A channel fingerprint generation method for trajectory distillation discrete flow matching
By employing a channel fingerprint generation method based on trajectory distillation and discrete flow matching, and utilizing a latent space coding-decoding model and a discrete flow matching generation model, the high-fidelity and low-step inference problems of channel fingerprint generation in wireless communication systems are solved, achieving low-latency and high-efficiency channel fingerprint generation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- YANCHENG TEACHERS UNIV
- Filing Date
- 2026-03-20
- Publication Date
- 2026-06-16
AI Technical Summary
In existing wireless communication systems, generative deep learning suffers from problems such as high inference overhead, high latency, error accumulation, and complex training and supervision construction when generating channel fingerprints, making it difficult to achieve a balance between high fidelity and fewer inference steps.
A trajectory distillation and discrete flow matching approach is adopted. Through a latent space coding and decoding model and a discrete flow matching generation model, a trajectory distillation pairing dataset is constructed using a pre-trained teacher generation model. Supervised training is carried out in conjunction with Euler step indexing to achieve high-fidelity generation with fewer steps.
It achieves high-fidelity generation with very few steps, reduces inference latency, improves the stability and training efficiency of the generation process, reduces error accumulation, and is suitable for deployment in real-time wireless communication systems.
Smart Images
Figure CN122227271A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of wireless communication and relates to a channel fingerprint generation method in a wireless communication system, particularly a channel fingerprint generation method based on trajectory distillation and discrete flow matching with few steps. Background Technology
[0002] In wireless communication systems, the channel varies with factors such as user location, propagation environment, and scatterer distribution. To support applications such as beam management, resource scheduling, location sensing, and channel digital twins, engineering often uses the channel statistical characteristics at a certain location as a "channel fingerprint," and constructs a fingerprint database or fingerprint map in the form of "location-fingerprint" for querying, matching, or inference during operation.
[0003] Compared to instantaneous channel states, statistical channel fingerprints typically involve statistically aggregating channel observations over a specific time, frequency, and sample range. This filters out fast-changing random fluctuations while retaining long-term stable features related to propagation geometry, time delay structure, and angular energy distribution. Therefore, they are more suitable for characterizing location-dependent electromagnetic environments and serving subsequent communication decisions. In practice, a common approach is to first transform the spatial frequency domain channel to a structured domain such as the beam delay domain. Utilizing the sparse or clustered characteristics of this domain, the corresponding domain coefficients (e.g., power distribution or energy distribution) are then statistically averaged to obtain the fingerprint representation. Subsequent processing, such as normalization, logarithmic transformation, or thresholding, can be performed according to application requirements to enhance numerical stability and facilitate storage or adaptation to learning models.
[0004] Existing technologies for mapping location to statistical channel fingerprints mainly fall into two categories. One category comprises traditional model-driven or interpolation-regression methods, which rely on measurement acquisition and fitting to extrapolate fingerprints spatially. While relatively straightforward, these methods often require frequent updates when fingerprint dimensionality is high, structure is complex, or the environment changes, leading to high costs and strong dependence on environmental stability. The other category is data-driven methods, which use deep networks to learn the location-to-fingerprint mapping. These methods can improve expressive power to some extent, but may still be limited in high-dimensional structural detail recovery, scene generalization, and uncertainty characterization.
[0005] In recent years, generative deep learning has been used for distributed modeling and data synthesis of wireless channel fingerprints to improve generation quality and coverage. However, existing generative schemes still have significant shortcomings in engineering deployment. First, inference overhead and latency are high; methods such as diffusion typically require multiple iterations to achieve high fidelity, which is not conducive to real-time or edge deployment. Second, error accumulation is prone to occur with few generation steps. When the number of inference steps is compressed to a small number, the inconsistency between the training objective and the discrete solver inference is more likely to lead to error accumulation, resulting in unstable generation quality. Third, the construction cost of conditional supervision is high. Conditional generation training often requires constructing effective pairings of "noise-target samples" or trajectory supervision. If the pairing strategy is complex or unstable, it will further limit the usability and repeatability of few-step generation.
[0006] Therefore, a technical solution for wireless communication systems is still needed that, while using statistical channel fingerprints as the generation object, simultaneously achieves high fidelity and fewer inference steps, reduces inference latency and error accumulation risks, and improves the stability of training supervision construction and engineering deployment. Summary of the Invention
[0007] Purpose of the invention: The purpose of this invention is to provide a statistical channel fingerprint generation method for wireless communication systems, which can achieve high-fidelity generation of location-related statistical channel fingerprints under few-step inference conditions, thereby reducing generation latency and computational overhead, and improving the stability of the few-step generation process and the consistency of training and inference.
[0008] Technical Solution: To achieve the above-mentioned objective, this invention provides a channel fingerprint generation method for trajectory distillation discrete flow matching, comprising the following steps:
[0009] Obtain the training dataset, which includes multiple locations and their corresponding statistical channel fingerprints;
[0010] Construct and train a latent space coding and decoding model. After training, fix the encoder and decoder parameters, and use the encoder to convert the statistical channel fingerprints in the training set into a set of latent variables.
[0011] Using a pre-trained, location-conditional teacher generation model, one or more sets of paired samples from the starting latent variable to the ending latent variable are generated for each location, forming a trajectory distillation paired dataset.
[0012] Training a discrete flow matching generation model includes: setting the number of Euler steps in the inference phase; dividing the normalized time interval of the generation process into equal parts, with each part corresponding to one Euler integral update; constructing a discrete flow matching network, which is a conditional generation network used to obtain the velocity vector prediction value updated by the Euler integral for the corresponding Euler step based on the latent variable state, position conditions, and Euler step index; constructing training samples according to the Euler step index and supervising them; during each training iteration, sampling paired samples from the trajectory distillation paired dataset, randomly extracting the Euler step index, and constructing intermediate latent variable states between the starting latent variable and the ending latent variable based on the extracted Euler step index; using the intermediate latent variable state, position conditions, and Euler step index as network inputs; and using the target displacement vector obtained by the difference between the ending latent variable and the starting latent variable of the paired sample as a supervision signal to supervise the velocity vector prediction value output by the network.
[0013] Given a target location, random noise is sampled as an initial latent variable. The Euler integral is updated sequentially by the set number of Euler steps. At each step, the velocity vector prediction value is obtained through a trained discrete flow matching network. The latent variable is updated according to the step size, and finally the destination latent variable is obtained.
[0014] Input the endpoint latent variable into the decoder to obtain the statistical channel fingerprint.
[0015] Preferably, the construction of the statistical channel fingerprint in the training dataset includes the following steps:
[0016] Acquire channel observation data of a wireless communication system at different locations;
[0017] The channel observation data is subjected to structured domain transformation to obtain a structured representation with sparse or clustered features;
[0018] Statistical convergence is performed on the structured representation to obtain a channel fingerprint in statistical form that characterizes the location-related propagation properties.
[0019] Furthermore, the construction step of the statistical channel fingerprint also includes: performing subsequent processing on the statistical channel fingerprint according to application requirements, wherein the subsequent processing includes one or more of normalization, logarithmic transformation, truncation, thresholding, scaling or quantization.
[0020] As a preferred option, the latent space coding and decoding model is a variational autoencoder, in which the encoder maps the statistical channel fingerprint to low-dimensional latent variables, and the decoder reconstructs the statistical channel fingerprint from the latent variables; the dimension of the latent variables is set according to the fingerprint dimension and compression requirements.
[0021] Preferably, the teacher generation model is a diffusion-type generation model or a flow model, and its generation process is a deterministic process or a process with randomness; the generated paired samples include starting latent variables, ending latent variables and corresponding positional conditions, and are cached offline as a trajectory distillation paired dataset for subsequent training.
[0022] Preferably, the Euler step number K ranges from 2 to 10; the normalized time interval of the generation process is divided into K equal parts, each part corresponds to one Euler integral update, and the Euler step index k is specified to be from 0 to K-1 in sequence, and supervision is performed on the same set of Euler step indices during the training phase.
[0023] Preferably, the discrete flow matching network adopts a conditionally enhanced multi-scale convolutional network structure, with a U-shaped backbone. Position conditional embedding and Euler step index embedding are injected into the multi-layer residual module, and the output end obtains the velocity vector prediction value through the output head.
[0024] Preferably, when constructing intermediate latent variable states between the starting latent variable and the ending latent variable based on the extracted Euler step index, linear interpolation is performed between the starting latent variable and the ending latent variable according to the normalization progress corresponding to the Euler step index to obtain the intermediate latent variable states.
[0025] The present invention also provides a computer system, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the channel fingerprint generation method for trajectory distillation discrete flow matching.
[0026] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the channel fingerprint generation method for trajectory distillation discrete flow matching.
[0027] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:
[0028] 1. High-fidelity generation with few steps, significantly improving generation accuracy: This invention constrains the generation process to a deterministic and easily learnable macrostep objective through discrete training aligned with trajectory distillation and inference solutions, enabling high-fidelity generation accuracy to be maintained even with very few inference steps. Experiments show that the normalized mean square error of this invention in few inference steps can match and slightly exceed that of a diffusion baseline model with hundreds of steps.
[0029] 2. Significantly reduced inference latency, meeting real-time deployment requirements: This invention maps high-dimensional statistical channel fingerprints to low-dimensional latent variables through a latent space coding-decoding model, and combines it with few-step Euler integral updates, increasing the backbone network forward computation overhead by only one step per step. Compared to the diffusion baseline model that requires hundreds of steps of reverse generation, this significantly reduces computational overhead and generation waiting time, meeting the low-latency generation requirements of applications such as channel digital twins and real-time beam management in wireless communication systems.
[0030] 3. Enhanced consistency between training and inference, reduced accumulation of discretization errors, and improved stability: This invention directly uses the same Euler step index as the inference stage for supervised optimization during the training phase, enabling the velocity vector to be explicitly optimized on the macrostep grid. This fundamentally avoids discretization errors and few-step degradation caused by the inconsistency between continuous training objectives and discrete inference solutions, making it more suitable for scenarios with very few steps of inference.
[0031] 4. Smoother convergence, lower error, and improved training reliability: This invention obtains deterministic paired samples from the starting latent variable to the ending latent variable through trajectory distillation, providing clear geometric constraints for velocity field learning. The training process is easier to converge, the steady-state error is lower, and it can maintain a smoother and lower error curve in the later stage of training, demonstrating a structural advantage that is more favorable for generating high-precision data in fewer steps.
[0032] 5. The supervised construction can be cached offline, improving training efficiency and reproducibility. This invention uses an offline pre-trained teacher-generated model to construct deterministic paired samples, making the supervision signal fixed and reusable, reducing the computational burden and instability caused by online pairing, thereby improving overall training efficiency and experimental reproducibility. Attached Figure Description
[0033] Figure 1 This is a flowchart of the overall process of the method of the present invention.
[0034] Figure 2 Flowchart for constructing and offline caching of trajectory distillation pairing data.
[0035] Figure 3 Flowchart for training a model for discrete flow matching.
[0036] Figure 4 A flowchart is generated for few-step reasoning and Euler integral updates.
[0037] Figure 5 The graph shows the normalized mean square error as a function of the number of Euler integration steps (inference steps) K.
[0038] Figure 6 The graph shows the inference time difference as a function of the number of Euler integration steps (inference steps) K. Detailed Implementation
[0039] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. Those skilled in the art should understand that equivalent substitutions or modifications made to the structures, steps, and parameters in the embodiments without departing from the spirit and scope of the present invention should fall within the scope of protection of the present invention.
[0040] Terminology convention: In this invention, K represents the number of Euler integral updates in a few-step inference, also known as the Euler steps or inference steps. k represents the step index of the Euler integral update, with k taking values from 0 to K-1.
[0041] like Figure 1 As shown in the figure, the channel fingerprint generation method for trajectory distillation discrete flow matching provided by the present invention mainly includes the following steps:
[0042] S1. Obtain the training dataset, which includes multiple locations and their corresponding statistical channel fingerprints;
[0043] S2. Construct and train the latent space coding and decoding model. After training, fix the encoder and decoder parameters, and use the encoder to convert the statistical channel fingerprints in the training set into a set of latent variables.
[0044] S3. Using a pre-trained teacher generation model conditioned on location, generate one or more sets of paired samples from the starting latent variable to the ending latent variable for each location to form a trajectory distillation paired dataset.
[0045] S4. Training a discrete flow matching generation model, including: setting the number of Euler steps in the inference phase, dividing the normalized time interval of the generation process into equal parts, with each part corresponding to one Euler integral update; constructing a discrete flow matching network, wherein the network is a conditional generation network, used to obtain the velocity vector prediction value of the corresponding Euler step update based on the latent variable state, position condition and Euler step index; constructing training samples according to the Euler step index and supervising them, in each training iteration, sampling paired samples from the trajectory distillation paired dataset, randomly extracting the Euler step index, and constructing an intermediate latent variable state between the starting latent variable and the ending latent variable based on the extracted Euler step index, using the intermediate latent variable state, position condition and Euler step index as network input, and using the target displacement vector obtained by the difference between the ending latent variable and the starting latent variable of the paired sample as a supervision signal to supervise the velocity vector prediction value output by the network;
[0046] S5. Given the target position, sample random noise as the initial latent variable, and sequentially perform Euler integral updates for the set number of Euler steps. At each step, obtain the velocity vector prediction value through the trained discrete flow matching network, update the latent variable according to the step size, and finally obtain the endpoint latent variable.
[0047] S6. Input the endpoint latent variable into the decoder to obtain the statistical channel fingerprint.
[0048] This embodiment provides a channel fingerprint generation method based on trajectory distillation and discrete flow matching. Using location conditions as input, it rapidly generates statistical channel fingerprints from random noise within the latent space. Stable supervision is constructed through trajectory distillation, and alignment optimization is performed during the training phase using the Euler integral update method from the inference phase. This enables the generation of high-fidelity statistical channel fingerprints even with few inference steps.
[0049] In some possible implementations, step S1, constructing the statistical channel fingerprints in the training dataset, may include the following steps:
[0050] S11. Obtain channel observation data of the wireless communication system at different locations.
[0051] S12. Perform structured domain transformation on the channel observation data to obtain a structured representation with sparsity or clustering characteristics.
[0052] S13. Perform statistical convergence on the structured representation to obtain a channel fingerprint in statistical form that characterizes the location-related propagation characteristics.
[0053] In specific applications, in step S11, the observation data can originate from simulation, field acquisition, or measurements on an experimental platform. In step S12, the channel observation data undergoes structured domain transformation processing, converting the original channel representation into a structured representation with sparse or clustered characteristics. This structured representation includes at least the beam domain, the time delay domain, or a combination of both. In step S13, the statistical convergence includes operations such as statistical averaging of the energy or power distribution to extract more stable long-term features.
[0054] Furthermore, the construction of the statistical channel fingerprint may further include step S14, which involves subsequent processing of the statistical channel fingerprint according to application requirements. The subsequent processing includes one or more of the following: normalization, logarithmic transformation, truncation, thresholding, scaling, or quantization, to enhance numerical stability and adapt to the input of the learning model.
[0055] After constructing the statistical channel fingerprint, a sample organization format is established, and the coordinate information or location code of each location is matched one-to-one with its statistical channel fingerprint to form a training dataset and a test dataset.
[0056] In practical applications, the location can be two-dimensional coordinates, grid number, location embedding vector, or it can be fused with cell identifier, array configuration parameters, or other auxiliary information. Subsequent processing of statistical channel fingerprints can be adjusted according to the numerical range, sparsity, or robustness requirements of different applications.
[0057] In some possible implementations, step S2, latent space representation learning and fingerprint compression, may include the following steps:
[0058] S21. Construct a latent variable encoding and decoding model consisting of an encoder and a decoder to compress high-dimensional statistical channel fingerprints into low-dimensional latent variable representations, and reconstruct statistical channel fingerprints from latent variables.
[0059] S22. Use the training dataset from step S1 to train the latent variable encoding and decoding model so that it satisfies the preset prior constraints while ensuring reconstruction accuracy, thereby making the latent variable space suitable for sampling and generation from random noise.
[0060] S23. After training, the encoder and decoder parameters are fixed, and the encoder is used to convert the statistical channel fingerprints in the training set into a set of latent variables, which will then be used as the learning object for the subsequent generative model. The trained decoder serves as the restoration module in the inference phase, converting the generated latent variables into statistical channel fingerprints as output.
[0061] In practical applications, the encoder and decoder can adopt a convolutional network structure or its equivalent structure. The dimension of the latent variables can be set to a range of several tens to several hundred dimensions according to the statistical channel fingerprint dimension and compression requirements, in order to strike a trade-off between reconstruction accuracy and generation efficiency.
[0062] In some possible implementations, step S3, the construction and offline caching of teacher-generated model trajectory distillation pairing data, may include the following steps:
[0063] S31. Prepare a teacher generation model, which is a location-based generative model capable of generating latent variable target samples corresponding to a given location from random noise. The teacher generation model can be obtained through pre-training and used as a fixed model in this step.
[0064] S32. For each location condition, repeat the following procedure to generate multiple sets of paired samples:
[0065] S321. Sample a set of random noise latent variables as the starting point for generation;
[0066] S322. Input the starting latent variable and the location condition into the teacher generation model, and obtain the corresponding ending latent variable through the teacher generation process;
[0067] S323. Combine the starting latent variable and the ending latent variable under the given location conditions into a paired sample and record the corresponding location condition information.
[0068] S33. Summarize the pairing samples under all location conditions to form a trajectory distillation pairing dataset, and cache it offline for repeated training and experiment reproduction, thereby reducing the computational cost and instability caused by online pairing.
[0069] Optionally, to enhance diversity and coverage, multiple sets of start-point and end-point paired samples can be generated for the same location conditions.
[0070] A detailed example of the above-mentioned trajectory distillation pairing data construction and offline caching process can be found in [link to documentation]. Figure 2 .like Figure 2 As shown, the above trajectory distillation pairing data construction and offline caching can be implemented according to the following algorithm flow.
[0071] Algorithm 1: Trajectory matching and offline caching algorithm based on pre-trained teacher model
[0072] Input: Set of training locations Pre-trained teacher model diffusion steps Number of trajectories corresponding to each location condition ; Indicates the first Each training location condition This represents the total number of training location conditions.
[0073] Output: Trajectory distillation pairing dataset , representing a trajectory distillation paired ternary dataset consisting of a starting latent variable, an ending latent variable, and a location condition.
[0074] step:
[0075] Step 1: Initialize an empty trajectory distillation pairing dataset ;
[0076] Step 2: Based on the preset diffusion scheduling parameters Calculate the corresponding and cumulative coefficient And pre-calculate the correlation coefficients corresponding to each diffusion step;
[0077] Step 3: Traverse the training location set Each positional condition in ;
[0078] Step 4: Based on the current location conditions Repeat execution Sub-independent trajectory sampling;
[0079] Step 5: In the During the next sampling, random noise latent variables are sampled from the Gaussian prior distribution. And use it as the latent variable of the starting point of the trajectory;
[0080] Step 6: Assign the starting latent variable the initial diffusion state of the teacher generation process;
[0081] Step 7: From the diffusion step index arrive The reverse generation process of the teacher model is executed sequentially.
[0082] Step 8: At each diffusion step, utilize the teacher model The noise term is predicted based on the current location conditions and the current diffusion state, and the diffusion state is updated with the posterior mean based on the noise prediction result to obtain the state corresponding to the next diffusion step.
[0083] Step 9: After the reverse generation process reaches the last diffusion step, the teacher model is used to perform terminal denoising on the current state to obtain the terminal latent variable corresponding to the trajectory. ;
[0084] Step 10: Set the starting latent variables Endpoint latent variables and the corresponding positional conditions A trajectory pairing sample is formed and written into the trajectory distillation pairing dataset. ;
[0085] Step 11: Repeat steps 5 through 10 until the current position conditions are met. Trajectory pairing sample construction;
[0086] Step 12: Repeat steps 3 to 11 until all position conditions in the training position set have been matched for trajectory.
[0087] Step 13: Output trajectory distillation paired dataset .
[0088] In the above steps, , , This represents the diffusion scheduling parameters and their corresponding coefficients used in the reverse generation process of the teacher model. The specific values can be set according to the usual diffusion model settings, and will not be elaborated in this embodiment.
[0089] In some possible implementations, step S4, training the discrete flow matching generation model, may include the following steps:
[0090] S41. Determine the number of Euler steps and the range of Euler step indices: Set the number of Euler steps used in the inference phase to K. Divide the normalized time interval of the generation process into K equal parts, each part corresponding to one Euler integral update, and specify that the Euler step index k is from 0 to K-1. During the training phase, supervision is performed on the same set of Euler step indices to ensure consistency between training and inference.
[0091] S42. Construct a discrete flow matching network and define its inputs and outputs: Construct a discrete flow matching network, which is a conditional generation network that receives at least the following inputs: (1) the current latent variable state; (2) the position condition information or its embedded representation; and (3) the embedded representation of the Euler step index k. The network outputs the velocity vector prediction value for the k-th Euler integral update. In practical applications, the network can adopt a conditionally enhanced convolutional network structure, such as a multi-scale network structure based on a U-shaped backbone, and inject position condition embeddings and Euler step index embeddings into the multi-layer residual module. The output end uses an output head to obtain the velocity vector prediction value.
[0092] S43. Construct training samples based on Euler step index and perform supervision: During each training iteration, sample several paired samples from the trajectory distillation paired cache dataset. Randomly select an Euler step index k for each paired sample, and construct the corresponding intermediate latent variable state between the starting latent variable and the ending latent variable based on the Euler step index. Use the intermediate latent variable state, position conditions, and Euler step index as network input. Use the target displacement vector obtained by the difference between the ending latent variable and the starting latent variable of the paired sample as the supervision signal to supervise the predicted velocity vector output by the network.
[0093] S44. Loss Calculation and Parameter Update: Calculate the error between the network output and the supervision signal. The error can be measured using mean squared error or its equivalent form. Backpropagation is then performed to update the network parameters. Training metrics can be recorded simultaneously, and checkpoints can be saved.
[0094] S45. Iterative training until convergence: Repeat the above process until the termination condition is met to obtain the trained discrete flow matching generation model.
[0095] Optionally, in step S43, based on the normalized progress corresponding to the extracted Euler step index, linear interpolation is performed between the starting latent variable and the ending latent variable to obtain the intermediate latent variable state. Let the starting latent variable be... The endpoint latent variable is If the number of Euler steps is K and the Euler step index is k, then the normalization progress is... The corresponding intermediate latent variable state is Furthermore, the target vector is determined by the difference between the endpoint latent variable and the starting latent variable, i.e. ,in It is used to characterize the target displacement from the starting latent variable to the ending latent variable, and serves as a supervision signal to supervise the predicted velocity vector output by the network.
[0096] In the optional implementation, the Euler step number K can be set from 2 to 10 according to the target latency and accuracy requirements, and multiple K configurations can be used for training during the training phase, or a unified model can be used to adapt to multiple K configurations.
[0097] A detailed example of the training process for the discrete flow matching generation model described above can be found here. Figure 3 .like Figure 3 As shown, the training process of the discrete flow matching generation model can be implemented according to the following algorithm flow.
[0098] Algorithm 2: Trajectory Distillation Discrete Flow Matching Generation Model Training Algorithm
[0099] Input: Trajectory distillation pairing dataset Training batch size Euler steps K.
[0100] Output: Parameters of the trained discrete flow matching network .
[0101] step:
[0102] Step 1: Initialize the parameters of the discrete flow matching network and learning rate ;
[0103] Step 2: If the model does not converge, repeat the following training process;
[0104] Step 3: Distilling the pairing dataset from the trajectory A batch of paired samples was sampled to obtain batch data. ; Indicates the training batch size. Indicates the sample index within the batch. Indicates the first The starting latent variable for each paired sample Indicates the first The endpoint latent variable of each paired sample. Indicates the first Location conditions of paired samples;
[0105] Step 4: For each paired sample in the batch data, uniformly draw an Euler step index from 0 to K-1. ;
[0106] Step 5: Based on the extracted Euler step index Calculate the corresponding normalized progress ;
[0107] Step 6: Latent variables at the starting point With endpoint latent variables Between these points, linear interpolation is performed according to the normalization progress to obtain the corresponding intermediate latent variable states. ;
[0108] Step 7: Calculate the target displacement vector corresponding to the paired sample, that is, subtract the starting point latent variable from the ending latent variable to obtain the target displacement vector. ;
[0109] Step 8: Set the intermediate latent variable state Euler step index and location conditions Input the discrete flow matching network to obtain the corresponding velocity vector prediction value. ;
[0110] Step 9: Calculate the average error loss between the predicted velocity vector and the target displacement vector in the current batch of data;
[0111] Step 10: Adjust the network parameters according to the error loss. Perform reverse propagation and updates;
[0112] Step 11: Repeat steps 3 to 10 until the termination condition is met, and output the parameters of the trained discrete flow matching network. .
[0113] In some possible implementations, step S5, where the few-step inference generates latent variables, may include the following steps:
[0114] S51. Under given location conditions, sample random noise latent variables as the initial state for generation.
[0115] S52. Divide the normalized time interval of the generation process into K equal parts, and use the length of each part as the single-step step size.
[0116] S53. Perform the k-th Euler integral update sequentially from 0 to K-1 according to the Euler step index k:
[0117] S531. Input the current latent variable state, position condition embedding, and Euler step index k into the discrete flow matching generation model to obtain the velocity vector prediction value for the kth Euler integral update.
[0118] S532. The latent variable state is updated using the Euler explicit integral update rule, which involves scaling the velocity vector predicted in this step by the single-step size and then superimposing it onto the current latent variable to obtain the state of the next latent variable. The update formula is as follows: ,in, This represents the state of the latent variables before the k-th Euler integral update. This represents the predicted velocity vector output by the discrete flow matching network at the k-th Euler integral update. Indicates the step size in a single step.
[0119] S533. Continue execution by using the updated latent variables as input for the next Euler integral update.
[0120] S54. After completing K Euler integral updates, the endpoint latent variable is obtained as the generation result.
[0121] A detailed example of the above few-step reasoning and Euler integral update generation process can be found in [link to relevant documentation]. Figure 4 .like Figure 4 As shown, the above-mentioned few-step channel fingerprint generation process can be implemented according to the following algorithm flow.
[0122] Algorithm 3, Few-Step Channel Fingerprint Generation Algorithm
[0123] Input: Position conditions to be generated The trained discrete flow matching network The trained latent variable decoder Euler steps K.
[0124] Output: Generated statistical channel fingerprint .
[0125] step:
[0126] Step 1: Sample random noise latent variables And use it as the initial state for generation;
[0127] Step 2: Calculate the single-step size based on the Euler step number K;
[0128] Step 3: Let the Euler step index k start from 0;
[0129] Step 4: Set the current latent variable state Location conditions And Euler step index k-input discrete flow matching network , thus obtaining the predicted velocity vector value for the k-th Euler integral update;
[0130] Step 5: Update the current latent variable state using the Euler explicit integration update rule to obtain the next latent variable state;
[0131] Step 6: Determine whether the current Euler step index k has reached K-1;
[0132] If not, increment the Euler step index by 1 and return to step 4; if yes, proceed to step 7.
[0133] Step 7: Flatten the latent variables of the endpoint after K Euler integral updates;
[0134] Step 8: Input the flattened endpoint latent variable into the trained latent variable decoder. The generated statistical channel fingerprint is obtained. ;
[0135] Step 9: Output the statistical channel fingerprint .
[0136] In some possible implementations, in step S6, the endpoint latent variables obtained in step S5 are input into the decoder trained in step S2 to generate the corresponding statistical channel fingerprint. If subsequent processing was performed on the fingerprint in step S1, then corresponding consistency processing is performed in the output stage to ensure that the output fingerprint meets the format requirements for system storage, display, or downstream tasks.
[0137] Optionally, random sampling can be repeated multiple times for the same location conditions to obtain diverse results, or multiple generated results can be integrated to improve stability.
[0138] The resulting statistical channel fingerprint can be used for fingerprint database completion, fingerprint map construction, channel digital twin update, perception-assisted communication, or other upper-layer tasks in wireless communication.
[0139] In practical applications, the steps involved in the channel fingerprint generation method, such as statistical channel fingerprint construction, latent space representation learning, teacher model trajectory distillation pairing and offline caching, discrete flow matching generation model training, few-step inference to generate latent variables, and decoding output statistical channel fingerprints, can be completed on the same computing device or distributed across different computing devices. For example, training and caching can be completed on the server side, and few-step inference generation and decoding output can be completed on the edge.
[0140] The effectiveness of the present invention is verified through specific experiments. The experiments were conducted on the QuaDRiGa channel simulation platform under the 3GPP 38.901 UMa NLoS scenario. The system center frequency was 4.8GHz, the base station used a 16×8 uniform planar array, the user terminal was a single-antenna device, and the subcarrier spacing was 30kHz. The target coverage area was 100×100m, and the grid was divided at a resolution of 1m, resulting in 10201 sets of location-statistical channel fingerprint samples, of which 9180 sets were used for training and 1021 sets were used for testing. The experiments used normalized mean square error and inference delay as evaluation metrics, and compared them with the latent space diffusion model and the accelerated latent space diffusion model. The latent space diffusion model uses a deterministic sampling strategy, generating high-fidelity statistical channel fingerprints through 500 steps of latent space reverse denoising. The accelerated latent space diffusion model represents the generation process as a non-Markov diffusion process and reduces inference overhead by performing inference on subsequences of the original sampling trajectory without retraining the model. Figure 5 and Figure 6 As shown, experimental results demonstrate that, in terms of generation accuracy, the proposed solution maintains a low and stable normalized mean square error across different Euler steps, achieving an overall accuracy level comparable to high-step inference of the latent space diffusion model and outperforming the accelerated latent space diffusion model. This indicates that the proposed solution effectively suppresses error accumulation in short-step inference. Regarding inference latency, the latency of the proposed solution increases approximately linearly with the number of Euler steps. It achieves significantly lower latency overhead than the high-step inference of the latent space diffusion model even with fewer steps. For example, with 4 Euler steps, the inference latency is approximately 0.14 s, compared to approximately 5.46 s latency for 500 steps of the latent space diffusion model, representing an acceleration of approximately 39 times. This verifies that the proposed solution possesses a significant low-latency advantage while maintaining generation accuracy.
[0141] It should be noted again that the implementation of the present invention does not depend on a specific model structure, data representation, or parameter configuration. In some optional implementation methods and variations:
[0142] (1) The teacher model is replaceable: the teacher generation model can be a diffusion generation model, a flow model or other conditional generation model. The teacher generation process can be a deterministic process or a process with randomness.
[0143] (2) Latent variable model can be replaced: the latent variable encoding and decoding model can be a variational autoencoder or other equivalent structure, and the encoder and decoder can adopt convolutional network, residual network or equivalent structure.
[0144] (3) Location conditions can be expanded: In addition to two-dimensional coordinates, location conditions can also include grid number, cell identifier, array configuration parameters and other auxiliary information. Location conditions can be directly input or first mapped to an embedded representation before input.
[0145] (4) Optional follow-up processing: Normalization, logarithmic transformation, thresholding and other follow-up processing of statistical channel fingerprints are optional steps, which can be selected and combined according to the numerical range, sparsity or robustness requirements of different applications.
[0146] (5) Euler steps are configurable: The number of Euler steps K can be configured from 2 to 10 according to the accuracy and latency requirements. Different models can be trained for different K or a unified model can be used to adapt multiple K.
[0147] The present invention also provides a computer system including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the steps of the channel fingerprint generation method for trajectory distillation discrete flow matching.
[0148] This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the channel fingerprint generation method for trajectory distillation discrete flow matching.
[0149] The program code used to implement the method of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the steps of the method of the present invention to be performed. The program code can be executed entirely on the machine, partially on the machine, partially on the machine and partially on a remote machine as a standalone software package, or entirely on a remote machine or server. All aspects not detailed in this invention are well-known to those skilled in the art.
[0150] The technical means disclosed in this invention are not limited to those disclosed in the above embodiments, but also include technical solutions composed of any combination of the above technical features. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications are also considered within the scope of protection of this invention.
Claims
1. A method for generating channel fingerprints using trajectory distillation discrete flow matching, characterized in that, Includes the following steps: Obtain the training dataset, which includes multiple locations and their corresponding statistical channel fingerprints; Construct and train a latent space coding and decoding model. After training, fix the encoder and decoder parameters, and use the encoder to convert the statistical channel fingerprints in the training set into a set of latent variables. Using a pre-trained, location-conditional teacher generation model, one or more sets of paired samples from the starting latent variable to the ending latent variable are generated for each location, forming a trajectory distillation paired dataset. Training a discrete flow matching generation model includes: setting the number of Euler steps in the inference phase; dividing the normalized time interval of the generation process into equal parts, with each part corresponding to one Euler integral update; constructing a discrete flow matching network, which is a conditional generation network used to obtain the velocity vector prediction value updated by the Euler integral for the corresponding Euler step based on the latent variable state, position conditions, and Euler step index; constructing training samples according to the Euler step index and supervising them; during each training iteration, sampling paired samples from the trajectory distillation paired dataset, randomly extracting the Euler step index, and constructing intermediate latent variable states between the starting latent variable and the ending latent variable based on the extracted Euler step index; using the intermediate latent variable state, position conditions, and Euler step index as network inputs; and using the target displacement vector obtained by the difference between the ending latent variable and the starting latent variable of the paired sample as a supervision signal to supervise the velocity vector prediction value output by the network. Given a target location, random noise is sampled as an initial latent variable. The Euler integral is updated sequentially by the set number of Euler steps. At each step, the velocity vector prediction value is obtained through a trained discrete flow matching network. The latent variable is updated according to the step size, and finally the destination latent variable is obtained. Input the endpoint latent variable into the decoder to obtain the statistical channel fingerprint.
2. The channel fingerprint generation method for trajectory distillation discrete flow matching according to claim 1, characterized in that, The construction of statistical channel fingerprints in the training dataset includes the following steps: Acquire channel observation data of a wireless communication system at different locations; The channel observation data is subjected to structured domain transformation to obtain a structured representation with sparse or clustered features; Statistical convergence is performed on the structured representation to obtain a channel fingerprint in statistical form that characterizes the location-related propagation properties.
3. The channel fingerprint generation method for trajectory distillation discrete flow matching according to claim 2, characterized in that, The construction steps of the statistical channel fingerprint also include: performing subsequent processing on the statistical channel fingerprint according to application requirements, wherein the subsequent processing includes one or more of the following: normalization, logarithmic transformation, truncation, thresholding, scaling or quantization.
4. The channel fingerprint generation method for trajectory distillation discrete flow matching according to claim 1, characterized in that, The latent space coding and decoding model described above is a variational autoencoder, in which the encoder maps the statistical channel fingerprint to low-dimensional latent variables, and the decoder reconstructs the statistical channel fingerprint from the latent variables; the dimension of the latent variables is set according to the fingerprint dimension and compression requirements.
5. The channel fingerprint generation method for trajectory distillation discrete flow matching according to claim 1, characterized in that, The teacher generation model is a diffusion-type generation model or a flow model, and its generation process is a deterministic process or a process with randomness. The generated paired samples include the starting latent variable, the ending latent variable, and the corresponding location conditions, and are cached offline as a trajectory distillation paired dataset for subsequent training.
6. The channel fingerprint generation method for trajectory distillation discrete flow matching according to claim 1, characterized in that, The number of Euler steps K is a preset positive integer, ranging from 2 to 10; the normalized time interval of the generation process is divided into K equal parts, each part corresponds to one Euler integral update, and the Euler step index k is specified to be 0 to K-1 in sequence, and supervision is performed on the same set of Euler step indices during the training phase.
7. The channel fingerprint generation method for trajectory distillation discrete flow matching according to claim 1, characterized in that, The discrete flow matching network adopts a conditionally enhanced multi-scale convolutional network structure, with a U-shaped backbone. Position conditional embedding and Euler step index embedding are injected into the multi-layer residual module, and the output end obtains the velocity vector prediction value through the output head.
8. The channel fingerprint generation method for trajectory distillation discrete flow matching according to claim 1, characterized in that, When constructing intermediate latent variable states between the starting latent variable and the ending latent variable using the extracted Euler step index, linear interpolation is performed between the starting latent variable and the ending latent variable according to the normalization progress corresponding to the Euler step index to obtain the intermediate latent variable states.
9. A computer system comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is executed by the processor, it implements the steps of the channel fingerprint generation method for trajectory distillation discrete flow matching according to any one of claims 1-8.
10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the channel fingerprint generation method for trajectory distillation discrete flow matching according to any one of claims 1-8.