A method and system for locating multimodal signals from unmanned aerial vehicles (UAVs) based on a world model.
By encoding the UAV's motion state and radio frequency spectrum into the latent space, and utilizing a time-series predictive network and a cross-entropy planner, the problems of high noise, error drift, and insufficient planning in UAV signal source tracing methods are solved, achieving high-success-rate autonomous source tracing control in complex environments.
Patent Information
- Application Number
- CN202611412828.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-11
- Publication Date
- 2026-11-10
AI Technical Summary
Existing UAV signal sourcing methods suffer from high observation noise and significant error drift under low-cost hardware conditions, making it difficult to stably guide UAVs close to targets. Furthermore, they lack explicit modeling of environmental dynamics and utilization of temporal information, resulting in unstable control processes and limited planning capabilities.
A multimodal signal source-finding method based on a world model is adopted. By uniformly encoding the UAV motion state, low-precision azimuth observation and radio frequency spectrum into the latent space, rolling inference is performed using a time series prediction network, and action sequences are screened by a cross-entropy planner, thus achieving autonomous source-finding control with a high success rate.
In complex electromagnetic environments, it is noise resistant and fault tolerant, has forward-looking planning capabilities, improves the robustness and success rate of UAV signal sourcing, and has low deployment costs.
Smart Images

Figure CN123067535A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the fields of unmanned aerial vehicle (UAV) autonomous navigation, radio signal processing, and artificial intelligence control technology, specifically relating to a UAV multimodal signal source tracing method and system based on a world model. Background Technology
[0002] In applications such as radio surveillance, emergency search and rescue, locating illegal emission sources, electromagnetic environment sensing, post-disaster communication restoration, and inspection of key areas, the use of drone platforms to quickly search for and approach target signal sources is of great significance. Compared to ground equipment, drones have advantages such as high maneuverability, good line-of-sight, and wide coverage, enabling them to perform flexible aerial source-finding missions in complex environments.
[0003] Existing UAV radio frequency sourcing methods mainly include the following categories: (1) Hardware-based direction finding methods: The angle of arrival information is obtained by array antenna, monopulse antenna or multi-channel phase difference measurement device, and then the direction or position of the signal source is estimated by combining geometric relationships. This type of method usually relies on high-cost radio frequency hardware, and is prone to problems such as large direction finding errors, angle jumps or even short-term failures under multipath, obstruction and low signal-to-noise ratio conditions.
[0004] (2) Signal strength gradient-based method: The UAV is guided to move towards the target area by receiving the trend of signal strength, signal-to-noise ratio or energy change. This method has low hardware requirements, but the local gradient is easily affected by noise, shadow fading, multipath reflection and UAV motion disturbance, which leads to unstable control strategy, or even getting stuck in local optima or oscillation.
[0005] (3) Methods based on end-to-end reinforcement learning or supervised learning: These methods directly map from observation to action. Although they can achieve good results in specific tasks, they often lack explicit modeling of environmental dynamics and have limited planning capabilities. When observation noise increases or sensors fail for a short time, the robustness of the strategy is insufficient.
[0006] The main problems currently are: when using a low-cost direction-finding module alone, the observation noise is large and the error drift is significant, making it difficult to stably guide the UAV to continuously approach the target; when using spectrograms or signal strength alone for control, it is difficult to fully utilize timing information and historical trajectories, and the control process is prone to jitter; most existing methods lack a unified framework that can simultaneously integrate multimodal sensor inputs, learn the evolution of the environment, and perform forward-looking action planning. For example, Chowdhury et al. proposed a typical reinforcement learning source-finding method based on received signal strength (RSS) in their paper MMU Chowdhury, F. Erden and I. Guvenc, "RSS-Based Q-Learning for Indoor UAV Navigation," MILCOM 2019 - 2019 IEEE Military Communications Conference (MILCOM), Norfolk, VA, USA, 2019, pp. 121-126. This method uses the signal strength value received by the UAV as the state input and reward basis for the Q-learning algorithm, guiding the UAV to autonomously navigate to the vicinity of the wireless signal source in an indoor environment where GPS is denied, without requiring any prior environmental information or GPS coordinates. However, this method has the following obvious technical defects: First, this method only uses the instantaneous RSS scalar as the state input, without utilizing the time-frequency structure information of the signal or integrating azimuth observation information, resulting in low state perception dimensionality. In complex electromagnetic environments or multipath interference scenarios, the RSS gradient is highly susceptible to shadow fading and noise contamination, and the control strategy is prone to jitter or even failure. Second, this method adopts a tabular Q-learning framework in a discrete action space, lacking the ability to model historical time-series information. Control decisions rely only on the current single-frame observation, making multi-step look-ahead planning impossible, and its robustness is insufficient when observations experience short-term failures or sudden changes. Third, this method does not construct any explicit predictive model of the dynamic evolution of the environment, and cannot pre-enact the future effects of candidate action sequences without executing real actions, thus limiting planning efficiency and task success rate. The above shortcomings fully demonstrate that single RSS modal information is insufficient to support high-performance UAV signal sourcing tasks under low-cost, high-noise conditions.
[0007] Therefore, there is an urgent need for a new method for UAV signal sourcing that, while maintaining low-cost hardware deployment capability, integrates coarse orientation information with high-dimensional radio frequency spectrum information to build a time-series prediction capability for the interaction process between the UAV and the signal source, and achieves robust and high-success-rate closed-loop autonomous sourcing control through planning algorithms. Summary of the Invention
[0008] To overcome the problems of unstable direction finding, insufficient spectrum utilization, and lack of temporal prediction and look-ahead planning capabilities in existing UAV signal source finding technologies, this invention provides a UAV multimodal signal source finding method and system based on a world model. This method and system encode the UAV's motion state, low-precision azimuth observation, and radio frequency spectrum into a unified latent space. It then uses a temporal prediction network based on action conditionation to continuously extrapolate future latent states and combines this with a cross-entropy planner to filter candidate action sequences, thereby achieving a high-success-rate autonomous approach to the target signal source. This method is particularly suitable for scenarios where low-cost direction finding sensors suffer from high noise, intermittent failures, or complex electromagnetic disturbances.
[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for locating multimodal signals from unmanned aerial vehicles (UAVs) based on a world model includes the following steps; S1. Construct a two-dimensional signal source-finding environment for UAVs and collect multimodal observation data; S2. The multimodal observation data is uniformly encoded to construct a hidden space state sequence representing the current environmental state; the velocity information, coarse orientation observation information and spectrum information are input into the multimodal encoder to obtain the hidden state representation at the current moment, and a historical hidden state sequence is constructed using a sliding time window. S3. Construct a world model predictor based on action conditionation; using the historical hidden state sequence and the random velocity action sequence sampled by the cross-entropy planner in a Gaussian distribution space with a specific mean and standard deviation as input, predict the hidden state at the next moment, and further realize multi-step rolling prediction. S4. Construct an auxiliary evaluation head to perform physical semantic estimation on the hidden state; estimate the target-related signal strength index from the hidden state using a signal-to-noise ratio estimation head, and estimate the target direction features at the next moment from the hidden state using a azimuth estimation head. S5. Action sequence planning based on world model and auxiliary evaluation head; S6. Execute the optimal action and perform closed-loop iterative control; the UAV executes the optimal action obtained from the planning at the current moment, acquires new multimodal observations, updates the historical hidden state sequence, and repeats steps S2 to S6 until the UAV enters the neighborhood of the target signal source or reaches the termination condition.
[0010] In step S1, the multimodal observation data includes at least: the current speed information of the UAV, the coarse azimuth observation information, and the spectrum information obtained by transforming the received IQ signal; The coarse azimuth observation information is provided by a low-cost direction finding module, which outputs the coarse angle of arrival information of the target signal, and the coarse angle of arrival information includes Gaussian noise, random disturbances and / or failure and loss phenomena. The spectrum information is obtained by short-time Fourier transform of the received IQ signal, and global decibel interval clipping and fixed range normalization are used to preserve the absolute energy difference information between different observation times.
[0011] S1 specifically includes the following: S1-1. Construct the UAV motion state vector. At time t, the UAV's true velocity is denoted as: in, and These represent the velocity components of the UAV along the x-axis and y-axis of the two-dimensional plane, respectively. After normalizing the velocity to eliminate the difference in velocity dimensions between different platforms, we obtain: in, The maximum speed allowed for the drone; S1-2. Obtaining rough azimuth observation information. Let the true geometric azimuth angle of the signal source relative to the UAV be: in, Indicates the location of the target signal source. Indicates the location of the drone; Considering the measurement errors, random jumps, and short-term failures inherent in low-cost direction-finding modules in practical applications, the coarse azimuth observation angle in this invention is expressed as: in, This represents the noise term for the coarse azimuth observation; at certain times, the coarse azimuth observation may randomly fail, at which point... Replaced by random angle perturbation or default value; The rough orientation observations are encoded into sine and cosine forms to avoid angle errors. The problem of jumps at π and π: S1-3. Construct the attitude observation vector by concatenating the normalized velocity vector with the coarse azimuth observation vector to form the attitude observation vector: This attitude observation vector contains both the current motion state of the UAV and directional guidance information from the low-precision direction finding module; S1-4. Acquire radio frequency signals and generate IQ observations. Assume the UAV's azimuth is... The azimuth of the target signal source is The distance between the two is: According to the law of signal propagation loss, the received signal-to-noise ratio at time t is expressed as: in, Indicates reference distance The signal-to-noise ratio at point n, where n is the path loss exponent; Based on the relative radial velocity between the UAV and the signal source, a Doppler frequency shift is introduced, and the radial velocity is set as... The Doppler frequency shift is then expressed as: in, Where c is the carrier frequency and c is the speed of light; Based on the above signal-to-noise ratio and Doppler frequency shift, the receiver generates the corresponding IQ sampling signal: in, For the target signal component, This is a noise component; S1-5. Observe the spectrum generated by time-frequency transformation. Perform a short-time Fourier transform on the IQ sampled signal to obtain the time-frequency spectrum matrix: Take its amplitude and convert it to the logarithmic decibel domain: in, To prevent extremely small positive numbers with a denominator of zero in logarithmic calculations; The spectrogram was then resampled to obtain a fixed-size two-dimensional spectrogram: S1-6. The spectrogram is normalized over a fixed interval. To preserve comparable absolute intensity differences between different times, this invention does not use frame-by-frame minimum-maximum normalization, but instead employs global decibel range clipping and normalization. The lower limit of the global clipping is set as... The upper limit is The normalized spectrum is then: Here, clip() represents the clipping operator; Therefore, the multimodal observations at time t can be written as: .
[0012] In S2, the multimodal encoder includes an attitude coding branch, a spectrum coding branch, and a fusion mapping branch; the attitude coding branch is used to encode the UAV's speed and coarse orientation observation, the spectrum coding branch is used to encode the two-dimensional spectrum map, and the fusion mapping branch is used to project the above two types of features into a unified latent space. S2 specifically includes the following: S2-1. Establish the attitude coding branch and convert the attitude observation vector... Inputting the pose encoding network yields pose feature representations: in, For attitude encoding function; S2-2. Establish a spectrum coding branch and normalize the spectrum graph. Inputting the spectrum into the spectrum coding network yields the spectrum feature representation: in, For spectral coding functions; S2-3. Perform multimodal feature fusion, concatenating pose features with spectral features: Then, by projecting the fusion mapping network onto the unified latent space, the current hidden state is obtained: in, D represents the dimension of the latent space; S2-4. Construct a historical hidden state window sequence. In order to characterize the temporal dependency in the UAV source-finding process, select the hidden states of the most recent L consecutive time steps to form a historical hidden state sequence: in, The historical hidden state sequence serves as the input to the world model predictor, used to infer the future state evolution under the influence of actions.
[0013] In step S3, the world model predictor employs a temporal network with action-conditional modulation. The input consists of a sequence of historical hidden states and the current candidate action, while the output is the hidden state at the next time step. Preferably, the output uses residual prediction, i.e., predicting the hidden state increment and superimposing it with the current hidden state to obtain the hidden state at the next time step. The world model predictor uses a Transformer structure with causal temporal modeling capabilities and achieves conditional modulation through action embedding.
[0014] S3 specifically includes the following: S3-1. Define the current action input. At time t, the candidate actions of the UAV are denoted as: in, As a normalized control vector, the range of values for each component can be limited to a preset action range; S3-2. Establish an action-conditional world model, which includes the historical hidden state sequence. and current candidate actions Input the world model predictor to obtain the predicted hidden state value for the next time step: Where F() represents the world model predictor; S3-3. The latent state increment is predicted using residuals. The world model predictor does not directly predict the complete latent state at the next time step, but rather predicts the change in the latent state: Therefore, we can conclude that: Using residual prediction helps improve training stability and makes the predictor closer to the identity mapping in the early stages of training. S3-4. Utilizing temporal networks to model historical dependencies, the predictor sequentially models the hidden states at each time step in the historical hidden state sequence and introduces action embeddings as conditional information. Let the action embeddings be: The predictor is then represented as: in, An action embedding function; S3-5. Construct a multi-step rolling prediction process. When it is necessary to extrapolate the future K-step states, let: Then, by incorporating it into a new history window, we get: And calculate recursively: By analogy, the future multi-step hidden state trajectory can be obtained: .
[0015] In S4, the auxiliary evaluation head includes a signal-to-noise ratio estimation head and a azimuth estimation head: The signal-to-noise ratio (SNR) estimation head is used to predict the SNR at the current or next moment from the hidden state; the orientation estimation head is used to predict the sine-cosine representation of the target's relative orientation from the hidden state; and the physical semantic estimation of the hidden state is performed by the auxiliary evaluation head to obtain the SNR and target orientation prediction results.
[0016] S4 specifically includes the following: S4-1. Establish a signal-to-noise ratio (SNR) estimation head. Input the current or future hidden state into the SNR estimation head to obtain the predicted SNR: in, This represents the signal-to-noise ratio estimation function; S4-2. Establish a target direction estimation head, input the hidden state into the direction estimation head, and obtain the sine and cosine representation of the target direction: Similarly, for predicting the hidden state in the future, we get: S4-3. Convert the direction prediction into a two-dimensional direction vector. To facilitate alignment calculations with the action direction, rearrange the predicted direction representation into a two-dimensional direction vector: S4-4. Construct an auxiliary supervision target. During the training phase, the signal-to-noise ratio and coarse orientation information from the actual next-time observations are used to supervise the auxiliary evaluation head; let the actual next-time signal-to-noise ratio be... The true direction is represented as: The corresponding monitoring losses can then be written as: Through the aforementioned auxiliary supervision, the hidden state becomes more physically interpretable and task-relevant.
[0017] In step S5, the cross-entropy method is used to randomly sample multiple sets of random velocity action sequences in the Gaussian distribution space with initialized mean and standard deviation as candidate action sequences. The corresponding future hidden state trajectory is obtained through multi-step latent space deduction. A comprehensive evaluation function is constructed based on the prediction signal-to-noise ratio gain, the degree of alignment between action and orientation, and the smoothness of action. The candidate action sequences are screened and optimized to select elite action sequences with high evaluation function scores. The mean and standard deviation of the elite action sequences are then used as the mean and standard deviation of the Gaussian distribution space for the next round of action sampling. This process is repeated for several rounds to gradually lock the optimal Gaussian distribution sampling space. Finally, the optimal action is obtained by sampling in the optimal space. The comprehensive evaluation function includes: the absolute value of the predicted signal-to-noise ratio, the increment of the predicted signal-to-noise ratio relative to the previous moment, the consistency between the action direction and the predicted target direction, and the smoothing penalty term for the changes in adjacent actions. An action sequence evaluation function is constructed based on the future hidden state trajectory. The cross-entropy planner is used to gradually converge to obtain the optimal action distribution through multiple rounds of sampling, elite sample selection, and mean and variance updates. The action at the first moment is then output as the control command.
[0018] S5 specifically includes the following: S5-1. Define a candidate action sequence. Let the planning time domain length be H. Then a set of candidate action sequences is represented as: Where i represents the i-th candidate sample; S5-2, Using a world model to extrapolate future hidden state trajectories; For each candidate action sequence, combine it with the historical hidden state sequence The future hidden state trajectory can be obtained by recursion: Furthermore, the corresponding prediction signal-to-noise ratio and prediction direction are obtained: S5-3. Construct an absolute return term for signal-to-noise ratio, and accumulate the predicted signal-to-noise ratio for each future step using a discount factor: in, Discount factor; S5-4. Constructing the signal-to-noise ratio increment benefit term: To encourage the UAV to continuously improve signal quality along the direction approaching the target, define the signal-to-noise ratio increment term: in, It can be estimated from the current hidden state; S5-5, Construct an alignment item between action and target direction, including candidate actions. Normalized to a unit direction vector: in, To prevent the minimum value where the denominator is zero, the action is... With the predicted target direction Consistency can be represented by the inner product as: When the two directions are in the same direction, this term takes a larger value; S5-6. Construct an action smoothing penalty term. To avoid severe jitter in the control sequence, define a penalty for the difference between adjacent actions: S5-7. Construct a comprehensive evaluation function, combining the above items to obtain the total score function for the candidate action sequence: in, These are the weighting coefficients for each evaluation item; S5-8. The cross-entropy method is used to optimize the action distribution. After initializing the mean and variance of the action sequence sampling distribution, multiple rounds of iterative sampling are performed. Each iteration includes: (1) Sample N candidate action sequences from the current Gaussian distribution; (2) Calculate the score for each action sequence. ; (3) Select the top E group elite samples with the highest scores; (4) Update the mean and variance based on the elite sample; Let the elite sample set of the k-th iteration be . Then the mean and variance are updated as follows: After multiple iterations, the optimal action sequence distribution is obtained; S5-9. Output the current control action, and take the action corresponding to the mean of the optimized action sequence at the first moment as the current action executed by the UAV: in, This represents the first term of the optimized mean action sequence.
[0019] In step S6, the system executes the optimal action and performs closed-loop updates until the signal source approach task is completed. This step is used to apply the planned control actions to the UAV platform and form a closed-loop autonomous source-finding control through environmental feedback.
[0020] S6 specifically includes the following: S6-1. Input the optimal action. The optimal action obtained from the action sequence evaluation function in step S5 is used as the input. The input is sent to the drone's motion control module to obtain the actual speed: And according to the time step Update drone location: S6-2. Acquire multimodal observations for the next time step. After executing the action, reacquire the observations for the next time step: And repeat the observation construction process in step S1; S6-3. Update the hidden state sequence, encoding the next time step observation as follows: And update the history hidden state window: S6-4. Determine the mission termination condition. Let the distance between the UAV and the target signal source be: The source tracing is considered successful when the following equation is satisfied: in, A preset success distance threshold is set. If the maximum number of steps or other preset termination conditions are reached, the current round of the task ends; otherwise, return to step S2 and continue to execute the next round of closed-loop planning and control. S6-5. A complete closed-loop source-finding process is formed. Through a cyclical process of "observation—encoding—prediction—evaluation—planning—execution—re-observation," this invention achieves UAV multimodal closed-loop signal source-finding based on a world model. Since each round of control makes decisions based on historical hidden states and future rolling projection results, it can better adapt to complex situations such as direction-finding noise, multipath disturbances, and short-term sensor failures compared to control methods that rely solely on current instantaneous observations.
[0021] A world model-based UAV multimodal signal source tracing system includes a multimodal observation construction module, a multimodal coding module, a world model prediction module, an auxiliary evaluation module, an action planning module, and a closed-loop control execution module. Multimodal observation construction module: Located at the front end of the system, it is used to collect the current speed information, coarse azimuth observation information and radio frequency IQ sampling signals of the UAV, and further generate a normalized spectrum diagram; this module is responsible for unifying the raw flight status information and radio frequency observation information into an input format that can be processed by the neural network.
[0022] Multimodal coding module: used to extract and fuse features from attitude vectors and spectrograms, and output a hidden state representation in a unified latent space; its function is to compress heterogeneous observations under different modalities into unified low-dimensional state variables, providing a state expression basis for subsequent dynamic prediction.
[0023] The world model prediction module receives historical hidden state sequences and candidate actions, and outputs the predicted hidden states for the next time step or multiple future time steps. This module is a key part of the entire system for achieving forward-looking planning, and is used to learn the potential dynamic patterns in the source-finding process. The auxiliary evaluation module includes a signal-to-noise ratio estimation head and a direction estimation head, which are used to recover interpretable task semantic indicators from the hidden state; its function is to provide the planner with the physical meaning quantification basis required for action evaluation.
[0024] Action planning module: Based on the cross-entropy method, it samples, scores, filters and iteratively optimizes multiple sets of candidate action sequences, and outputs the current optimal control action; this module is responsible for simulating and deciding the future action effect within the world model.
[0025] Closed-loop control execution module: This module sends the current optimal action output by the planner to the UAV platform, and after execution, re-collects observations, updates the status, and triggers the next round of planning; this module ensures that the system can achieve online autonomous source finding based on real-time feedback.
[0026] The beneficial effects of this invention are: (1) Achieving multimodal fusion source finding: This invention does not rely solely on a single sensing method, but rather uses a branched multimodal encoder to independently extract features from low-precision azimuth observations (encoded as sine and cosine attitude vectors) and high-dimensional radio frequency spectrograms (time-frequency features extracted by a convolutional network). These features are then projected onto a unified latent space via a fusion mapping network to obtain a joint latent state representation. Azimuth observations provide low-dimensional directional guidance information, while the spectrograms provide signal energy distribution and Doppler structure information. The two complement each other in the latent space: when azimuth observations are unreliable due to noise or failure, the signal energy change trend in the spectrograms can still provide effective constraints for source finding; conversely, when the spectrograms are blurred due to weak signals, azimuth observations can provide coarse-grained directional compensation, thereby fully leveraging the complementary advantages of the two types of information.
[0027] (2) It has the ability to predict the hidden space time series. This invention learns the dynamic relationship between UAV actions, environmental state and future observations through the world model. It can complete the future state deduction without directly reconstructing high-dimensional observations, thereby improving planning efficiency.
[0028] (3) It has noise resistance and fault tolerance capabilities. Since the low-cost direction finding module has large noise, jumps and failures, the present invention achieves noise resistance and fault tolerance in two aspects: First, the multimodal encoder constructs a historical hidden state window sequence from the hidden states of L consecutive time moments. This window naturally plays a time-series averaging effect on the azimuth observation jumps or loss at a single time moment, avoiding the direct transmission of single-frame noise to control decisions; Second, the world model predictor uses the historical hidden state sequence + action as input and predicts the hidden state of the next time moment in the form of residuals. Its prediction results do not completely depend on the azimuth observation input at the current time moment, but are inferred by comprehensively considering the historical evolution law. Thus, when the azimuth observation of the current frame experiences random failure or large jumps, it can still rely on the direction and intensity information accumulated in the historical sequence to give a reasonable state estimate, realize the implicit correction of coarse azimuth observation, and maintain high robustness in complex electromagnetic environments.
[0029] (4) Possessing forward-looking closed-loop planning capability, this invention utilizes the cross-entropy method (CEM) to iteratively optimize multi-step candidate action sequences in the latent space: In each round of planning, N groups of candidate action sequences of length H are sampled from a Gaussian distribution parameterized by mean and standard deviation. A world model is used to perform multi-step latent state rolling deduction for each sequence, and an auxiliary evaluation head is used to decode each latent state into predicted signal-to-noise ratio and target orientation. Based on this, a comprehensive evaluation function score is calculated, including absolute SNR gain, incremental SNR gain, alignment of action with target orientation, and action smoothing penalty. Subsequently, several elite samples with the highest scores are selected, the mean and variance of the Gaussian distribution are updated, and the next round of sampling begins. After several iterations, the sampling distribution converges to the high-score region. The first action of the final mean sequence is taken as the current execution instruction, and replanning is performed at the next moment, forming a rolling time-domain optimization closed loop. Compared to the greedy control method that relies solely on real-time observation, this method can more stably guide the UAV towards the signal source.
[0030] (5) The deployment cost is low and the adaptability is strong. This invention does not rely on high-cost large-scale array direction finding equipment. It can achieve high success rate source finding by using only low-cost coarse direction finding input and conventional radio frequency receiving front end. It is suitable for deployment on small UAV platforms. Attached Figure Description
[0031] Figure 1 This is a flowchart of a UAV multimodal signal source tracing method based on a world model according to the present invention.
[0032] Figure 2 This is a schematic diagram of the overall structure of the multimodal encoder, world model predictor, and planner of the present invention.
[0033] Figure 3 This is a schematic diagram illustrating the statistical results of success rate, signal-to-noise ratio gain, and final distance in the closed-loop evaluation of this invention. Detailed Implementation
[0034] The present invention will now be described in further detail with reference to the accompanying drawings.
[0035] like Figure 1 As shown, this invention discloses a multimodal signal source tracing method for unmanned aerial vehicles (UAVs) based on a world model. This embodiment uses a UAV autonomously approaching a single target radio frequency signal source in a two-dimensional planar scene as an example to verify the proposed method. The method mainly includes six key steps: multimodal observation construction, latent space state encoding, action-conditional world model prediction, auxiliary physical semantic evaluation, action planning based on cross-entropy, and closed-loop execution control. The specific implementation steps are as follows: S1. Construct a two-dimensional signal source-finding environment for UAVs and collect multimodal observation data; S2. Unify the encoding of multimodal observation data to construct a hidden space state sequence that represents the current environmental state; S3. Construct a world model predictor based on action conditionation, using historical hidden state sequences and candidate actions as inputs to predict the hidden state at the next moment, and further perform multi-step rolling prediction. S4. Construct an auxiliary evaluation head to perform physical semantic estimation on the hidden state and obtain the corresponding signal-to-noise ratio and target direction prediction results; S5. Based on the world model and auxiliary evaluation head, perform action sequence planning and select the optimal action at the current moment; S6. Execute the optimal action and perform closed-loop iterative control until the UAV enters the neighborhood of the target signal source.
[0036] The core idea of this invention is to avoid greedy control based solely on instantaneous direction-finding results or signal strength at a specific moment. Instead, it first maps the UAV's motion state, coarse azimuth observations, and radio frequency spectrum into a low-dimensional latent space. Then, it learns the evolutionary relationship of "historical state - action - future state" through a world model and performs forward-looking planning based on multi-step deduction results. Because this method explicitly utilizes historical time-series information and action influences, it can more effectively suppress noise, jumps, and intermittent failures caused by low-cost direction-finding modules compared to source-finding methods that rely solely on current observations.
[0037] Furthermore, in S1, a two-dimensional motion environment for the UAV is first constructed. The target signal source is assumed to be located at a fixed position in the two-dimensional plane, and the UAV moves according to control actions at discrete moments. At each moment, the UAV can acquire multimodal observations related to its current state, including: UAV velocity information, coarse azimuth observation information provided by a low-cost direction-finding module, and spectrogram information obtained by transforming the IQ sampled signal at the receiving end.
[0038] The purpose of this step is to acquire complementary observations that reflect both the "target direction trend" and the "target signal strength characteristics" at the lowest possible hardware cost. Specifically, coarse orientation observations provide low-dimensional direction cues but are susceptible to significant noise and random failures; the spectrum analysis, a high-dimensional radio frequency observation, can reflect the target signal energy and Doppler variation characteristics to some extent; and velocity information reflects the UAV's current motion inertia and control state. These three elements together form the foundational input for subsequent world model learning.
[0039] Specifically, step S1 includes the following: S1-1. Construct the UAV's velocity state. At time t, the UAV's true velocity on the two-dimensional plane is denoted as: in, These represent the velocity components of the UAV along the x-axis and y-axis, respectively. To eliminate the dimensional differences in maximum maneuverability between different platforms, they are represented by the maximum permissible speed. After normalization, we get: The normalized velocity vector is used to represent the current motion trend and control state.
[0040] S1-2. Obtain rough azimuth observation information. Let the true geometric azimuth angle of the target signal source relative to the UAV be: in, Indicates the location of the target signal source. This indicates the current location of the drone.
[0041] Considering the potential for measurement errors, random jumps, and short-term failures in low-cost direction-finding modules during practical applications, the coarse azimuth observation angle in this invention is expressed as follows: in, This represents the azimuth observation noise term. At certain times, the coarse azimuth observation may randomly fail; in such cases, it can be replaced by a random perturbation angle, a fixed default value, or an invalid flag.
[0042] To avoid abrupt changes in angle at the boundary between -π and π, the coarse azimuth observation is converted into a sine-cosine form: This representation is more suitable for neural network modeling and also facilitates subsequent similarity calculations with the direction of action.
[0043] S1-3, Constructing the attitude observation vector The normalized velocity vector is concatenated with the azimuth observation vector to form the attitude observation vector: This attitude observation vector contains both the current motion state and coarse orientation guidance information, providing a compact, low-dimensional input for the subsequent encoder.
[0044] S1-4. Acquire radio frequency signals and generate IQ observations. Furthermore, consider the distance between the UAV and the target signal source as follows: According to the path loss model, the received signal-to-noise ratio at time t can be expressed as: in, The signal-to-noise ratio at the reference distance is given, and n is the path loss exponent. For reference distance.
[0045] In a preferred embodiment, the Doppler frequency shift caused by the relative radial velocity between the UAV and the signal source is also considered. Let the radial velocity be... Then the Doppler frequency shift can be expressed as: ; in, denoted as carrier frequency, and c as the speed of light.
[0046] Accordingly, the receiving end generates the corresponding IQ sampling signal at time t: in, Represents the target signal component. This represents the noise component. The IQ sampled signal can be generated by a software-defined radio front-end, an RF receiver, or a simulation environment.
[0047] S1-5. Generate a spectrum observation: Perform a short-time Fourier transform (STFT) on the above IQ sampled signals to obtain the time-spectrum matrix: Taking its amplitude and mapping it to the logarithmic decibel domain, we get: in, To prevent extremely small positive numbers from occurring in logarithmic operations, the resulting time-spectrum matrix is then resampled or interpolated to obtain a fixed-size two-dimensional spectrum. This ensures that the dimensions of the input neural network are consistent at different time steps.
[0048] S1-6. The spectrogram is normalized within a fixed interval. Unlike frame-by-frame minimum-maximum normalization, this invention employs a global decibel interval clipping and normalization strategy to preserve comparable absolute intensity differences between different time steps. Let the lower limit of clipping be... The upper limit is The normalized spectrum is then represented as: Here, clip() represents the clipping operator. Thus, the complete multimodal observation at time t can be written as: Furthermore, in step S2, the multimodal observations obtained in step S1 are input into the multimodal encoder and mapped to hidden state representations in a unified latent space. Since attitude observations are low-dimensional vectors, while spectrograms are high-dimensional two-dimensional tensors, there are significant differences between the two in terms of data structure, statistical distribution, and information density. Therefore, this invention adopts a branched coding structure to extract features of the two modalities separately, and then unifies them into the same latent space through fusion mapping.
[0049] like Figure 2 As shown, the multimodal encoder of this invention mainly includes three parts: an attitude coding branch, a spectrum coding branch, and a fusion mapping branch. The attitude coding branch is used to encode speed and coarse orientation observations; the spectrum coding branch is used to encode a two-dimensional spectrogram; and the fusion mapping branch is responsible for splicing and fusing the two types of features and projecting them into the latent space to obtain the current state representation.
[0050] Specifically, step S2 includes the following: S2-1. Establishing the attitude coding branch attitude observation vector Inputting the pose encoding network yields pose feature representations: ; in, ( ) denotes the attitude coding function. In a preferred embodiment, the attitude coding network may employ a multilayer perceptron (MLP), comprising several linear layers, nonlinear activation layers, and normalization layers. Due to the low input dimension, this branch focuses on nonlinearly extracting the combination relationship between velocity and orientation information.
[0051] S2-2. Establish a spectrum coding branch and normalize the spectrum graph. Inputting the spectrum into the spectrum coding network yields the spectrum feature representation: ; in, ( ) represents the spectrum coding function. In a preferred embodiment, the spectrum coding network may adopt a convolutional neural network structure, including convolutional layers, pooling layers, activation layers and fully connected projection layers, to extract features such as energy distribution, local texture, spectral peak structure and Doppler shift in the spectrum.
[0052] S2-3. Perform multimodal feature fusion, concatenating pose features with spectral features: The hidden state at the current moment is obtained by projecting it onto the unified latent space through a fusion mapping network: in, D represents the latent space dimension. This represents the fusion mapping function. The hidden state is not directly equivalent to a specific physical quantity, but rather a jointly compressed representation of the "UAV motion state—directional cues—spectral observations" in the current environment. Through training, the hidden state can gradually learn key potential factors relevant to the mission.
[0053] S2-4. Construct a historical hidden state window sequence. In order to characterize the temporal dependencies in the source-finding process, select the hidden states of the most recent L consecutive time steps to form a historical hidden state sequence: ; in, .
[0054] This historical window sequence can characterize the recent trends in UAV status and observations, such as whether the signal continues to strengthen, whether the azimuth observation is stable, and whether the spectral structure has shifted. Compared to single-moment input, the historical hidden state sequence is more beneficial for world models to make robust predictions about the future.
[0055] Furthermore, in S3, a world model predictor based on action conditionation is constructed. The task of this module is to learn the mapping relationship from "historical hidden state sequence + current candidate action" to "next hidden state", and to recursively deduce the trajectory of future hidden states in multiple steps during the planning phase.
[0056] The world model in this invention does not directly predict in the original high-dimensional observation space, but rather performs dynamic modeling in the latent space. This significantly reduces the difficulty and computational complexity of prediction, while preserving the state evolution patterns strongly correlated with the control task. Preferably, the world model employs a Transformer structure with causal temporal modeling capabilities and achieves conditional modulation through action embedding.
[0057] Specifically, step S3 includes the following: S3-1. Define the current action input. At time t, the candidate actions of the UAV are denoted as: in, This represents a normalized control vector, whose components can be constrained within a preset action range. This action can represent the desired velocity increment, the desired planar movement direction, or a two-dimensional planar control command mapped by the controller.
[0058] S3-2. Establishing a world model of action-conditionalization Historical hidden state sequence With current candidate actions Input the world model predictor to obtain the predicted hidden state value for the next time step: Where F() represents the world model predictor.
[0059] S3-3. Using residuals to predict hidden state increments, to improve training stability, preferably, the world model does not directly output the complete hidden state for the next time step, but instead outputs the hidden state increment: Therefore, we get: This form of residual prediction makes it easier for the model to learn the patterns of local state changes, and is especially suitable for control tasks where state changes are relatively smooth between adjacent time steps.
[0060] S3-4. Using temporal networks to model historical dependencies, in a preferred embodiment, actions are first embedded to obtain: in, ( ) represents the action embedding function. Then, the action embedding is injected as conditional information into the temporal prediction network. The predictor can be represented as: in, ( ) represents a Transformer prediction network with action modulation. This network learns the potential evolution pattern of the observation sequence by sequentially modeling the hidden states at each time point in the historical window. At the same time, it introduces action embedding, enabling the model to perceive "how the future state will change if a certain action is performed".
[0061] S3-5. Constructing a multi-step rolling prediction process often requires extrapolating the future K-step states during planning. Let the prediction result for the first step be: ; Then it is incorporated into the new history window, resulting in: ; This rolling prediction mechanism enables the present invention to pre-imagine future state changes in the latent space without actually performing actions, thus providing a basis for subsequent action planning.
[0062] Furthermore, in S4, to enhance the physical interpretability and task relevance of the hidden states, an auxiliary evaluation head is constructed outside the world model. This auxiliary evaluation head is used to recover semantic quantities directly relevant to the task from the hidden states, including target-related signal strength indices and target orientation features.
[0063] Unlike relying solely on black-box hidden state scoring, this invention maps the hidden states to an interpretable physical semantic space using a signal-to-noise ratio (SNR) estimation head and a location estimation head. This allows for the introduction of additional supervision during training and also enables the direct construction of action scoring functions using these predictions during the planning phase.
[0064] Specifically, step S4 includes the following: S4-1. Establish a signal-to-noise ratio (SNR) estimation head. Input the current or future hidden state into the SNR estimation head to obtain the predicted SNR: Similarly, for predicting the hidden state at future moments, we can obtain: in, ( ) represents the signal-to-noise ratio estimation function. The estimation result can represent the signal quality level at the current moment, and can also be used to reflect whether future actions will bring the drone closer to the target signal source.
[0065] S4-2. Establish a target direction estimation head, input the hidden state into the direction estimation head, and obtain the sine and cosine representation of the target direction: For predicting the hidden state of the future, we have: S4-3. Convert to a two-dimensional direction vector. To facilitate alignment calculations with the action direction, the direction prediction result is used as a two-dimensional direction vector, i.e.: ; In practical applications, it can also be standardized: S4-4. Construct an auxiliary supervision target. During the training phase, the signal-to-noise ratio and coarse orientation information from the actual next-time observations can be used to supervise the auxiliary evaluation head. Let the actual next-time signal-to-noise ratio be... The true direction is represented as: The auxiliary loss can then be written as: By introducing the aforementioned auxiliary supervision, the hidden state not only serves the prediction of the next state, but is also encouraged to explicitly encode the "strength" and "direction" semantics directly related to the source-finding task, thereby improving the planning effect and robustness.
[0066] Furthermore, in S5, the cross-entropy method is used to sample and optimize the future H-step action sequence in the action space. For each candidate action sequence, a multi-step latent space deduction is first performed using a world model, then an auxiliary evaluation head is used to obtain the corresponding future signal-to-noise ratio and target direction prediction results, and finally the action sequence is scored based on the comprehensive evaluation function.
[0067] The essence of this step is to transform the traditional "execute then observe" control method into a "simulate within the model first, then select the optimal action" planning method. This can significantly improve the control stability of UAVs in complex noise environments and the success rate of approaching targets.
[0068] Specifically, step S5 includes the following: S5-1. Define the candidate action sequence. Let the planning time domain length be H. Then the i-th candidate action sequence is represented as: Where i represents the candidate sample number.
[0069] S5-2. Using a world model to extrapolate future hidden state trajectories, for each set of candidate action sequences, combine the current historical hidden state window. The future hidden state trajectory can be obtained by recursion: Furthermore, the corresponding predicted signal-to-noise ratio sequence and direction sequence are obtained through an auxiliary evaluation head: S5-3. Constructing the absolute benefit term for signal-to-noise ratio: To encourage drones to move closer to areas with high signal quality, a discounted weighted sum is applied to the predicted signal-to-noise ratio for each future step. in, This is the discount factor.
[0070] S5-4. Construct the signal-to-noise ratio increment benefit term. To further encourage continuous improvement in signal quality, the signal-to-noise ratio increment term is defined as follows: in, It can be estimated from the current hidden state.
[0071] S5-5. Construct an action-target direction alignment term, normalizing candidate actions into unit direction vectors: The consistency between the direction of the action and the predicted target direction can be represented by the inner product as follows: This value is larger when the direction of the action is the same as the direction of the target; otherwise, it decreases.
[0072] S5-6. Construct a motion smoothing penalty term. To prevent drastic jittering of control commands in adjacent time steps, the motion smoothing penalty term is defined as follows: .
[0073] S5-7. Construct a comprehensive evaluation function, combining the above items to obtain the total score function for the candidate action sequence: in, , , , The weighting coefficients for each evaluation item are used. In the preferred embodiment, additional scoring factors such as motion amplitude constraints, boundary penalties, and flight safety constraints can also be added according to task requirements.
[0074] S5-8. The action distribution is optimized using the cross-entropy method. After initializing the mean and variance of the action sequence sampling distribution, multiple rounds of iterative optimization are performed using the cross-entropy method. Each iteration includes the following process: (1) Sample N candidate action sequences from the current Gaussian distribution; (2) Calculate the score for each action sequence using a world model and an auxiliary evaluation head; (3) Select the top E group elite samples with the highest scores; (4) Update the mean and variance of the action distribution based on the elite sample.
[0075] Let the elite sample set in the k-th iteration be . Then its update can be expressed as: After several iterations, the action sampling distribution gradually converges towards the high-scoring region.
[0076] S5-9. Output the current control action. After optimization, take the first item in the average action sequence as the action to be executed by the UAV at the current moment: ,in, This represents the action corresponding to the optimized mean action sequence at the first time step. Only the first action is executed, and then the observation, encoding, and planning are repeated at the next time step. This approach constitutes a typical rolling time-domain optimization control (Model Predictive Control, MPC) closed-loop mechanism.
[0077] Furthermore, in S6, the current optimal action planned in step S5 is input into the UAV motion control module to drive the UAV to complete the position update; then, new multimodal observations are collected again, and steps S2 to S6 are repeated until the source finding success condition or the termination condition is met.
[0078] This step ensures that the entire system is not planned offline in a one-time process, but can continuously correct subsequent actions based on real feedback after execution, thus having better environmental adaptability and anti-interference ability.
[0079] Specifically, step S6 includes the following: S6-1. Execute the current optimal action. Input the data into the drone's motion control module to obtain the actual speed: in, ( ) indicates that the control execution mapping function is used. Then, based on the time step... Update drone location: S6-2. Acquire multimodal observations for the next time step. After executing the action, reacquire the observations for the next time step: The observation construction process in step S1 is repeated, including velocity update, coarse azimuth observation acquisition, IQ sampling generation, and spectrogram transformation.
[0080] S6-3. Update the hidden state sequence, encoding the next time step observation as follows: Where E() represents the multimodal encoder. Then update the history hidden state window: .
[0081] S6-4. Determine the mission termination condition. Let the distance between the UAV and the target signal source be: When the following conditions are met: When the source tracing is successful, the following is determined: A preset success distance threshold is set. In addition, if the current step count reaches the maximum number of iterations, the drone exceeds the boundary of the mission area, or other preset termination conditions are met, the current round of the mission ends; otherwise, return to step S2 and continue to execute the next round of "encoding-prediction-evaluation-planning-execution" closed-loop control.
[0082] S6-5. Forming a complete closed-loop source-finding process: Through the above-described cyclic process, this invention forms a closed-loop source-finding framework of "observation construction—latent space encoding—world model prediction—auxiliary semantic evaluation—action planning—execution control—re-observation update". Compared with control methods that rely solely on instantaneous direction-finding angles or instantaneous intensity changes, this invention can comprehensively utilize multimodal information and historical temporal dependencies, maintaining good control stability and a high source-finding success rate even in complex environments with high direction-finding noise, strong multipath interference, and significant spectral observation fluctuations.
[0083] like Figure 2 As shown, the UAV multimodal signal source tracing system based on the world model proposed in this invention mainly includes six parts in terms of structure: multimodal observation construction module, multimodal coding module, world model prediction module, auxiliary evaluation module, action planning module, and closed-loop control execution module.
[0084] Multimodal observation construction module: Located at the front end of the system, it is used to collect the current speed information, coarse azimuth observation information and radio frequency IQ sampling signals of the UAV, and further generate a normalized spectrum diagram; this module is responsible for unifying the raw flight status information and radio frequency observation information into an input format that can be processed by the neural network.
[0085] Multimodal coding module: used to extract and fuse features from attitude vectors and spectrograms, and output a hidden state representation in a unified latent space; its function is to compress heterogeneous observations under different modalities into unified low-dimensional state variables, providing a state expression basis for subsequent dynamic prediction.
[0086] The world model prediction module receives historical hidden state sequences and candidate actions, and outputs the predicted hidden states for the next time step or multiple future time steps. This module is a key part of the entire system for achieving forward-looking planning, and is used to learn the potential dynamic patterns in the source-finding process. The auxiliary evaluation module includes a signal-to-noise ratio estimation head and a direction estimation head, which are used to recover interpretable task semantic indicators from the hidden state; its function is to provide the planner with the physical meaning quantification basis required for action evaluation.
[0087] Action planning module: Based on the cross-entropy method, it samples, scores, filters and iteratively optimizes multiple sets of candidate action sequences, and outputs the current optimal control action; this module is responsible for simulating and deciding the future action effect within the world model.
[0088] Closed-loop control execution module: This module sends the current optimal action output by the planner to the UAV platform, and after execution, re-collects observations, updates the status, and triggers the next round of planning; this module ensures that the system can achieve online autonomous source finding based on real-time feedback.
[0089] To verify the effectiveness of the proposed method, the source-finding task was tested in a two-dimensional simulation environment, and the average success rate, average signal-to-noise ratio gain, and final approach distance of the UAV were statistically analyzed in multiple sets of experiments. The results are as follows: Figure 3 As shown.
[0090] In the experiment, the angle observation noise of the low-cost direction-finding module was set to Gaussian noise with a standard deviation of 5°–15°, and random jumps and a short-term failure probability of 5% were introduced with a probability of 10%–20% to simulate the typical error characteristics of low-precision direction-finding hardware in actual deployment scenarios. Simultaneously, the receiver generated corresponding IQ observations based on the path loss model and Doppler effect, and converted them into a spectrum input model via short-time Fourier transform. Statistical results under the above noise conditions show that the method of this invention exhibits significant advantages in the following aspects: (1) Under noise conditions with a large standard deviation of azimuth observation noise, the source finding success rate of the method of the present invention can still be maintained at 75%, indicating that the present invention has a stronger resistance to coarse direction finding noise. (2) In the event of a short-term failure of azimuth observation, due to the temporal smoothing and predictive capabilities of the historical hidden state sequence and the world model, the trajectory offset of the method of the present invention is controlled to within 10% of the target distance on average compared with the case without failure.
[0091] (3) In a complex electromagnetic environment with the combined disturbance of multipath fading and shadow fading, the average signal-to-noise ratio gain of the method of the present invention reaches about 21dB, indicating that the multi-mode fusion mechanism of the present invention can make full use of the spectrum information to compensate for the instability of low-precision azimuth information, thereby maintaining high overall robustness under the combined noise conditions.
[0092] In summary, this invention, through a unified framework of multimodal fusion, world model prediction, auxiliary physical semantic evaluation, and cross-entropy planning, achieves superior success rate, trajectory stability, and signal-to-noise ratio gain compared to the comparative methods under the aforementioned simulation conditions. This verifies the effectiveness of the method in enabling UAVs to robustly and autonomously approach target radio frequency signal sources under low-cost hardware conditions, and is applicable to various application scenarios such as radio monitoring, emergency communication support, investigation of illegal emission sources, and post-disaster search and rescue.
Claims
1. A method for locating multimodal signals from unmanned aerial vehicles (UAVs) based on a world model, characterized in that, Includes the following steps; S1. Construct a two-dimensional signal source-finding environment for the UAV and collect multimodal observation data, including the UAV's current speed information, coarse azimuth observation information, and spectrum information obtained by transforming the received IQ signal. Represent the multimodal observation data as a multimodal observation consisting of attitude observation vector and normalized spectrum. S2. The multimodal observations are uniformly encoded to construct a hidden space state sequence representing the current environmental state; the velocity information, coarse orientation observation information and spectrum information are input into the multimodal encoder to obtain the hidden state representation at the current moment, and a historical hidden state sequence is constructed using a sliding time window. S3. Construct a world model predictor based on action conditionation; using the historical hidden state sequence and the random velocity action sequence sampled by the cross-entropy planner in a Gaussian distribution space with a specific mean and standard deviation as input, predict the hidden state at the next moment and realize multi-step rolling prediction. S4. Construct an auxiliary evaluation head to perform physical semantic estimation on the hidden state; estimate the target-related signal strength index from the hidden state using a signal-to-noise ratio estimation head, and estimate the target direction features at the next moment from the hidden state using a azimuth estimation head. S5. Action sequence planning based on world model and auxiliary evaluation head; S6. Execute the optimal action and perform closed-loop iterative control; the UAV executes the optimal action obtained from the planning at the current moment, acquires new multimodal observations, updates the historical hidden state sequence, and repeats S2 to S6 until the UAV enters the neighborhood of the target signal source or reaches the termination condition.
2. The method for locating multimodal signals from unmanned aerial vehicles based on a world model according to claim 1, characterized in that, The S1 The coarse azimuth observation information is provided by a low-cost direction finding module, which outputs the coarse angle of arrival information of the target signal, and the coarse angle of arrival information includes Gaussian noise, random disturbances and / or failure and loss phenomena. The spectrum information is obtained by short-time Fourier transform of the received IQ signal, and global decibel interval clipping and fixed range normalization are used to preserve the absolute energy difference information between different observation times.
3. A method for locating multimodal signals from unmanned aerial vehicles (UAVs) based on a world model, as described in claim 1 or 2, characterized in that... S1 specifically includes the following: S1-1. Construct the UAV motion state vector. At time t, the UAV's true velocity is denoted as: in, and These represent the velocity components of the UAV along the x-axis and y-axis of the two-dimensional plane, respectively. After normalizing the velocity to eliminate the difference in velocity dimensions between different platforms, we obtain: in, The maximum speed allowed for the drone; S1-2. Let the true geometric azimuth angle of the signal source relative to the UAV be: in, Indicates the location of the target signal source. Indicates the location of the drone; The rough azimuth observation angle is expressed as: in, This represents the noise term for the coarse azimuth observation; at certain times, the coarse azimuth observation may randomly fail, at which point... Replaced by random angle perturbation or default value; The rough orientation observations are encoded into sine and cosine forms to avoid angle errors. The problem of jumps at π and π: S1-3. Construct the attitude observation vector by concatenating the normalized velocity vector with the coarse azimuth observation vector to form the attitude observation vector: This attitude observation vector contains both the current motion state of the UAV and directional guidance information from the low-precision direction finding module; S1-4. Acquire radio frequency signals and generate IQ observations. Assume the UAV's azimuth is... The azimuth of the target signal source is The distance between the two is: According to the law of signal propagation loss, the received signal-to-noise ratio at time t is expressed as: in, Indicates reference distance The signal-to-noise ratio at point n, where n is the path loss exponent; Based on the relative radial velocity between the UAV and the signal source, a Doppler frequency shift is introduced, and the radial velocity is set as... The Doppler frequency shift is then expressed as: in, Where c is the carrier frequency and c is the speed of light; Based on the above signal-to-noise ratio and Doppler frequency shift, the receiver generates the corresponding IQ sampling signal: in, For the target signal component, This is a noise component; S1-5. Observe the spectrum generated by time-frequency transformation. Perform a short-time Fourier transform on the IQ sampled signal to obtain the time-frequency spectrum matrix: Take its amplitude and convert it to the logarithmic decibel domain: in, To prevent extremely small positive numbers with a denominator of zero in logarithmic calculations; The spectrogram was then resampled to obtain a fixed-size two-dimensional spectrogram: S1-6. Normalize the spectrum over a fixed interval, and then perform global decibel range clipping and normalization. Let the lower limit of the global clipping be... The upper limit is The normalized spectrum is then: Here, clip() represents the clipping operator; Therefore, the multimodal observations at time t can be written as: 。 4. The method for locating multimodal signals from unmanned aerial vehicles based on a world model according to claim 3, characterized in that, In S2, the multimodal encoder includes an attitude coding branch, a spectrum coding branch, and a fusion mapping branch; the attitude coding branch is used to encode the UAV's speed and coarse orientation observation, the spectrum coding branch is used to encode the two-dimensional spectrum map, and the fusion mapping branch is used to project the above two types of features into a unified latent space. S2 specifically includes the following: S2-1. Establish the attitude coding branch and convert the attitude observation vector... Inputting the pose encoding network yields pose feature representations: in, For attitude encoding function; S2-2. Establish a spectrum coding branch and normalize the spectrum graph. Inputting the spectrum into the spectrum coding network yields the spectrum feature representation: in, For spectral coding functions; S2-3. Perform multimodal feature fusion, concatenating pose features with spectral features: Then, by projecting the hidden state onto the unified latent space through a fusion mapping network, the current hidden state is obtained: in, D represents the dimension of the latent space; S2-4. Construct a historical hidden state window sequence. In order to characterize the temporal dependency in the UAV source-finding process, select the hidden states of the most recent L consecutive time steps to form a historical hidden state sequence: in, The historical hidden state sequence serves as the input to the world model predictor, used to infer the future state evolution under the influence of actions.
5. The method for locating multimodal signals from unmanned aerial vehicles based on a world model according to claim 4, characterized in that, S3 specific Includes the following: S3-1. Define the current action input. At time t, the candidate actions of the UAV are denoted as: in, As a normalized control vector, the value range of each component is limited to a preset action range; S3-2. Establish an action-conditional world model, which includes the historical hidden state sequence. and current candidate actions Input the world model predictor to obtain the predicted hidden state value for the next time step: Where F() represents the world model predictor; S3-3. The latent state increment is predicted using residuals. The world model predictor does not directly predict the complete latent state at the next time step, but rather predicts the change in the latent state: Therefore, we can conclude that: Using residual prediction helps improve training stability and makes the predictor closer to the identity mapping in the early stages of training. S3-4. Utilizing temporal networks to model historical dependencies, the predictor sequentially models the hidden states at each time step in the historical hidden state sequence and introduces action embeddings as conditional information. Let the action embeddings be: The predictor is then represented as: in, An action embedding function; S3-5. Construct a multi-step rolling prediction process. When it is necessary to extrapolate the future K-step states, let: Then, by incorporating it into a new history window, we get: And calculate recursively: By analogy, the future multi-step hidden state trajectory can be obtained: 。 6. The method for locating multimodal signals from unmanned aerial vehicles based on a world model according to claim 5, characterized in that, In S4, the auxiliary evaluation head includes a signal-to-noise ratio estimation head and a azimuth estimation head: The signal-to-noise ratio (SNR) estimation head is used to predict the SNR at the current or next moment from the hidden state; the orientation estimation head is used to predict the sine-cosine representation of the target's relative orientation from the hidden state; and the physical semantic estimation of the hidden state is performed by the auxiliary evaluation head to obtain the SNR and target orientation prediction results. S4 specifically includes the following: S4-1. Establish a signal-to-noise ratio (SNR) estimation head. Input the current or future hidden state into the SNR estimation head to obtain the predicted SNR: in, This represents the signal-to-noise ratio estimation function; S4-2. Establish a target direction estimation head, input the hidden state into the direction estimation head, and obtain the sine and cosine representation of the target direction: Similarly, for predicting the hidden state in the future, we get: S4-3. Convert the direction prediction into a two-dimensional direction vector. To facilitate alignment calculations with the action direction, rearrange the predicted direction representation into a two-dimensional direction vector: S4-4. Construct an auxiliary supervision target. During the training phase, the signal-to-noise ratio and coarse orientation information from the actual next-time observations are used to supervise the auxiliary evaluation head; let the actual next-time signal-to-noise ratio be... The true direction is represented as: The corresponding monitoring losses can then be written as: Through the aforementioned auxiliary supervision, the hidden state becomes more physically interpretable and task-relevant.
7. The method for locating multimodal signals from unmanned aerial vehicles based on a world model according to claim 6, characterized in that, In step S5, the cross-entropy method is used to randomly sample multiple sets of random velocity action sequences in the Gaussian distribution space with initialized mean and standard deviation as candidate action sequences. The corresponding future hidden state trajectory is obtained through multi-step latent space deduction. A comprehensive evaluation function is constructed based on the prediction signal-to-noise ratio gain, the degree of alignment between action and orientation, and the smoothness of action. The candidate action sequences are screened and optimized to select elite action sequences with high evaluation function scores. The mean and standard deviation of the elite action sequences are then used as the mean and standard deviation of the Gaussian distribution space for the next round of action sampling. This process is repeated for several rounds to gradually lock the optimal Gaussian distribution sampling space. Finally, the optimal action is obtained by sampling in the optimal space. The comprehensive evaluation function includes: the absolute value of the predicted signal-to-noise ratio, the increment of the predicted signal-to-noise ratio relative to the previous moment, the consistency between the action direction and the predicted target direction, and the smoothing penalty term for the changes in adjacent actions. An action sequence evaluation function is constructed based on the future hidden state trajectory. The cross-entropy planner is used to gradually converge to obtain the optimal action distribution through multiple rounds of sampling, elite sample selection, and mean and variance updates. The action at the first moment is then output as the control command.
8. The method for locating multimodal signals from unmanned aerial vehicles based on a world model according to claim 7, characterized in that, S5 specifically includes the following: S5-1. Define a candidate action sequence. Let the planning time domain length be H. Then a set of candidate action sequences is represented as: Where i represents the i-th candidate sample; S5-2, Using a world model to extrapolate future hidden state trajectories; For each candidate action sequence, combine it with the historical hidden state sequence The future hidden state trajectory can be obtained by recursion: Furthermore, the corresponding prediction signal-to-noise ratio and prediction direction are obtained: S5-3. Construct an absolute return term for signal-to-noise ratio, and accumulate the predicted signal-to-noise ratio for each future step using a discount factor: in, Discount factor; S5-4. Constructing the signal-to-noise ratio increment benefit term: To encourage the UAV to continuously improve signal quality along the direction approaching the target, define the signal-to-noise ratio increment term: in, It can be estimated from the current hidden state; S5-5, Construct an alignment item between action and target direction, including candidate actions. Normalized to a unit direction vector: in, To prevent the minimum value where the denominator is zero, the action is... With the predicted target direction Consistency can be represented by the inner product as: When the two directions are in the same direction, this term takes a larger value; S5-6. Construct a motion smoothing penalty term and define the difference penalty between adjacent actions: S5-7. Construct a comprehensive evaluation function, combining the above items to obtain the total score function for the candidate action sequence: in, These are the weighting coefficients for each evaluation item; S5-8. The cross-entropy method is used to optimize the action distribution. After initializing the mean and variance of the action sequence sampling distribution, multiple rounds of iterative sampling are performed. Each iteration includes: (1) Sample N candidate action sequences from the current Gaussian distribution; (2) Calculate the score for each action sequence. ; (3) Select the top E group elite samples with the highest scores; (4) Update the mean and variance based on the elite sample; Let the elite sample set of the k-th iteration be . Then the mean and variance are updated as follows: After multiple iterations, the optimal action sequence distribution is obtained; S5-9. Output the current control action, and take the action corresponding to the mean of the optimized action sequence at the first moment as the current action executed by the UAV: in, This represents the first term of the optimized mean action sequence.
9. The method for locating multimodal signals from unmanned aerial vehicles based on a world model according to claim 8, characterized in that, S6 specifically includes the following: S6-1. Input the optimal action. The optimal action obtained from the action sequence evaluation function in step S5 is used as the input. The input is sent to the drone's motion control module to obtain the actual speed: And according to the time step Update drone location: S6-2. Acquire multimodal observations for the next time step. After executing the action, reacquire the observations for the next time step: And repeat the observation construction process in step S1; S6-3. Update the hidden state sequence, encoding the next time step observation as follows: And update the history hidden state window: S6-4. Determine the mission termination condition. Let the distance between the UAV and the target signal source be: The source tracing is considered successful when the following equation is satisfied: in, A preset success distance threshold is set. If the maximum number of steps or other preset termination conditions are reached, the current round of the task ends; otherwise, return to step S2 and continue to execute the next round of closed-loop planning and control. S6-5, Form a complete closed-loop source tracing process.
10. A world-model-based UAV multimodal signal sourcing system for implementing the method of any one of claims 1-9, characterized in that, It includes a multimodal observation construction module, a multimodal coding module, a world model prediction module, an auxiliary evaluation module, an action planning module, and a closed-loop control execution module; Multimodal observation construction module: Located at the front end of the system, it is used to collect the current speed information, coarse azimuth observation information and radio frequency IQ sampling signals of the UAV, and further generate a normalized spectrum diagram; this module is responsible for unifying the raw flight status information and radio frequency observation information into an input format that can be processed by the neural network; Multimodal coding module: used to extract and fuse features from pose vectors and spectrograms, and output a unified latent state representation in the latent space; World model prediction module: used to receive historical hidden state sequences and candidate actions, and output the hidden state prediction results for the next time step or multiple future time steps; The auxiliary evaluation module includes a signal-to-noise ratio estimation head and a direction estimation head, which are used to recover interpretable task semantic metrics from the hidden state. Action planning module: Based on the cross-entropy method, it samples, scores, filters and iteratively optimizes multiple sets of candidate action sequences, and outputs the current optimal control action; Closed-loop control execution module: Used to send the current optimal action output by the planner to the UAV platform, and after execution, re-collect observations, update the status, and trigger the next round of planning.
