Fall trajectory prediction method based on remote sensing and satellite communication positioning

Abstract

The present application relates to the technical field of computer system based on specific calculation model, disclose a kind of fall trajectory prediction method based on remote sensing and satellite communication positioning, comprising: constructing and parallel evolution one whole path hypothesis library;When receiving the delayed observation data carrying historical timestamp, extract historical state and asynchronously update the global belief weight of trajectory hypothesis;After updating, calculate the belief validity index to diagnose model health degree;When judging that belief failure occurs, discard all invalid hypotheses and reconstruct the hypothesis library based on delayed observation data, the present application separates state evolution and belief update, realizes the lossless fusion to sparse and out-of-order observation data, by introducing belief validity check and hypothesis library reconstruction mechanism, make the calculation model have self-diagnosis and fault recovery ability, solve the problem of systematic belief failure caused by physical prior model failure.

Description

Technical Field

[0001] This invention relates to a method for predicting water-fall trajectory based on remote sensing and satellite communication positioning, belonging to the technical field of computer system technology based on a specific computational model. Background Technology

[0002] Currently, probabilistic reasoning methods based on specific computational models, such as particle filtering or Bayesian networks, are commonly used to fuse prior physical models with real-time observation data to achieve continuous tracking and prediction of target states. However, existing search and rescue systems, when integrating different technical modules, also reveal limitations in probabilistic reasoning. For example, Chinese invention patent CN111538059B discloses an adaptive fast dynamic positioning system and method based on an improved Boltzmann machine. This method involves obtaining the initial location information of a person who has fallen into the water through the BeiDou positioning system and running a maritime person-in-the-water location prediction model to predict their position. The drift trajectory is then used to plan the UAV path using an improved Boltzmann machine. This type of scheme is essentially a predictive sequential task flow, focusing on planning the UAV path in the back end rather than estimating the probabilistic state in the front end. The drift model is used to extrapolate after receiving the initial position. However, this scheme does not disclose a mechanism to handle sparse or out-of-order observation data that may be received after trajectory prediction and during UAV execution, which carries historical timestamps. This design makes it lack the ability to fuse subsequent observation data. Once the initial prediction is biased or misses key delay calibration information, it is difficult to correct and is not suitable for complex real-world data environments.

[0003] However, in specific applications such as maritime emergency response, the data streams on which the above-mentioned computational models rely differ from the ideal real-time and continuous assumptions. The data streams typically exhibit two characteristics: first, the time intervals between observation points are extremely large, i.e., the data is extremely sparse; second, there is an unpredictable delay between the time of receiving the observation data and the time of its generation, i.e., the data stream is out of order. When such sparse and out-of-order data streams are input into existing sequential processing computational models, they will be unable to operate effectively in principle.

[0004] Therefore, the technical problem to be solved by this invention is how to provide a computational model processing method that avoids the irreversible loss of information caused by real-time resampling and enables the probabilistic inference model to effectively process sparse and out-of-order data. Summary of the Invention

[0005] This invention provides a method for predicting water-fall trajectory based on remote sensing and satellite communication positioning. Its main purpose is to solve the problems of existing probabilistic inference models, which rely on real-time resampling and sequential processing mechanisms, and have problems such as zero confidence of calculation assumptions and inability to utilize delayed data when dealing with sparse and disordered observation data.

[0006] To achieve the above objectives, this invention provides a method for predicting water-fall trajectory based on remote sensing and satellite communication positioning, comprising the following steps:

[0007] Step 101: Construct and maintain a full path hypothesis library containing N independent trajectory hypotheses, where N is the preset total number of trajectory hypotheses, and each trajectory hypothesis has its own historical state sequence and global belief weight.

[0008] Step 102: Based on the physical prior model and the stochastic process model, drive the parallel evolution of all N independent trajectory hypotheses in the full path hypothesis library, and record their complete historical state sequences after evolution.

[0009] Step 103, Receive the historical timestamp at the current moment. When extracting delayed observation data, each independent trajectory hypothesis from the full path hypothesis library is extracted. The historical states corresponding to the historical states in the historical state sequence at each time point are used to calculate the likelihood between the delayed observation data and the historical states, and the global belief weights of each independent trajectory hypothesis are updated using Bayesian methods based on the likelihood.

[0010] Step 104: After step 103, calculate a belief effectiveness index based on the updated global belief weights assumed by N independent trajectories.

[0011] Step 105: When the belief validity index reaches the preset model failure threshold, it is determined that the model has failed. Then, all N independent trajectory hypotheses in the full path hypothesis library are discarded, and N new trajectory hypotheses are re-initialized and generated based on the delayed observation data to construct the reconstructed full path hypothesis library.

[0012] Preferably, the step of calculating the belief effectiveness index in step 104 specifically includes: calculating the information entropy of the global belief weights or the effective particle number of the global belief weights based on the updated N independent trajectory assumptions, and the preset model failure threshold is a preset information entropy threshold or a preset effective particle number threshold.

[0013] Preferably, the effective number of particles is calculated using the following formula: ,in The effective number of particles, The total number of preset trajectory assumptions. For the index of the trajectory hypothesis, For the first The global belief weights are normalized based on the assumption of independent trajectories.

[0014] Preferably, the method further includes a delayed resampling step: when it is determined in step 104 that the belief validity index has not reached the preset model failure threshold, the delayed resampling step is performed after step 103; the delayed resampling step resamples the entire path hypothesis library based on the updated global belief weights of N independent trajectory hypotheses to generate N new trajectory hypotheses to replace the entire path hypothesis library.

[0015] Preferably, the physical prior model in step 102 includes: a regional ocean model system for providing background ocean current field data, a third-generation wave model for providing wind-generated wave field, and a target drift model for calculating the drift of the target under the combined action of background ocean current field data and wind-generated wave field.

[0016] Preferably, the delayed observation data in step 103 includes: data acquired via a satellite communication positioning module, characterizing the target's position at historical timestamps. The step of calculating the likelihood in step 103 includes: calculating the likelihood based on the Euclidean distance between the location data of the point base position and the historical state.

[0017] Preferably, the delayed observation data further includes: acquiring field-based remote sensing data characterizing the distribution area of ​​potential targets through remote sensing images; the step of calculating the likelihood in step 103 further includes: when the delayed observation data is field-based remote sensing data, determining that each independent trajectory is assumed to be at a historical timestamp. Whether the historical state falls within the region represented by the field-based remote sensing data is used as the basis for calculating the field-based likelihood, which is then used to update the global belief weights.

[0018] Preferably, the step of updating the global belief weights in step 103 includes: multiplying the likelihood calculated based on delayed observation data and historical states with the existing global belief weights of the independent trajectory hypothesis before step 103 to obtain the updated global belief weights; and normalizing the updated global belief weights of the N independent trajectory hypotheses.

[0019] Preferably, step 105, which involves re-initializing and generating N new trajectory hypotheses based on delayed observation data, includes: taking the location of the delayed observation data as the center and combining it with the preset observation uncertainty of the delayed observation data, generating N new historical states through resampling; using the N new historical states as the initial states of the N new trajectory hypotheses and resetting their global belief weights.

[0020] Preferably, the method further includes a trajectory prediction output step: when the belief validity index is determined not to have reached the preset model failure threshold in step 104, the trajectory prediction output step is executed; the trajectory prediction output step includes: based on the updated global belief weights of N independent trajectory hypotheses and the current states of the N independent trajectory hypotheses, performing a weighted average or extracting the current state of the trajectory hypothese with the largest global belief weight to generate the final trajectory prediction result.

[0021] Compared with the prior art, the beneficial effects of the present invention are:

[0022] 1. By constructing and maintaining a full-path hypothesis library containing complete historical state sequences, the state evolution process and belief update process of the probabilistic inference model are separated. When delayed observation data carrying arbitrary historical timestamps are received, the method does not need to backtrack or recompile the calculation. Instead, it directly extracts the historical state corresponding to the historical timestamp of each trajectory hypothesis in the hypothesis library and updates its global belief weight based on this. This mechanism enables the probabilistic inference model to perform lossless fusion of asynchronous and out-of-order observation data, avoiding the irreversible loss of historical possibilities caused by the resampling step in the existing sequential processing model.

[0023] 2. By preserving the complete path of all trajectory assumptions in the parallel evolution step and performing delayed resampling only after asynchronous belief updates, the mechanism changes the way the probabilistic inference model operates under sparse observation data. This mechanism allows the particle cloud to fully diverge during the long prediction phase between two sparse observation points to encapsulate the uncertainty of the physical model, rather than prematurely discarding low-probability paths through resampling. When new observation data arrives, the system uses this data to update the beliefs of the entire path, ensuring that computational resources can eventually converge to credible assumptions verified by observation data, thus avoiding the weight degradation problem that the standard model is prone to cause due to scale mismatch under sparse data.

[0024] 3. Based on the above mechanism, by introducing a belief validity verification step after asynchronous belief update, the global belief weight vector generated by the computational model is reused as a meta-information source to diagnose whether the physical prior model on which the probabilistic inference model depends has failed. When it is determined that belief failure has occurred (i.e. all hypotheses are seriously in conflict with the observation data), the hypothesis library reconstruction step provides a risk-avoidance reconstruction mechanism, that is, discarding all current failed hypotheses and forcibly resynchronizing based on the latest observation data. This collaborative mechanism enables the probabilistic inference model to have self-diagnosis and fault recovery capabilities, transforming it from an executor of blindly following the physical model into a computational system with critical reflection capabilities, solving the problem of the model encountering unforeseen sudden operating conditions. Attached Figure Description

[0025] Figure 1 This is a flowchart of the prediction method for belief verification and reconstruction of the present invention;

[0026] Figure 2 This is a comparison chart of RMSE between the experimental group and the control group under the model failure scenario of this invention;

[0027] Figure 3 This is a closed-loop state flow diagram of the method of the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be described in detail below. However, it should be understood that the embodiments described in this specification are only for explaining the present invention and are not intended to limit the scope of protection of the present invention.

[0029] This invention provides a method for predicting water-fall trajectory based on remote sensing and satellite communication positioning. In step 101, a method is constructed and maintained that includes... The entire path hypothesis library is established based on the assumption of independent trajectories. This step is the initialization phase of the probabilistic inference model. The total number of preset trajectory assumptions, whose values ​​are: For example, it is necessary to balance the prediction resolution required by the computational model with the computing resources available to the system; the range of values ​​can be set as follows: to The data structure of the full-path hypothesis library is used to store... Each trajectory hypothesis has its own independent historical state sequence and global belief weights. The historical state sequence is used to losslessly record the trajectory hypothesis from its initial time. Up to the current moment The complete path, data nodes can include status information such as timestamps, latitude and longitude locations, and velocity components, global belief weights. ( The global belief weights of all trajectory hypotheses are the global confidence metric assigned to each trajectory hypothesis by the model. All were set to In step 102, the system drives all hypotheses in the full-path hypothesis library based on the physical prior model and the stochastic process model. Each independent trajectory hypothesis evolves in parallel, recording its complete historical state sequence after evolution. This step is the prediction phase of the model. No resampling is performed during the evolution process. The physical prior model may include the Regional Ocean Model System (ROMS) for providing background ocean current field data, the Third Generation Wave Model (SWAN) for providing wind-generated wave field data, and the target drift model for calculating the drift of the target under the combined action of the above fields. The stochastic process model, within the framework, is used to characterize the microscopic stochastic drift or model uncertainty not covered by the physical prior model. During parallel evolution, each trajectory hypothesis independently calculates its state at the next moment based on the physical prior model and the stochastic process model. The system appends this new state to its respective historical state sequence.

[0030] In step 103, when at the current time Received with historical timestamp When the delayed observation data (where The model initiates an asynchronous belief update. This step is the model update phase, and its core lies in separating state evolution from belief updates to process out-of-order data. This step extracts each (the first) hypothesis from the full path hypothesis library. strip) The assumption of independent trajectories The historical state corresponding to the historical state sequence at time 1 Then, calculate the delayed observation data. Historical state of this trajectory likelihood Finally, based on likelihood For the first Global belief weights of the independent trajectory hypothesis A Bayesian update is performed; the likelihood calculation method depends on the data type of the delayed observation data. When the delayed observation data is point base location data obtained through a satellite communication positioning module, the likelihood... The calculation can be based on the point base location. With historical status The Euclidean or Gaussian distance is used; the closer the distance, the higher the likelihood. The higher the likelihood, the more the model preprocesses the field-based remote sensing data, which is acquired through remote sensing images and characterizes the distribution area of ​​potential targets, converting it into a spatial likelihood field. At this point, the likelihood is... The calculation is transformed into a field-point likelihood relation, that is, to determine the historical state. Does it fall within the spatial likelihood field? Within the represented region, the field basis likelihood is calculated based on this; the Bayesian update of the global belief weights can specifically be: the likelihood calculated based on delayed observation data and historical states is then applied. , and this article (Article) strip) The independent trajectory assumption pre-existing global belief weights before step 103 Multiply the results to obtain the updated global belief weights. The model is applicable to all Assuming the trajectory has completed this multiplication update, it also... The updated global belief weights based on the assumption of independent trajectories Normalization is performed to make the sum of all weights equal to In step 104, immediately after the asynchronous belief update in step 103, the model performs a self-diagnosis, i.e., based on... The updated global belief weight vector based on the assumption of independent trajectories The calculation of the belief validity index involves reusing the model's belief weight distribution as meta-information to diagnose whether the physical prior model upon which the model relies has failed. Specifically, the calculation of the belief validity index may include: calculating the information entropy of the global belief weight vector or calculating the effective particle count of the global belief weights. ; number of effective particles The standardized statistical tools available are used to calculate this using the following formula: ,in, The effective number of particles, The total number of preset trajectory assumptions. For the index of the trajectory hypothesis, For the first The global belief weights are normalized based on the assumption of independent trajectories.

[0031] In step 105, the model performs decision-making and recovery actions upon belief failure. The system compares the belief validity index calculated in step 104 with a preset model failure threshold, which can correspond to a preset information entropy threshold. or a preset effective particle number threshold When the model judges the validity index of beliefs (in terms of...) For example, touching (e.g.) When a preset model failure threshold is reached, the model determines that belief failure has occurred, meaning that the physical prior model upon which the model relies has failed. Once belief failure is determined, the model immediately triggers a hypothesis base reconstruction: discarding all hypotheses in the entire hypothesis base. The model assumes independent trajectories; then, it re-initializes and generates data based on the delayed observations that have just arrived and caused the belief to fail. A new trajectory hypothesis is proposed to construct a reconstructed full-path hypothesis library. This re-initialization step can be specifically as follows: using the location of the delayed observation data as the center and combining the pre-defined observation uncertainties of the delayed observation data, a resampling method is used to generate... A new historical state; will A new historical state as The initial state of a completely new trajectory hypothesis is determined, and its global belief weights are reset (reset to...). The model resumes the parallel evolution of step 102 from this hypothesis base, which is resynchronized with the latest evidence. Furthermore, the method of this invention also includes a mode where the model operates under the conventional path of valid beliefs. When, in step 104, the model determines that the belief validity index has not reached a preset model failure threshold, the model determines that the system is healthy and then executes the conventional delayed resampling step or trajectory prediction output step. The trajectory prediction output step can be initiated by the user at any time. Trigger, model based The updated global belief weights based on the assumption of independent trajectories ,and Each independent trajectory assumes its own current state. A weighted average is performed to generate the final trajectory prediction result representing the probability density distribution; a delayed resampling step is performed by the model after step 103 (asynchronous belief update) and when the belief is determined to be valid. This step is based on... The updated global belief weights based on the assumption of independent trajectories Resampling of the full path hypothesis base to generate A new trajectory hypothesis is used to replace the full path hypothesis library.

[0032] Example 1: In a maritime emergency search and rescue scenario, the search and rescue coordination system at the current moment... Faced with the following data conditions: the system in At that moment, satellite positioning beacon data of the target that had fallen into the water was received for the first time, but the historical timestamp of this data was missing. This indicates that it originated from Previous Hours; thereafter, the system remained in operation for a long time. No new data was received in the hourly forecast until At the hour mark, the system received a second set of observation data, but this data had a different historical timestamp. for This is a delayed and out-of-order data set, and the system uses the method of this invention to... Received at any time carrying historical timestamps When the data is received, the computational model immediately executes step 101, based on the data. The location of the data and the pre-defined observation uncertainties are used to construct a system that includes... A complete path hypothesis library for each independent trajectory assumption; subsequently, the computational model executes step 102, driving this... The trajectory assumption, based on a prior physical model, in this case the ROMS current and SWAN wave model, and a stochastic process model, is derived from... Parallel evolution begins at time 1 At each hourly interval, and by fully recording the historical state sequence of each trajectory, the computational model does not perform resampling during this evolutionary phase, allowing... The probability cloud of the trajectory hypothesis in to Extended over the hours.

[0033] exist At the hour, when the computational model receives the second data carrying the historical timestamp... When dealing with delayed observation data, the model does not perform any state backtracking or model replay, but directly initiates the asynchronous belief update in step 103; the model then traverses all hypotheses in the full path hypothesis library. The trajectory assumption, for each trajectory Extracting its state from its recorded historical state sequence The historical state corresponding to a given moment The computational model calculates this. Compared with the second set of observation data Likelihood between According to the Bayesian update rule, the likelihood is... With this trajectory Existing global belief weights Multiply the results to obtain the updated global belief weights; after this asynchronous belief update step is completed, those... The historical state of a moment Compared with the second set of delayed observation data The likelihood of a trajectory that deviates significantly from the assumed trajectory. Approaching zero, resulting in its updated global belief weight. It also approaches zero; while the global belief weights of those trajectory assumptions whose historical paths match this delayed data... The model then receives an improvement; the computational model then executes step 104 to calculate the updated... Effective particle number of global belief weights for trajectory hypothesis Assuming the calculation result This value is greater than the preset model failure threshold. The computational model determines that the belief is valid and the physical prior model is not invalid. Therefore, the computational model does not execute the reconstruction step 105, but instead executes the trajectory prediction output step. The system is based on this... The updated global belief weights based on the trajectory assumption and each at the current moment Hourly state The weighted average is then calculated; the final output of the calculation model to the search and rescue coordinator is a probability density distribution map calibrated with historical evidence, which provides data support for subsequent search and rescue decisions.

[0034] Example 2: This example verifies the collaborative effectiveness of the belief validity verification in step 104 and the hypothesis base reconstruction in step 105 of the method of the present invention on a probabilistic reasoning simulation test platform. The experiment sets up a control group and an experimental group, both used to simulate trajectory prediction of a target falling into the water at sea. The control group uses a standard particle filter calculation model, which performs a conventional real-time resampling step when the particle weight decreases, but does not include the belief failure judgment in step 104 and the hypothesis base reconstruction mechanism in step 105. The experimental group uses the complete method as described in the specific implementation, including all logic from steps 101 to 105. Both calculation models run in the same simulation environment, with the initial state defined as follows: The total number of preset trajectory assumptions is set to [value]. The preset model failure threshold used for the experimental group, i.e., the effective particle number threshold. , set as The simulation platform generates realistic trajectories, which in turn generate a set of sparse and disordered observation data. These observation data are then overlaid. The standard deviation of Gaussian noise in the ocean; in the simulation environment, it is set at... Hours later, the physical prior models invoked by both computational models failed, exhibiting a failure to predict strong local vortices appearing in the actual trajectory. This caused the drift path provided by the physical prior model to begin to deviate systematically from the actual trajectory. to During this period, both computational models incorporated observational data, and the root mean square error (RMSE) between the predicted and actual trajectories remained within a certain range. Below nautical miles, in After time step, due to the failure of the physical prior model, the two computational models... The trajectory is assumed to deviate from the true trajectory under the drive of step 102.

[0035] exist At the hour, the observation data carrying that time timestamp arrives. At this point, the location of the observation data corresponds to almost all of the data in the two computational models. The current states of all trajectory hypotheses are quite far apart; for the control group, the model calculates the likelihood of all trajectory hypotheses and finds that all weights All approach zero. The value dropped sharply to The control group model then triggers its real-time resampling step, but because it resamples from a completely invalid hypothesis library, the newly generated... The trajectory assumptions still cluster on incorrect physical prior model paths, resulting in a belief failure that is irreversible; for the experimental group, the model also calculates the likelihood of all trajectory assumptions. All values ​​approach zero, and their values ​​are calculated in step 104. The value also dropped to The model will immediately Value and preset model failure threshold Comparison; due to The model determines that a threshold has been reached, indicating a belief failure; the model immediately initiates step 105, discarding all current beliefs. The failure hypothesis, and based on this The location of the observed data at a given time, generating A completely new trajectory hypothesis was proposed, the hypothesis library was reconstructed, and the two computational models were... The RMSE comparison data is then shown in Table 1.

[0036] Table 1: Comparison of Root Mean Square Error (RMSE) of Models After Failure of Physical Prior Models

[0037]

[0038] The data in Table 1 shows that, to During this period, the RMSE of both models increased synchronously due to the failure of the physical prior model; At a certain point, the control group's model experienced belief failure, with errors continuing to diverge, indicating the model had failed; the experimental group... At that moment, through the synergistic effect of steps 104 and 105, after determining that the belief has failed ( Perform hypothesis base refactoring, its RMSE is Then converged back sea ​​mile (in (At any given moment), to restore effective tracking of the true trajectory.

[0039] Example 3: This example combines Figures 1 to 3 This section explains the method for predicting water-fall trajectories based on remote sensing and satellite communication positioning, such as... Figure 1As shown, the process begins in step 101, which constructs a full-path hypothesis library to initialize the historical states and belief weights of N trajectories. Step 102 proceeds to parallel evolution, driving the evolution of N hypotheses based on a physical prior model. Step 103 receives delayed or out-of-order observation data, which can originate from delayed observation data point bases from satellite communication and positioning modules or delayed observation data field bases from remote sensing images. After receiving the data, step 103 performs asynchronous belief updates, i.e., extracting historical states and updating global belief weights. Step 104 then performs belief validity verification, calculating belief validity indices, such as the number of effective particles. Step 105 determines belief failure based on whether this index reaches the model failure threshold. If the belief is determined to be invalid (yes), the hypothesis library reconstruction in step 105 is performed, discarding invalid hypotheses and reconstructing based on new data. After reconstruction, the process returns to step 102. If the belief is determined to be valid (no), delayed resampling or trajectory prediction output is performed to output prediction results or optimize the hypothesis library, ultimately obtaining the trajectory prediction results as shown in the probability density distribution diagram.

[0040] like Figure 2 As shown, the vertical axis of the graph represents RMSE in nautical miles, ranging from 0 to 9, and the horizontal axis represents time in hours, ranging from 2.0 to 5.5. The graph contains two curves: the dashed line represents the RMSE of the control group, whose RMSE value increases from 0.42 nautical miles at 2.0 hours to 4.31 nautical miles at 3.5 hours, and then continues to increase to 8.03 nautical miles at 5.5 hours; the solid line represents the RMSE of the experimental group, whose RMSE value is close to that of the control group from 2.0 to 3.5 hours, but after 3.5 hours, the RMSE of the experimental group drops rapidly, falling to 0.39 nautical miles at 4.5 hours, and remaining at 0.47 nautical miles at 5.5 hours. Figure 3 As shown, the system starts from standby state and enters the initialization state after receiving the first observation data, i.e., step 101, to build the hypothesis base. After the construction is completed, the system enters the stable prediction state, i.e., step 102, parallel evolution. When delayed / out-of-order observation data is received, the system enters the update and verification state, i.e., steps 103 and 104. In this state, a judgment is made. If the judgment is that the belief is valid, then output / resampling is performed and the stable prediction state is returned. If the judgment is that the belief is invalid, then the system enters the fault recovery state, i.e., step 105, to reconstruct the hypothesis base. After the reconstruction is completed, the system returns to the stable prediction state again.

[0041] Example 4: Before applying the computational model to actual trajectory prediction tasks, it faces key engineering challenges, namely the effectiveness of step 104 belief validity verification and step 105 hypothesis base reconstruction. These challenges highly depend on two core parameters: the preset model failure threshold (based on the effective particle number threshold). For example, the pre-set observation uncertainty (based on the reconstructed sampling standard deviation) used during reinitialization in step 105 is compared with that used during reinitialization. For example, these parameters were only defined as presets in previous technical solutions, lacking a systematic and reproducible basis for determination. To eliminate this uncertainty, this embodiment adopts a systematic two-stage engineering calibration procedure. This procedure is executed on a historical backtesting dataset containing real trajectories and corresponding observation data. The simulation environment of this dataset allows for calibration at specific times. Introducing known physical prior models of failure; in calibration At that time, the model was The trajectory assumes a total number of runs, system testing. The range to be tested, this range is In each simulation Model failure is introduced at various times, and the model is statistically analyzed under different conditions. Under candidate values, the failure detection rate is the rate at which the model correctly identifies a failure after the model has failed, and the false alarm rate is the rate at which the model is incorrectly identified as a failure in a control run where the model has not failed; this procedure selects values ​​that result in a failure detection rate greater than [value missing]. And the false alarm rate is lower than of Value, such as (Right now ), which serves as the preset model failure threshold for step 104.

[0042] exist Once determined, the procedure continues to calibrate the reconstructed sampling standard deviation in step 105. This parameter is defined so that after the belief fails, the model is regenerated using Gaussian sampling, centered on the location of the observed data. The standard deviation used for each historical state; determine The process includes: during the simulation run, the system uses predetermined... exist Then step 105 is triggered, using each of the values ​​within the test range respectively. Candidate values ​​are reconstructed; the system measurement model is used differently each time. After reconstructing the candidate values, the predicted RMSE converges to a stable level (e.g., Recovery time required (nautical miles) This procedure selects the option that maximizes the mean recovery time. Minimized Value, such as As the preset observation uncertainty parameter in step 105; the calculation model finally obtains the parameter set determined by this procedure ( This ensures the effectiveness of the self-diagnosis in step 104 and the fault recovery mechanism in step 105 in actual deployment.

[0043] Example 5: The likelihood on which the computational model of the method of the present invention depends when performing asynchronous belief update in step 103. This requires a calibrated observation model; this embodiment provides a standardized engineering calibration procedure for determining the key parameters of the observation model, to calculate the likelihood of the positioning data of the base point location in step 103; this procedure will use the likelihood function Defined as a Gaussian probability density function, its core parameter is the standard deviation of the observation noise. To determine The system obtains offline data including The calibration dataset of the group data, in which Not less than This dataset contains positioning data collected by the satellite communication positioning module applied in this method at known fixed points or along ground-based real trajectories synchronously recorded by a high-precision differential global positioning system; the procedure calculates this... The Euclidean distance error between the satellite communication positioning module data and its corresponding ground-based real trajectory data is used to form an error residual sequence. Finally, the procedure calculates the statistical standard deviation of this error residual sequence and determines this value as... nautical miles, defined as the standard deviation of observation noise used in the likelihood calculation step 103 of the computational model. .

[0044] In step 102, when driving the parallel evolution of the full-path hypothesis library, the method of this invention relies not only on the physical prior model but also on a stochastic process model to characterize uncertainties not covered by the physical prior model. This embodiment provides a calibration procedure for determining the parameters of the stochastic process model. This procedure specifically defines the stochastic process model within the framework as a discrete-time Gaussian random walk process applied to the trajectory velocity, with the core parameter being the diffusion coefficient characterizing the amplitude of the random disturbance. To determine The system obtains offline data including A validation dataset of historical real drift trajectories, in which Not less than Such as buoy data from satellite tracking; the procedure in the simulation environment uses only a physical prior model and does not impose any stochastic processes on it. The procedure involves replaying and predicting historical trajectories; calculating the point-by-point residual sequence between the predicted trajectory of the physical prior model and the historical actual trajectory; and finally, analyzing the statistical characteristics of the residual sequence's evolution over time, such as the mean square displacement, to fit and determine the diffusion coefficient of the Gaussian random walk process. ;Should The values ​​are set in the computational model for the parallel evolution of step 102 to ensure The degree of divergence of the trajectory hypothesis matches the inherent prediction error of the physical prior model.

[0045] Example 6: In step 103 of the method of the present invention, the computational model is configured to fuse field-based remote sensing data. This faces technical challenges in engineering implementation, namely, the need for standardized and reproducible procedures to convert the original remote sensing images into field-based likelihoods that can be directly used by the model for Bayesian updates. This embodiment provides a specific calibration and processing procedure for synthetic aperture radar (SAR) image data. This procedure is executed offline before model deployment and is used to determine the relevant parameters of the constant false alarm rate (CFAR) detector and its likelihood transformation function. This procedure obtains data including... ( A calibration dataset of SAR image samples, containing sample areas of known targets in the water (such as life rafts) and pure sea clutter background areas; the procedure is based on the CFAR detection algorithm (data preprocessor), by adjusting its detection threshold. The calibration dataset is processed until its detection probability for known targets is reached. Not less than Record the false alarm probability in the pure sea clutter region at this time. Therefore Value (e.g.) (This serves as the lower bound for the background likelihood of the model) When the model is running online, when Receives historical timestamps at all times When using field-based remote sensing data (SAR images), the model calls the calibrated ( The CFAR detection algorithm is used to process the SAR image and generate a georeferenced target probability map. This probability diagram constitutes the spatial likelihood field in the specific implementation method. When the model performs the asynchronous belief update in step 103, for the first... The trajectory assumption is in The historical state of a moment (its coordinates are) The model calculates its field basis likelihood by incorporating a function that integrates the lower bound of the background likelihood. This calculation is performed in the following way: This procedure ensures that when the model fuses field-based data, even if the trajectory assumption falls in a region where no target is detected, the global belief weights remain unchanged. It will not be updated to zero, but rather to the calibrated false alarm rate. The corresponding non-zero lower limit value.

[0046] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0047] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for predicting waterfall trajectories based on remote sensing and satellite communication positioning, characterized in that, Includes the following steps: Step 101: Construct and maintain a full path hypothesis library containing N independent trajectory hypotheses, where N is the preset total number of trajectory hypotheses, and each trajectory hypothesis has its own historical state sequence and global belief weight. Step 102: Based on the physical prior model and the stochastic process model, drive the parallel evolution of all N independent trajectory hypotheses in the full path hypothesis library, and record their complete historical state sequences after evolution. Step 103, Receive the historical timestamp at the current moment. When extracting delayed observation data, each independent trajectory hypothesis from the full path hypothesis library is extracted. The historical states corresponding to the historical states in the historical state sequence at each time point are used to calculate the likelihood between the delayed observation data and the historical states, and the global belief weights of each independent trajectory hypothesis are updated using Bayesian methods based on the likelihood. Step 104: After step 103, calculate a belief effectiveness index based on the updated global belief weights assumed by N independent trajectories. Step 105: When the belief validity index reaches the preset model failure threshold, it is determined that the model has failed. Then, all N independent trajectory hypotheses in the full path hypothesis library are discarded, and N new trajectory hypotheses are re-initialized and generated based on the delayed observation data to construct the reconstructed full path hypothesis library. And, the step of calculating the belief effectiveness index in step 104 specifically includes: calculating the global belief weight based on the updated N independent trajectory assumptions, calculating the information entropy of the global belief weight or calculating the effective particle number of the global belief weight, and the preset model failure threshold is the preset information entropy threshold or the preset effective particle number threshold.

2. The method for predicting waterfall trajectories based on remote sensing and satellite communication positioning according to claim 1, characterized in that, The effective number of particles is calculated using the following formula: ,in The effective number of particles, The total number of preset trajectory assumptions. For the index of the trajectory hypothesis, For the first The global belief weights are normalized based on the assumption of independent trajectories.

3. The method for predicting waterfall trajectories based on remote sensing and satellite communication positioning according to claim 1, characterized in that, The method also includes a delayed resampling step: when it is determined in step 104 that the belief validity index has not reached the preset model failure threshold, a delayed resampling step is performed after step 103; the delayed resampling step resamples the entire path hypothesis library based on the updated global belief weights of N independent trajectory hypotheses to generate N new trajectory hypotheses.

4. The method for predicting waterfall trajectories based on remote sensing and satellite communication positioning according to claim 1, characterized in that, The physical prior models in step 102 include: a regional ocean model system for providing background ocean current field data, a third-generation wave model for providing wind-generated wave field, and a target drift model for calculating the drift of a target under the combined action of background ocean current field data and wind-generated wave field.

5. The method for predicting waterfall trajectories based on remote sensing and satellite communication positioning according to claim 1, characterized in that, The delayed observation data in step 103 includes: data acquired via a satellite communication positioning module, characterizing the target's position at historical timestamps. The step of calculating the likelihood in step 103 includes: calculating the likelihood based on the Euclidean distance between the location data of the point base position and the historical state.

6. The method for predicting water-fall trajectory based on remote sensing and satellite communication positioning according to claim 5, characterized in that, The delayed observation data also includes: acquiring field-based remote sensing data characterizing the distribution area of ​​potential targets through remote sensing images; the step of calculating the likelihood in step 103 further includes: when the delayed observation data is field-based remote sensing data, determining the assumption of each independent trajectory at the historical timestamp. Whether the historical state falls within the region represented by the field-based remote sensing data, and calculate the field-based likelihood based on this.

7. The method for predicting waterfall trajectories based on remote sensing and satellite communication positioning according to claim 1, characterized in that, Step 103, which involves Bayesian updating of the global belief weights, includes: multiplying the likelihood calculated based on delayed observation data and historical states with the existing global belief weights of the N independent trajectory hypotheses before step 103 to obtain the updated global belief weights; and normalizing the updated global belief weights of the N independent trajectory hypotheses.

8. The method for predicting waterfall trajectories based on remote sensing and satellite communication positioning according to claim 1, characterized in that, Step 105, which involves re-initializing and generating N new trajectory hypotheses based on delayed observation data, includes: taking the location of the delayed observation data as the center and combining it with the preset observation uncertainty of the delayed observation data, generating N new historical states through resampling; using the N new historical states as the initial states of the N new trajectory hypotheses and resetting their global belief weights.

9. The method for predicting waterfall trajectories based on remote sensing and satellite communication positioning according to claim 1, characterized in that, The method also includes a trajectory prediction output step: when the belief validity index does not reach the preset model failure threshold in step 104, the trajectory prediction output step is executed; the trajectory prediction output step includes: based on the updated global belief weights of N independent trajectory hypotheses and the current states of the N independent trajectory hypotheses, a weighted average is performed or the current state of the trajectory hypothesis with the largest global belief weight is extracted to generate the final trajectory prediction result.

Images