A method and system for generating USBL data based on state simulation

Abstract

The present application relates to underwater acoustic positioning simulation technical field, especially in kind based on state simulation USBL data generation method and system, including obtaining trajectory parameters, USBL receiving end position, scene label and random seed, according to scene label to load scene parameter group;Generate underwater target true value track, calculate real geometric relationship and generate USBL simulation observation value of superimposed observation error;Based on scene parameter group, calculate loss probability and loss of lock probability and generate candidate observation state;The candidate observation state is input into three-state observation state machine, and the final observation state is output according to state holding time and recovery condition;Generate explanatory label and carry out position solution and state statistics, export simulation data set.The present application can generate USBL observation value simultaneously with the generation of three-state observation state, explanatory label and state statistical result, which provides data support for underwater positioning.

Description

Technical Field

[0001] This invention relates to the field of underwater acoustic positioning simulation technology, and in particular to a USBL data generation method and system based on stateful simulation. Background Technology

[0002] Ultra-short baseline (USBL) underwater acoustic positioning systems estimate the position of an underwater target relative to a receiver by measuring parameters such as propagation time, azimuth, and pitch angle between the acoustic head array and the underwater transponder. They are widely used in underwater robot navigation and underwater target localization. With the rise of supervised learning tasks such as state recognition and observation reliability assessment, constructing simulation data generation methods capable of reproducing the USBL observation link has become fundamental for model training and algorithm verification. Existing technologies are mainly divided into two categories: one is geared towards real measurement systems, aiming to improve the accuracy of actual measurement positioning; the other is geared towards simulation training environments, aiming to reproduce the measurement link or navigation scenario.

[0003] Regarding the improvement of positioning accuracy, existing technologies still have the following shortcomings. First, the simulation output is mostly limited to distance, angle or error statistics, lacking sample-level observation states, making it difficult to support supervised learning tasks such as state recognition. For example, patent application number CN114608567A discloses a USBL positioning method under small pitch angle conditions. It proposes a coordinate calculation and accuracy improvement scheme for small pitch angle conditions, but it focuses on the correction of calculation errors in the real measurement link, does not involve the generation of simulation observation data, and cannot output supervised information such as sample-level observation states and failure causes. Regarding the construction of simulation environments, patent CN112614412A proposes a joint simulation method for marine environment and underwater positioning and navigation. It reproduces navigation training scenarios by coupling marine environmental fields with acoustic currents. However, its simulation revolves around the environmental field and navigation process. Observation degradation is only reflected as error changes. Anomalies such as point loss and lock loss are treated as single-point random events. It lacks time-series modeling of continuous degradation, continuous failure, and delayed recovery processes, and does not output three-state observation states and interpretive labels. At present, a USBL data generation method and system based on stateful simulation is needed. Summary of the Invention

[0004] To address the technical problem of poor adaptability of underwater target positioning systems caused by the lack of sample-level observation states and the inability to simulate continuous degradation and phased recovery processes in existing USBL simulation technology, this invention provides a USBL data generation method and system based on state-based simulation.

[0005] In a first aspect, the present invention provides a USBL data generation method based on stateful simulation, which adopts the following technical solution: A USBL data generation method based on stateful simulation includes: Obtain trajectory parameters, USBL receiver position, scene label and random seed, and load the corresponding scene parameter group according to the scene label. The scene parameter group includes environmental parameters, anomaly model parameters, lockout parameters and state machine parameters. The true trajectory of the underwater target is generated based on the trajectory parameters. The true distance, true azimuth and true pitch angle are calculated based on the geometric relationship between the underwater target and the USBL receiver. The USBL simulation observation values ​​with superimposed observation errors are generated. Based on the scene parameter set, candidate observation states are generated according to propagation distance, pitch angle, signal-to-noise ratio, loss events, lock-out events, and abnormal events. Among them, the loss event is the event that no usable USBL observation is obtained at the current sampling time, the lock-out event is the event that the acoustic tracking loses lock or exceeds the maximum working range, and the abnormal events include multipath events, outlier events, and burst events. Input the candidate observation state into the three-state observation state machine, and output the final observation state according to the state holding time and recovery conditions; Interpretive labels are generated based on observation errors and final observation states. The location of USBL simulation observations is calculated, and the state transition relationships and state durations are statistically analyzed. A simulation dataset containing observations, observation states, interpretive labels, state statistics, and dataset indexes is then exported.

[0006] Furthermore, the step of loading the corresponding scene parameter group based on the scene tag includes inputting the scene tag into a scene mapping function and outputting a scene parameter vector through the scene mapping function. The scene mapping function establishes a mapping relationship between the scene tag and the scene parameter vector through a preset table, a configuration dictionary, or parameter generation rules. The scene parameter vector is represented as follows: , in, For scene tags, This is the mapping function from scene labels to scene parameter vectors. For scene tags The corresponding scene parameter vector, For sea state intensity, For multipath strength, For sound speed deviation, For environmental noise gain, For the abnormal model parameter group, For the unlock parameter group, This is the state machine parameter set.

[0007] Furthermore, the generation of USBL simulated observations with superimposed observation errors includes determining the target motion form based on trajectory type, duration, spatial scale, and depth parameters, and initializing the trajectory generation process based on a trajectory random seed to generate a true sequence of target positions; The true distance, true azimuth, and true elevation angle of the target relative to the USBL receiver are calculated in the true value sequence of the target position at each sampling time to obtain the true geometric observations. The signal-to-noise ratio (SNR) is estimated based on the actual distance and pitch angle. The estimated SNR is then used to generate observation errors, which are then superimposed onto the actual geometric observations to obtain USBL simulation observations.

[0008] Further, the step of generating observation errors based on signal-to-noise ratio estimates includes calculating the fundamental error based on reference noise and range-related noise, amplifying and modulating the fundamental error using a low pitch angle error amplification factor, a signal-to-noise ratio error amplification factor, and anomaly event error amplification factor, determining the standard deviations of range observation errors, azimuth observation errors, and pitch angle observation errors, and generating corresponding observation errors based on each standard deviation. The standard deviation of the range observation error is expressed as: , in, This represents the standard deviation of the distance observation error. The standard deviation of the noise level from the reference level. The slope of the distance-related noise. This is the error amplification factor caused by low pitch angles. This is the amplification factor of the signal-to-noise ratio on the distance error. This is the error amplification factor for abnormal events triggered by outlier events or sudden events; The standard deviations of the azimuth and elevation observation errors are respectively expressed as: , , in, and These are the standard deviations of the azimuth observation error and the standard deviations of the elevation observation error, respectively. and These are the reference noise standard deviations for azimuth and elevation angles, respectively. and These are the slope terms for azimuth and elevation errors as a function of distance, respectively. and These are the amplification factors of the signal-to-noise ratio for the azimuth and elevation angle errors, respectively. The corresponding point-by-point random observation error is generated based on the standard deviation of each observation error, and is expressed as follows: , in, Indicates the observation dimension as distance, azimuth, or elevation. For distance observation dimension, For azimuth observation dimension, For the perspective of pitch angle observation dimension, Indicates time Standard deviation of observation error for the corresponding observation dimension Let be a standard normal random variable generated based on an observational random seed. For a moment Point-by-point random observation error corresponding to the observation dimension.

[0009] Furthermore, the generation of candidate observation states includes calculating the observation loss probability based on the actual distance, pitch angle, signal-to-noise ratio, sea state intensity, and sudden events, and comparing the observation loss probability with a random number to determine the loss event; The upper limit of the range is calculated based on the probability of loss of lock and capped. The probability of loss of lock is calculated using a double exponential function based on the capped distance. When the maximum working range limit is enabled and the actual distance exceeds the maximum working range, the probability of loss of lock is set to 1. The loss of lock event is determined based on the probability of loss of lock and the corresponding sample is entered into the invalid candidate logic. The probability of losing the lock is expressed as: , , in, This is the capping distance used for calculating the probability of loss of lock. , , and These represent the coefficients of the lock loss probability model. Let be the probability of losing the lock at time t. For the amplitude limiting function, This is an over-range hard limit indication. For indicator functions, The true distance at time t This is the maximum working range.

[0010] Furthermore, the generation of candidate observation states also includes generating candidate invalidity determination quantities based on the comparison results of loss events, lock-out events, quality scores, signal-to-noise ratio, loss probability, and lock-out probability with invalidity state thresholds; When the candidate invalidity determination quantity is not valid, a candidate degradation determination quantity is generated based on the comparison results of multipath events, burst events, outlier events, quality scores, signal-to-noise ratio, loss probability, and lockout probability with the degradation state threshold. Among them, multipath events are events that generate additional distance and angle biases, outlier events are events that amplify the observation error of a single sampling moment by a preset multiple, and burst events are events that cause abnormal disturbances at multiple consecutive sampling moments. When the candidate invalidity criterion is true, the invalid state is taken as the candidate observation state; when the candidate invalidity criterion is false but the candidate degradation criterion is true, the degradation state is taken as the candidate observation state; when neither the candidate invalidity criterion nor the candidate degradation criterion is true, the normal state is taken as the candidate observation state. The candidate invalidity criterion is expressed as follows: , in, Let be the candidate invalidation criterion at time t. For the lost event, This is a lockout event. To rate the quality, This is the signal-to-noise ratio estimate. To observe the probability of loss, This represents the probability of losing the lock. , , and These are the quality, signal-to-noise ratio, loss probability, and lockout probability thresholds corresponding to the invalid state; The candidate degradation determination quantity is expressed as: , in, For a moment Candidate degradation criteria. , and These are multipath events, burst events, and outlier events. Represents logical OR, This represents logical AND. Indicates logical NOT. , , and These are the quality score, signal-to-noise ratio, loss probability, and lockout probability threshold corresponding to the degraded state.

[0011] Furthermore, the three-state observation state machine introduces state machine memory variables, state holding time, and recovery clearing sample number. The current output state is determined based on the candidate observation state, the final observation state at the previous sampling time, and the state machine memory variables, so that the degenerate state and invalid state form a continuous state segment with a phased recovery path. The state machine memory variables include the remaining holding length, the number of consecutive clearing samples, the current failure cause, and the current degradation source. The three-state observation state machine is represented as follows: , in, The final observed state at time t, and Let represent the state machine memory variables at time t and the previous sampling time, respectively. For state machine functions, To maintain length in the degenerate state, To preserve the length of the invalid state, The number of samples to be cleared to restore to the degraded state, The number of samples to be cleared to restore to normal. This is a candidate observation state.

[0012] Furthermore, the step of determining the current output state based on the candidate observation state, the final observation state at the previous sampling time, and the state machine memory variables specifically involves: when When the state is invalid, output the invalid state and keep the remaining length set to 0. The sample count is continuously cleared to zero, and the current failure cause and current degradation source are recorded. when When it is in a degenerate state, if If the state is invalid and the remaining length is greater than 0, then maintain the invalid output and decrease the remaining length. If the state is otherwise, output the degenerate state and keep the remaining length set to 0. ; when In the normal state, the maintenance period continues until the end of the maintenance period. Output: After the retention period ends, the number of continuously removed samples reaches [a certain value]. The system returned to normal after a period of time, with the number of samples continuously cleared reaching [a certain threshold]. However, it was not achieved. It will eventually revert to its degenerate state; when The normal state and the previous state If the state is already normal, then output "normal state".

[0013] Furthermore, the explanatory labels include report accuracy labels, failure cause labels, degradation source labels, and state duration labels; The report accuracy label is calculated based on the standard deviation of the distance error and the error contribution after the azimuth and pitch angle errors are converted to the distance scale. The failure cause label is determined based on the over-range hard limit indication, the loss of lock event, the missing event, the quality score, the signal-to-noise ratio, the loss probability, and the loss of lock probability. The degradation source label is determined based on the multipath event, the sudden event, the outlier event, the missing event, the loss of lock event, the quality score, the signal-to-noise ratio, the loss probability, and the loss of lock probability. The state duration label is determined based on the number of continuous sampling points in the final observation state.

[0014] Secondly, a USBL data generation system based on stateful simulation includes: The input unit is used to obtain trajectory parameters, USBL receiver position, scene label and random seed, and load the corresponding scene parameter group according to the scene label; The observation generation unit is used to generate the true trajectory of the underwater target based on the trajectory parameters, calculate the true distance, true azimuth and true pitch angle based on the geometric relationship between the underwater target and the USBL receiver, and generate USBL simulation observation values ​​with superimposed observation errors. The degradation determination unit is used to calculate the observation quality score, loss probability and lockout probability based on the scene parameter set, and set the lockout probability to 1 when the actual distance exceeds the maximum working range; The state machine unit is used to generate candidate observation states based on observation quality score, signal-to-noise ratio, loss probability, lockout probability, loss event, lockout event, and abnormal events. The abnormal events include multipath events, outlier events, and burst events. The candidate observation states are input into the three-state observation state machine, and the final observation state is output based on the state holding time and recovery conditions. The output unit is used to generate interpretive labels based on observation errors and final observation states, perform position calculations on USBL simulation observations, perform statistics on state transition relationships and state durations, and export a simulation dataset containing observations, observation states, interpretive labels, state statistics data, and dataset indexes.

[0015] In summary, the present invention has the following beneficial technical effects: 1. This invention uses a scene mapping function to uniformly map scene labels into scene parameter vectors that include sea state intensity, multipath intensity, sound speed deviation, environmental noise gain, anomaly model parameter group, lockout parameter group, and state machine parameter group. This enables synchronous linkage between observation degradation, lockout determination, and state evolution with scene labels, and allows for the batch construction of USBL simulation datasets with different degradation distributions.

[0016] 2. This invention generates observation errors based on signal-to-noise ratio estimates and amplifies and modulates the basic error by using low pitch angle error amplification factor, signal-to-noise ratio error amplification factor, and abnormal event error amplification factor. This enables the observation error to adapt to changes in propagation distance, pitch angle, signal-to-noise ratio, and abnormal events, making the generated USBL simulation observations closer to the error distribution of the real observation link.

[0017] 3. This invention achieves explicit expression of boundary range degradation and over-range failure by capping the actual distance and explicitly calculating the unlock probability using a double exponential function. Combined with the hard cutoff of the maximum working range, the unlock probability of over-range samples is set to 1 and directly enters the invalid candidate logic. This avoids the ambiguity in availability determination caused by approximating the over-range state with increased error in the prior art.

[0018] 4. This invention introduces a three-state observation state machine with state machine memory variables, state retention time, and recovery clearing sample number. The current output state is determined based on the combination relationship between the candidate observation state and the final observation state at the previous sampling time. This realizes the generation of continuous state segments of degraded and invalid states, as well as the time-series modeling of invalid states recovering to normal states in stages through degraded states. This makes the temporal evolution behavior of simulated observations closer to the continuous degradation, continuous failure, and delayed recovery process in the real USBL observation link.

[0019] 5. This invention provides interpretable supervisory information for observation reliability assessment and state identification models by outputting explanatory labels such as report accuracy, failure cause, degradation source and state duration while generating observation values, and by exporting simulation datasets hierarchically after statistically analyzing state transition relationships and state durations. This allows the generated data to be directly used for model training, failure sample verification and positioning algorithm validation, thereby reducing the cost of acquiring measured sea trial data. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall process of a USBL data generation method based on stateful simulation according to an embodiment of the present invention.

[0021] Figure 2 This is a schematic diagram of the overall structure of the USBL data generation system based on stateful simulation according to an embodiment of the present invention; Figure 3 This is a schematic diagram of scene-level data flow according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the three-state observation state machine according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the state transition matrix according to an embodiment of the present invention; Figure 6 This is a schematic diagram of the state duration distribution according to an embodiment of the present invention; wherein, Figure 6 (a) in the text represents the valid state. Figure 6 (b) in the diagram represents the degenerate state. Figure 6 (c) in the text represents an invalid state.

[0022] Figure 7 This is a schematic diagram of the state timeline and observation report accuracy in an embodiment of the present invention. Detailed Implementation

[0023] The present invention will be further described in detail below with reference to the accompanying drawings.

[0024] Example 1: Refer to Figure 1 This embodiment of a USBL data generation method based on stateful simulation includes: S1. Obtain trajectory parameters, USBL receiver position, scene label and random seed, and load the corresponding scene parameter group according to the scene label. The scene parameter group includes environmental parameters, anomaly model parameters, lockout parameters and state machine parameters. S2. Generate the true trajectory of the underwater target based on the trajectory parameters, calculate the true distance, true azimuth and true pitch angle based on the geometric relationship between the underwater target and the USBL receiver, and generate the USBL simulation observation values ​​with superimposed observation errors. S3. Based on the scene parameter group, candidate observation states are generated according to the propagation distance, pitch angle, signal-to-noise ratio, lost events, lost lock events, and abnormal events. Among them, lost events are events in which no usable USBL observations are obtained at the current sampling time, lost lock events are events in which acoustic tracking loses lock or exceeds the maximum working range, and abnormal events are a collective term for multipath events, outlier events, and sudden events. S4. Input the candidate observation state into the three-state observation state machine, and output the final observation state according to the state holding time and recovery conditions. S5. Generate interpretive labels based on observation errors and final observation states, calculate the location of USBL simulation observations, and perform statistics on state transition relationships and state durations. Export a simulation dataset containing observations, observation states, interpretive labels, state statistics, and dataset indexes.

[0025] Specifically, a USBL data generation method based on stateful simulation includes the following: like Figure 1 As shown, S1 first obtains the simulation input parameters, which include trajectory parameters, USBL receiver position, scene label, trajectory random seed, and observation random seed. The trajectory parameters include trajectory type, simulation duration, number of sampling points, spatial scale, and depth parameters. The trajectory type is used to determine the motion form of the underwater target, and the spatial scale and depth parameters are used to limit the spatial range of the target's motion. The USBL receiver position is used to determine the observation geometric reference. The trajectory random seed is used to initialize the generation process of the true trajectory, and the observation random seed is used to initialize the generation process of the observation error. By separating the trajectory random seed and the observation random seed, the same trajectory can correspond to multiple different sets of observations, while ensuring the reproducibility of the simulation results.

[0026] like Figure 3As shown, after obtaining the simulation input parameters, the corresponding scene parameter group is loaded according to the scene label. The scene label is not a single noise switch, but an index identifier used to map a set of environmental parameters, anomaly model parameters, lockout parameters, and state machine parameters. Specifically, the scene label is input into the scene mapping function. The scene mapping function pre-establishes the mapping relationship between scene labels and scene parameter vectors through a preset table, configuration dictionary, or parameter generation rules. After inputting the scene label, it outputs the corresponding scene parameter vector. The scene parameter vector uniformly organizes sea state intensity, multipath intensity, sound speed deviation, environmental noise gain, anomaly model parameter group, lockout parameter group, and state machine parameter group, represented as: , in, For scene tags, This is the mapping function from scene labels to scene parameter vectors. For scene tags The corresponding scene parameter vector, For sea state intensity, For multipath strength, For sound speed deviation, For environmental noise gain, For the abnormal model parameter group, For the unlock parameter group, This is the state machine parameter set.

[0027] Through the aforementioned scene mapping function, a single scene label can load multiple parameter sets in tandem: environmental parameters such as sea state intensity, multipath intensity, sound velocity deviation, and environmental noise gain are used to control the generation of observation errors; the anomaly model parameter set is used to control the generation of lost events, sudden events, outlier events, and multipath events; the lockout parameter set is used to control the lockout probability model coefficients, maximum working range, and hard cutoff switch; and the state machine parameter set is used to control the candidate state determination threshold, state holding time, and number of samples to be cleared during recovery. Thus, the scene, observation degradation, lockout, and state evolution change synchronously with the scene label, avoiding inconsistencies caused by isolated parameter configurations.

[0028] In one specific implementation, the scene labels include normal scene, low signal-to-noise ratio scene (low_snr), multipath-dominated scene (multipath_heavy), and long-range scene (long_range). Examples of environmental parameters corresponding to each scene label are shown in Table 1 below: Table 1 Scene Preset Parameters nominal 2.0 0.35 0.0m / s 0.0dB Normal scenario, used as a benchmark sample low_snr 4.0 0.45 -3.0m / s 5.0dB Low signal-to-noise ratio scenes are more prone to degradation. multipath_heavy 3.0 0.80 1.5m / s 2.0dB In multipath-dominated scenarios, degradation segments are more pronounced. long_range 2.5 0.50 -1.0m / s 1.5dB In long-distance scenarios, it is more prone to lock loss and exceeding the range. Among them, the unlocking parameter group for long-distance scenarios includes the unlocking model enable switch, maximum working range of 285.0m, hard range cutoff switch, unlocking probability model coefficients, and probability calculation range upper limit of 800.0m; the state machine parameter group includes degradation quality threshold, invalid quality threshold, degradation signal-to-noise ratio threshold, invalid signal-to-noise ratio threshold, degradation hold length, invalid hold length, number of samples restored to normal clearing, and number of samples restored to degradation clearing. Each threshold and hold recovery parameter is configured according to the scenario label, as shown in Tables 2 and 3. Table 2 shows the configuration of each threshold and scene label. nominal 0.76 0.20 18.0dB 10.0dB low_snr 0.86 0.28 22.0dB 14.0dB multipath_heavy 0.84 0.24 19.0dB 11.0dB long_range 0.82 0.22 20.0dB 12.0dB Table 3 Retains Recovery Parameters and Scene Label Configurations nominal 2 1 2 1 low_snr 3 2 3 2 multipath_heavy 4 2 3 2 long_range 3 2 3 2 The above methods enable low signal-to-noise ratio (SNR) scenes and multipath-dominated scenes to exhibit higher degradation sensitivity and longer degradation retention time compared to normal scenes. The loaded scene parameter set serves as the unified parameter basis for subsequent ground truth observation generation, degradation determination, and state machine determination.

[0029] S2. Generate the true underwater target trajectory based on the trajectory parameters obtained in step S1. Specifically, determine the target motion type based on the trajectory type, which includes circular, figure-eight, and spiral diving trajectories; limit the spatial range and depth variation range of the target motion based on spatial scale and depth parameters; determine the time discrete form of the trajectory based on the simulation duration and the number of sampling points; and initialize the trajectory generation process based on a trajectory random seed to generate the true target position sequence. In one specific embodiment, the number of sampling points is determined based on the simulation duration, expressed as: , in, Indicates the number of sampling points. Indicates the duration of the simulation. This indicates rounding to the nearest integer. The number of sampling intervals is the integer part of the simulation duration, and the number of sampling points is the number of intervals plus one, to ensure that the beginning and end endpoints of the true trajectory are sampled. This setting results in an equivalent sampling rate of approximately 1Hz under the default simulation duration, which conforms to the common low-frequency update characteristics of USBL and facilitates the generation of batch datasets.

[0030] After generating the target position ground truth sequence, the true geometric observations of the target relative to the USBL receiver are calculated at each sampling time. Specifically, the true distance is determined by the spatial Euclidean distance between the target position and the USBL receiver position, and the true azimuth and true elevation angles are determined by solving the spatial geometric relationship between the target and the USBL receiver. The true geometric observations provide a benchmark for subsequent observation generation and error assessment.

[0031] After obtaining the actual geometric observations, the signal-to-noise ratio (SNR) is estimated based on the actual distance and pitch angle. The SNR estimate is based on the baseline SNR at the reference distance, minus the propagation loss and absorption loss terms that increase with the actual distance, and then adding low pitch angle penalties, sudden event penalties, sea state penalties, and environmental noise penalties. The calculated result is then clipped to obtain the estimated SNR value, expressed as: , in, This is the signal-to-noise ratio estimate. For the actual distance, For the true pitch angle, The reference signal-to-noise ratio at the reference distance. For reference distance, This is used to avoid undefined or over-amplified terms when the distance is too small. The absorption attenuation coefficient, This is a penalty for low pitch angles. For sudden events at time t, For sea condition penalties, For environmental noise penalties, For the amplitude limiting function, and These are the lower and upper limits of the signal-to-noise ratio (SNR), respectively. The sea state penalty and environmental noise penalty are determined by the sea state intensity and environmental noise gain in the scene parameter group loaded in step S1, causing the SNR estimate to change in conjunction with the scene.

[0032] After obtaining the signal-to-noise ratio (SNR) estimate, the standard deviation of the observation error is determined based on the SNR estimate and geometric conditions. Specifically, the fundamental error is calculated based on the reference noise and range-dependent noise. The fundamental error is then amplified and modulated using a low pitch angle error amplification factor, a SNR error amplification factor, and anomaly event error amplification factor to obtain the standard deviation of the range observation error, expressed as: , in, This represents the standard deviation of the distance observation error. The standard deviation of the noise level from the reference level. The slope of the distance-related noise. This is the error amplification factor caused by low pitch angles. This is the amplification factor of the signal-to-noise ratio on the distance error. This is the error amplification factor for outlier events triggered by abnormal or sudden events. The impact of multipath events is reflected through additional distance and angle offsets.

[0033] The aforementioned error model parameters can be obtained through equipment nominal accuracy, calibration tests, or statistical analysis of residuals from actual flight tests. Specifically, the difference between the USBL measurement value and the true synchronous reference value can be calculated, and the standard deviations of the distance, azimuth, and pitch angle residuals within each bin can be calculated based on the propagation distance, absolute pitch angle, and signal-to-noise ratio. Then, the reference noise, range correlation slope, and various amplification factors can be fitted using the weighted least squares method or the maximum likelihood method. When no actual measurement data is available, initial values ​​can be set according to the equipment's technical specifications, and Monte Carlo simulation can be used to ensure that the standard deviations of the residuals in each bin fall within the target range. The parameters obtained through the above calibration are stored according to the equipment model and scenario parameter group, and can be recalibrated.

[0034] In one embodiment, the low pitch angle intermediate value and the low pitch angle error amplification factor are respectively expressed as: , , in, For low pitch reference angles, The maximum magnification intensity parameter at low pitch angles. Let be the midpoint of the low pitch angle at time t. This is the low pitch angle error amplification factor at time t. Let be the true pitch angle at time t. It is a pitch angle protection value greater than 0, used to prevent division by zero when the actual pitch angle is close to 0; This is a limiting function used to restrict the calculation results to a range of 1 to 4. The standard deviation of the distance observation error is calculated using... The square root is modulated to reduce the amplification of the range term by low elevation angles; the standard deviation of the azimuth and elevation angle observation errors is adopted. Modulation is performed. When the absolute pitch angle is not less than the reference angle, the intermediate value of the low pitch angle is 1; when the absolute pitch angle is less than the reference angle, the intermediate value of the low pitch angle increases as the absolute pitch angle decreases, and is limited to the range of 1 to 4.

[0035] The signal-to-noise ratio error amplification factor can be expressed as: , in, Let be the signal-to-noise ratio error amplification factor for the observation dimension x at time t. For the corresponding observation dimension The signal-to-noise ratio sensitivity coefficient, The reference signal-to-noise ratio at the reference distance. Here is the signal-to-noise ratio estimate at time t. for Corresponding observation dimensions: distance, azimuth, or elevation. Error amplification factor for abnormal events in normal samples. Set to 1; a preset amplification factor can be used when an outlier or sudden event is triggered. In this embodiment, multipath events are modeled by adding distance and angle biases, the multipath biases being determined by the multipath bias mean and random perturbations. As a set of reproducible examples, Take 32dB, Take 8°, Take 1.6, Take 0.8 m, Take 0.004 m / m, Take 0.55, Take 4.0, where, The signal-to-noise ratio sensitivity coefficient is for the distance observation dimension. The above values ​​are used to illustrate the parameter acquisition and calculation process and do not constitute a limitation on other device or scene parameters.

[0036] The standard deviations of azimuth and elevation observation errors are determined using the same modulation method as the standard deviation of range observation errors. Specifically, the basic error is constructed using the angle reference noise and the slope of the angle error increasing with distance. This is then modulated by a low elevation error amplification factor, a signal-to-noise ratio error amplification factor, and an abnormal event error amplification factor, and is expressed as follows: , , in, This represents the standard deviation of the azimuth observation error. The standard deviation of the elevation angle observation error. and These are the reference noise standard deviations for azimuth and elevation angles, respectively. and These are the slope terms of the angle error as a function of distance. and These are the amplification factors of the signal-to-noise ratio for the azimuth and pitch angle errors, respectively, with the angle error term calculated in radians.

[0037] The azimuth and elevation angle error parameters are determined using the same binning residual statistics and fitting method as the distance error parameters. During calculation, the standard deviation of the angle error can first be obtained in degrees, then converted to radians using the following formula before being used in position back projection and report accuracy calculations: , in, The standard deviation of the angular observation error expressed in radians. The standard deviation of angular observation error, expressed in degrees, is the standard deviation of azimuth reference noise, which is a set of repeatable real-world examples. azimuth distance slope Standard deviation of pitch angle reference noise Pitch angle, distance, slope Signal-to-noise ratio sensitivity coefficients for azimuth and elevation observation dimensions and All During fitting, the standard deviations and growth slopes are constrained to be no less than [a certain value]. The measured residual standard deviation within each distance, pitch angle, and signal-to-noise ratio bin is used as the fitting target, thus clarifying the parameter source and reproduction steps.

[0038] The corresponding point-by-point random observation error is generated based on the standard deviation of each observation error, and is expressed as follows: , in, These represent the observation dimensions of distance, azimuth, and elevation, respectively. Indicates time Standard deviation of observation error for the corresponding observation dimension Let be a standard normal random variable generated based on an observational random seed. This represents the corresponding point-by-point random observation error.

[0039] After determining the standard deviation of each observation error, the corresponding distance error, azimuth error, and elevation error are generated based on the standard deviation of each observation error. These errors are then superimposed onto the actual geometric observations to obtain the distance observation, azimuth observation, elevation observation, and propagation time observation. Specifically, the observation errors also include fixed bias and random walk. Fixed bias and random walk are generated based on the observation random seed, ensuring that the observation errors contain both point-by-point random components and time-dependent components. The propagation time observation is determined based on the distance observation and the sound speed with a sound speed deviation, which is determined by the scene parameter set loaded in step S1. Multipath events, sudden events, and outlier events are generated under the control of the anomaly model parameter set loaded in step S1. Outlier events and sudden events amplify the corresponding observation errors through anomaly event error amplification factors, while multipath events are reflected through additional distance and angle biases, ensuring that the USBL simulation observations simultaneously reflect the effects of geometric conditions, scene degradation, and anomaly events.

[0040] S3. Based on the true distance, true pitch angle, signal-to-noise ratio estimate obtained in step S2 and the scene parameter group loaded in step S1, calculate the observation loss probability and determine the loss event at each sampling time, calculate the lock loss probability and determine the lock loss event, and then comprehensively generate candidate invalid decision quantity and candidate degradation decision quantity to determine the candidate observation state.

[0041] Specifically, based on the actual distance, pitch angle, estimated signal-to-noise ratio, sea state intensity, environmental noise gain, and sudden events, the probability of observation loss at each sampling time is calculated and expressed as: , in, Indicates time The probability of observation loss, Indicates the base loss probability. to These represent the weights of distance, low pitch angle, low signal-to-noise ratio, sea state, and unforeseen events on the probability of loss, respectively. Indicates time The actual distance This represents the reference distance used in calculating the probability of loss. This means normalizing the distance to the range of 0 to 1. This indicates the low pitch angle magnification term. This represents the low signal-to-noise ratio loss term. Indicates the quantity of emergency events. The upper limit of the loss probability is represented, which is set to 0.95 in one implementation. The low signal-to-noise ratio loss term is determined based on the signal-to-noise ratio estimate obtained in step S2, and the sea state term is determined by the sea state intensity in the scene parameter group, so that the loss probability changes in conjunction with the scene and geometric conditions. After obtaining the observation loss probability, the observation loss probability is compared with a random number generated based on the observation random seed. When the random number is less than the observation loss probability, it is determined that a loss event has occurred at the corresponding sampling time.

[0042] The aforementioned low pitch angle amplification term and low signal-to-noise ratio loss term can be determined by the following formulas: , , in, This is the low pitch angle amplification term at time t, which shares the low pitch angle intermediate and pitch angle protection terms with the observation error model. , Let be the low signal-to-noise ratio loss term at time t, which monotonically increases as the signal-to-noise ratio decreases. As a set of reproducible real-world examples, the basic loss probability... Take 0.03 as the reference distance. Take 120 m as the distance weight. Set the weight to 0.20 for low pitch angles. Set to 0.08, low signal-to-noise ratio weight. Take 0.18, sea state weight Set the weight to 0.02 for sudden events. Set to 0.12, the upper limit of the probability of loss. Take 0.95.

[0043] The aforementioned loss probability parameters can be calibrated using measured samples with loss labels or target simulation statistics. Specifically, empirical loss rates are statistically analyzed grouped by distance, absolute pitch angle, signal-to-noise ratio, sea state intensity, and sudden events. Using these empirical loss rates as the objective, non-negative constrained least squares method, binomial maximum likelihood method, or equivalent methods are employed to estimate the loss probability. and to Then, verify the predicted loss rate, monotonicity, and upper limit constraint on independent samples. When measured samples are lacking, initial values ​​can be set according to the equipment range boundary and the expected loss rate, and iteratively adjusted through Monte Carlo experiments with fixed observation random seeds to obtain a repeatable sequence of loss events with the same input parameters and random seeds.

[0044] Secondly, the probability of lock loss is calculated and a lock loss event is determined. The upper limit of the range is calculated based on the actual distance using the lock loss probability, and then capped. The lock loss probability is calculated using a double exponential function based on the capped distance, and the range is limited. When the maximum working range limit is enabled and the actual distance exceeds the maximum working range, the over-range hard limit indicator is set, and the lock loss probability is directly set to 1, causing the corresponding sample to enter the invalid candidate logic. The lock loss probability is expressed as: , , The over-range hard cutoff rule is expressed as: , in, This is the capping distance used for calculating the probability of loss of lock. , , and These represent the coefficients of the lock loss probability model. Let be the probability of losing the lock at time t. For the amplitude limiting function, This is an over-range hard limit indication. For indicator functions, The true distance at time t For maximum working range, Calculate the upper limit of the range for the probability of loss of lock. It is an exponential function.

[0045] The unlock probability model coefficients, maximum operating range, and hard cutoff switch are all determined by the unlock parameter set loaded in step S1. A capping process is used to prevent the unlock probability from increasing infinitely in the over-range interval; a double exponential function is used to characterize the nonlinear growth of the unlock probability with distance. When the over-range hard limit indicator is set, instead of approximating the over-range state by increasing noise, the unlock probability is set to 1, explicitly expressing boundary range degradation and over-range failure. After obtaining the unlock probability, it is determined whether an unlock event has occurred at the corresponding sampling time based on the unlock probability.

[0046] Next, the quality score and abnormal events are determined. The quality score is calculated based on the propagation distance, pitch angle, signal-to-noise ratio estimate, sea state intensity, and environmental noise, and is used to characterize the quality of observation at the current sampling time. Multipath events are events that generate additional distance and angle biases, outlier events are events that amplify the observation error at a single sampling time by a preset factor, and sudden events are events that cause abnormal disturbances at multiple consecutive sampling times. These three are collectively referred to as abnormal events, and are generated under the control of the abnormal model parameter group loaded in step S1. Each event participates in the candidate state determination in the form of a binary indicator.

[0047] In one embodiment, propagation distance, pitch angle, signal-to-noise ratio, sea state intensity, and environmental noise are first applied to the probabilities of loss, lockout, multipath propagation, and sudden events, respectively. Then, a quality score is generated from the event indication, to avoid the same adverse factor being weighted repeatedly in both the quality score and the event probability. The quality score can be expressed as: , in, , , , and These are indicators for sudden events, outlier events, multipath events, loss events, and unlocked events, respectively. The algorithm represents a logical OR, and the weights can be set according to the observation availability at different times, and calibrated through a state label confusion matrix or the target state proportion. Lost events and unlocked events are not considered abnormal events; all event indicators can be 1 at the same sampling time, and when a candidate invalidity determination is valid, it is preferentially determined to be an invalid state.

[0048] Finally, candidate invalidation criteria are generated based on the comparison results of the loss event, the unlocking event, the quality score, the signal-to-noise ratio estimate, the loss probability, the unlocking probability, and the invalid state threshold, and are expressed as: , in, Let be the candidate invalidation criterion at time t. For the lost event, This is a lockout event. To rate the quality, This is the signal-to-noise ratio estimate. To observe the probability of loss, This represents the probability of losing the lock. , , and These are the threshold values ​​for quality, signal-to-noise ratio, loss probability, and lockout probability corresponding to the invalid state, respectively. When any invalid condition is met, the candidate invalidity determination value is set to 1.

[0049] The invalid state threshold and degraded state threshold satisfy the following order constraints based on the index direction: lower quality scores and signal-to-noise ratios indicate worse observations, therefore the invalid state quality threshold is lower than the degraded state quality threshold, and the invalid state signal-to-noise ratio threshold is lower than the degraded state signal-to-noise ratio threshold; higher loss probability and unlock probability indicate worse observations, therefore the invalid state probability threshold is higher than the corresponding degraded state probability threshold. This order constraint ensures that the degraded interval lies between the normal interval and the invalid interval.

[0050] The threshold can be initially determined based on the minimum available signal-to-noise ratio, maximum operating range, and effective measurement rate given in the equipment manual. Then, the conditional distribution of each indicator is calculated using calibration data with status labels, and the threshold is determined through grid search, receiver operating characteristic curves, or Monte Carlo statistics, constrained by the invalid sample false positive rate, the normal sample false positive rate, and the proportion of target states. Different scenarios are calibrated separately, but the above-mentioned order constraints are maintained.

[0051] Taking a normal scenario as an example, the quality score thresholds for the degraded / invalid state can be set to 0.76 / 0.20, the signal-to-noise ratio thresholds can be set to 18 dB / 10 dB, the loss probability thresholds can be set to 0.07 / 0.82, and the lockout probability thresholds can be set to 0.18 / 0.98. This set of thresholds prioritizes outliers, multipath events, or sudden events in the degradation candidate logic, while loss events, lockout events, or continuous risk indicators that reach the invalidity boundary are placed in the invalidity candidate logic.

[0052] When the candidate invalidity criterion is not valid, a candidate degradation criterion is generated based on the comparison results of multipath events, sudden events, outlier events, quality scores, signal-to-noise ratio estimates, loss probabilities, and unlock probabilities with the degradation state threshold, and is expressed as: , in, Indicates time Candidate degradation criteria. Indicates multipath events, Indicates an emergency. Indicates an outlier event. Indicates quality score. Represents logical OR, This represents logical AND. Indicates logical NOT. , , and These represent the quality, signal-to-noise ratio, loss probability, and lockout probability thresholds corresponding to the degraded state, respectively. The invalid state threshold and the degraded state threshold are both determined by the state machine parameter group loaded in step S1 according to the scene label, so that the sensitivity of the candidate state judgment changes with the scene linkage.

[0053] In low signal-to-noise ratio (SNR) scenarios, the quality score thresholds for degraded / invalid states can be set to 0.86 / 0.28, the SNR thresholds to 22 dB / 14 dB, the loss probability thresholds to 0.05 / 0.70, and the unlock probability thresholds to 0.16 / 0.92. In multipath-signal scenarios, the above four thresholds can be set to 0.84 / 0.24, 19 dB / 11 dB, 0.06 / 0.75, and 0.18 / 0.95, respectively.

[0054] In long-distance scenarios, the quality score thresholds for degraded / invalid states can be set to 0.82 / 0.22, the signal-to-noise ratio thresholds to 20dB / 12dB, the loss probability thresholds to 0.05 / 0.68, and the unlock probability thresholds to 0.10 / 0.85. This scenario also allows setting a maximum working range and a hard cutoff for exceeding the range; for example, when the maximum working range is set to 285 m, the unlock probability will be set to 1 if the actual distance exceeds this value.

[0055] The thresholds mentioned above are all reproducible real-world examples, and other calibrated thresholds are not restricted. After the candidate states are determined, a time hysteresis is further formed by maintaining the number of samples in the degenerate and invalid states and restoring the number of samples to be cleared, so that a single-point threshold exceeding the limit will not cause high-frequency jitter in the final state; the threshold is responsible for spatial classification, while the maintain and restore parameters are responsible for temporal continuity, and both can be calibrated separately.

[0056] Candidate observation states are determined based on the values ​​of the candidate invalidity criterion and the candidate degradation criterion. :when When, the invalid state is taken as the candidate observation state; when and When, the degenerate state is taken as the candidate observation state; when and When the normal state is selected, it is taken as the candidate observation state. The candidate observation state serves as the input to the three-state observation state machine in step S4.

[0057] like Figure 4 As shown, in step S4, the candidate observation states generated in step S3 are input into the three-state observation state machine, and the final observation state is output according to the state holding time and recovery conditions. The state set of the three-state observation state machine includes normal state, degenerate state, and invalid state, represented as: , in, This corresponds to the normal state, indicating that observations are normal and usable. The corresponding degenerate state indicates a decrease in observation quality, but it can still be used for position calculation. The corresponding invalid state indicates that the observation is unusable for location calculation. By extending the observation validity from a binary representation to a three-state representation, degraded samples and invalid samples can be distinguished.

[0058] The three-state observation state machine introduces state machine memory variables, state holding time, and the number of samples to be cleared upon recovery. The state machine memory variables include the remaining holding length, the number of consecutively cleared samples, the current cause of failure, and the current source of degradation. The remaining holding length records the number of sampling points that need to be maintained for the current state; the number of consecutively cleared samples records the number of sampling points where the candidate observation state is consecutively better than the current output state; and the current cause of failure and the current source of degradation record the basis for state determination. The three-state observation state machine determines the current output state based on the candidate observation state, the final observation state at the previous sampling time, and the state machine memory variables, expressed as: , in, The final observed state at time t, and Let represent the state machine memory variables at time t and the previous sampling time, respectively. For state machine functions, To maintain length in the degenerate state, To preserve the length of the invalid state, The number of samples to be cleared to restore to the degraded state, The number of samples to be cleared to restore to the normal state, the state retention length, and the number of samples to be cleared are all determined by the state machine parameter group loaded in step S1 according to the scenario label.

[0059] State machine function Based on candidate observation status Final observation state compared to the previous sampling time Based on the combination relationships, the current output state is determined according to the following rules. : (1) If the candidate state If it is an invalid state, then the current output state is... Set to invalid state, keep the remaining length set to The sample count is continuously cleared to zero, and the current failure cause and degradation source are recorded.

[0060] (2) If the candidate state It is a degenerate state and the previous state If the state is invalid, first check if the remaining length of the invalid state has ended; if the remaining length is still greater than 0, continue to output invalid states and decrement the remaining length; if the holding period has ended and the number of continuously cleared samples has reached a certain threshold... Then output the degenerate state and set the degenerate state to maintain the remaining length. Otherwise, continue outputting invalid states.

[0061] (3) If the candidate state It is a degenerate state and the previous state If it is not an invalid state, output the degenerate state, record the current source of degeneration, and set the remaining length of the degenerate state to be [value missing]. .

[0062] (4) If the candidate state The normal state and the previous state If the state is invalid, the invalid state maintenance requirement is satisfied first; after the maintenance period ends, if the number of continuously removed samples reaches... If the number of samples continuously cleared reaches a certain threshold, the system will return to normal. But it has not yet been achieved. If the state is negative, it will first revert to the degenerate state; otherwise, it will remain in the invalid state.

[0063] (5) If the candidate state The normal state and the previous state If the state is in a degenerate state, the requirement for maintaining the degenerate state must be met first; after the maintenance period ends, if the number of continuously removed samples reaches... If it recovers to the normal state, it will continue to remain in the degenerate state.

[0064] (6) If the candidate state The normal state and the previous state If the state is already normal, output "normal state" and clear the cause of failure and set the source of degradation to "normal source".

[0065] After each sampling point completes its state determination, the state duration, current failure reason, and current degradation source are updated synchronously, correspondingly outputting the final observation state `validity_state`, state duration `state_duration_samples`, failure reason `failure_reason`, and degradation source `degradation_source`. Through the transfer of state machine memory variables between adjacent sampling times, degraded and invalid states are no longer independent random points, but rather continuous temporal states with hold times, recovery lags, and phased recovery paths. This makes the temporal evolution behavior of the simulated observations more closely resemble the continuous degradation, continuous failure, and delayed recovery processes in a real USBL observation link.

[0066] S5. Generate explanatory labels based on the observation error generated in step S2 and the final observation status output in step S4. The explanatory labels include report accuracy label, failure cause label, degradation source label and status duration label. The generation methods of each type of label are explained below.

[0067] The report accuracy label is calculated based on the error contribution after converting the azimuth and elevation errors to the distance scale, and is expressed as follows: , in, Indicates time The report accuracy label, Indicates time The standard deviation of the distance error Indicates time The standard deviation of the azimuth error, Indicates time Standard deviation of pitch angle error, with cap Indicates time The observation distance or the measurement distance used to report accuracy estimates, and The lateral or spatial error contribution is represented by the angle error converted to the distance scale. The angle term is in radians. By transforming the observation error model into a sample-level report accuracy label, each sampling moment has a quantitative indicator that characterizes the reliability of the positioning, which can be directly used as supervisory information for observation reliability modeling.

[0068] The failure cause label is determined by a failure cause selection function based on the over-range hard limit indication, lock-out event, loss event, quality score, signal-to-noise ratio estimate, loss probability, and lock-out probability, and is expressed as follows: , The values ​​for the failure cause label include over-range, acoustic lockout, loss, low quality, low signal-to-noise ratio, risk of loss, and risk of lockout. The failure cause selection function determines a unique failure cause from the triggering factors according to the priority of each failure factor, so that over-range hard limit, risk of loss, and risk of lockout can all be included in the interpretive label.

[0069] The degradation source label is determined by the degradation source selection function based on the over-range hard limit indication, multipath event, sudden event, outlier event, loss event, lockout event, signal-to-noise ratio estimate, quality score, loss probability, and lockout probability, and is expressed as follows: , in, Indicates time The failure reason label, Indicates time The source of degradation label, The function to indicate the cause of failure is selected. Indicates the function for selecting the source of degradation; This is an over-range hard limit indication. This is a lockout event. For the lost event, To rate the quality, For a moment The signal-to-noise ratio estimate, For a moment The probability of loss, For a moment The probability of losing the lock, This is a multipath event. For emergencies, This is an outlier event.

[0070] The degradation source label takes values ​​including multipath, burst, outlier, loss risk, lockout risk, low signal-to-noise ratio, low quality, and recovery margin, with the recovery margin used to label samples in the phased recovery process. The state duration label is determined based on the number of continuous sampling points of the final observation state and is updated synchronously with the point-by-point determination of the three-state observation state machine. Through the failure cause label and degradation source label, the training samples not only contain state information on whether the observation is usable, but also information on the cause and extent of observation degradation, forming an interpretable supervision signal.

[0071] After generating interpretive labels, the position is calculated from the USBL simulation observations. Specifically, the original back-projected position is obtained by back-projecting the distance, azimuth, and elevation observations. For samples whose final observation state is normal or degraded, position estimates are output. For samples whose final observation state is invalid, the position estimate is marked as null, while the original back-projected position and interpretive labels are retained for subsequent visualization and analysis of failed samples. The calculation results, along with the solver identifier, are recorded. Whether the position estimate is null indicates whether the sample participated in the effective positioning output.

[0072] like Figure 5 As shown, the state transition relationships and state durations for each scene sample are then statistically analyzed. Specifically, the state transition matrix, state duration distribution, state proportion, state summary, and maximum continuous duration are statistically analyzed between the final observed states at adjacent time points, and a state transition diagram, duration distribution diagram, and state timeline diagram are generated. This allows the temporal state behavior of the scene samples to be visually verified. The state transition matrix includes the transition probability between states and the number of times the transition occurs. Let represent the conditional probability of transitioning from state i to state j. This indicates the number of times the transition probability occurs.

[0073] Finally, the simulation dataset is exported. The simulation dataset is organized in a four-layer structure: the geometric truth layer stores the target position truth value and sampling time; the observation layer stores distance observations, azimuth observations, elevation observations, and propagation time observations; the state and interpretation layer stores the final observation state, report accuracy label, failure cause label, degradation source label, and state duration label; and the statistics and organization layer stores the state transition matrix, state duration distribution, state summary, and manifest index. Each scene sample is exported as a scene-level data package. The manifest index records scene parameters, sample quantity, and file index. The simulation dataset as a whole is organized into training set, validation set, test set, and out-of-distribution test set.

[0074] The generated simulation dataset, derived through the aforementioned layered export method, can be directly applied to underwater robot navigation, underwater target localization, and underwater acoustic measurement verification. State labels and interpretive labels are read as supervisory information to train the observation reliability assessment model and state recognition model. State statistical results are used to verify the consistency between the sample distribution and real-world conditions. This allows for thorough verification of the USBL positioning algorithm and review of failed samples before sea trials, reducing the number of actual sea trials and lowering the R&D costs and engineering risks of the underwater positioning system.

[0075] Example 2: The difference between this example and Example 1 is that this example provides a simulation experiment of a USBL data generation method based on stateful simulation; This embodiment is based on the USBL data generation method based on state simulation described in Embodiment 1. Three sets of simulation experiments are designed to verify the timing modeling capability of the three-state observation state machine, the triggering behavior of explicit unlocking and range hard cutoff mechanisms, and the engineering rationality of basic observation errors. Typical output file organization and scalable implementation methods are also given.

[0076] I. Experimental Setup like Figure 6 , Figure 7 As shown, the simulation duration was set to 240.0 s, the number of sampling points was 241, and the equivalent sampling rate was approximately 1 Hz. Trajectory types included circular trajectory, figure-eight trajectory, and spiral descent trajectory; scene labels included normal scene, low signal-to-noise ratio scene, and multipath scene; for each trajectory and scene combination, 5 sets of trajectory random seeds and 3 sets of observation random seeds were set to ensure the statistical validity of the experimental results. Figure 7 The upper layer represents the evolution of the three-state validity over time, while the lower layer represents the observation report accuracy for each sampling point.

[0077] II. Verification of the Three-State Observation State Machine This experiment was conducted to verify the impact of the three-state observation state machine on the temporal evolution behavior of the observation link. The experimental group enabled the three-state observation state machine, while the control group disabled the state maintenance and recovery logic, only outputting the point-by-point candidate observation state, with all other parameter configurations remaining the same.

[0078] Experimental results show that the average duration of invalid states was 2.17 sampling points when the state machine was enabled, compared to 1.25 sampling points in the control group; the state transition rate was 0.2175 when the state machine was enabled, compared to 0.2224 in the control group; and the sample proportions of normal, degenerate, and invalid states were 0.1934, 0.6056, and 0.2010 when the state machine was enabled, compared to 0.2200, 0.6538, and 0.1262 in the control group.

[0079] The results above show that after enabling the three-state observation state machine, the average duration of invalid states increases significantly, the proportion of invalid state samples increases, and the proportion of degenerate state samples is adjusted accordingly. This indicates that the three-state observation state machine changes the temporal evolution behavior of the observation link through state holding time and recovery clearing sample number, so that degenerate and invalid states form continuous state segments and reflect a phased recovery process, rather than independent random points.

[0080] III. Loss of Lock and Range Mechanism Trigger Verification This set of experiments was used to verify the triggering behavior of the explicit loss-of-lock probability model and the maximum working range hard cutoff. The trajectory types included circular trajectories and spiral descent trajectories, with spatial scales of 60 m, 140 m, 240 m, and 320 m, respectively. For each combination, three sets of trajectory random seeds and three sets of observation random seeds were set. The explicit loss-of-lock model and the maximum working range limit were enabled, with the maximum working range set to 285.0 m. The loss-of-lock trigger rate was statistically analyzed according to distance intervals and the maximum scale scenario.

[0081] Experimental results show that the unlock trigger rate is 1.0000 in the distance range of 300 m to 400 m; 0.4524 in the distance range of 200 m to 300 m; and 0.8744 in the largest scale scenario.

[0082] The results above show that after enabling the distance-dependent loss-of-lock probability and setting the maximum working range, the simulator can generate partial loss-of-lock near the boundary range and form a stable loss-of-lock output in the over-range range. The triggering behavior of the explicit loss-of-lock model and the range hard cutoff rule is self-consistent, and both boundary range degradation and over-range failure are explicitly expressed.

[0083] IV. Verification of the magnitude of basic observation errors This group of experiments was used to confirm that the generated USBL observation errors were within a reasonable engineering range. The reference group used 18 sets of reference scene data from an open-source C++ USBL simulator, covering near-range, medium-range, long-range, low signal-to-noise ratio, multipath, and over-range conditions; the experimental group used the method of this invention to generate USBL observations under corresponding geometric and scene conditions. This group of experiments only used the open-source C++ simulator as a reference source for the magnitude of the error, and the evaluation metric was the root mean square value of the three-dimensional position error, which could be jointly calculated by both parties.

[0084] Experimental results show that the average three-dimensional root mean square error of the reference simulator on the evaluable sample is 0.8050 m, while the average three-dimensional root mean square error of the simulator of this invention is 0.7662 m under the same error evaluation caliber. The difference between the two is about 0.0388 m, which is in the same meter-level error order.

[0085] As can be seen from the above results, the basic USBL observation error generated by this invention does not deviate from the reasonable range of existing device-level simulation results, and the generation of observation values ​​has engineering rationality and can serve as a reliable basis for generating state-based simulation data.

[0086] V. Typical Output Files In this embodiment, a scene sample package includes at least the output files shown in the table below: Table 4 Output Files *_gt.csv True value of target trajectory *_usbl.csv USBL observations, three-state observations, and interpretive labels *_solver_output.csv Location estimation, original backprojection location and solver identifier *_validity_transitions.csv State transition matrix *_validity_durations.csv State duration distribution *_validity_summary.csv State Summary *_validity_transition_matrix.png State transition diagram *_validity_duration_distribution.png Duration distribution map *_validity_timeline.png State Timeline manifest.csv Scene parameters, number of samples, and file index Through the above file organization method, the ground truth data, observation data, solution data, state statistics data, image results and index information of each scene sample are completely corresponding, which facilitates batch sample management, verification and model training division.

[0087] VI. Scalable Implementation Methods This embodiment can be further extended to include platform attitude changes, installation angle errors, boom offsets, receiver depth changes, non-uniform sound velocity profiles, ocean current disturbances, continuous degradation segments induced by platform maneuvers, RTT / TDOA intermediate observations, and state statistics aggregated by scene labels. All of the above extensions can be implemented as supplementary scene parameters, observation model parameters, or output fields, without changing the core structure of this invention: scene-driven, state-based observation, interpretive labeling, state statistics, and training dataset export.

[0088] In summary, the simulation results of this embodiment show that the method can generate USBL simulation data with continuous temporal states, explicit unlocking expressions, and reasonable engineering error levels, and simultaneously output three-state observation states, interpretive labels, state statistics results, and dataset indexes, which can directly support engineering applications such as underwater positioning algorithm verification, observation reliability modeling, and state recognition model training.

[0089] Example 3: As Figure 2As shown, this embodiment provides a USBL data generation system based on stateful simulation, including an input unit, an observation generation unit, a degradation determination unit, a state machine unit, and an output unit. The units are connected in sequence to form a complete data flow from scene input to dataset output.

[0090] The input unit is used to acquire trajectory parameters, USBL receiver position, scene labels, and random seeds. Based on the scene labels, it loads the corresponding scene parameter set. Specifically, the input unit includes a trajectory and scene input module. The trajectory and scene input module receives trajectory type, simulation duration, number of sampling points, spatial scale, depth parameters, USBL receiver position, scene labels, trajectory random seeds, and observation random seeds. It also maps scene labels to scene parameter vectors through a scene mapping function. The scene parameter vectors uniformly organize environmental parameters, anomaly model parameters, lockout parameters, and state machine parameters. The loaded scene parameter set is output to the observation generation unit, degradation judgment unit, and state machine unit as the unified parameter basis for each unit.

[0091] The observation generation unit is used to generate the true trajectory of the underwater target and the USBL simulation observation values. Specifically, the observation generation unit includes a truth generation module and a USBL observation generation module. The truth generation module generates a true sequence of target positions based on trajectory type, duration, spatial scale, and depth parameters, and calculates the true distance, true azimuth, and true pitch angle of the target relative to the USBL receiver at each sampling time to obtain the true geometric observations. The USBL observation generation module estimates the signal-to-noise ratio based on the true geometric observations and generates observation errors. The observation errors are then superimposed on the true geometric observations to generate distance observations, azimuth observations, pitch angle observations, and propagation time observations.

[0092] The degradation determination unit is used to calculate the observation quality score, loss probability, and lockout probability based on the scene parameter set, and sets the lockout probability to 1 when the actual distance exceeds the maximum working range. Specifically, the degradation determination unit includes a degradation modeling module and a lockout determination module. The degradation modeling module calculates the observation quality score and observation loss probability based on the propagation distance, pitch angle, signal-to-noise ratio, sea state intensity, environmental noise, and multipath probability, and determines the loss event based on the observation loss probability. The lockout determination module calculates the lockout probability through a double exponential function after capping the actual distance. When the hard range cutoff is enabled and the actual distance exceeds the maximum working range, the over-range hard limit indicator is set and the lockout probability is set to 1, causing the corresponding sample to enter the invalid candidate logic.

[0093] The state machine unit includes a three-state machine module, which is used to generate candidate observation states and output the final observation state. Specifically, the state machine module generates candidate observation states based on the observation quality score, loss probability, lockout probability, loss event, lockout event, and multipath event, outlier event, and sudden event among the abnormal events output by the degradation judgment module. The candidate observation states are input into the three-state observation state machine. The three-state observation state machine introduces state machine memory variables, state holding time, and recovery and clearing sample number. The current output state is determined based on the combination relationship between the candidate observation state and the final observation state at the previous sampling time. The final observation state, which is in a normal state, a degraded state, or an invalid state, is output, and the state duration, current failure cause, and current degradation source are updated synchronously.

[0094] The output unit is used to generate interpretive labels, calculate locations, perform state statistics, and export datasets. Specifically, the output module includes an interpretive label generation module, a solution output module, a state statistics module, and a data export module. The interpretive label generation module generates report accuracy labels, failure cause labels, degradation source labels, and state duration labels based on observation errors and the final observed state. The solution output module obtains the original back-projection position based on the back projection of USBL simulation observations, outputs position estimates for normal and degraded state samples, and marks the position estimate as null for invalid state samples while retaining the original back-projection position. The state statistics module calculates the state transition matrix, state duration distribution, state proportion, state summary, and maximum continuous duration, and generates state transition diagrams, duration distribution diagrams, and state timeline diagrams. The data export module organizes and exports the files of the geometric truth layer, observation layer, state and interpretation layer, and statistics and organization layer according to scene sample packages, records scene parameters, sample quantity, and file index through manifest index, and divides them into training set, validation set, test set, and out-of-distribution test set.

[0095] The USBL data generation system based on stateful simulation in this embodiment generates three-state observations, interpretive labels, and state statistics simultaneously while generating USBL simulation observations through the collaborative work of the input unit, observation generation unit, degradation judgment unit, state machine unit, and output unit. The output is a simulation dataset that can be directly used for model training, verification, and review.

[0096] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A USBL data generation method based on stateful simulation, characterized in that, include: Obtain trajectory parameters, USBL receiver position, scene label and random seed, and load the corresponding scene parameter group according to the scene label. The scene parameter group includes environmental parameters, anomaly model parameters, lockout parameters and state machine parameters. The true trajectory of the underwater target is generated based on the trajectory parameters. The true distance, true azimuth and true pitch angle are calculated based on the geometric relationship between the underwater target and the USBL receiver. The USBL simulation observation values ​​with superimposed observation errors are generated. Based on the scene parameter set, candidate observation states are generated according to propagation distance, pitch angle, signal-to-noise ratio, loss events, lock-out events, and abnormal events. Among them, the loss event is the event that no usable USBL observation is obtained at the current sampling time, the lock-out event is the event that the acoustic tracking loses lock or exceeds the maximum working range, and the abnormal events include multipath events, outlier events, and burst events. Input the candidate observation state into the three-state observation state machine, and output the final observation state according to the state holding time and recovery conditions; Interpretive labels are generated based on observation errors and final observation states. The location of USBL simulation observations is calculated, and the state transition relationships and state durations are statistically analyzed. A simulation dataset containing observations, observation states, interpretive labels, state statistics, and dataset indexes is then exported.

2. The USBL data generation method based on stateful simulation according to claim 1, characterized in that, The step of loading the corresponding scene parameter group based on the scene tag includes inputting the scene tag into a scene mapping function and outputting a scene parameter vector through the scene mapping function. The scene mapping function establishes a mapping relationship between scene tags and scene parameter vectors through a preset table, a configuration dictionary, or parameter generation rules. The scene parameter vector is represented as follows: ， in, For scene tags, This is the mapping function from scene labels to scene parameter vectors. For scene tags The corresponding scene parameter vector, For sea state intensity, For multipath strength, For sound speed deviation, For environmental noise gain, For the abnormal model parameter group, For the unlock parameter group, This is the state machine parameter set.

3. The USBL data generation method based on stateful simulation according to claim 1, characterized in that, The USBL simulation observations that generate superimposed observation errors include determining the target motion form based on trajectory type, duration, spatial scale, and depth parameters, and initializing the trajectory generation process based on a trajectory random seed to generate a true sequence of target positions. The true distance, true azimuth, and true elevation angle of the target relative to the USBL receiver are calculated in the true value sequence of the target position at each sampling time to obtain the true geometric observations. The signal-to-noise ratio (SNR) is estimated based on the actual distance and pitch angle. The estimated SNR is then used to generate observation errors, which are then superimposed onto the actual geometric observations to obtain USBL simulation observations.

4. The USBL data generation method based on stateful simulation according to claim 3, characterized in that, The process of generating observation errors based on signal-to-noise ratio estimates includes calculating the fundamental error based on reference noise and range-related noise, amplifying and modulating the fundamental error using a low pitch angle error amplification factor, a signal-to-noise ratio error amplification factor, and anomaly event error amplification factor, determining the standard deviations of range observation errors, azimuth observation errors, and pitch angle observation errors, and generating corresponding observation errors based on each standard deviation. The standard deviation of the range observation error is expressed as follows: ， in, This represents the standard deviation of the distance observation error. The standard deviation of the noise level from the reference level. The slope of the distance-related noise. This is the error amplification factor caused by low pitch angles. This is the amplification factor of the signal-to-noise ratio on the distance error. This is the error amplification factor for abnormal events triggered by outlier events or sudden events; The standard deviations of the azimuth and elevation observation errors are respectively expressed as: ， ， in, and These are the standard deviations of the azimuth observation error and the standard deviations of the elevation observation error, respectively. and These are the reference noise standard deviations for azimuth and elevation angles, respectively. and These are the slope terms for azimuth and elevation errors as a function of distance, respectively. and These are the amplification factors of the signal-to-noise ratio for the azimuth and elevation angle errors, respectively. The corresponding point-by-point random observation error is generated based on the standard deviation of each observation error, and is expressed as follows: ， in, Indicates the observation dimension as distance, azimuth, or elevation. For distance observation dimension, For azimuth observation dimension, For the perspective of pitch angle observation dimension, Indicates time Standard deviation of observation error for the corresponding observation dimension Let be a standard normal random variable generated based on an observational random seed. For a moment Point-by-point random observation error corresponding to the observation dimension.

5. The USBL data generation method based on stateful simulation according to claim 1, characterized in that, The process of generating candidate observation states includes calculating the observation loss probability based on the actual distance, pitch angle, signal-to-noise ratio, sea state intensity, and sudden events, and comparing the observation loss probability with a random number to determine the loss event. The upper limit of the range is calculated based on the probability of loss of lock and capped. The probability of loss of lock is calculated using a double exponential function based on the capped distance. When the maximum working range limit is enabled and the actual distance exceeds the maximum working range, the probability of loss of lock is set to 1. The loss of lock event is determined based on the probability of loss of lock and the corresponding sample is entered into the invalid candidate logic. The probability of losing the lock is expressed as: ， ， in, This is the capping distance used for calculating the probability of loss of lock. , , and These represent the coefficients of the lock loss probability model. Let be the probability of losing the lock at time t. For the amplitude limiting function, This is an over-range hard limit indication. For indicator functions, The actual distance at time t. This is the maximum working range.

6. The USBL data generation method based on stateful simulation according to claim 5, characterized in that, The generation of candidate observation states also includes generating candidate invalidity determination quantities based on the comparison results of loss events, lock-out events, quality scores, signal-to-noise ratio, loss probability, and lock-out probability with invalidity state thresholds; When the candidate invalidity determination quantity is not valid, a candidate degradation determination quantity is generated based on the comparison results of multipath events, burst events, outlier events, quality scores, signal-to-noise ratio, loss probability, and lockout probability with the degradation state threshold. Among them, multipath events are events that generate additional distance and angle biases, outlier events are events that amplify the observation error of a single sampling moment by a preset multiple, and burst events are events that cause abnormal disturbances at multiple consecutive sampling moments. When the candidate invalidity criterion is true, the invalid state is taken as the candidate observation state; when the candidate invalidity criterion is false but the candidate degradation criterion is true, the degradation state is taken as the candidate observation state; when neither the candidate invalidity criterion nor the candidate degradation criterion is true, the normal state is taken as the candidate observation state. The candidate invalidity criterion is expressed as follows: ， in, Let be the candidate invalidation criterion at time t. For the lost event, This is a lockout event. To rate the quality, This is the signal-to-noise ratio estimate. To observe the probability of loss, This represents the probability of losing the lock. , , and These are the quality, signal-to-noise ratio, loss probability, and lockout probability thresholds corresponding to the invalid state; The candidate degradation determination quantity is expressed as: ， in, For a moment Candidate degradation criteria. , and These are multipath events, burst events, and outlier events. Represents logical OR, AND represents logical AND, NOT represents logical NOT. , , and These are the quality score, signal-to-noise ratio, loss probability, and lockout probability threshold corresponding to the degraded state.

7. The USBL data generation method based on stateful simulation according to claim 1, characterized in that, The three-state observation state machine introduces state machine memory variables, state holding time, and recovery clear sample number. The current output state is determined based on the candidate observation state, the final observation state at the previous sampling time, and the state machine memory variables, so that the degenerate state and invalid state form a continuous state segment with a phased recovery path. The state machine memory variables include the remaining holding length, the number of consecutive clear samples, the current failure cause, and the current degradation source. The three-state observation state machine is represented as follows: ， in, The final observed state at time t, and Let represent the state machine memory variables at time t and the previous sampling time, respectively. For state machine functions, To maintain length in the degenerate state, To preserve the length of the invalid state, The number of samples to be cleared to restore to the degraded state, The number of samples to be cleared to restore to normal. This is a candidate observation state.

8. The USBL data generation method based on stateful simulation according to claim 7, characterized in that, The process of determining the current output state based on the candidate observation state, the final observation state at the previous sampling time, and the state machine memory variables is as follows: when When the state is invalid, output the invalid state and keep the remaining length set to 0. The sample count is continuously cleared to zero, and the current failure cause and current degradation source are recorded. when When it is in a degenerate state, if If the state is invalid and the remaining length is greater than 0, then maintain the invalid output and decrease the remaining length. If the state is otherwise, output the degenerate state and keep the remaining length set to 0. ; when In the normal state, the maintenance period continues until the end of the maintenance period. Output: After the retention period ends, the number of continuously removed samples reaches [a certain value]. The system returned to normal after a period of time, with the number of samples continuously cleared reaching [a certain threshold]. However, it was not achieved. It will eventually revert to its degenerate state; when The normal state and the previous state If the state is already normal, then output "normal state".

9. The USBL data generation method based on stateful simulation according to claim 1, characterized in that, The explanatory labels include report accuracy labels, failure cause labels, degradation source labels, and state duration labels; The report accuracy label is calculated based on the standard deviation of the distance error and the error contribution after the azimuth and pitch angle errors are converted to the distance scale. The failure cause label is determined based on the over-range hard limit indication, the loss of lock event, the missing event, the quality score, the signal-to-noise ratio, the loss probability, and the loss of lock probability. The degradation source label is determined based on the multipath event, the sudden event, the outlier event, the missing event, the loss of lock event, the quality score, the signal-to-noise ratio, the loss probability, and the loss of lock probability. The state duration label is determined based on the number of continuous sampling points in the final observation state.

10. A USBL data generation system based on stateful simulation, executing the method of claim 1, characterized in that, include: The input unit is used to obtain trajectory parameters, USBL receiver position, scene label and random seed, and load the corresponding scene parameter group according to the scene label; The observation generation unit is used to generate the true trajectory of the underwater target based on the trajectory parameters, calculate the true distance, true azimuth and true pitch angle based on the geometric relationship between the underwater target and the USBL receiver, and generate USBL simulation observation values ​​with superimposed observation errors. The degradation determination unit is used to calculate the observation quality score, loss probability and lockout probability based on the scene parameter set, and set the lockout probability to 1 when the actual distance exceeds the maximum working range; The state machine unit is used to generate candidate observation states based on observation quality score, signal-to-noise ratio, loss probability, lockout probability, loss event, lockout event, and abnormal events. The abnormal events include multipath events, outlier events, and burst events. The candidate observation states are input into the three-state observation state machine, and the final observation state is output based on the state holding time and recovery conditions. The output unit is used to generate interpretive labels based on observation errors and final observation states, perform position calculations on USBL simulation observations, perform statistics on state transition relationships and state durations, and export a simulation dataset containing observations, observation states, interpretive labels, state statistics data, and dataset indexes.

Images