Marine launching platform attitude sensing method based on machine learning and particle filtering

Abstract

The invention discloses a sea launching platform attitude sensing method based on machine learning and particle filtering, and the method comprises the steps: constructing a sea condition and sensor state vector; generating an attitude dynamics residual sequence; training an improved Diffusion model of the physical constraint; deploying the trained improved Diffusion model to an edge computing device, and outputting attitude increment prior distribution, observation error observation likelihood distribution and a score function through a few-step sampling mode; obtaining a particle prediction set; obtaining a weighted particle set; generating a resampled particle set; and calculating a weighted attitude estimation mean value and variance based on the resampled particle set, and outputting an offshore platform attitude estimation result, a particle diversity index and a marginal likelihood trend. According to the method, directional sampling of particles on the SO (3) manifold along a high-probability region is realized, particle degradation is remarkably reduced, and the sampling efficiency and estimation precision in a high-dimensional complex environment are improved.

Description

Technical Field

[0001] This invention relates to the field of marine launch platform technology, and in particular to a method for attitude perception of marine launch platforms based on machine learning and particle filtering. Background Technology

[0002] With the continuous development of marine engineering, high-precision attitude estimation and dynamic control of offshore platforms under various disturbance environments have become key requirements for ensuring platform safety and operational efficiency.

[0003] Traditional extended Kalman filtering and unscented Kalman filtering generally assume that the system state and observation noise follow a Gaussian distribution. This assumption makes it difficult to accurately characterize the non-Gaussianity and multimodal distribution characteristics of attitude changes under strong sea conditions and multi-source sensor anomalies. When the platform encounters complex scenarios, existing filtering algorithms often fail to accurately capture the tail, outliers and multimodal structures of the actual distribution, and are prone to estimation divergence and short-term instability.

[0004] While existing particle filtering methods possess the ability to model non-Gaussian and nonlinear phenomena, they are prone to particle degradation in high-dimensional attitude manifolds and when observational information is scarce, leading to a rapid decrease in the effective particle number and low particle sampling efficiency. Furthermore, most existing particle filters employ fixed proposal distributions and static noise assumptions, lacking the ability to adapt to changes in sea state, sensor health status, and observational anomalies. This results in a mismatch between model parameters and the actual operating environment, leading to decreased estimation accuracy over long-term operation and increasing the burden of manual maintenance and system tuning. Summary of the Invention

[0005] One objective of this invention is to propose an attitude perception method for a marine launch platform based on machine learning and particle filtering. This invention enables directional sampling of particles along high-probability regions on the SO(3) manifold, significantly reducing particle degradation and improving sampling efficiency and estimation accuracy in high-dimensional complex environments.

[0006] An attitude perception method for a marine launch platform based on machine learning and particle filtering according to an embodiment of the present invention includes:

[0007] Collect and preprocess multi-source raw observation datasets to construct sea state and sensor state vectors;

[0008] Based on the rigid body six-degree-of-freedom kinematic model, an attitude dynamics model of the offshore platform is established on the Lie group SO(3) manifold, and the attitude increment residual is predicted by using the sea state and sensor state vectors to generate the attitude dynamics residual sequence.

[0009] The improved Diffusion model with physical constraints is trained by using sea state and sensor state vectors, attitude dynamics residual sequences, and synchronized multi-source raw observation datasets as conditional inputs. This allows the improved Diffusion model to simultaneously learn the prior distribution of the attitude increment of the offshore platform and the observation likelihood distribution of the multi-source observation error.

[0010] The trained improved Diffusion model is deployed to an edge computing device, and the pose increment prior distribution, observation error observation likelihood distribution, and score function are output through a few-step sampling method.

[0011] The particle set of the adaptive particle filter is initialized based on the prior distribution of attitude increment, and the particle propagation along the high-probability manifold direction of attitude is guided by the scoring function. At the same time, the attitude dynamics model of the offshore platform is used for particle prediction update to obtain the particle prediction set.

[0012] The particle weights of the predicted particle set are calculated by using the observation error likelihood distribution, real-time multi-source observation, and robust loss function, thus obtaining a weighted particle set.

[0013] The effective number of particles in the weighted particle set is calculated. When the effective number of particles is lower than the threshold, an adaptive resampling mechanism is triggered. The particle number, resampling frequency, process noise covariance parameter and observation noise covariance parameter are adaptively adjusted according to the marginal likelihood value and observation residual statistics to generate a resampled particle set.

[0014] The weighted mean and variance of attitude estimation are obtained based on the resampled particle set, and the attitude estimation results of the offshore platform, particle diversity index and marginal likelihood trend are output.

[0015] Optionally, the construction of the sea state and sensor state vector includes:

[0016] Data from offshore platform inertial measurement units, global navigation satellite systems, visual sensors, wave radar, and meteorological buoys are collected and combined with collection timestamps to generate a multi-source raw observation dataset. The multi-source raw observation dataset is then synchronized, denoised, has outlier removed, and missing data is filled in to construct sea state and sensor state vectors.

[0017] Optionally, the generation of the attitude dynamics residual sequence includes:

[0018] Using the attitude dynamics prediction model, the unit quaternion at the current moment is calculated based on the unit quaternion at the previous moment, the angular velocity vector at the previous moment, and the sampling period parameter. The predicted unit quaternion at the current moment represents the attitude prediction result of the offshore platform based on the dynamics model.

[0019] The observation unit quaternion estimated by fusing multiple source raw observation datasets and the current time unit quaternion predicted by the attitude dynamics prediction model are multiplied and conjugately operated to calculate the attitude dynamics residual quaternion.

[0020] The attitude dynamics residual quaternion is transformed into a three-dimensional rotation vector through a logarithmic mapping operation from Lie group to Lie algebra.

[0021] Repeat the steps to calculate the three-dimensional rotation vectors at all times in the time dimension, and combine the three-dimensional rotation vectors at all times in time order to form an attitude dynamics residual sequence.

[0022] Optionally, the improved Diffusion model includes:

[0023] For each sampling moment, the sea state, sensor state vector, attitude dynamics residual vector, and multi-source raw observation dataset that has been synchronously processed are merged to generate a conditional input sample. Conditional input samples are collected at all moments within a certain period of time to form a training dataset.

[0024] For each moment in the training dataset, the attitude increment quaternion is extracted as a parameter to describe the attitude change of the offshore platform. The attitude dynamics residual vector is restored to the attitude increment quaternion through the transformation relationship from Lie group to Lie algebra. At the same time, the multi-source observation error vector at the corresponding moment is calculated as a parameter to describe the observation residual. The attitude increment quaternion and the multi-source observation error vector are concatenated to form the target variable.

[0025] For each target variable, Gaussian noise is added step by step according to the set number of diffusion steps and noise scheduling parameters to obtain the diffusion state at each step;

[0026] Construct a scoring network with conditional input, and obtain the score matching loss by calculating the mean square error between the output of the scoring network and the actual added noise;

[0027] During training, a physical constraint regularization term is introduced, which is calculated by the squared difference between the magnitude of the quaternion of the attitude increment at all times in the statistical training dataset and 1.

[0028] The score matching loss is weighted and summed with the physical constraint regularization term to obtain the total loss function. The parameters of the improved Diffusion model are then updated by iteratively optimizing the total loss function using the stochastic gradient descent algorithm, resulting in the trained improved Diffusion model.

[0029] Optionally, the step of using the offshore platform attitude dynamics model for particle prediction updates to obtain a particle prediction set includes:

[0030] The current conditional input sample is fed into the trained improved Diffusion model to generate the current pose increment prior distribution and score function;

[0031] Set the total number of particles, initialize the particle set of the adaptive particle filter, and initialize the particle state starting from the particle attitude state of the previous moment. By mapping the change trend of the score function in the attitude space to the Lie group space after weighted superposition with the diffusion noise, and combining it with the particle attitude state of the previous moment according to the quaternion multiplication rule, the particle attitude state of the current moment is obtained.

[0032] The particle attitude state obtained by guided propagation based on the scoring function is weighted and fused with the attitude prediction result based on the dynamic model using the unit quaternion at the current time, and then normalized to obtain the particle prediction set.

[0033] Optionally, the step of calculating the particle weights of the particle prediction set using observation error likelihood distribution, real-time multi-source observation, and robust loss function includes:

[0034] For each particle in the particle prediction set, the particle's attitude state is mapped to a particle observation value using the observation model. The particle observation value is then subtracted from the real-time multi-source observation vector to obtain the observation residual vector.

[0035] The improved Diffusion model, which has been trained, is invoked. Based on the conditional input samples at the current time, the observation likelihood distribution of the observation error is obtained. The observation residual vector of the particle is substituted into the observation likelihood distribution model of the observation error to obtain the basic weight of the particle.

[0036] A robust loss function is introduced for the observation residual vector of each particle. The robust loss function judges whether the absolute value of each component of the observation residual is less than a threshold. When the absolute value of the observation residual is less than the threshold, the squared loss is used; otherwise, the linear loss is used minus a constant. The loss results of all components are summed to obtain the total robust loss value of the particle. The negative number of the total robust loss value is then exponentially calculated to obtain the robust scaling factor. The base weights are multiplied by the robust scaling factor to obtain the weighted particle weights.

[0037] The weighted particle weights of all particles are normalized. After normalization, each particle is associated with its particle attitude state and the corresponding normalized particle weight, forming a weighted particle set.

[0038] Optionally, generating the resampled particle set includes:

[0039] Calculate the number of effective particles at the current moment based on the normalized weights of all particles in the weighted particle set;

[0040] Set an effective particle count threshold. When the effective particle count at the current moment is less than the effective particle count threshold, trigger the adaptive resampling mechanism. The adaptive resampling mechanism uses the system resampling method to resample the particle set and obtain a new particle set.

[0041] The marginal likelihood value for the current fusion cycle is calculated by multiplying and summing the normalized weights of all particles with the probability density of the observation likelihood distribution of the observation error.

[0042] A dynamic adjustment function based on the marginal likelihood trend is constructed, which calculates the particle number adjustment and the resampling frequency adjustment respectively.

[0043] The total number of particles should be adjusted based on the current fusion status and historical statistical results. The adjusted number of particles should be limited to the lower limit and the upper limit of the particle number.

[0044] Update the particle process noise covariance matrix and the observation noise covariance matrix;

[0045] The newly generated set of particles is used as the resampled set of particles output in the current fusion cycle.

[0046] Optionally, the output of the offshore platform attitude estimation results, particle diversity index, and marginal likelihood trend includes:

[0047] Calculate the weighted mean attitude estimate of all particles in the resampled particle set, and normalize the weighted summation result to obtain the weighted mean attitude estimate.

[0048] By squaring the geodesic distance between the attitude state of each particle and the weighted attitude estimation mean, multiplying the squared distance value by the normalized particle weight of the particle, and summing the results for all particles, the weighted variance of the resampled particle set on the SO(3) manifold is obtained.

[0049] Calculate the diversity index of the particle set in the current fusion cycle;

[0050] By multiplying the normalized particle weight of each particle by the probability density of the observation error likelihood distribution of the particle, and summing the results for all particles, the marginal likelihood trend of the current fusion cycle is obtained.

[0051] The beneficial effects of this invention are:

[0052] (1) This invention combines an improved Diffusion model with an adaptive particle filter in the field of attitude fusion for offshore platforms, which significantly improves the robustness, accuracy and adaptability of the system under strong non-Gaussian sea conditions and multi-source observation anomalies. By proposing a new attitude estimation framework that deeply integrates the Diffusion model based on physical constraints and conditions with the adaptive particle filter algorithm, the dependence of traditional Bayesian filtering on Gaussian assumptions and fixed noise models is effectively alleviated. By introducing an improved Diffusion model based on sea conditions and sensor states, the attitude increment distribution and multi-source observation error distribution of offshore platforms are dynamically modeled. The score function and uncertainty output by the Diffusion model are used to shape the proposal distribution and particle weights of the particle filter in real time, so that the particles are oriented to sample along the high-probability region on the SO(3) manifold, which significantly reduces particle degradation and improves the sampling efficiency and estimation accuracy in high-dimensional complex environments.

[0053] (2) This invention realizes online adaptive control of particle filters and dynamic optimization of noise parameters, which significantly improves the system's adaptive capability and maintenance efficiency during long-term operation. The adaptive resampling and noise covariance online adjustment mechanism can automatically adjust the number of particles, resampling frequency, process noise and observation noise parameters according to the current effective number of particles, marginal likelihood trend and observation residual statistics. It does not require frequent manual intervention and can effectively cope with seasonal changes in sea state, changes in platform load and sensor performance degradation, as well as dynamic changes in environment and equipment, ensuring that the system always operates under the optimal parameter configuration, significantly extending the system maintenance cycle and reducing operation and maintenance costs. Attached Figure Description

[0054] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0055] Figure 1 This is a flowchart of an attitude perception method for a marine launch platform based on machine learning and particle filtering proposed in this invention. Detailed Implementation

[0056] Example 1: Reference Figure 1 An attitude perception method for a marine launch platform based on machine learning and particle filtering includes:

[0057] Collect and preprocess multi-source raw observation datasets to construct sea state and sensor state vectors;

[0058] In this embodiment, the construction of sea state and sensor state vectors includes:

[0059] Data from offshore platform inertial measurement units, global navigation satellite systems, visual sensors, wave radar, and meteorological buoys are collected and combined with collection timestamps to generate a multi-source raw observation dataset. The multi-source raw observation dataset is then synchronized, denoised, has outlier removed, and missing data is filled in to construct sea state and sensor state vectors.

[0060] Based on the rigid body six-degree-of-freedom kinematic model, an attitude dynamics model of the offshore platform is established on the Lie group SO(3) manifold, and the attitude increment residual is predicted by using the sea state and sensor state vectors to generate the attitude dynamics residual sequence.

[0061] In this embodiment, generating the attitude dynamics residual sequence includes:

[0062] Using the attitude dynamics prediction model, the unit quaternion at the current moment is calculated based on the unit quaternion at the previous moment, the angular velocity vector at the previous moment, and the sampling period parameter. The predicted unit quaternion at the current moment represents the attitude prediction result of the offshore platform based on the dynamics model.

[0063] ;

[0064] in, Indicates the posture from the previous moment and angular velocity Predicted current pose The sampling period is This represents the exponential mapping operation from Lie algebras to Lie groups. It represents the product of quaternions.

[0065] The attitude state of the offshore platform at the current moment is represented as a unit quaternion. The unit quaternion includes a scalar part and three vector parts. The unit quaternion is used to describe the attitude state of the offshore platform on the SO(3) manifold. The unit quaternion needs to satisfy the constraint that the modulus is equal to 1.

[0066] The six-degree-of-freedom motion state of the offshore platform under the disturbance of wind, waves and current is modeled by rigid body dynamics to obtain the angular velocity components in three directions, namely roll angular velocity, pitch angular velocity and yaw angular velocity. The angular velocity components in the three directions are combined into an angular velocity vector, which describes the rate of rotation of the offshore platform on the SO(3) manifold.

[0067] The observation unit quaternion estimated by fusing multiple source raw observation datasets and the current time unit quaternion predicted by the attitude dynamics prediction model are multiplied and conjugately operated to calculate the attitude dynamics residual quaternion.

[0068] ;

[0069] in, Represents the attitude dynamics residual quaternion. The observation pose is estimated by fusing multi-source raw observation datasets after synchronous processing. To predict attitude for the dynamic model, The conjugate quaternion is represented by the attitude dynamics residual quaternion, which reflects the relative rotational relationship between the observation results and the dynamic prediction results on the SO(3) manifold.

[0070] The attitude dynamics residual quaternion is transformed into a three-dimensional rotation vector through a logarithmic mapping operation from Lie group to Lie algebra.

[0071] The three-dimensional rotation vector describes the deviation between attitude prediction and observation in rotation space in a minimal parameter form.

[0072] Repeat the steps to calculate the three-dimensional rotation vectors at all times in the time dimension, and combine the three-dimensional rotation vectors at all times in time order to form an attitude dynamics residual sequence.

[0073] The attitude dynamics residual sequence represents the dynamic characteristics of the deviation between the attitude prediction and actual observation of an offshore platform at different times.

[0074] The improved Diffusion model with physical constraints is trained by using sea state and sensor state vectors, attitude dynamics residual sequences, and synchronized multi-source raw observation datasets as conditional inputs. This allows the improved Diffusion model to simultaneously learn the prior distribution of the attitude increment of the offshore platform and the observation likelihood distribution of the multi-source observation error.

[0075] In this embodiment, the Diffusion model is improved, including:

[0076] For each sampling moment, the sea state, sensor state vector, attitude dynamics residual vector, and multi-source raw observation dataset that has been synchronously processed are merged to generate a conditional input sample. Conditional input samples are collected at all moments within a certain period of time to form a training dataset.

[0077] For each moment in the training dataset, the attitude increment quaternion is extracted as a parameter to describe the attitude change of the offshore platform. The attitude dynamics residual vector is restored to the attitude increment quaternion through the transformation relationship from Lie group to Lie algebra. At the same time, the multi-source observation error vector at the corresponding moment is calculated as a parameter to describe the observation residual. The attitude increment quaternion and the multi-source observation error vector are concatenated to form the target variable.

[0078] For each target variable, Gaussian noise is added step by step according to the set number of diffusion steps and noise scheduling parameters to obtain the diffusion state at each step;

[0079] The diffusion state is used to allow the model to learn the pose increment quaternion and the statistical distribution of the multi-source observation error vector under multi-step noise conditions during training.

[0080] Construct a scoring network with conditional input, and obtain the score matching loss by calculating the mean square error between the output of the scoring network and the actual added noise;

[0081] The scoring network takes the current diffusion state, the number of diffusion steps, and the corresponding conditional input samples as input and outputs the estimated noise components as output. The score matching loss measures the model’s ability to learn noise under different noise intensities for attitude increment quaternions and multi-source observation error vectors.

[0082] During training, a physical constraint regularization term is introduced, which is calculated by the squared difference between the magnitude of the quaternion of the attitude increment at all times in the statistical training dataset and 1.

[0083] The physical constraint regularization term is used to constrain the magnitude of the attitude increment quaternion to be close to 1.

[0084] The score matching loss is weighted and summed with the physical constraint regularization term to obtain the total loss function. The parameters of the improved Diffusion model are then updated by iteratively optimizing the total loss function using the stochastic gradient descent algorithm, resulting in the trained improved Diffusion model.

[0085] The trained improved Diffusion model is deployed to an edge computing device, and the pose increment prior distribution, observation error observation likelihood distribution, and score function are output through a few-step sampling method.

[0086] The particle set of the adaptive particle filter is initialized based on the prior distribution of attitude increment, and the particle propagation along the high-probability manifold direction of attitude is guided by the scoring function. At the same time, the attitude dynamics model of the offshore platform is used for particle prediction update to obtain the particle prediction set.

[0087] In this embodiment, the attitude dynamics model of the offshore platform is used for particle prediction updates to obtain a particle prediction set, including:

[0088] The current conditional input sample is fed into the trained improved Diffusion model to generate the current pose increment prior distribution and score function;

[0089] The attitude increment prior distribution represents the probability distribution of attitude changes of an offshore platform, and the scoring function represents the trend of the attitude increment prior distribution in the attitude space.

[0090] Set the total number of particles, initialize the particle set of the adaptive particle filter, and initialize the particle state starting from the particle attitude state of the previous moment. By mapping the change trend of the score function in the attitude space to the Lie group space after weighted superposition with the diffusion noise, and combining it with the particle attitude state of the previous moment according to the quaternion multiplication rule, the particle attitude state of the current moment is obtained.

[0091] ;

[0092] in, Step size factor For particle diffusion noise, Given the current particle state, This represents the particle state at the previous moment. This represents the quaternion representing the attitude increment at the current moment. To improve the final parameters of the Diffusion model.

[0093] The particle attitude state obtained by guided propagation based on the scoring function is weighted and fused with the attitude prediction result based on the dynamic model using the unit quaternion at the current time, and then normalized to obtain the particle prediction set.

[0094] The particle weights of the predicted particle set are calculated by using the observation error likelihood distribution, real-time multi-source observation, and robust loss function, thus obtaining a weighted particle set.

[0095] In this embodiment, the particle weights of the particle prediction set are calculated using observation error likelihood distribution, real-time multi-source observation, and robust loss function, including:

[0096] For each particle in the particle prediction set, the particle's attitude state is mapped to a particle observation value using the observation model. The particle observation value is then subtracted from the real-time multi-source observation vector to obtain the observation residual vector.

[0097] The observation model is used to map the particle's attitude state to inertial measurement unit (IMU) observations, global navigation satellite system (GNSS) observations, and visual sensor observations. Specifically, it includes: based on the sensor installation parameters of the offshore platform, the sensor measurement principle, and the observation geometry, a system identification method based on sensor output and known attitude data is used to establish a mapping relationship from attitude quaternions to the outputs of each sensor. The mapping relationship remains consistent during the fusion process. During fusion, for each particle, the particle's attitude state is substituted into the observation model, and the theoretical observation value of the particle is calculated through the mapping relationship. The theoretical observation values ​​correspond to IMU observations, GNSS observations, and visual sensor observations, respectively.

[0098] The improved Diffusion model, which has been trained, is invoked. Based on the conditional input samples at the current time, the observation likelihood distribution of the observation error is obtained. The observation residual vector of the particle is substituted into the observation likelihood distribution model of the observation error to obtain the basic weight of the particle.

[0099] The base weights are used to measure the probability that the observation residuals of a particle will occur under the current observation conditions.

[0100] A robust loss function is introduced for the observation residual vector of each particle. The robust loss function judges whether the absolute value of each component of the observation residual is less than a threshold. When the absolute value of the observation residual is less than the threshold, the squared loss is used; otherwise, the linear loss is used minus a constant. The loss results of all components are summed to obtain the total robust loss value of the particle. The negative number of the total robust loss value is then exponentially calculated to obtain the robust scaling factor. The base weights are multiplied by the robust scaling factor to obtain the weighted particle weights.

[0101] ;

[0102] in, For robust loss function, For the first Dimensional residual components, The threshold value is used.

[0103] The weighted particle weights of all particles are normalized. After normalization, each particle is associated with its particle attitude state and the corresponding normalized particle weight, forming a weighted particle set.

[0104] The effective number of particles in the weighted particle set is calculated. When the effective number of particles is lower than the threshold, an adaptive resampling mechanism is triggered. The particle number, resampling frequency, process noise covariance parameter and observation noise covariance parameter are adaptively adjusted according to the marginal likelihood value and observation residual statistics to generate a resampled particle set.

[0105] In this embodiment, generating the resampled particle set includes:

[0106] Calculate the number of effective particles at the current moment based on the normalized weights of all particles in the weighted particle set;

[0107] The effective particle count is used to measure the representative number of particles with different weights in a particle set. The effective particle count is obtained by summing the squares of the normalized weights of all particles and taking the reciprocal.

[0108] Set an effective particle count threshold. When the effective particle count at the current moment is less than the effective particle count threshold, trigger the adaptive resampling mechanism. The adaptive resampling mechanism uses the system resampling method to resample the particle set and obtain a new particle set.

[0109] The marginal likelihood value for the current fusion cycle is calculated by multiplying and summing the normalized weights of all particles with the probability density of the observation likelihood distribution of the observation error.

[0110] Marginal likelihood measures the overall degree of match between the current particle set and the real-time observation.

[0111] A dynamic adjustment function based on the marginal likelihood trend is constructed, which calculates the particle number adjustment and the resampling frequency adjustment respectively.

[0112] A dynamic adjustment function based on marginal likelihood trend is constructed. The dynamic adjustment function calculates the particle number adjustment by comparing the marginal likelihood value of the current fusion cycle with the expected marginal likelihood value. The particle number adjustment is equal to the difference between the expected marginal likelihood value and the current marginal likelihood value multiplied by the particle number adjustment ratio coefficient. The dynamic adjustment function calculates the resampling frequency adjustment by comparing the current effective particle number with the expected effective particle number. The resampling frequency adjustment is equal to the difference between the expected effective particle number and the current effective particle number multiplied by the resampling frequency adjustment ratio coefficient.

[0113] The total number of particles should be adjusted based on the current fusion status and historical statistical results. The adjusted number of particles should be limited to the lower limit and the upper limit of the particle number.

[0114] Update the particle process noise covariance matrix and the observation noise covariance matrix;

[0115] The particle process noise covariance matrix is ​​obtained by adding the particle attitude state variance multiplied by the noise adjustment factor to the particle process noise covariance matrix at the previous time step. The observation noise covariance matrix is ​​obtained by adding the observation residual covariance multiplied by the noise adjustment factor to the observation noise covariance matrix at the previous time step. The particle attitude state variance and the observation residual covariance are both the variances or covariances of the particle set and the observation residual set.

[0116] The newly generated set of particles is used as the resampled set of particles output in the current fusion cycle.

[0117] The weighted mean and variance of attitude estimation are obtained based on the resampled particle set, and the attitude estimation results of the offshore platform, particle diversity index and marginal likelihood trend are output.

[0118] In this embodiment, the output of the offshore platform attitude estimation results, particle diversity index, and marginal likelihood trend includes:

[0119] Calculate the weighted mean attitude estimate of all particles in the resampled particle set, and normalize the weighted summation result to obtain the weighted mean attitude estimate.

[0120] The weighted attitude estimation mean is used to measure the overall attitude estimation result of the current fusion cycle. The weighted attitude estimation mean is obtained by multiplying the attitude state of each particle by its normalized particle weight and then summing the results.

[0121] By squaring the geodesic distance between the attitude state of each particle and the weighted attitude estimation mean, multiplying the squared distance value by the normalized particle weight of the particle, and summing the results for all particles, the weighted variance of the resampled particle set on the SO(3) manifold is obtained.

[0122] Weighted variance is used to measure the uncertainty of the attitude estimation results in the current fusion cycle.

[0123] Calculate the diversity index of the particle set in the current fusion cycle;

[0124] The diversity index is used to measure the uniformity of the weight distribution of a particle set and reflects the diversity of the particle set. The diversity index is obtained by taking the logarithm of the normalized particle weight of each particle, multiplying the normalized particle weight by its logarithm, summing the results for all particles and taking the negative number.

[0125] By multiplying the normalized particle weight of each particle by the probability density of the observation error likelihood distribution of the particle, the results of all particles are accumulated to obtain the marginal likelihood trend of the current fusion cycle.

[0126] Marginal likelihood trend is used to measure the overall degree of fit between the current particle set and the real-time observation.

[0127] Example 2: In the actual operating environment of a certain offshore platform, the platform is equipped with multi-source observation equipment such as an inertial measurement unit, a global navigation satellite system, and visual sensors. During the platform's operation, a total of 132,000,000 frames of raw multi-source observation datasets were collected, including IMU acceleration and angular velocity data, GNSS positioning information, visual sensor attitude estimates, and sea state parameters from meteorological buoys. All data were automatically aligned and stored on a local industrial server through a high-precision time synchronization system.

[0128] During the platform's stable operation phase, the system trained the Diffusion model on 40,000,000 synchronous observation samples. The input features were: effective wave height 0.9m, main wave period 9.6s, IMU zero-bias drift rate 0.002~0.008rad / s, and visual occlusion rate 0.8%~3.1%. During training, the mean absolute value of the platform attitude dynamics residual sequence was 0.11rad. After 72 rounds of model convergence, the root mean square error of the Diffusion model on the training set was 0.023rad, and the root mean square error on the test set was 0.026rad. In this phase, the adaptive particle filter was initialized with 768 particles, and the mean particle weighted entropy distribution was 3.48 bits.

[0129] During the subsequent environmental abrupt change phase, the system detected a sharp increase in the significant wave height to 4.8m, and the main wave period fluctuation range narrowed to 3.4s~4.1s. The zero-bias drift rate in the IMU sampling data surged to 0.027rad / s, the visual sensor frame loss rate jumped to 19.7%, and the GNSS instantaneous signal loss rate reached 12%. The proportion of observational anomalies in the synchronous observation samples reached 4.5%, and the mean absolute value of the attitude dynamics residual sequence rose to 0.42rad.

[0130] The method of the present invention behaves as follows at this stage:

[0131] After detecting the observational anomaly, the Diffusion model dynamically adjusted the observational likelihood distribution, causing the effective particle number of the particle filter to drop to 52. The system then automatically increased the particle size to 2048, adaptively improving the process noise variance to 0.013 rad² and the observation noise variance to 0.019 rad².

[0132] The weighted variance of the particle prediction ensemble was controlled below 0.25 rad², the mean of the marginal likelihood trend curve decreased to 0.19, and the standard deviation of the fluctuation was 0.022.

[0133] The weighted attitude estimation mean was compared with the true value of multi-source observation fusion, and the absolute deviation was 0.34°, with a maximum peak value of no more than 0.81°.

[0134] The particle diversity index (weighted entropy) remained stable above 2.94 bits, without any short-term divergence.

[0135] No human intervention occurred throughout the entire process. The anomaly detection system output a health report every hour, and all reports showed "stable".

[0136] For ease of comparison, both the traditional particle filter algorithm and the unscented Kalman filter algorithm were implemented simultaneously. Under the same dataset and the same sensor anomaly conditions, the traditional particle filter showed that during the anomaly observation phase, the number of effective particles dropped to a minimum of 21, the particle weighted variance increased to 0.72 rad², the mean marginal likelihood trend decreased to 0.08, the cumulative short-term divergence duration was 4.7 minutes, and the average absolute deviation of the weighted attitude estimation mean from the true value was 1.03°, with a maximum deviation as high as 2.26°. During the period of overlapping anomalies from multiple sources, the unscented Kalman filter algorithm showed 8 instantaneous jumps in attitude estimation, with the maximum duration of a single divergence reaching 41 seconds, and an average deviation of 1.49° in the mean estimation.

[0137] After the platform entered the equipment aging and sensor performance degradation stage, the IMU zero-bias drift rate fluctuation range increased to 0.032~0.045 rad / s, the GNSS satellite loss rate reached 21.2%, and the visual obstruction rate exceeded 9.8%. The proportion of observational anomalies in the synchronous observation samples increased to 11.8%, and the mean absolute value of the attitude dynamics residual sequence reached 0.67 rad. The performance data of the system of this invention during this stage are as follows:

[0138] The minimum effective particle number of the system is 61, the particle size is dynamically adjusted in the range of 1152~3072, the process noise variance is automatically increased to 0.018 rad², and the observation noise variance is adjusted to 0.029 rad².

[0139] The weighted attitude estimation mean was compared with the observed true value. The absolute deviation was 0.18° and the maximum deviation was 0.47°. After 72 consecutive hours without human intervention, all health reports were "normal".

[0140] The particle diversity index has a minimum of 2.66 bits and an average of 3.21 bits. The marginal likelihood trend curve has a mean of 0.15 and a standard deviation of 0.014.

[0141] Comparative data shows that traditional particle filtering, during equipment aging, reduces the effective particle count to as low as 16, with a short-term attitude estimation divergence duration of 6.2 minutes, a mean attitude estimation deviation of 0.81°, and a maximum deviation of 1.53°. Unscented Kalman filtering requires multiple manual adjustments and calibrations of noise parameters, totaling 8 manual interventions, and has a mean attitude estimation deviation of 1.21°.

[0142] Throughout the entire process, the platform collected over 130 million frames of synchronized samples. The statistical results of the deviation between the mean and true values ​​of the final fused pose estimation output, and the key performance indicators of the method of this invention and the traditional method are shown in Table 1 below:

[0143] Table 1 Key performance indicators of the method of the present invention and the traditional method

[0144] Operating condition phase method Absolute deviation of the mean of attitude estimation (°) Maximum deviation (°) Particle diversity index (bit) Cumulative divergence duration (min) Number of manual interventions Normal operation This invention 0.05 0.13 3.53 0 0 Particle filtering 0.11 0.31 2.48 0 1 Unmarked Kalman 0.19 0.41 1.33 1.7 2 Environmental mutation This invention 0.34 0.81 2.94 0 0 Particle filtering 1.03 2.26 1.22 4.7 2 Unmarked Kalman 1.49 3.47 0.98 7.2 5 Equipment aging This invention 0.18 0.47 3.21 0 0 Particle filtering 0.81 1.53 1.49 6.2 2 Unmarked Kalman 1.21 2.18 0.85 9.3 8

[0145] As can be seen from Table 1, even under complex environments such as extreme winds and waves, abnormal observations, and aging equipment, the method of the present invention can still rely on the dynamic distribution modeling capability of the improved Diffusion model and the real-time adjustment mechanism of the adaptive particle filter.

[0146] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for attitude perception of a marine launch platform based on machine learning and particle filtering, characterized in that, include: Collect and preprocess multi-source raw observation datasets to construct sea state and sensor state vectors; An attitude dynamics model of an offshore platform is established based on a rigid body six-degree-of-freedom kinematic model, and the attitude increment residual is predicted by using sea state and sensor state vectors to generate an attitude dynamics residual sequence. An improved Diffusion model for training physical constraints; The trained improved Diffusion model is deployed to an edge computing device, and the pose increment prior distribution, observation error observation likelihood distribution, and score function are output through a few-step sampling method. The particle set of the adaptive particle filter is initialized based on the prior distribution of attitude increment, and the particle propagation along the high-probability manifold direction of attitude is guided by the scoring function. At the same time, the attitude dynamics model of the offshore platform is used for particle prediction update to obtain the particle prediction set. The particle weights of the predicted particle set are calculated by using the observation error likelihood distribution, real-time multi-source observation, and robust loss function, thus obtaining a weighted particle set. The effective number of particles in the weighted particle set is calculated. When the effective number of particles is lower than the threshold, an adaptive resampling mechanism is triggered. The particle number, resampling frequency, process noise covariance parameter and observation noise covariance parameter are adaptively adjusted according to the marginal likelihood value and observation residual statistics to generate a resampled particle set. The weighted mean and variance of attitude estimation are obtained based on the resampled particle set, and the attitude estimation results of the offshore platform, particle diversity index and marginal likelihood trend are output.

2. The attitude perception method for a marine launch platform based on machine learning and particle filtering according to claim 1, characterized in that, The construction of the sea state and sensor state vector includes: Data from offshore platform inertial measurement units, global navigation satellite systems, visual sensors, wave radar, and meteorological buoys are collected and combined with collection timestamps to generate a multi-source raw observation dataset. The multi-source raw observation dataset is then synchronized, denoised, has outlier removed, and missing data is filled in to construct sea state and sensor state vectors.

3. The attitude perception method for a marine launch platform based on machine learning and particle filtering according to claim 2, characterized in that, The generated attitude dynamics residual sequence includes: Using the attitude dynamics prediction model, the unit quaternion at the current moment is calculated based on the unit quaternion at the previous moment, the angular velocity vector at the previous moment, and the sampling period parameter. The predicted unit quaternion at the current moment represents the attitude prediction result of the offshore platform based on the dynamics model. The observation unit quaternion estimated by fusing multiple source raw observation datasets and the current time unit quaternion predicted by the attitude dynamics prediction model are multiplied and conjugately operated to calculate the attitude dynamics residual quaternion. The attitude dynamics residual quaternion is transformed into a three-dimensional rotation vector through a logarithmic mapping operation from Lie group to Lie algebra. Repeat the steps to calculate the three-dimensional rotation vectors at all times in the time dimension, and combine the three-dimensional rotation vectors at all times in time order to form an attitude dynamics residual sequence.

4. The attitude perception method for a marine launch platform based on machine learning and particle filtering according to claim 3, characterized in that, The improved Diffusion model includes: For each sampling moment, the sea state, sensor state vector, attitude dynamics residual vector, and multi-source raw observation dataset that has been synchronously processed are merged to generate a conditional input sample. Conditional input samples are collected at all moments within a certain period of time to form a training dataset. For each moment in the training dataset, the attitude increment quaternion is extracted as a parameter to describe the attitude change of the offshore platform. The attitude dynamics residual vector is restored to the attitude increment quaternion through the transformation relationship from Lie group to Lie algebra. At the same time, the multi-source observation error vector at the corresponding moment is calculated as a parameter to describe the observation residual. The attitude increment quaternion and the multi-source observation error vector are concatenated to form the target variable. For each target variable, Gaussian noise is added step by step according to the set number of diffusion steps and noise scheduling parameters to obtain the diffusion state at each step; Construct a scoring network with conditional input, and obtain the score matching loss by calculating the mean square error between the output of the scoring network and the actual added noise; During training, a physical constraint regularization term is introduced, which is calculated by the squared difference between the magnitude of the quaternion of the attitude increment at all times in the statistical training dataset and 1. The score matching loss is weighted and summed with the physical constraint regularization term to obtain the total loss function. The parameters of the improved Diffusion model are then updated by iteratively optimizing the total loss function using the stochastic gradient descent algorithm, resulting in the trained improved Diffusion model.

5. The attitude perception method for a marine launch platform based on machine learning and particle filtering according to claim 1, characterized in that, The process of using the offshore platform attitude dynamics model for particle prediction updates to obtain a particle prediction set includes: The current conditional input sample is fed into the trained improved Diffusion model to generate the current pose increment prior distribution and score function; Set the total number of particles, initialize the particle set of the adaptive particle filter, and initialize the particle state starting from the particle attitude state of the previous moment. By mapping the change trend of the score function in the attitude space to the Lie group space after weighted superposition with the diffusion noise, and combining it with the particle attitude state of the previous moment according to the quaternion multiplication rule, the particle attitude state of the current moment is obtained. The particle attitude state obtained by guided propagation based on the scoring function is weighted and fused with the attitude prediction result based on the dynamic model using the unit quaternion at the current time, and then normalized to obtain the particle prediction set.

6. The attitude perception method for a marine launch platform based on machine learning and particle filtering according to claim 5, characterized in that, The calculation of particle weights for the particle prediction set using observation error likelihood distribution, real-time multi-source observation, and robust loss function includes: For each particle in the particle prediction set, the particle's attitude state is mapped to a particle observation value using the observation model. The particle observation value is then subtracted from the real-time multi-source observation vector to obtain the observation residual vector. The improved Diffusion model, which has been trained, is invoked. Based on the conditional input samples at the current time, the observation likelihood distribution of the observation error is obtained. The observation residual vector of the particle is substituted into the observation likelihood distribution model of the observation error to obtain the basic weight of the particle. A robust loss function is introduced for the observation residual vector of each particle. The robust loss function judges whether the absolute value of each component of the observation residual is less than a threshold. When the absolute value of the observation residual is less than the threshold, the squared loss is used; otherwise, the linear loss is used minus a constant. The loss results of all components are summed to obtain the total robust loss value of the particle. The negative number of the total robust loss value is then exponentially calculated to obtain the robust scaling factor. The base weights are multiplied by the robust scaling factor to obtain the weighted particle weights. The weighted particle weights of all particles are normalized. After normalization, each particle is associated with its particle attitude state and the corresponding normalized particle weight, forming a weighted particle set.

7. The attitude perception method for a marine launch platform based on machine learning and particle filtering according to claim 6, characterized in that, The generated resampled particle set includes: Calculate the number of effective particles at the current moment based on the normalized weights of all particles in the weighted particle set; Set an effective particle count threshold. When the effective particle count at the current moment is less than the effective particle count threshold, trigger the adaptive resampling mechanism. The adaptive resampling mechanism uses the system resampling method to resample the particle set and obtain a new particle set. The marginal likelihood value for the current fusion cycle is calculated by multiplying and summing the normalized weights of all particles with the probability density of the observation likelihood distribution of the observation error. A dynamic adjustment function based on the marginal likelihood trend is constructed, which calculates the particle number adjustment and the resampling frequency adjustment respectively. The total number of particles should be adjusted based on the current fusion status and historical statistical results. The adjusted number of particles should be limited to the lower limit and the upper limit of the particle number. Update the particle process noise covariance matrix and the observation noise covariance matrix; The newly generated set of particles is used as the resampled set of particles output in the current fusion cycle.

8. The attitude perception method for a marine launch platform based on machine learning and particle filtering according to claim 7, characterized in that, The output of the offshore platform attitude estimation results, particle diversity index, and marginal likelihood trend includes: Calculate the weighted mean attitude estimate of all particles in the resampled particle set, and normalize the weighted summation result to obtain the weighted mean attitude estimate. By squaring the geodesic distance between the attitude state of each particle and the weighted attitude estimation mean, multiplying the squared distance value by the normalized particle weight of the particle, and summing the results for all particles, the weighted variance of the resampled particle set on the SO(3) manifold is obtained. Calculate the diversity index of the particle set in the current fusion cycle; By multiplying the normalized particle weight of each particle by the probability density of the observation error likelihood distribution of the particle, and summing the results for all particles, the marginal likelihood trend of the current fusion cycle is obtained.

Images