Six-degree-of-freedom microgravity quality evaluation method based on task phase driving

Abstract

This invention relates to a six-DOF microgravity quality assessment method based on mission phase-driven approaches, belonging to the field of spacecraft control and ground simulation. It constructs a multi-physics fusion six-DOF microgravity quality index, divides the on-orbit service mission phase using a finite state machine and matches dynamic weights and assessment thresholds, combines digital twins to achieve virtual-real consistency verification, employs rapid independent component analysis and extended state observers to complete disturbance decoupling and contribution rate quantification, and uses a graph neural network long short-term memory network model and Shapley sum interpretation algorithm to achieve pre-experiment performance prediction and in-experiment anomaly diagnosis. This invention addresses the technical problems of existing microgravity ground simulation assessments, such as the disconnect between static thresholds and on-orbit mission requirements, the inability of a single index to cover multi-DOF coupling characteristics, the lack of virtual-real consistency verification, the inability to quantify and attribute disturbance sources, over-reliance on expert experience, and post-event offline assessment.

Description

Technical Field

[0001] This invention relates to a six-degree-of-freedom microgravity quality assessment method based on mission phase drive, belonging to the field of aircraft control and ground simulation. Background Technology

[0002] As on-orbit servicing missions advance (such as non-cooperative target acquisition, on-orbit refueling, and module replacement), higher demands are placed on the fidelity of ground simulation environments. Currently, six-degree-of-freedom full-physics simulation systems have become core equipment for simulating microgravity environments and verifying complex spacecraft interaction dynamics and control algorithms. However, in the field of experimental performance evaluation, industry evaluation methods generally lag behind the development of hardware testing systems. Existing evaluation methods mostly remain at the stage of single-index, post-event offline analysis, and black-box trial and error relying on expert experience; existing systems generally lack comprehensive online quantitative dimensions of multi-physics fields and cannot provide dynamic fidelity constraint boundaries in conjunction with specific on-orbit mission phases; at the same time, limited by complex pipeline disturbances, air film nonlinearity, and structural modal disturbances, existing technologies cannot quantify the contribution rate of specific physical disturbance sources and cannot clarify the fundamental problem of the difference between ground physical simulation test results and the actual on-orbit mission environment. Therefore, establishing a dynamic performance evaluation method that integrates multi-dimensional indicators, mission phase-driven approaches, digital twin verification, and AI online diagnosis has become an urgent need and industry consensus for the development of microgravity ground simulation technology.

[0003] Existing technologies suffer from several limitations. A single or limited set of indicators cannot simultaneously reflect six-degree-of-freedom disturbances, translational-rotational coupling, and contact transients, leading to situations where the overall average of the experiment meets the standards but critical phases are distorted. Evaluations are often conducted offline after the fact, making it impossible to determine the effectiveness of the experiment during its execution. This results in wasted resources from invalid experiments and difficulty in timely correction. Thresholds are typically static, end-to-end limits that cannot be adjusted according to changes in mission phase (approach / capture / fueling / separation) and operational variables (distance, relative velocity, contact state, valve action, etc.), thus failing to reflect the true requirements of on-orbit service missions at different stages. Furthermore, there is a lack of verification for consistency between virtual and real-world data, making it impossible to provide credible quantifiable evidence that ground results can be extrapolated to on-orbit behavior. The sources of disturbances are unclear, and their contributions cannot be quantified, making it impossible to directly guide targeted improvements to hardware systems or control parameters. Finally, there is a lack of data-driven prediction and diagnosis, relying too heavily on expert experience and hindering the transition from expert evaluation to intelligent prediction and diagnosis. Summary of the Invention

[0004] This invention aims to solve the problems of existing six-degree-of-freedom microgravity ground simulation evaluations, such as the one-sidedness of a single index, the lack of verification of consistency between virtual and real data, unclear sources of disturbance, unquantifiable contributions, and over-reliance on expert experience. It proposes a six-degree-of-freedom microgravity quality evaluation method based on mission phase driving.

[0005] The technical solution of the present invention:

[0006] A six-degree-of-freedom microgravity quality assessment method based on mission phase drive is implemented using a six-degree-of-freedom full-physics ground simulation system. The six-degree-of-freedom full-physics ground simulation system includes a six-degree-of-freedom air-bearing platform, a high-frequency sensor network, a host computer mission control terminal, and a performance evaluation server. The method includes the following steps:

[0007] S1. Construct a six-degree-of-freedom microgravity mass index 6D-MQI based on multi-physics field fusion. Obtain real-time state data of the air-bearing platform through a sensor network, calculate multi-dimensional feature parameters based on a sliding time window, and input them into the multi-dimensional state fusion equation to output the six-degree-of-freedom microgravity mass index in real time.

[0008] S2. Implement dynamic performance monitoring driven by task phase, establish a task phase recognition model based on finite state machine, divide the continuous time axis of on-orbit service tasks into multiple discrete task phases, match the corresponding dynamic weights and evaluation thresholds for different task phases, perform state verification in real time, issue test interruption commands and generate abnormal logs for microgravity quality deviation scenarios.

[0009] S3. Based on digital twin, complete the virtual-real consistency verification, control the physical system and digital twin to solve in parallel under the same clock source, collect the measured state vector of the physical system and the ideal zero-gravity state vector of the digital twin, calculate the dynamic difference degree, and determine the on-orbit credibility of the ground test results based on the preset tolerance upper limit.

[0010] S4. Online decoupling and contribution rate quantification of multi-source disturbances: triggered when the state verification in step S2 fails or the dynamic difference in step S3 exceeds the limit, extracting multi-source sensor data during abnormal periods, completing the decoupling and separation of multi-source disturbances and the estimation of nonlinear residual disturbances, and calculating and outputting the variance contribution rate of each disturbance source.

[0011] S5. Based on machine learning, performance prediction and anomaly diagnosis are achieved. A GNN-LSTM multi-task prediction model is built by calling the historical test database. Before the test, the performance is pre-verified based on the current test plan, and high-risk tests are intercepted. During the test, anomalies are monitored in real time and the anomaly attribution is completed through the SHAP algorithm. After the test, the database is updated and the model is iteratively optimized.

[0012] Specifically, the length is set as The sliding time window is used to set the sampling step size. Within each sliding window, the characteristic parameters of translational dimension, rotational dimension, spatial coupling effect, and internal force circulation dimension are calculated. The characteristic parameters are then substituted into the multidimensional state fusion equation to output the six-degree-of-freedom microgravity quality index.

[0013] The translational dimension feature parameters include:

[0014] Triaxial residual translational acceleration vector The corresponding root mean square value of translational acceleration and the use of the Welch method to After performing a discrete Fourier transform, the vertical power spectral density integral within the 0.1Hz to 10Hz frequency band is: ,in, Let be the translational acceleration in the x-axis direction at time t. Let be the translational acceleration in the y-axis direction at time t. Let be the translational acceleration in the z-axis direction at time t. Sliding time window length, For signal frequency, Let be the power spectral density function of the z-axis translational acceleration;

[0015] The rotation dimension feature parameters include:

[0016] Triaxial residual angular acceleration vector The corresponding root mean square value of rotational angular acceleration ,in, The residual angular acceleration vector on the roll axis. The residual angular acceleration vector on the pitch axis. Let $\mathbf$ be the residual angular acceleration vector on the yaw axis and the magnitude of the attitude drift rate vector per unit time during the phase without control torque input. ,in Set the initial Euler angles for the window. The Euler angle at the end of the window; the characteristic parameter of the spatial coupling effect is the Z-axis translational acceleration. With pitch acceleration Pearson correlation coefficient The characteristic parameter of the internal force circulation dimension is the internal force circulation index. ,in The number of active actuators inside the air flotation platform. The nominal thrust output by the i-th actuator. This represents the total external environmental load on the air-bearing platform. To ensure the effectiveness of the internal force circulation index, the amount is kept to a minimum;

[0017] The multidimensional state fusion equation is:

[0018] ;

[0019] in, to These are the basic weight coefficients determined by the analytic hierarchy process. to As a dimensional unification scale factor, all physical parameters are normalized to the 0~1 range, achieving unified quantification of features across all dimensions.

[0020] Specifically, step S1 further includes performing dimensionless scoring on the six physical characteristic parameters to generate a microgravity quality characterization index between 0 and 1. The specific steps are as follows:

[0021] The six physical feature parameters are converted into scores between 0 and 1 using a negative exponential function, generating a feature vector:

[0022] ;

[0023] Based on the current mission phase, set the mission phase accordingly. Changing diagonal adjustment matrix :

[0024] ;

[0025] in, The task phase dynamic weight adjustment coefficient is used to dynamically normalize the basic weight vector. This yields the final dynamic weight vector:

[0026] ;

[0027] Calculate the microgravity mass characterization index at the current moment:

[0028] ;

[0029] The The score is a composite score between 0 and 1. The closer the score is to 1, the closer the ground simulation environment is to the real microgravity environment in space.

[0030] Specifically, step S2 is as follows:

[0031] Based on the task phase identification model of the finite state machine (FSM), let the total system running time axis be... The task sequence variable issued by the host computer is The state machine is based on the Euclidean distance between the end effector of the air-bearing platform and the target aircraft. The contact force sensor status, locking command signal of the docking mechanism, and unlocking command signal divide the continuous time axis into proximity segments. Capture segment Refueling section Separation section Four discrete phases;

[0032] The system constructs a time-varying dynamic threshold function for different task phases. and dynamic weight adjustment matrix Within each sampling step, a status check is performed. If the microgravity quality continues to deviate from the fidelity requirement of the corresponding task phase, an experiment interruption command is issued and an anomaly log is generated.

[0033] Specifically, in step S2, the rules for dynamic weight adjustment and threshold setting for different task phases are as follows:

[0034] when At this point, the weights of the root mean square component of translational acceleration and the integral component of low-frequency power spectrum are increased to twice the basic weights, and the lower limit threshold for the quality index is set at this time. ;

[0035] when At this point, the weights of the root mean square component of the rotational angular acceleration and the attitude drift rate component are increased to three times the base weights, and the lower limit threshold for the quality index is set at this time. ;

[0036] when At this time, the translational-rotational cross-interference coefficient and the vertical residual acceleration are set as rigid penalty terms, and the lower limit threshold for the quality index at this time is set as follows. ;

[0037] when At this point, the weights of both the root mean square components of translational residual acceleration and rotational residual angular acceleration are simultaneously increased to three times their base weights. The minimum passing threshold for the quality index at this time is set as follows: .

[0038] Specifically, the state verification formula in step S2 is:

[0039] ;

[0040] Establish time delay tolerance variable If it appears consecutively Duration The system determines that the current microgravity quality has deviated from the fidelity requirement of the corresponding mission phase, issues an experiment interruption command, and generates an anomaly log with a timestamp.

[0041] Specifically, in step S2, the specific determination rule for task phase division based on the finite state machine is: a preset relative distance threshold. ,

[0042] when When the contact force is zero, it is determined to be the approach segment. ;

[0043] when If the lock signal has not yet been triggered, it is determined to be a capture segment. ;

[0044] When the docking mechanism lock signal takes effect and the fluid transfer command is initiated, it is determined to be the filling section. ;

[0045] When the docking mechanism unlocking and separation signal is triggered, and When the interval gradually increases from 0, it is determined to be a separation segment. .

[0046] Specifically, in step S3, the measured state vector output by the physical air-float platform is acquired. Synchronously record the ideal zero-gravity state vector output by the digital twin. According to the integration period The formula for calculating dynamic variability is:

[0047] ;

[0048] In the formula, To calibrate a positive definite weighted diagonal matrix based on the noise covariance of each sensor measurement, the system sets an upper limit for the variability tolerance. When detected If the physical system's operating state deviates significantly from the ideal space environment, the ground verification results lose credibility.

[0049] Specifically, in step S4, the points before and after the point of anomaly are extracted. Multi-source sensor data within a time period constitute the sensor observation matrix. ,in Number of sensor channels The total number of sampling points; a linear unmixing model is constructed using the Fast Independent Component Analysis (FastICA) algorithm:

[0050]

[0051] In the formula, For the separation matrix, The decoupled independent physical interference source matrix is ​​divided into specific physical interference terms by extracting the time-domain waveform features and spectral distribution features of each independent signal component.

[0052] An extended state observer (ESO) is constructed to estimate the sum of nonlinear residual disturbances that cannot be directly separated online. The equation of motion for the single-axis air-bearing platform is given by... , To construct a third-order ESO for the total disturbance term, which includes unmodeled system dynamics and external disturbances:

[0053] ;

[0054] ;

[0055] ;

[0056] ;

[0057] in, For ESO-tracked system position observations, For system velocity observations tracked by ESO, The total system disturbance observations approximated by ESO. For observation error, It is a non-linear power factor. , For the threshold of the linear segment of a nonlinear function, , The system's position and velocity were tracked separately. Approximating the total system disturbance , , , The observer gain constant is... It is a nonlinear continuous power function;

[0058] Calculate the variance of each explicit physical component and the observer-estimated component within the diagnostic window. Total variance Calculate and output the variance contribution rate of each physical source. .

[0059] Specifically, in step S5, the historical test database called by the system contains Group of historical test samples, each sample group structure is defined as follows: ,in , This is the static configuration parameter vector before system testing. It is a set of instructions for time-series task control; This is the sequence of actual performance indicators obtained in step one. The sequence of virtual-real consistency dynamic difference generated in step three; This is the variance contribution rate matrix of each physical interference source obtained in step four;

[0060] Constructing a multi-task prediction model based on graph neural network (GNN) and long short-term memory network (LSTM) The network weight parameter set is During the model training phase, a multi-objective joint optimization strategy is adopted, with the objective function being... Defined as the weighted mean squared error loss between the predicted sequence, the true sequence, and the attribution results:

[0061] ;

[0062] In the formula, , and These are the model's predicted performance index, variance, and disturbance contribution rate, respectively. This refers to the multi-task loss weighting coefficient; The regularization coefficient is used.

[0063] Before conducting a new physical experiment, enter the configuration parameters for the current experiment plan. With the planned instruction set The expected performance curve is obtained through feedforward calculation. Check whether it satisfies the requirements in all set task phases. and If the predicted curve is at risk of falling below the expected threshold or being severely distorted, an interception warning will be issued and the experiment will be refused to start; if the prediction passes and the experiment starts, the monitoring residual at the current moment will be calculated in real time. Extract the standard deviation of the residual distribution of the training set. ,when If an unmodeled anomaly is detected during the experiment, the Shapley algorithm and the SHAP algorithm are invoked to calculate the static configuration parameter vector. Each feature element in the current residual Deviation in marginal contribution value :

[0064] ;

[0065] In the formula, For the set of all feature parameters, Not including the first A subset of features The expected output function of the model under a specific feature subset; according to Sort the values ​​in descending order of their absolute values, and combine them with the predicted values. The system outputs the most likely physical cause of the anomaly. After the experiment is completed, the newly generated data sequence is encapsulated together with the SHAP attribution label and updated to the historical experiment database. This triggers backpropagation to optimize the model parameters, enabling the method to self-iterate and evolve.

[0066] The beneficial effects of this invention are:

[0067] 1. By transforming static single thresholds into task phase-driven dynamic thresholds, the problem of the disconnect between evaluation metrics and actual on-orbit requirements is solved. This invention innovatively introduces a finite state machine (FSM) for task phase identification and segmentation, enabling the system to dynamically adjust the weights and thresholds of various physical metrics in 6D-MQI according to the task sequence and operating conditions.

[0068] 2. A dual verification system of multidimensional fusion index (6D-MQI) and virtual-real consistency (DDM) was constructed. The 6D-MQI constructed in this invention integrates translation, rotation and coupling effects of 6 degrees of freedom, and introduces high-fidelity digital twin synchronous calculation. By calculating the dynamic difference degree (DDM), credible quantitative evidence that the results of ground test can be extrapolated to the real space environment is provided.

[0069] 3. Online decoupling and quantitative attribution of multi-source physical disturbances have been achieved. Existing equipment monitoring systems can only issue over-limit alarms when abnormal observation data is detected. This invention introduces a decoupling architecture based on FastICA and Extended State Observer (ESO), which can decompose residual disturbances into explicit physical terms while issuing alarms, and directly output the variance contribution rate (VCR), providing direct and quantitative guidance for platform hardware tuning and feedforward compensation of control laws.

[0070] 4. Intelligent-driven feedforward prediction and online autonomy are achieved. Compared with the traditional analytic hierarchy process and expert scoring, this invention introduces neural networks and the SHAP attribution algorithm. This not only enables performance prediction based on historical configurations before the start of physical experiments, avoiding wasteful experiments, but also allows the system to capture unmodeled anomalies in real time during experiments. The algorithm automatically locks the configuration parameters that cause the anomalies, giving the microgravity assessment system the ability to continuously iterate and self-evolve. Attached Figure Description

[0071] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0072] Example 1:

[0073] This embodiment provides a six-degree-of-freedom microgravity quality assessment method based on mission phase-driven principles, implemented using a six-degree-of-freedom full-physics ground simulation system. This physical simulation system includes a six-degree-of-freedom air-bearing platform, a high-frequency sensor network, a host computer mission control terminal, and a performance evaluation server.

[0074] Through five collaborative steps, a closed-loop performance evaluation architecture is presented. (1) Construction of the six-degree-of-freedom microgravity quality index (6D-MQI) system to provide a unified quantitative standard for the fusion of multi-dimensional data at the bottom layer of the entire system; (2) Online monitoring of mission phase, as the core step of this application, calls the quantitative results of step one and assigns a dynamic pass line in combination with the mission time sequence; (3) Verification of the consistency between digital twin and reality, to verify at the macro level whether the conclusions obtained from the above steps can truly represent the space environment; (4) Disturbance source diagnosis, to provide a physical-level directional solution when the evaluation results alarm; (5) AI-driven predictive performance and autonomy, to fully learn historical accumulated data and perform intelligent diagnosis and driving of the above steps.

[0075] The five steps described above are structured as follows in terms of time series and data flow: Step 5 utilizes historical experimental data to perform initial state assessment and prediction before the experiment begins, determining whether to run the physical experiment. After the physical experiment begins, Step 1 uses high-frequency sampling of the kinematic and dynamic parameters of the experimental system, employing multidimensional filtering and statistical algorithms to calculate quantitative indicators representing the microgravity quality of the system in real time. Step 2 receives the task sequence from the host computer, divides the continuous physical time axis into discrete task phases, and matches time-varying constraint boundaries to the quantitative indicators calculated in Step 1, performing online state judgment. Simultaneously, Step 3 runs a high-fidelity dynamic model in an independent computing node, inputs control commands from the same source as the physical system, generates a reference state matrix under ideal zero gravity conditions, and performs a virtual-real consistency check by solving the Euclidean norm distance between the measured state matrix and the reference state matrix. When the online judgment in Step 2 fails or the norm distance in Step 3 exceeds the limit, the system triggers Step 4. Step 4 performs reverse decoupling operations on the raw data from the underlying sensors, calculates the variance contribution rate of each physical interference source, and outputs quantitative attribution results. After a single trial, all state matrices and evaluation sequences generated in the first four steps are returned to the model database in step five. The network weight parameters are updated through the backpropagation algorithm to achieve closed-loop evolution of the system's evaluation capability.

[0076] The specific steps include:

[0077] Step 1: Construct and calculate the six-DOF microgravity mass index of multiphysics fusion in real time.

[0078] The performance evaluation server acquires real-time status data of the air-floating platform in the natural coordinate system through a sensor network, and establishes a length of The sliding time window, with the sampling step size set to The system calculates the feature parameters in the following dimensions within each sliding window.

[0079] First, calculate the root mean square value of the residual acceleration in the translational dimension and the integral of the low-frequency power spectral density. Let the triaxial residual translational acceleration vector at the center of mass of the air-bearing platform be: .

[0080] Root mean square value of translational acceleration .

[0081] Using the Welch method Perform a discrete Fourier transform to obtain the vertical power spectral density. To quantify low-frequency environmental disturbances, the system calculates the power spectrum integral over the 0.1Hz to 10Hz frequency band:

[0082] ;

[0083] Next, calculate the root mean square value of the residual angular acceleration and the attitude drift rate in the rotational dimension. Let the three-axis residual angular acceleration vector be... .

[0084] Root mean square value of rotational angular acceleration .

[0085] The system records the initial Euler angles during the phase without control torque input. Euler angle at the end of the window Calculate the magnitude of the attitude drift rate vector per unit time. .

[0086] Calculate the cross-interference coefficient, which represents the spatial coupling effect of the system, by calculating the Z-axis translational acceleration. With pitch acceleration Pearson correlation coefficient To quantify the degree of parasitic torque caused by the misalignment of the system's center of mass with the air-bearing rotation center:

[0087] ;

[0088] Finally, the internal force circulation index is calculated. The air-floating platform is equipped with... The first proactive implementing agency, the The nominal thrust output by each actuator The total external environmental load on the air-floating platform is The calculation formula is: .

[0089] The system substitutes the above characteristic parameters into the multidimensional state fusion equation and outputs a six-DOF microgravity mass index:

[0090] ;

[0091] in, to These are the basic weight coefficients determined by the analytic hierarchy process. to The 6D-MQI index serves as a standardized scaling factor, providing a unified quantitative standard for system performance evaluation.

[0092] Step 2: Implement dynamic performance monitoring driven by mission critical phases

[0093] The performance evaluation server establishes a task phase identification model based on a finite state machine (FSM). Let the total system runtime axis be... The task sequence variable issued by the host computer is The state machine is based on the Euclidean distance between the end effector of the air-bearing platform and the target aircraft. The contact force sensor status and the locking / unlocking command signal of the docking mechanism divide the continuous time axis into four discrete phases.

[0094] The preset relative distance threshold represents a critical operational envelope distance between the spacecraft's end effector and the target spacecraft.

[0095] when Furthermore, when the contact force is zero, the two parties have not yet entered the danger zone of physical contact, which is a process of relative posture approaching, and is judged as the approach phase. ;

[0096] when If the lock signal has not yet been triggered, the region is determined to be within the envelope of physical contact, collision, or capture action, and is thus classified as a capture segment. ;

[0097] When the docking mechanism lock signal takes effect and the fluid transfer command is initiated, the system is in a state of strong rigid connection and mass jump, which is determined to be the filling section. ;

[0098] When the docking mechanism unlocks and separates, the relative distance is... When the range gradually increases from 0, the system is in a transient unloading impact state, which is determined to be a separation segment. .

[0099] The system constructs a time-varying dynamic threshold function for different task phases. and dynamic weight adjustment matrix .

[0100] when At this point, the system increases the weights of the root mean square component of translational acceleration and the integral component of low-frequency power spectrum to twice the base weights, and sets the lower limit threshold for the quality index at this time as follows: ;

[0101] when At this point, the system increases the weights of the root mean square component of the rotational angular acceleration and the attitude drift rate component to three times the base weights, and sets the minimum passing threshold for the quality index at this time as follows: ;

[0102] when At this time, the system sets the translational-rotational cross-interference coefficient and the vertical residual acceleration as rigid penalty terms, and sets the lower limit threshold for the quality index at this time as follows: .

[0103] when At this time, due to the transient disturbance caused by the release of mechanical constraints, the system simultaneously increases the weights of the root mean square components of translational residual acceleration and rotational residual angular acceleration to three times the base weights, strictly controlling the system's ability to suppress transient unloading forces. The lower limit threshold for the quality index at this time is set as follows: .

[0104] The performance evaluation server at each sampling step Internal execution status verification formula:

[0105] ;

[0106] The system sets up a time delay tolerance variable. If it occurs consecutively Duration The system determines that the current microgravity quality has deviated from the fidelity requirement of the corresponding mission phase. At this time, the control system directly issues an experiment interruption command and generates an abnormal log with a timestamp to cut off the continued accumulation of invalid data.

[0107] Step 3: Verification of Virtual-Real Consistency Based on High-Fidelity Digital Twin

[0108] The physical system and digital twin are computed in parallel under the strict same clock source. The system acquires the measured state vectors, including pose, velocity, and contact force, output by the physical air-bearing platform. Synchronously record the ideal zero-gravity state vector output by the digital twin. The system operates on an integral period basis. Calculate dynamic variability:

[0109] ;

[0110] In the formula, To calibrate a positive definite weighted diagonal matrix based on the noise covariance of each sensor measurement, the system sets an upper limit for the variability tolerance. When detected If the physical system's operating state deviates significantly from the ideal space environment, the ground verification results lose credibility.

[0111] Step 4: Online decoupling of multi-source disturbances and measurement of variance contribution

[0112] This step is executed only when the default condition in step two or three is triggered. The system extracts information before and after the anomaly occurrence point. Multi-source sensor data within a time period constitute the sensor observation matrix. ,in Number of sensor channels This represents the total number of sampling points.

[0113] The system utilizes the Fast Independent Component Analysis (FastICA) algorithm to construct a linear unmixing model:

[0114]

[0115] In the formula, For the separation matrix, To obtain the decoupled independent physical interference source matrix, the system extracts the time-domain waveform characteristics and spectral distribution characteristics of each independent signal component and divides them into specific physical interference items, such as low-frequency ground micro-vibrations transmitted by the factory foundation, high-frequency disturbances in air film thickness caused by pressure fluctuations in the air bearing, and other physical interference items in the ground test process.

[0116] The system constructs an extended state observer (ESO) to estimate the sum of nonlinear residual disturbances that cannot be directly separated online. Let the single-axis motion equation of the air-bearing platform be... ,in This represents the total disturbance term, including unmodeled system dynamics and external disturbances. The system is constructed using a third-order ESO:

[0117] ;

[0118] ;

[0119] ;

[0120] ;

[0121] in, , The system's position and velocity were tracked separately. Approximating the total system disturbance , , , The observer gain constant is... It is a nonlinear continuous power function.

[0122] The system calculates the variance of each definite physical component and the observer-estimated component within the diagnostic window. Total variance The system calculates and outputs the variance contribution rate of each physical source. It provides precise quantitative results, offering data input for structural tuning or feedforward torque compensation in control systems.

[0123] Step 5: Performance Prediction and Anomaly Diagnosis Based on Machine Learning

[0124] The system accesses the historical test database, which contains... Group of historical test samples, each sample group structure is defined as follows: ,in , This is a vector of static configuration parameters for the system before testing. It is saved according to the actual experimental requirements, such as the total load of the air-bearing platform, the initial value of the three-dimensional coordinates of the centroid, and the ambient temperature and humidity. It is a set of instructions for time-series task control; This is the sequence of actual performance indicators obtained in step one. The sequence of virtual-real consistency dynamic difference generated in step three; This is the variance contribution rate matrix of each physical interference source obtained in step four.

[0125] The system constructs a multi-task prediction model based on graph neural networks (GNNs) and long short-term memory networks (LSTMs). The network weight parameter set is During the model training phase, a multi-objective joint optimization strategy is adopted, with the objective function being... Defined as the weighted mean squared error loss between the predicted sequence, the true sequence, and the attribution results:

[0126] ;

[0127] In the formula, , and These are the model's predicted performance index, variance, and disturbance contribution rate, respectively. , , This refers to the multi-task loss weighting coefficient; Regularization coefficient

[0128] Before conducting a new physical test on the air flotation platform, the operator inputs the configuration parameters for the current test plan. With the planned instruction set The system derives the expected performance curve through feedforward calculation. Check whether it satisfies the requirements in all set task phases. and If the predicted curve is at risk of falling below the expected threshold or being severely distorted between the predicted and actual curves, the system will output an interception prompt on the operating system interface and refuse to start the experiment.

[0129] If the prediction is successful and the experiment is initiated, the system will calculate the monitoring residual at the current moment in real time:

[0130] The system extracts the standard deviation of the residual distribution in the training set. When the conditions are met If an unmodeled anomaly occurs during the experiment, the system will call the Shapley-Sum Interpretation (SHAP) algorithm to calculate the static configuration parameter vector. Each feature element in the current residual Deviation in marginal contribution value :

[0131] ;

[0132] In the formula, For the set of all feature parameters, Not including the first A subset of features This is the expected output function of the model under a specific subset of features. The system is based on... Sort the values ​​in descending order of their absolute values, and combine them with the predicted values. The system outputs the most likely physical cause of the anomaly. After the experiment is completed, the system encapsulates the newly generated data sequence along with SHAP attribution labels, updates the historical experiment database, triggers backpropagation to optimize model parameters, and realizes the self-iteration and evolution of the entire microgravity quality assessment method.

[0133] Based on the above five steps, a unified definition is provided for the six-DOF microgravity mass characterization index based on mission phase drive, denoted as [insert definition here]. This index combines six physical indicators with the current experimental phase to arrive at a comprehensive score between 0 and 1. The closer the score is to 1, the closer the ground simulation environment is to the real microgravity environment in space. The specific scoring rules are summarized as follows:

[0134] (1) A unified score is assigned to the six physical characteristics. Due to the inconsistency of dimensions, a negative exponential function is used to convert the six physical quantities collected in step one into scores between 0 and 1, generating a feature vector:

[0135] ;

[0136] Here arrive It is an adjustment coefficient that maps various physical quantities to a reasonable score range.

[0137] (2) Dynamically set adjustment coefficients according to the current task stage. In different task stages (approach, capture, refueling, separation), the system has different concerns regarding the performance of the air-floating platform. Therefore, a coefficient is set based on the current task stage. Changing diagonal adjustment matrix :

[0138] ;

[0139] The specific adjustment rules are as follows:

[0140] In the approaching section ( The focus is on position tracking and low-frequency stability, with translational and low-frequency correlation coefficients being considered. , Set it to 2, and the rest to 1;

[0141] In the capture segment ( The focus is on the attitude interference resistance at the moment of collision, and the rotation correlation coefficient is used. , Set it to 3, and the rest to 1;

[0142] In the refueling section ( The focus is on the swaying and parasitic torque caused by changes in mass. , Set the larger value as the penalty coefficient, and set the rest to 1;

[0143] In the separation section ( The focus is on the instantaneous impact during the separation process, specifically the core coefficients of translation and rotation. , Set it to 3, and the rest to 1.

[0144] (3) Dynamic weight normalization. The basic weight vector set in step one. To ensure that the final calculated total score does not exceed the limit, the weights are adjusted and scaled proportionally to obtain the final dynamic weight vector: .

[0145] (4) Calculate the final quality characterization index. Multiply the normalized dynamic weight vector by the eigenvector to obtain the microgravity quality characterization coefficient at the current moment: .

[0146] Example 2:

[0147] This embodiment uses a "non-cooperative target capture experiment" as a background to extract the moment of contact and collision ( The data at any given time were used to perform calculations and simulations to verify the core beneficial effects of the present invention.

[0148] 1. Setting up basic experimental data

[0149] Before the experiment started, the system's six-dimensional basic weight vector was set as follows:

[0150] ;

[0151] exist At a certain moment, the air-bearing platform's robotic arm collides physically with the target. Data collected by the underlying sensors and mapped using a dimensionless negative exponential function generates a six-dimensional microgravity feature mapping vector for the current moment. .

[0152] Since the collision mainly causes an instantaneous twisting of the attitude, while the translation of the center of mass is less affected, the scores for each dimension after mapping are as follows (out of 1.0):

[0153] Translational acceleration score: 0.90 (Excellent)

[0154] Low-frequency PSD score: 0.85 (Excellent)

[0155] Rotational angular acceleration score: 0.40 (Poor, instantaneous impact occurred)

[0156] Attitude drift rate score: 0.50 (below average, attitude deflection occurred)

[0157] Cross-coupling score: 0.80 (Good)

[0158] Internal force circulation score: 0.90 (Excellent)

[0159] That is, the input feature vector is: Meanwhile, it is assumed that the lower limit threshold for the capture segment test is set to... .

[0160] 2. Dynamic evaluation method derivation

[0161] This system incorporates a finite state machine to identify the current state as "capture segment" (…). The system automatically applies adjustment rules: focusing on the attitude anti-interference capability at the moment of collision, and adjusting the rotation correlation coefficient. , Set it to 3, and the rest to 1.

[0162] Step 1: Generate Attention Matrix and Weight Amplification

[0163] ;

[0164] The amplified weights are ;

[0165] Step 2: Dynamic Weight Adaptive Normalization

[0166] The amplified total weight is The system performs forced normalization to obtain the dynamic weight vector of the current phase:

[0167] ;

[0168] Step 3: Calculate the microgravity quality characterization index

[0169] Multiply the new dynamic weights by the feature vector:

[0170] ;

[0171] Step 4: State Verification and Closed-Loop Decision

[0172] Execution status verification formula: ;

[0173] Assessment conclusion: Due to The system determines that the microgravity quality is substandard at the current moment. The system immediately triggers an alarm, cuts off the accumulation of invalid data, and activates the underlying multi-source perturbation decoupling algorithm for attribution diagnosis.

[0174] However, in traditional evaluation methods, the evaluation weights do not change with the task stage, and the evaluation system directly uses the inner product summation of the basic weights and the feature vectors: Because the translational score was extremely high, it masked the distortion in the rotational dimension, resulting in a higher final score. The system will determine that the test is qualified at that moment, which could easily lead to test data with serious attitude distortion being mistakenly considered credible, causing subsequent on-orbit risks.

Claims

1. A six-degree-of-freedom microgravity quality assessment method based on task phase-driven operation, implemented using a six-degree-of-freedom full-physics ground simulation system, wherein the six-degree-of-freedom full-physics ground simulation system includes a six-degree-of-freedom air-bearing platform, a high-frequency sensor network, a host computer task control terminal, and a performance evaluation server, characterized in that... The method includes the following steps: S1. Construct a six-degree-of-freedom microgravity mass index 6D-MQI based on multi-physics field fusion. Obtain real-time state data of the air-bearing platform through a sensor network, calculate multi-dimensional feature parameters based on a sliding time window, and input them into the multi-dimensional state fusion equation to output the six-degree-of-freedom microgravity mass index in real time. S2. Implement dynamic performance monitoring driven by task phase, establish a task phase recognition model based on finite state machine, divide the continuous time axis of the task into multiple discrete task phases, match the corresponding dynamic weights and evaluation thresholds for different task phases, perform state verification in real time, issue test interruption commands and generate abnormal logs for microgravity quality deviation scenarios. S3. Based on digital twin, complete the virtual-real consistency verification, control the physical system and digital twin to solve in parallel under the same clock source, collect the measured state vector of the physical system and the ideal zero-gravity state vector of the digital twin, calculate the dynamic difference degree, and determine the on-orbit credibility of the ground test results based on the preset tolerance upper limit. S4. Online decoupling and contribution rate quantification of multi-source disturbances: triggered when the state verification in step S2 fails or the dynamic difference in step S3 exceeds the limit, extracting multi-source sensor data during abnormal periods, completing the decoupling and separation of multi-source disturbances and the estimation of nonlinear residual disturbances, and calculating and outputting the variance contribution rate of each disturbance source. S5. Based on machine learning, performance prediction and anomaly diagnosis are achieved. A GNN-LSTM multi-task prediction model is built by calling the historical test database. Before the test, the performance is pre-verified based on the current test plan, and high-risk tests are intercepted. During the test, anomalies are monitored in real time and the anomaly attribution is completed through the SHAP algorithm. After the test, the database is updated and the model is iteratively optimized.

2. The six-degree-of-freedom microgravity quality assessment method based on task phase drive according to claim 1, characterized in that, Step S1 specifically involves: The length is set as The sliding time window is used to set the sampling step size. Within each sliding window, the characteristic parameters of translational dimension, rotational dimension, spatial coupling effect, and internal force circulation dimension are calculated. The characteristic parameters are then substituted into the multidimensional state fusion equation to output the six-degree-of-freedom microgravity quality index. The translational dimension feature parameters include: Triaxial residual translational acceleration vector The corresponding root mean square value of translational acceleration and the use of the Welch method to After performing a discrete Fourier transform, the vertical power spectral density integral within the 0.1Hz to 10Hz frequency band is: ;in, Let be the translational acceleration in the x-axis direction at time t. Let be the translational acceleration in the y-axis direction at time t. Let be the translational acceleration in the z-axis direction at time t. Sliding time window length, For signal frequency, Let be the power spectral density function of the translational acceleration in the z-direction; The rotation dimension feature parameters include: Triaxial residual angular acceleration vector The corresponding root mean square value of rotational angular acceleration ,in, The residual angular acceleration vector on the roll axis. The residual angular acceleration vector on the pitch axis. Let $\mathbf$ be the residual angular acceleration vector on the yaw axis and the magnitude of the attitude drift rate vector per unit time during the phase without control torque input. ,in Set the initial Euler angles for the window. The Euler angle at the end of the window; the characteristic parameter of the spatial coupling effect is the Z-axis translational acceleration. With pitch acceleration Pearson correlation coefficient The characteristic parameter of the internal force circulation dimension is the internal force circulation index. ,in The number of active actuators inside the air flotation platform. The nominal thrust output by the i-th actuator. This represents the total external environmental load on the air-bearing platform. To ensure the effectiveness of the internal force circulation index, the amount is kept to a minimum; The multidimensional state fusion equation is: ； in, to These are the basic weight coefficients determined by the analytic hierarchy process. to As a scale factor for dimensional unification, all physical parameters are normalized to the range of 0 to 1, thereby achieving unified quantification of features across all dimensions.

3. The six-degree-of-freedom microgravity quality assessment method based on task phase drive according to claim 2, characterized in that, Step S1 further includes dimensionless scoring of the six physical characteristic parameters to generate a microgravity quality characterization index between 0 and 1. The specific steps are as follows: The six physical feature parameters are converted into scores between 0 and 1 using a negative exponential function, generating a feature vector: ； Based on the current mission phase, set the mission phase accordingly. Changing diagonal adjustment matrix : ； in, The task phase dynamic weight adjustment coefficient is used to dynamically normalize the basic weight vector. This yields the final dynamic weight vector: ； Calculate the microgravity mass characterization index at the current moment: ； The The score is a composite score between 0 and 1. The closer the score is to 1, the closer the ground simulation environment is to the real microgravity environment in space.

4. The six-degree-of-freedom microgravity quality assessment method based on mission phase drive according to claim 3, characterized in that, Step S2 specifically involves: Based on the task phase identification model of the finite state machine (FSM), let the total system running time axis be... The task sequence variable issued by the host computer is The state machine is based on the Euclidean distance between the end effector of the air-bearing platform and the target aircraft. The contact force sensor status, locking command signal of the docking mechanism, and unlocking command signal divide the continuous time axis into proximity segments. Capture segment Refueling section Separation section Four discrete phases; The system constructs a time-varying dynamic threshold function for different task phases. and dynamic weight adjustment matrix Within each sampling step, a status check is performed. If the microgravity quality continues to deviate from the fidelity requirement of the corresponding task phase, an experiment interruption command is issued and an anomaly log is generated.

5. The six-degree-of-freedom microgravity quality assessment method based on task phase drive according to claim 4, characterized in that, In step S2, the rules for dynamic weight adjustment and threshold setting for different task phases are as follows: when At this point, the weights of the root mean square component of translational acceleration and the integral component of low-frequency power spectrum are increased to twice the basic weights, and the lower limit threshold for the quality index is set at this time. ; when At this point, the weights of the root mean square component of the rotational angular acceleration and the attitude drift rate component are increased to three times the base weights, and the lower limit threshold for the quality index is set at this time. ; when At this time, the translational-rotational cross-interference coefficient and the vertical residual acceleration are set as rigid penalty terms, and the lower limit threshold for the quality index at this time is set as follows. ; when At this point, the weights of both the root mean square components of translational residual acceleration and rotational residual angular acceleration are simultaneously increased to three times their base weights. The minimum passing threshold for the quality index at this time is set as follows: .

6. The six-degree-of-freedom microgravity quality assessment method based on mission phase drive according to claim 5, characterized in that, The state verification formula in step S2 is: ； Establish time delay tolerance variable If it appears consecutively Duration The system determines that the current microgravity quality has deviated from the fidelity requirement of the corresponding mission phase, issues an experiment interruption command, and generates an anomaly log with a timestamp.

7. The six-degree-of-freedom microgravity quality assessment method based on mission phase drive according to claim 6, characterized in that, In step S2, the specific determination rule for task phase division based on the finite state machine is: a preset relative distance threshold. , when When the contact force is zero, it is determined to be the approach segment. ; when If the lock signal has not yet been triggered, it is determined to be a capture segment. ; When the docking mechanism lock signal takes effect and the fluid transfer command is initiated, it is determined to be the filling section. ; When the docking mechanism unlocking and separation signal is triggered, and When the interval gradually increases from 0, it is determined to be a separation segment. .

8. The six-degree-of-freedom microgravity quality assessment method based on task phase drive according to claim 7, characterized in that, In step S3, the measured state vector output by the physical air-float platform is acquired. Synchronously record the ideal zero-gravity state vector output by the digital twin. According to the integration period The formula for calculating dynamic variability is: ； In the formula, To calibrate a positive definite weighted diagonal matrix based on the noise covariance of each sensor measurement, the system sets an upper limit for the variability tolerance. When detected If the physical system's operating state deviates significantly from the ideal space environment, the ground verification results lose credibility.

9. The six-degree-of-freedom microgravity quality assessment method based on mission phase drive according to claim 8, characterized in that, In step S4, before and after extracting the point of anomaly occurrence... Multi-source sensor data within a time period constitute the sensor observation matrix. ,in Number of sensor channels The total number of sampling points; a linear unmixing model is constructed using the Fast Independent Component Analysis (FastICA) algorithm: ； In the formula, For the separation matrix, The decoupled independent physical interference source matrix is ​​divided into specific physical interference terms by extracting the time-domain waveform features and spectral distribution features of each independent signal component. An extended state observer (ESO) is constructed to estimate the sum of nonlinear residual disturbances that cannot be directly separated online. The equation of motion for the single-axis air-bearing platform is given by... , To construct a third-order ESO for the total disturbance term, which includes unmodeled system dynamics and external disturbances: ； ； ； ； in, For ESO-tracked system position observations, For system velocity observations tracked by ESO, The total system disturbance observations approximated by ESO. For observation error, It is a non-linear power factor. , For the threshold of the linear segment of a nonlinear function, , The system's position and velocity were tracked separately. Approximating the total system disturbance , , , The observer gain constant is... It is a nonlinear continuous power function; Calculate the variance of each explicit physical component and the observer-estimated component within the diagnostic window. Total variance Calculate and output the variance contribution rate of each physical source. .

10. The six-degree-of-freedom microgravity quality assessment method based on mission phase drive according to claim 9, characterized in that, In step S5, the historical test database called by the system contains Group of historical test samples, each sample group structure is defined as follows: ,in , This is the static configuration parameter vector before system testing. It is a set of instructions for timing task control; This is the sequence of actual performance indicators obtained in step one. The sequence of virtual-real consistency dynamic difference generated in step three; This is the variance contribution rate matrix of each physical interference source obtained in step four; Constructing a multi-task prediction model based on graph neural network (GNN) and long short-term memory network (LSTM) The network weight parameter set is During the model training phase, a multi-objective joint optimization strategy is adopted, with the objective function being... Defined as the weighted mean squared error loss between the predicted sequence, the true sequence, and the attribution results: ； In the formula, , and These are the model's predicted performance index, variance, and disturbance contribution rate, respectively. This refers to the multi-task loss weighting coefficient; The regularization coefficient is used. Before conducting a new physical experiment, enter the configuration parameters for the current experiment plan. With the planned instruction set The expected performance curve is obtained through feedforward calculation. Check whether it satisfies the requirements in all set task phases. and If the predicted curve is at risk of falling below the expected threshold or being severely distorted, an interception warning will be issued and the experiment will be refused to start; if the prediction passes and the experiment starts, the monitoring residual at the current moment will be calculated in real time. Extract the standard deviation of the residual distribution of the training set. ,when If an unmodeled anomaly is detected during the experiment, the Shapley algorithm and the SHAP algorithm are invoked to calculate the static configuration parameter vector. Each feature element in the current residual Deviation in marginal contribution value : ； In the formula, For the set of all feature parameters, Not including the first A subset of features The expected output function of the model under a specific feature subset; according to Sort the values ​​in descending order of their absolute values, and combine them with the predicted values. The system outputs the most likely physical cause of the anomaly. After the experiment is completed, the newly generated data sequence is encapsulated together with the SHAP attribution label and updated to the historical experiment database. This triggers backpropagation to optimize the model parameters, enabling the method to self-iterate and evolve.

Images