CN-121977536-A - Nacelle rapid self-alignment method for multi-sensor joint estimation

Abstract

The invention discloses a nacelle rapid self-alignment method for multi-sensor joint estimation, which relates to the technical field of safety management and comprises the following steps: i, initializing a multisource sensor and collecting environmental information; II, constructing an adaptive filtering structure and estimating the initial attitude of the nacelle; III, analyzing causal relation of the multisource observation data and evaluating credibility of the causal relation; IV, correcting the gesture of the nacelle and eliminating redundant information in the multi-source observation data; the invention can obviously shorten the whole alignment time, improve the system starting efficiency, avoid the influence of single sensor failure or noise increase on the system, thereby improving the data fusion effect, dynamically adjust the filtering structure according to the actual running condition, thereby improving the adaptability to complex environment, identifying the nonlinear noise mode which is difficult to find by the traditional statistical method, stripping abnormal data by the noise index positioning and data reconstruction mode, and improving the quality of the observed data.

Inventors

• She Panpan
• CHEN YIXIN

Assignees

• 南京易信同控制设备科技有限公司

Dates

Publication Date

20260505

Application Date

20260408

Claims (6)

1. 1. A method for rapid self-alignment of a pod for joint estimation of multiple sensors, comprising the steps of: I. Initializing multisource sensors and collecting environmental information, namely after the nacelle system is started, carrying out initialization configuration of the multisensor system, establishing a unified time synchronization mechanism, synchronously starting each sensing device to collect multisource observation data in real time, and classifying the current motion environment through real-time reasoning to generate an initial environment state label; II. Constructing a self-adaptive filter structure and estimating the initial attitude of the nacelle, namely constructing a data filter structure adapting to the current condition according to an initial environmental state label, dynamically adjusting various parameters according to the environmental label, and then calculating an initial estimated value of the attitude angle of the nacelle; III, analyzing causal relation of the multisource observation data and evaluating credibility of the causal relation, namely after initial attitude estimation of the nacelle is completed, analyzing dynamic association relation among various sensors and establishing a credibility weight matrix of the corresponding dynamic sensors; IV, correcting the nacelle gesture and eliminating redundant information in the multi-source observation data, namely after the data reliability assessment is completed, iteratively correcting the nacelle gesture estimation value based on the multi-source observation data and the reliability weight matrix, generating a stable gesture resolving result, extracting topological structure information of the multi-source observation data on different scales after the stable nacelle gesture estimation is obtained, and identifying and stripping noise structures in the observation data; v, constructing a simulation model of the nacelle, performing alignment optimization based on the simulation model, namely constructing a nacelle simulation model after data optimization processing is completed, simulating alignment behaviors of a real nacelle system under various conditions, synchronously operating the nacelle digital simulation model and the real nacelle system, simulating various alignment paths by randomly perturbing different error parameters, and selecting an optimal alignment path by utilizing a countermeasure optimization strategy.
2. 2. The method for rapid pod self-alignment based on multi-sensor joint estimation according to claim 1, wherein the specific steps of classifying the current motion environment by real-time reasoning in the step I to generate the initial environment state label are as follows: S1.1, powering on each sensing device according to a preset power-on sequence, reading a basic state from each sensing device, detecting and extracting a plurality of groups of key quantities from the basic state, simultaneously recording abnormal values in the detected key quantities, isolating corresponding sensors, and calculating health scores of each sensor according to preset weighting indexes; S1.2, setting a time reference node, then sending a time request by a main device, recording local time stamps when each sensor receives and sends the time request, interactively recording four time stamps of local sending, local receiving, remote receiving and remote sending each time, calculating offset and delay of communication of the corresponding sensor according to each time stamp of the interactive record so as to acquire offset estimation of a local clock of each sensor and a unified time reference, fitting short-term clock drift parameters of each sensor, and generating a corresponding time correction function; S1.3, placing the nacelle on a plane for 60S, continuously collecting original observation data of each sensor when the nacelle is in a static state, respectively calculating sample mean values of an accelerometer and a gyroscope to serve as initial estimation of static deviation, generating a corresponding deviation table according to a preset temperature section, writing the initial estimation deviation into sensor operation parameters, and recording uncertainty of the initial estimation deviation, namely corresponding sensor variance; S1.4, mapping an original timestamp of each sensor to a unified reference time by using a time correction function, then inserting sensor values with different sampling rates or sudden frame loss onto a public time grid by using linear interpolation, calculating interpolation error indexes and recording after interpolation is completed, carrying out framing treatment on observation data of various sensors by using a short-time sliding window with the window length of 1S and the step of 0.2S, calculating an original feature vector of each frame, carrying out dimension-by-dimension standardization treatment on each original feature vector, and carrying out dimension compression; S1.5, inputting the processed feature vectors into a preloaded lightweight neural network, wherein the lightweight neural network acquires the un-normalized logits vectors through forward reasoning layer by layer on the received feature vectors, then obtains probability distribution of each type of environment through softmax processing, calculates entropy as confidence measure at the same time, outputs corresponding environment state labels according to a preset confidence threshold value, and packages the environment state labels with confidence and feature window time stamps.
3. 3. The method for rapid self-alignment of a pod with joint estimation by multiple sensors according to claim 2, wherein the specific steps of constructing a data filtering structure adapted to the current conditions according to the initial environmental state label in step II, dynamically adjusting various parameters according to the environmental label, and then calculating the initial estimated value of the pod attitude angle are as follows: S2.1, receiving generated environment state labels, online selecting a group of basic filtering parameter templates comprising a basic process noise matrix, a basic observation noise matrix and a filter type identifier for each type of environment state labels, constructing a noise parameter set according to the environment state labels, then selecting a filter form according to the environment state labels and the current computing capacity constraint, configuring the dimensionality of a state vector based on the selected filter form, and initializing a corresponding priori covariance structure; S2.2, mapping each noise parameter in the noise parameter set into process noise and measurement noise corresponding to the selected filter, storing the process noise and the measurement noise in filter configuration, selecting a corresponding short-time window according to an environment state label after the filter is started, carrying out low-pass average on original samples of the accelerometer according to the selection, and taking an average vector after the low pass as an observed gravity direction estimation vector; S2.3, projecting a gravity direction estimation vector to a machine system, acquiring initial values of pod roll and pitch angles through inverse trigonometric operation, carrying out boundary cutting on the initial values, simultaneously calculating uncertainty approximation of the corresponding initial values according to sample covariance linear propagation in a window, calculating weighted average of angular velocities of a gyroscope in the same time window, constructing a base axis direction vector under pod coordinates according to the acquired initial values of the roll and pitch angles, and acquiring a pod heading angle initial value by utilizing the weighted average of the angular velocities of the gyroscope; S2.4, mapping a standard Euler angle to a quaternion, mapping a roll angle initial value, a pitch angle initial value and a course angle initial value to quaternion gesture priori, constructing a state priori according to uncertainty approximation of roll and pitch angles and gyroscope average variance, establishing a corresponding priori covariance matrix, and scaling gesture components in the prior covariance matrix; S2.5, reading gyro instantaneous measurement in a current sampling period, carrying out short-time average on the gyro instantaneous measurement to obtain a corresponding angular velocity vector, integrating the angular velocity vector into corresponding gesture increment in a sampling interval by utilizing small-angle approximate mapping, updating quaternion prior according to the generated gesture increment, and updating prior covariance matrix according to process noise; S2.6, constructing a linearization form of a corresponding measurement model according to the selected observation type, calculating predictive measurement from the current quaternion prior to the observation space, acquiring a corresponding measurement residual based on the calculated predictive measurement, and updating and correcting the linearization form, the measurement residual and the measurement noise.
4. 4. A nacelle rapid self-alignment method for joint estimation of multiple sensors according to claim 3, wherein the specific steps of establishing a corresponding dynamic sensor reliability weight matrix are as follows: S3.1, intercepting a time window to be analyzed of an observation channel of each sensor on a unified time reference, carrying out trend removal and normalization processing on each time sequence in each time window to be analyzed, carrying out differential operation on each time sequence, calculating an autocorrelation coefficient of the time sequence after the differential operation, considering that the time window is approximately stable if the attenuation of the autocorrelation exceeds a preset threshold value in preset lag time, and otherwise, carrying out windowed resampling on the time window; S3.2, respectively constructing a limited model only containing the past items of the target itself and a complete model simultaneously containing the past items of the target itself and the past items of the candidate observation for each pair of the target sequence and the candidate observation sequence, estimating each model coefficient by adopting a least square method, acquiring a corresponding residual sequence to calculate residual energy, and simultaneously taking the residual energy as a model fitting error measure; S3.3, constructing a standardized causal strength index by using residual energy, performing F-type significance test on the index in each time window, recording a significance result as short-time causal judgment of the time window, adopting a convergence cross mapping method, predicting candidate sequence values by using neighbor points of candidate observation sequences in phase space reconstruction of a target sequence, comparing the candidate sequence values with a base line of rearrangement test, and judging that nonlinear causal relation exists and taking a prediction correlation coefficient of convergence cross mapping as nonlinear causal score if the predicted candidate sequence values are higher than the base line and are located in a preset stable threshold value along with the increase of the number of samples; S3.4, repeatedly carrying out residual energy calculation, saliency assessment and nonlinear causal inspection on a plurality of continuous overlapping time windows to obtain linear causal strength, corresponding saliency and nonlinear causal score of each time window, calculating the mean value and variance of each index on a time sequence, calculating a corresponding stability score, and judging that causal relation among corresponding sensors is stable and consistent if the stability score is higher than a preset threshold; And S3.5, respectively carrying out normalization processing on the linear causal strength, the nonlinear causal score and the stability score according to preset weights for each candidate sensor channel, synthesizing all indexes after processing into corresponding comprehensive scores, normalizing the comprehensive scores of all sensor channels into weight vectors so as to obtain instantaneous credibility weights at all moments, carrying out smoothing processing on the instantaneous credibility weights by adopting exponential weighted moving average to obtain corresponding dynamic weights, and arranging the corresponding dynamic weights into a diagonal or generalized dynamic credibility weight matrix according to the channels.
5. 5. The method for rapid pod self-alignment based on multi-sensor joint estimation according to claim 1, wherein the specific steps of extracting topology information of multi-source observation data on different scales, and identifying and stripping noise structures in the observation data in step IV are as follows: S4.1, selecting a time window with preset length in a signal to be analyzed from observed data by taking a stable gesture interval as a reference, setting sampling point indexes in the time window, respectively splicing each time point with the observed data with a plurality of groups of delay moments to construct a combination point cloud, calculating a distance matrix between any two points in the combination point cloud, and constructing a scale set according to empirical distribution of the distance matrix; s4.2, on each scale, when the Euclidean distance between a group of point pairs in the joint point cloud is less than or equal to the current scale through a VR rule, connecting two points in the group of point pairs by edges to form a simple complex, recording the birth scale and the extinction scale of each topological feature in the coherent group from small to large according to the scale, and calculating the persistence of each topological feature based on the birth scale and the extinction scale; s4.3, calculating a corresponding threshold strategy according to the persistence statistical distribution of each topological feature, marking each topological feature as a structure or noise through the corresponding threshold strategy for each coherence dimension, recording the birth-death interval of each topological feature marked as noise and the coherence dimension corresponding to the birth-death interval, calculating a representative generator of each topological feature marked as noise, and mapping all vertex sets covered by the representative generator into a noise index set of an original time sequence sample; And S4.4, carrying out low-rank reconstruction through median replacement in a time neighborhood of each noise point in the noise index set within a preset radius range, replacing or smoothing a polluted sample, recording energy or variance change before and after replacement after each replacement, replacing the corresponding section of the noise index set in the original observed data with the reconstructed observed data, and recombining point clouds, and repeating noise identification and data replacement, and meanwhile generating a persistent bar graph.
6. 6. The method for rapid pod self-alignment based on joint estimation of multiple sensors according to claim 5, wherein the specific steps of simulating multiple alignment paths by randomly perturbing different error parameters in step V are as follows: S5.1, disassembling the physical behavior of the nacelle into a plurality of groups of sub-modules, setting minimum necessary state quantity for each sub-module and combining the sub-modules into a complete state vector of a nacelle simulation model, carrying out corresponding dynamics description by using a continuous time nonlinear state equation to establish a nacelle simulation model, establishing a corresponding random model for each error source required to be disturbed by the nacelle simulation model, determining distribution parameters by using engineering experience or priori calibration results, and assembling all disturbed parameters to form priori distribution of random parameter vectors; S5.2, generating a plurality of groups of parameter samples by using a random sampling method according to prior distribution, independently simulating each group of parameter samples through a nacelle simulation model, obtaining continuous dynamics and generating a discrete time track, inputting the parameter samples and initial conditions, keeping synchronization with a real system, simultaneously injecting real observation noise information and sensor sampling delay for each simulation, outputting a virtual observation sequence corresponding to the parameter samples, and recording the alignment error quantity and intermediate index of each sample on each discrete time grid point; S5.3, calculating final state attitude error, convergence time, overshoot and energy consumption or control command amplitude indexes of each round of simulation track, combining the indexes into total cost of single simulation, accumulating cost of each group of parameter samples in all time steps, and recording time when the first time reaches a preset error threshold; s5.4, establishing a cost set after multiple simulation of each parameter sample, calculating the performance indexes of a mean value, a variance, a quantile and an extremum to evaluate the expected performance and the tail risk, sequencing the simulated track each time according to the cost, the convergence speed and the tail robustness, identifying a plurality of groups of candidate optimal paths and a plurality of groups of robust paths, screening out non-compliant parameter samples according to preset engineering constraints, and generating a final recommended limited set; S5.5, after the physical pod runs the finite set and obtains corresponding real observation, carrying out likelihood assessment on corresponding parameter samples by utilizing the observation, calculating probability or approximate likelihood value of each parameter sample for generating the observation, generating corresponding posterior weights for the corresponding parameter samples, re-weighting parameter distribution and recommended paths in the finite set based on the posterior weights to generate posterior parameter estimation and updated risk assessment, and based on assessment results, preferentially sampling in corresponding areas in the next round of simulation.

Description

Nacelle rapid self-alignment method for multi-sensor joint estimation Technical Field The invention relates to the technical field of safety management, in particular to a nacelle rapid self-alignment method based on multi-sensor joint estimation. Background In modern airborne photoelectric systems, unmanned aerial vehicle reconnaissance systems and intelligent monitoring platforms, the stabilized nacelle system generally bears key functions such as target observation, attitude stabilization, direction indication and the like. The nacelle system typically incorporates an Inertial Measurement Unit (IMU) including gyroscopes, accelerometers, magnetometers, satellite navigation modules (such as GPS or beidou), vision sensors and other auxiliary measurement devices for acquiring attitude, angular velocity and position information of the platform in real time. When the system is started or repositioned, the relation between the inertial navigation system coordinate system and the geographic reference coordinate system needs to be determined through an initial alignment process, so that the accuracy of subsequent gesture calculation and navigation calculation is ensured. Conventional pod alignment methods rely primarily on inertial devices to perform static or inertial alignment of the measurements of the earth's gravity vector and the earth's rotational angular velocity. Such methods generally require the system to operate in a relatively stable environment and require relatively long alignment times to gradually reduce attitude errors. When the system is in a vibration environment, a mobile platform or a complex external interference environment, inertial measurement data is easily affected by noise, drift and external disturbance, so that the convergence speed of the alignment process is low, and even the problem of error accumulation occurs. There is therefore a need for a nacelle fast self-alignment method for multi-sensor joint estimation. The existing nacelle rapid self-alignment method has long integral alignment time, reduces system starting efficiency, has the influence of single sensor failure or noise increase on the system, reduces data fusion effect, cannot dynamically adjust a filtering structure according to actual running conditions, has poor adaptability to complex environments, cannot identify nonlinear noise modes which are difficult to find by the traditional statistical method, and reduces observed data quality. Disclosure of Invention The invention aims to solve the defects in the prior art and provides a nacelle rapid self-alignment method for multi-sensor joint estimation. The invention provides a nacelle rapid self-alignment method for multi-sensor joint estimation, which solves the technical problems by adopting the following technical scheme: I. Initializing multisource sensors and collecting environmental information, namely after the nacelle system is started, carrying out initialization configuration of the multisensor system, establishing a unified time synchronization mechanism, synchronously starting each sensing device to collect multisource observation data in real time, and classifying the current motion environment through real-time reasoning to generate an initial environment state label; II. Constructing a self-adaptive filter structure and estimating the initial attitude of the nacelle, namely constructing a data filter structure adapting to the current condition according to an initial environmental state label, dynamically adjusting various parameters according to the environmental label, and then calculating an initial estimated value of the attitude angle of the nacelle; III, analyzing causal relation of the multisource observation data and evaluating credibility of the causal relation, namely after initial attitude estimation of the nacelle is completed, analyzing dynamic association relation among various sensors and establishing a credibility weight matrix of the corresponding dynamic sensors; IV, correcting the nacelle gesture and eliminating redundant information in the multi-source observation data, namely after the data reliability assessment is completed, iteratively correcting the nacelle gesture estimation value based on the multi-source observation data and the reliability weight matrix, generating a stable gesture resolving result, extracting topological structure information of the multi-source observation data on different scales after the stable nacelle gesture estimation is obtained, and identifying and stripping noise structures in the observation data; v, constructing a simulation model of the nacelle, performing alignment optimization based on the simulation model, namely constructing a nacelle simulation model after data optimization processing is completed, simulating alignment behaviors of a real nacelle system under various conditions, synchronously operating the nacelle digital simulation model and the real nacelle system, simulating various ali