the following way: Where n is the number of sample points, N ZC Let u be the length of the ZC sequence, and u be the root index where u∈{1,...,N}. ZC -1}. All available preambles are generated using different root indices. The user randomly selects one for transmission on the physical random access channel. After passing through the satellite multipath fading channel, the preamble sequence received by the receiver is... Where L is the total number of paths, ω(n) is the noise, l is the l-th path to the destination, τ is the round-trip time, and h l and τ l ε represents the channel gain and relative path delay of the l-th path, and ε is the normalized frequency offset after normalizing the subcarrier spacing, including both integer and fractional parts. The PDP spectrum is generated by correlation operation between the received signal and the local root sequence with root index u, as follows:

[0075]

[0076] in(·) * For complex conjugate operations, m is the time index. Because this invention focuses on the mapping relationship between the PDP spectrum and the CFO, it assumes that the TA has been pre-compensated, the channel model only includes the line-of-sight channel with a channel gain of 1, and ignores the influence of noise. The received signal obtained at this time is: The generated PDP spectrum is as follows:

[0077]

[0078] Expanding the above equation reveals that when ε is an integer, the value of p(m) can only be 0 or 1, and the position where 1 appears is... When ε is a decimal, At this point, there is more than one non-zero position, and the largest peak value is no longer 1. In other words, when the frequency offset has a fractional part, the energy of the largest peak value will be distributed to other positions, and these distribution positions will only be... The peaks in the PDP spectrum are integer multiples of the sequence length N. Therefore, it can be concluded that under the influence of CFO, the locations where peaks appear in the PDP spectrum are finite and definite, and these peak locations are relative to the sequence length N. ZC The amount of data is extremely small, so the present invention can obtain PDP spectrum data that is sparse and regular.

[0079] However, in practical satellite communication systems, TA compensation errors and multipath effects exist, causing the peaks of PDP spectra generated by different frequency offsets to appear at the same location or overlap, which will have a certain impact on frequency offset estimation. To maintain a one-to-one correspondence between frequency offset and PDP spectrum, this invention designs a root sequence selection principle, as follows:

[0080] 1) The root index u is a prime number;

[0081] 2)

[0082] Where Δτ is the TA compensation error, τ L This represents the maximum multipath delay spread. z1 and z2 are two arbitrary integers.

[0083] Step 2: Construct the dataset based on the leading design principles:

[0084] For all available leader sequences, within a preset frequency offset range, generate datasets of PDP spectra generated at different frequency offsets, denoted by D:

[0085] D={(p1,ε1),(p2,ε2),...,(p G ,ε G )}

[0086] Here, G represents the size of the dataset. It should be noted that the dataset contains data for multiple channel models, p g The frequency offset value is ε g The generated PDP spectrum data, where g represents the g-th one, and g∈{1,2,...,G};

[0087] Step 3: For each root index, cluster the dataset using the semi-supervised K-means algorithm:

[0088] For each root index, a frequency offset estimation model based on a clustering neural network is established. Based on the analysis in the first step, the PDP spectral matrix exhibits sparsity and regularity under the influence of frequency offset. Furthermore, when the frequency offset has a fractional part, the position of the maximum peak remains unchanged compared to the PDP spectrum affected only by integer multiples of frequency offset. Based on these findings, this invention utilizes PDP spectral sample data generated by integer multiples of frequency offset as supervisory information to guide a semi-supervised K-means algorithm, clustering the entire dataset into several classes. Specifically, in the clustering initialization phase, the number of integer multiples of frequency offset and their corresponding samples are used as the K value and initial cluster centers, and the labels of these samples remain unchanged during the clustering process. After reaching the iteration termination condition, all samples will be clustered into K classes.

[0089] Step 4: Perform sparse dimensionality reduction on the clustered data to construct a new dataset:

[0090] Specifically, based on the analysis in the first step, the location of the peak is related to u, while N is related to the PDP spectral data. ZC For each feature, only a very small portion of the features are related to the frequency offset, therefore the sample features are sparse. To reduce the impact of redundant features on the data, this invention extracts only the features affected by the frequency offset from the PDP spectrum, obtaining a more accurate low-dimensional feature vector p'.

[0091] Step 5: Train a BP neural network using the new dataset to obtain the frequency offset estimation model:

[0092] The complexity of a backpropagation (BP) neural network is related to the length of the input feature vector, and using low-dimensional feature vectors can reduce the network's complexity. Next, we use feature vectors of size G from the k-th category... k The network is trained using a new dataset, where k∈[1,K]. Specifically, the data of this category is divided into training and test sets in an 8:2 ratio. The training set is used to train the network, and the test set is used to evaluate whether the network has good frequency offset estimation performance. The backpropagation algorithm is used to backpropagate the network error, and the network parameters are further tuned to obtain the optimal parameters to ensure that the network performance reaches the expected accuracy level.

[0093] Step 6: Estimate the frequency offset using the frequency offset estimation model based on a clustering neural network:

[0094] First, the target receives the preamble signal and the local root sequence, performing correlation operations to obtain the PDP spectrum. Based on its peak positions, the root index used by the user is determined. Then, the obtained PDP spectrum is used as an input vector and fed into a pre-trained frequency offset estimation model based on a clustering neural network under that root index. Specifically, the frequency offset category is estimated by calculating the minimum Euclidean distance between the maximum peak position of the target PDP spectrum and the maximum peak position of each cluster center. k∈[1,K], where (x1,n1) and (x k ,n k ) represent the maximum peak position of the target PDP spectrum and the maximum peak position of the k-th cluster center, respectively. Finally, after dimensionality reduction of the PDP spectrum, a complete estimate of the frequency offset is obtained through network regression. Target frequency offset The precise value is estimated as follows: Where f(·) is the mapping function between the PDP spectrum and CFO, p t 'This is the data after dimensionality reduction of the target PDP spectrum.' It is an estimate of the target frequency offset.

[0095] The frequency offset estimation system based on clustering neural networks provided in this embodiment of the invention includes:

[0096] The principle design module proposes a leader design principle to select root indices that are not affected by TA compensation errors and multipath effects within the frequency offset range set in the dataset, and generates all available leaders.

[0097] The dataset generation module is used to generate PDP spectrum datasets for all available root indices under a set frequency offset.

[0098] The model building module is used to build a comprehensive frequency offset estimation model based on a clustering neural network using the dataset of each root index. The model is trained offline and used to achieve efficient frequency offset estimation in the online stage.

[0099] To obtain complete frequency offset estimates under various channel models, an intelligent algorithm and frequency offset estimation are combined, and the mean square error performance of this algorithm is compared with that of traditional frequency offset estimation algorithms. By observing the simulation results, it can be concluded that the frequency offset estimation algorithm based on clustering neural networks proposed in this invention can significantly improve the frequency offset estimation performance under various channel models; the performance gain increases with the increase of dispersion; it can achieve complete frequency offset estimation; and it is robust to different channel models.

[0100] Two specific embodiments of the present invention are as follows:

[0101] Example 1:

[0102] The first step is to analyze the impact of different CFOs on the PDP spectrum and design a leader selection principle. For example, we can use simulation experiments to determine how changes in CFO affect the PDP spectrum, and then select the root index that is not affected by TA compensation error and maximum multipath delay spread based on the experimental results.

[0103] The second step is to generate a PDP spectrum dataset based on the selected root index. We can generate the corresponding PDP spectrum using physical models or experimental data.

[0104] The third step involves clustering the PDP spectral data for each root index using a semi-supervised K-means algorithm. This groups similar frequency offsets into the same class, improving the accuracy of frequency offset estimation.

[0105] The fourth step is to perform sparse dimensionality reduction on the clustered data to obtain a new dataset. This step can be achieved through Principal Component Analysis (PCA) or other dimensionality reduction methods.

[0106] The fifth step is to use a back propagation (BP) neural network to train the new dataset and obtain the frequency offset estimation model.

[0107] The sixth step is to estimate the frequency offset using the trained frequency offset estimation model. For example, we can input the test data into the model and then obtain the estimated value of the frequency offset.

[0108] Example 2:

[0109] Step 1: First, analyze the impact of different CFOs on the PDP spectrum through simulation or experimentation. Based on these results, we can design a leading selection principle, such as selecting root indices that are not affected by TA value compensation errors and multipath effects within the frequency offset range set in the dataset.

[0110] Step 2: After determining the leader selection principle, we generate the PDP spectrum dataset using the selected root index. This dataset will contain all possible frequency offsets and their corresponding PDP spectra.

[0111] Step 3: For the PDP spectral data generated for each root index, we use the spectral clustering algorithm for clustering. Spectral clustering is a graph-based clustering method that can find complex structures in the data, and therefore may be more suitable for processing our data than the K-means algorithm.

[0112] Step 4: To reduce computational complexity, we use the t-Distributed Stochastic Neighbor Embedding (t-SNE) algorithm to reduce the dimensionality of the clustered data. t-SNE is a non-linear dimensionality reduction method, which is particularly suitable for preserving the local structure of the data, and therefore may be more suitable for our task.

[0113] Step 5: Using the new dimensionality-reduced dataset, we train a Convolutional Neural Network (CNN) to obtain a frequency offset estimation model. Since CNNs have excellent spatial feature extraction capabilities, they may be more suitable for processing PDP spectral data than traditional backpropagation (BP) neural networks.

[0114] Step 6: Finally, we use the trained frequency offset estimation model to estimate the frequency offset. The PDP spectral data to be tested can be input into the model, and the model will output the corresponding frequency offset estimate.

[0115] The embodiments of the present invention have achieved some positive results during the research and development or use process, and have indeed great advantages compared with the prior art. The following content describes them in conjunction with the data, charts and other information of the experimental process.

[0116] The primary purpose of this simulation is to obtain an accurate estimate of frequency offset in 5G non-terrestrial scenarios. This invention utilizes a random access preamble to achieve this frequency offset estimation. The preamble sequence length N... ZC The value is 839, the root index u is 11, the normalized frequency offset range of the dataset is [-5, 5], the sampling interval is 0.001, the length of the cyclic prefix is ​​set to 8, and the number of neurons in the hidden layer of the BP neural network is set to 5.

[0117] Figure 3 , Figure 4The mean square error (MSE) performance curves of integer multiple frequency offset estimation and fractional multiple frequency offset estimation of the present invention and traditional methods in high-dynamic non-ground scenarios under different signal-to-noise ratios (SNRs) are given. In the simulation, the additive white Gaussian noise (AWGN) and the NTN-TDL-D channel model in 3GPP are used to simulate the line-of-sight scenarios with different user elevation angles, and the NTN-TDL-B channel model in 3GPP is used to simulate the non-line-of-sight scenarios. The so-called traditional method, Figure 3 refers to the integer multiple frequency offset estimation method based on conjugate symmetric ZC sequences in Figure 4 refers to the fractional multiple frequency offset estimation method based on cyclic prefix in Figure 3 、 Figure 4 It can be seen from both that, compared with the traditional method, the present invention can obtain better estimation performance than the traditional method under various channel models, and as the degree of dispersion increases, the obtained performance gain will also increase. Specifically, Figure 3 in , the method of the present invention can accurately estimate the integer multiple frequency offset when the signal-to-noise ratio is greater than -14 dB, Figure 4 in , the fractional multiple frequency offset estimation of the proposed method is more accurate under the same signal-to-noise ratio. In addition, the method of the present invention can not only estimate the integer multiple frequency offset, but also estimate the fractional multiple frequency offset, that is, it has comprehensive frequency offset estimation ability, which is the key advantage of this invention.

[0118] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware part can be implemented using dedicated logic; the software part can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. Those of ordinary skill in the art can understand that the above devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code is provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and their modules of the present invention can be implemented by hardware circuits of programmable hardware devices such as very large scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, or can be implemented by a combination of the above hardware circuits and software, such as firmware.

[0119] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A frequency offset estimation method based on clustering neural networks, characterized in that, The frequency offset estimation method based on clustering neural networks selects root indices that are not affected by timing advance TA value compensation errors and multipath effects within the frequency offset range set in the dataset, according to the designed preamble design principle. All available preamble sequences are generated from different root indices. Then, power delay spectrum (PDP) datasets are generated for all available root indices under the set frequency offset. Finally, a frequency offset estimation model based on clustering neural networks is constructed using the datasets of each root index. The model is trained offline to achieve efficient frequency offset estimation in the online stage. The frequency offset estimation method based on clustering neural networks includes the following steps: Step 1: Analyze the impact of different carrier frequency offsets (CFOs) on the PDP spectrum, and design the preamble selection principle based on TA compensation error, maximum multipath delay spread, and PDP spectrum peak position; Step 2: Construct the PDP spectral dataset based on the leading design principles; Step 3: For each root index, based on the impact of CFO on the PDP spectrum obtained in Step 1, a semi-supervised K-means algorithm is proposed to cluster the generated PDP spectrum data: Step 4: Combining the impact of CFO on PDP spectrum obtained in Step 1, perform sparse dimensionality reduction on the clustered data to construct a new dataset; Step 5: Train a backpropagation BP neural network using the new dataset to obtain the frequency offset estimation model; Step 6: Estimate the frequency offset based on the frequency offset estimation model based on clustering neural network; The first step involves generating a 5G random access preamble from a ZC sequence. The ZC sequence is generated as follows: ,in For the number of sample points, The length of the ZC sequence, It is the root index and Using different root indices, all available preambles are generated. The user randomly selects one and transmits it on the physical random access channel. After passing through the satellite multipath fading channel, the preamble sequence received by the receiver is... ,in It is the total number of paths. It's noise. It is the first one to arrive Path, It is the round-trip time delay. and It is the first Channel gain and relative path delay for each path, It is the normalized frequency offset after normalizing the subcarrier spacing, including an integer part and a fractional part; it is obtained by using the received signal and the root index. The local root sequence correlation operation generates the PDP spectrum as follows: ， in This is a complex conjugate operation. This is a time index. Assuming the transmitted sequence has undergone TA pre-compensation, the channel model only includes the line-of-sight channel with a channel gain of 1, and the noise term has no effect on the position of the PDP spectrum peak, neglecting the noise term, the received signal is: The generated PDP spectrum is as follows: ， Expanding the above equation, we find that when... When it is an integer, The value can only be 0 or 1, and the position where 1 appears is... At the point where integer multiples of frequency offset exist, the peak position is unique and fixed; when When it is a decimal, = At this point, there is more than one non-zero position, and the largest peak value is no longer 1. In other words, when the frequency offset has a fractional part, the energy of the largest peak value will be distributed to other positions, and these distributions will only occur in certain locations. Integer multiples of; Design a root sequence selection principle as follows: Root Index It is a prime number; ； in To compensate for the error of the TA, For maximum multipath delay spread, , Let be any two integers.

2. The frequency offset estimation method based on clustering neural networks as described in claim 1, characterized in that, The second step, for all available root sequences, generates a dataset of PDP spectra at different frequency offsets within a preset frequency offset range. To indicate: ， in, This represents the size of the dataset, which contains data from multiple channel models. The frequency offset value is The generated PDP spectral data, Representing the One, of which ; The third step involves establishing a frequency offset estimation model based on a clustering neural network for each root index. The PDP spectrum sample data generated by integer multiples of frequency offset is used as supervisory information to guide the semi-supervised K-means algorithm, clustering the entire dataset into several classes. Specifically, during the initialization phase, the number of integer multiples of frequency offset and the samples generated by integer multiples of frequency offset are used as the K value and initial cluster centers, and the labels of these samples do not change during the clustering process. After the iteration termination condition is met, all samples will be clustered into K categories.

3. The frequency offset estimation method based on clustering neural networks as described in claim 1, characterized in that, The fourth step, based on the analysis in the first step, extracts only the features affected by frequency offset in the PDP spectrum to obtain a more accurate low-dimensional feature vector. ; The fifth step uses the first The size of each category is The new dataset is fed into the BP neural network for training. Specifically, the data of this category is divided into training and test sets in an 8:2 ratio. The training set is used to train the network, and the test set is used to evaluate whether the network has good frequency offset estimation performance. The backpropagation algorithm is used to backpropagate the network error and further tune the network parameters to obtain the optimal parameters.

4. The frequency offset estimation method based on clustering neural networks as described in claim 1, characterized in that, The sixth step specifically refers to: First, the target receives the preamble signal and the local root sequence, performing correlation operations to obtain the PDP spectrum. Based on its peak positions, the root index used by the user is determined. Then, the obtained PDP spectrum is used as an input vector and fed into a pre-trained frequency offset estimation model based on a clustering neural network under that root index. Specifically, the frequency offset category is estimated by calculating the minimum Euclidean distance between the maximum peak position of the target PDP spectrum and the maximum peak position of each cluster center. ,in, and These represent the maximum peak position and the first peak position of the target PDP spectrum, respectively. The maximum peak position of each cluster center is located; finally, the PDP spectrum is reduced in dimensionality and a complete estimate of the frequency offset is obtained through network regression, which is the target frequency offset. The precise value is estimated as follows: ,in, This is a mapping function between the PDP spectrum and the CFO. This is the data after dimensionality reduction of the target PDP spectrum. It is an estimate of the target frequency offset.

5. A computer device, characterized in that, The computer device includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, causes the processor to perform the frequency offset estimation method based on a clustering neural network as described in any one of claims 1 to 4.

6. A computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the frequency offset estimation method based on a clustering neural network as described in any one of claims 1 to 4.

7. An information data processing terminal, characterized in that, The information data processing terminal is used to implement the frequency offset estimation method based on clustering neural networks as described in any one of claims 1 to 4.

8. A frequency offset estimation system based on a clustering neural network, implementing the frequency offset estimation method based on any one of claims 1 to 4, characterized in that, The frequency offset estimation system based on clustering neural networks includes: The principle design module proposes a leader design principle to select root indices that are not affected by TA compensation errors and multipath effects within the frequency offset range set in the dataset, and generates all available leaders. The dataset generation module is used to generate PDP spectrum datasets for all available root indices under a set frequency offset. The model building module is used to build a frequency offset estimation model based on a clustering neural network using the dataset of each root index. The model is trained offline and used to achieve efficient frequency offset estimation in the online stage.

Images