Flight simulator motion platform health management system and method based on digital twinning

Abstract

The present application belongs to the technical field of flight simulator health management, and relates to a flight simulator motion platform health management system and method based on digital twinning, which solves the problem that the prior art cannot meet the high-reliability continuous operation guarantee and predictive precise maintenance requirements of the motion platform. The data perception layer collects operation data, and based on the operation data, the digital twinning model layer uses a physical mechanism model and an LSTM data-driven compensation network to calculate spatial pose residuals; the state evaluation and diagnosis layer calculates the physical performance degradation index of each actuator through inverse Jacobian dimensionality reduction mapping and decoupling correction; the decision optimization layer dynamically sets an early warning threshold based on the flight training subject, constructs a health degree-driven degradation operation envelope optimization model, and outputs the maximum available degradation envelope; in the human-computer interaction layer, the envelope is converted into a limiting instruction and executed, and a standardized maintenance work order is generated. The present application realizes predictive precise maintenance of the six-degree-of-freedom motion platform and guarantees high-reliability continuous operation.

Description

Technical Field

[0001] This invention belongs to the field of flight simulator health management technology, specifically relating to a health management system and method for a flight simulator motion platform based on digital twins. Background Technology

[0002] The six-degree-of-freedom motion platform of a full-motion flight simulator is a core piece of equipment for flight training, and its health status directly affects flight training safety and simulator availability. Currently, the health management of the six-degree-of-freedom motion platform mainly relies on the following technical means: 1. Regular maintenance and routine inspections, based on flight hours or calendar time, involve the periodic replacement, lubrication, and functional testing of critical components of the motion platform. This method is a passive maintenance approach, unable to adjust maintenance strategies according to the actual health status of the equipment, and is prone to "over-maintenance" or "under-maintenance."

[0003] 2. Built-in In-System Testing (BIT): During system power-on or operation, key parameters are collected by built-in sensors and compared with preset thresholds to trigger fault alarms. This method can only provide post-incident alarms and cannot provide early warnings or predict remaining lifespan.

[0004] 3. Single-parameter trend monitoring records the historical curves of key parameters for individual components to determine performance degradation trends. However, there is strong motion coupling between the actuators in a six-degree-of-freedom motion platform. Isolating and monitoring a single component can mask early fault characteristics, leading to delayed fault detection.

[0005] 4. Model-based self-diagnosis of control systems monitors the following error between actuator position commands and actual feedback. This method relies on a single error index, lacks the ability to comprehensively assess the overall health of the system, and the fixed threshold cannot adapt to the dynamic load differences in different flight training subjects.

[0006] Existing technologies generally have the following problems when dealing with the health management needs of a complex electromechanical-hydraulic system such as a six-degree-of-freedom motion platform: (1) data silos and information fragmentation, ignoring the strong motion coupling between actuators; (2) lack of deep integration of physical mechanisms and data-driven approaches, pure data-driven approaches are prone to false alarms, and pure model-driven approaches are not accurate enough; (3) static thresholds do not match dynamic operating conditions, and fixed thresholds cannot adapt to the dynamic load differences of different flight training subjects; (4) fault diagnosis is lagging behind, and early warning and remaining life prediction cannot be achieved; (5) lack of operable maintenance decision support, and the remaining available capacity in the "sick" state cannot be assessed. Summary of the Invention

[0007] To address the aforementioned problems in existing technologies, namely the masking of early faults due to strong motion coupling of multiple actuators in the health management of existing six-DOF motion platforms of full-motion flight simulators, the lack of deep integration between physical mechanisms and data-driven approaches, the mismatch between static thresholds and dynamic training subjects, the inability to achieve early warning and remaining life prediction due to lagging fault diagnosis, and the lack of executable degraded operation and maintenance decision support, thus failing to meet the requirements for high-reliability continuous operation assurance and predictive and accurate maintenance of motion platforms, this invention proposes a flight simulator motion platform health management system based on digital twins. This system includes: The data perception layer is configured to: collect real-time multi-source sensor data from the six-degree-of-freedom motion platform of the full-motion flight simulator, and obtain operational data through preprocessing; The digital twin model layer is configured to: based on the running data, calculate the theoretical driving force and the correction compensation force by mixing and performing forward evolution, and measure the spatial pose residual between the actual feedback pose and the theoretical expected pose; The state assessment and diagnosis layer is configured to: perform dimensionality reduction mapping on the spatial pose residuals, thereby obtaining the independent length unidirectional residuals and residual accumulation energy of each actuator; and decouple the residual accumulation energy by combining the real-time operating condition characteristics of the actuators to obtain the physical performance degradation index of each actuator. The decision optimization layer is configured as follows: a dynamic early warning threshold is set based on the characteristics of the current flight training subject; when the physical performance degradation index triggers the dynamic early warning threshold, boundary optimization is performed in combination with the multi-dimensional safety constraint set to solve and generate the maximum available degraded operation envelope under the current health state. The human-computer interaction layer is configured to: convert the maximum available degraded operating envelope into a limiting control command and send it to the underlying motion controller; and render the device's 3D visualization interface and generate maintenance work orders based on the physical performance degradation index.

[0008] The second aspect of this invention proposes a health management method for a flight simulator motion platform based on digital twins, and a health management system for a flight simulator motion platform based on digital twins. This method includes the following steps: S1 collects real-time multi-source sensor data from the six-degree-of-freedom motion platform of the full-motion flight simulator, and obtains operational data after preprocessing; S2, based on the running data, the theoretical driving force and the correction compensation force are mixed and calculated and forward evolution is performed to calculate the spatial pose residual between the actual feedback pose and the theoretical expected pose. S3, perform dimensionality reduction mapping on the spatial pose residual, and then obtain the independent length unidirectional residual and residual accumulation energy of each actuator; and combine the real-time operating condition characteristics of the actuator to decouple the residual accumulation energy to obtain the physical performance degradation index of each actuator. S4. Set a dynamic early warning threshold based on the characteristics of the current flight training subject; when the physical performance degradation index triggers the dynamic early warning threshold, perform boundary optimization by combining the multi-dimensional safety constraint set, and solve to generate the maximum available degraded operation envelope under the current health state. S5, convert the maximum available degraded operating envelope into a limiting control command and send it to the underlying motion controller; and render the device's three-dimensional visualization interface and generate a maintenance work order based on the physical performance degradation index.

[0009] The beneficial effects of this invention are: 1) By using spatial pose residual analysis and inverse Jacobian dimension reduction mapping, the pose residual of the six degrees of freedom coupling is decoupled into the independent length unidirectional residual of each actuator. Combined with the physical performance degradation index of decoupling correction of running speed and load force, the health status of each actuator is independently and accurately quantified, which overcomes the masking effect of strong motion coupling of multiple actuators on early faults, the fault warning time is advanced on average, and the false alarm rate is reduced. 2) Construct a hybrid driving architecture that deeply integrates the physical mechanism model and the LSTM data-driven compensation network at the force level. The mechanism model ensures physical interpretability, and the data-driven network compensates for model bias and unmodeled dynamics. With the help of adaptive Kalman filtering, the key parameters of the mechanism model are calibrated online to achieve high-fidelity mapping of the digital twin to the physical platform. This solves the problems of false alarms caused by pure data-driven and insufficient accuracy of pure model-driven, and enables accurate detection of early faults in the six-degree-of-freedom motion platform. 3) Establish a dynamic early warning threshold generation mechanism driven by flight training subject parameter table. Obtain the current training subject identifier from the instructor's station in real time. Dynamically adjust the health assessment threshold through a weighted combination of amplitude factor, frequency factor and load intensity. With the matching smoothing filtering and linear transition strategy, the health assessment standard adapts to the subject load characteristics, which solves the problem of false alarm / missed alarm caused by the mismatch between fixed threshold and dynamic working conditions. 4) Deep neural network multi-source fusion diagnosis that integrates wavelet packet frequency band energy / kurtosis features with LuGre friction parameters for online identification, combined with LSTM temporal degradation extraction and MC Dropout confidence interval calculation for remaining lifetime prediction, achieves accurate classification of fault modes and probabilistic estimation of remaining lifetime. 5) Construct a health-driven degradation operation envelope optimization model. The upper limits of multi-dimensional safety constraints such as speed, force, and current shrink exponentially with the decline of physical performance. Use a sequential quadratic programming algorithm to solve the optimal envelope boundary online to achieve a quantitative balance between safety and availability under the condition of "operating with defects". The three-level progressive degradation strategy avoids the rigid treatment of direct shutdown.

[0010] The entire chain forms an automated closed loop of "perception → modeling → evaluation → decision-making → interaction". Standardized maintenance work orders are automatically generated and connected to the enterprise asset management system, reducing spare parts inventory costs, improving equipment availability, shortening the mean time to repair, and meeting the needs of motion platform for high-reliability continuous operation and predictive precision maintenance. Attached Figure Description

[0011] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings.

[0012] Figure 1 This is an architecture diagram of the health management system for a flight simulator motion platform based on digital twins, as described in this invention.

[0013] Figure 2 This is a flowchart of the digital twin model layer of the health management system for a flight simulator motion platform based on digital twins, as described in this invention.

[0014] Figure 3 This is a flowchart of the decision optimization layer of the health management system for a flight simulator motion platform based on digital twins, as described in this invention.

[0015] Figure 4 This invention relates to a three-dimensional visualization interface for the human-computer interaction layer of a health management system for a flight simulator motion platform based on digital twins.

[0016] in, Figure 4 The diagram shows six actuators, labeled ①, ②, ③, ④, ⑤, and ⑥. In this 3D visualization interface, different colors represent the health status of each actuator: green represents a healthy state, yellow represents an attentive state, and red represents an alarm state. Specifically, actuators ①, ②, ④, and ⑥ are in a healthy state, actuator ③ is in an attentive state, and actuator ⑤ is in an alarm state. Detailed Implementation

[0017] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0018] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0019] To more clearly explain the health management system for a flight simulator motion platform based on digital twins of this invention, the following will be combined with... Figures 1 to 4 The steps in the embodiments of the present invention will be described in detail below.

[0020] The first embodiment of the present invention is a health management system for a flight simulator motion platform based on digital twins, see [link to relevant documentation]. Figure 1 The system comprises: a data perception layer, a digital twin model layer, a state assessment and diagnosis layer, a decision optimization layer, and a human-computer interaction layer; the data flow direction is: physical platform → data perception → digital twin model → state assessment and diagnosis → decision optimization → human-computer interaction; the layers are connected via an EtherCAT bus, and the data refresh rate is no less than 500Hz (motion control loop) and 100Hz (health management loop); In terms of hardware deployment, the data perception layer is configured with a real-time controller as the EtherCAT master station, responsible for hard real-time acquisition and distributed clock synchronization; the human-computer interaction layer is configured with a three-level computing unit consisting of edge computing nodes, a central health management server, and a motion control computer; the edge computing nodes use STM32H7 or Zynq FPGA (each actuator) for data preprocessing and local LSTM inference; the central health management server uses an Intel Xeon processor, 64GB RAM, and an RTX 4000 GPU, running the Linux RT operating system, and is responsible for digital twin modeling, global optimization, and RUL prediction; the motion control computer uses the PXIe platform, runs the VxWorks real-time operating system, and has a latency of less than 50μs with the server; the health management server adopts a 1+1 redundancy configuration, with a master-slave switching time of less than 1 second; when the server fails, the motion control system automatically reverts to a fixed threshold protection mode; The details of each layer are as follows: The data perception layer is configured to: collect real-time multi-source sensor data from the six-degree-of-freedom motion platform of the full-motion flight simulator, and obtain operational data through preprocessing; In this embodiment, the data perception layer deploys multiple types of sensors on the six-degree-of-freedom motion platform of the full-motion flight simulator. These sensors include displacement sensors, a six-degree-of-freedom pose measurement unit, pressure sensors, vibration sensors, temperature sensors, and strain gauges deployed at various actuators and key components. These sensors are synchronously activated to acquire data, and each sensor operates independently according to a preset sampling rate: 1kHz for displacement and pressure sensors, 200Hz for the six-degree-of-freedom pose measurement unit, 5kHz for the vibration sensor, 10Hz for the temperature sensor, and 100Hz for the strain gauges. All sensor signals are connected to the real-time controller via an EtherCAT bus, and a distributed clock protocol is used to achieve time synchronization between nodes with an accuracy better than ±1μs. The specific configuration is shown in Table 1 below. Table 1

[0021] The real-time controller aggregates data from all sensors, ensuring a data refresh rate of at least 500Hz for the motion control loop. It then passes data copies to the edge computing nodes via a shared memory mechanism (100Hz push cycle). The edge computing nodes (STM32H7 / Zynq FPGA) execute the following sequentially: Noise reduction: 4th order low-pass filter, cutoff frequency adaptive according to analysis purpose (1kHz / 200Hz). Outlier removal: Automatic identification using the 3σ criterion, with linear interpolation or replacement based on preserving previous values; Missing data completion: LSTM prediction model imputation (long-term loss) or cubic spline interpolation (short-term loss); To achieve spatial reference unification for multi-source heterogeneous sensor data, the following coordinate system is defined: Base coordinate system {B}: The origin is located at the geometric center of the lower platform, the X-axis is forward (heading +), the Y-axis is to the right, and the Z-axis is upward (vertical). This coordinate system is a globally fixed reference system. The coordinates of all actuator lower hinge points are fixed in this system. The spatial pose of the six-degree-of-freedom motion platform is ultimately described in this system. Platform coordinate system {P}: The origin is located at the geometric center of the upper platform, and it coincides with {B} at the initial moment. It moves with the platform. The coordinates of the hinge points of the six actuators are fixed in this system. The translational and rotational motion of the platform is described by the pose change of {P} relative to {B}. Sensor coordinate system {Si}: Each actuator has a built-in displacement sensor corresponding to a sensor coordinate system. The measurement axis is along the actuator axis, and the origin is located at the lower hinge point of the actuator. The change in rod length output by the displacement sensor is expressed in this system and is converted to the platform pose through forward kinematics. Pose measurement coordinate system {M}: The laser tracker or inertial measurement unit (IMU) is installed on the upper platform, and its output is the pose measurement value relative to the inertial space. This measurement value is transformed to the platform coordinate system {P} through the factory-calibrated rotation and translation matrix. Then, combined with the pose relationship between {P} and {B}, it is finally unified to the base coordinate system {B} to participate in the calculation of spatial pose residual. Coordinate unification transformation rule: The actuator displacement sensor measurement value is mapped from {Si} to {P} through the forward kinematics solution; the six-degree-of-freedom pose measurement value is transformed from {M} to {P} through the calibration matrix; all data are finally unified to {B} to form operation data with consistent spatial reference; it should be noted that the multi-sensor step-by-step coordinate unification transformation framework adopted is a mature existing technology, and will not be elaborated here. After this preprocessing, the running data is obtained and can be used by the digital twin model layer. Digital twin model layer, see Figure 2The configuration is as follows: based on the running data, the theoretical driving force and the correction compensation force are calculated in a hybrid manner and then evolved forward to calculate the spatial pose residual between the actual feedback pose and the theoretical expected pose. In this embodiment, the spatial pose residual between the actual feedback pose and the theoretical expected pose is calculated, specifically as follows: A1. The running data is input into the physical mechanism model and the data-driven compensation network respectively. The physical mechanism model calculates the theoretical driving force of each actuator based on the current platform motion state, and the data-driven compensation network generates the correction compensation force of each actuator based on the current and historical time-series sensing characteristics. The digital twin model layer includes a physical mechanism model, a data-driven compensation network, a hybrid twin fusion unit, and an online calibration unit. The physical mechanism model includes kinematic models, dynamic models, and electro-actuator models; The kinematic model is based on the Stewart platform configuration. The coordinates of the upper hinge points of the six actuators are fixed in the platform coordinate system {P}, and the coordinates of the lower hinge points are fixed in the base coordinate system {B}. The platform pose is represented by the position vector and Euler angles, the rotation matrix is ​​calculated by Euler angles, and the actuator length vector is obtained by calculating the platform position, rotation matrix, and coordinates of the upper and lower hinge points. ; The actual length is the Euclidean norm of the actuator length vector: ; In the formula, Let be the length vector of the i-th actuator; It is the Euclidean norm; This is the actual length of the i-th actuator; (i=1..6) are the coordinates of the upper hinge point, which are fixed in the platform coordinate system {P}. (i=1..6) represents the coordinates of the lower hinge point, which are fixed in the base coordinate system {B}. The platform pose is determined by the position vector. Euler angles (ZYX order) representation; rotation matrix By Euler angle calculate; Given the desired platform pose Calculate the expected length of each actuator. : First, the Jacobian matrix Linking platform speed with actuator speed: ,in: The actuator velocity vector; For platform speed, (where angular velocity is the angular velocity) Secondly, the reverse Jacobi is adopted. Used to map actuator length error to pose error: ; When the condition number of the Jacobian matrix When the configuration is determined to be close to a singular configuration, the residual analysis of the health management system is paused and updated, the previous valid value is used to maintain the value, and the event log is recorded. The dynamic model uses the Newton-Euler method to establish the dynamic equations of the upper platform. The driving force of each actuator along the axial direction is related to the force on the platform through the transpose of the Jacobian matrix: ; in: It is a 6×6 mass matrix, which includes the platform mass and moment of inertia; It is a 6×6 Coriolis force / centrifugal force matrix; It is a 6×1 gravity term vector; This refers to the driving force of each actuator along the axial direction; Based on the above dynamic equations, the physical mechanism model performs an inverse solution to the dynamic equations according to the current platform motion state (pose, velocity, acceleration) and the motion parameters of each actuator, to obtain the theoretical driving force along the axial direction of each actuator required to balance the current platform spatial motion parameters. This theoretical driving force This is the theoretical driving force vector output by the physical mechanism model; The electric actuator model is established for an electric cylinder, which consists of a servo motor, a ball screw pair, a cylinder body, and a built-in position sensor. It converts the motor's rotational motion into linear motion, including: Servo motor electromagnetic torque equation: ; in: The electromagnetic torque is (Nm). The motor torque constant (Nm / A); The current is the q-axis current (A). Equation of motor shaft rotation: ; in: The moment of inertia of the motor rotor (kg·m²) The motor rotation angle is in rad. The viscous damping coefficient is (Nm·s / rad). The load torque is (Nm). Ball screw transmission relationship: linear displacement of actuator With motor rotation angle Relationship: ; Actuator linear velocity With motor angular velocity Relationship: ; in: The lead of the lead screw (m / rev); Relationship between load torque and actuator output force: ; in: The efficiency of the lead screw drive is 0.90~0.95. The actuator outputs axial force (N); Actuator output force balance equation: ; in: External load force (N); Friction force (using the LuGre model); Equivalent mass of moving parts (kg); Friction model (LuGre model): ; ; ; in: The average deformation of the bristles (m); The actuator speed (m / s); The stiffness coefficient is (N / m). The damping coefficient is (N·s / m). The coefficient of viscous friction is denoted as N·s / m. Coulomb friction (N); Static friction (N); Stribeck's velocity (m / s); Electrical equations of an electric motor (dq coordinate system): ; ; in: The voltages are d-axis and q-axis (V). The currents are d-axis and q-axis (A). Stator resistance (Ω); For d-axis and q-axis inductance (H); ω is the electric angular velocity (rad / s). ( (Number of pole pairs of the motor); For permanent magnet flux linkage (Wb); The power and thermal model includes motor power calculation: ; Motor temperature rise (first-order thermal model): ; in: Temperature of the motor windings (°C); The thermal time constant (s); Thermal resistance (°C / W); For power loss; The key parameters and identifiable range of the mechanism model are shown in Table 2 below: Table 2

[0022] Degradation mechanism models include: Lead screw wear model: transmission efficiency With running time The index declined: ; in: For initial efficiency; The wear coefficient is (fitted from historical data).

[0023] Bearing fatigue model: The remaining bearing life is calculated based on the LP model. ; in: For the basic rated life (10 6 change); The basic rated dynamic load (N); For the equivalent dynamic load (N); Life index (3 for ball bearings, 10 / 3 for roller bearings); Actual remaining lifespan is adjusted according to health status: ; in: This is a performance degradation index; This is the load correction factor; Motor insulation aging model: Motor winding life decreases exponentially with increasing temperature (Arrhenius model): ; in: This refers to the remaining life of the insulation. Lifespan at rated temperature; Activation energy (eV); It is the gas constant; Rated temperature (K); The actual winding temperature (K); Grease aging model: friction (Coefficient of viscous friction) increases linearly with operating time: ; in: The initial viscous friction coefficient; This refers to the aging rate. The data-driven compensation network uses an LSTM network structure. The input vector contains multi-source sensor features from the current time and the historical time window (the past 50ms, corresponding to 50 samples @ 1kHz): displacement of each actuator. ,speed acceleration (Obtained from displacement difference), motor q-axis current (Reflecting output torque), motor d-axis current Servo driver command voltage Motor speed Motor winding temperature Vibration features (feature vectors after CNN dimensionality reduction) and platform pose residual history sequence The input dimension is 50 features × 50 time steps = 2500; The LSTM network structure consists of an input layer, a first LSTM layer, a second LSTM layer, a fully connected layer, and an output layer, all connected in series. The output tensor of the previous layer serves as the input tensor of the next layer. The workflow is as follows: The input layer receives a 2500-dimensional historical time-series feature tensor, where 50 feature channels are flattened into a one-dimensional vector at 50 sampling points along the time axis. This 2500-dimensional vector is then fed into the first LSTM layer. The first LSTM layer contains 128 LSTM units. Each unit performs gated cyclic processing on the input sequence at each time step: first, a forget gate determines the retention ratio of the memory units from the previous time step; then, an input gate determines the amount of candidate memory to be written at the current time step; finally, an output gate generates the hidden state at the current time step. This layer performs point-by-point cyclic processing on the 2500-dimensional feature sequence along the time axis, transforming the time step dimension... The first LSTM layer compresses the output dimension to a 128-dimensional hidden state vector (i.e., 128 dimensions per time step). The second LSTM layer also contains 128 LSTM units, using the 128-dimensional hidden state sequence output from the first LSTM layer as input. It employs the same gated recurrent mechanism for secondary temporal feature extraction, maintaining the 128-dimensional hidden state vector output, further capturing higher-level temporal dependencies. The fully connected layer receives the 128-dimensional vector output from the second LSTM layer, performs a linear transformation of the weight matrix and bias term, and then performs a nonlinear mapping using the ReLU activation function, transforming the 128-dimensional feature vector into a 64-dimensional feature vector. This layer enhances the network's ability to express complex nonlinear features. The output layer is a fully connected layer that receives the 64-dimensional feature vector, maps it to a 6-dimensional output vector through a linear activation function, corresponding to the correction compensation force Δf of the six actuators. i LSTM (i=1,2,...,6); The output is the corrected compensation force. The theoretical driving force output by the mechanism model is superimposed on the training method, which is offline pre-training plus online fine-tuning. The offline training uses historical operation data. Specifically, the offline training data comes from the 6-month historical operation records of the six-degree-of-freedom motion platform of the full-motion flight simulator, including normal operating condition data and 8 types of typical fault operating condition data. The fault operating condition data is obtained through fault injection, including internal leakage, external leakage, bearing wear, seal wear, sensor failure, lead screw wear, and motor demagnetization, with no less than 5,000 sets of samples collected for each. The training samples for online fine-tuning are the most recent 1,000 samples in the real-time operation data cache window, and the parameters are updated every 1 hour. The loss function is mean squared error. Online fine-tuning is performed every hour using the most recent 1000 samples for single-step gradient descent updates, with a learning rate of 0.0001. A2, apply the nonlinear constraint to the corrected compensation force to obtain the constrained corrected compensation force; the nonlinear constraint is specifically a multi-dimensional physical safety constraint, including amplitude constraint, energy constraint and power constraint; The amplitude limiting constraint specifically refers to limiting the amplitude of the corrective compensation force to no more than a preset percentage of the theoretical driving force amplitude. ,in, To correct the compensating force, (specific The theoretical driving force is used to ensure that the compensation force does not exceed 20% of the mechanistic force, thus avoiding the data model from "dominating" the physical system. The energy constraint specifically involves: discretely integrating the square of the correction compensation force amplitude over time; when the accumulated integral value exceeds the maximum compensation energy limit set by the system, performing amplitude attenuation processing on the correction compensation force. ,in, A preset maximum allowable compensation energy threshold (e.g., 100J) is used to prevent system drift caused by long-term energy injection into the data model. The power constraint specifically involves: real-time estimation of the additional power expected to be generated by each actuator due to the superimposed correction compensation force, ensuring that the expected additional power does not exceed a preset percentage threshold of the motor's rated power. For example, for an electric cylinder: ,in, The compensation power is 10% of the motor's rated power to prevent the compensation power from exceeding the motor's capacity. A3, the hybrid twin fusion unit superimposes the constrained correction compensation force component by component onto the theoretical driving force. Above, a mixed driving force is formed; ;in: The theoretical driving force calculated from the current state by the mechanistic model (dynamics + electric cylinder model); The compensation force output by the LSTM; A4. Substitute the hybrid driving force into the platform's forward dynamic differential equation and perform numerical integration to obtain the theoretical desired pose of the platform under the action of the hybrid driving force; subtract the theoretical desired pose from the actual feedback pose to generate a spatial pose residual containing six degrees of freedom. The digital twin model layer also includes an online calibration unit. The online calibration unit uses key physical parameters in the mechanism model as state variables, and uses an adaptive Kalman filter to dynamically update the state variables online. It also uses an adaptive algorithm to update the process noise covariance and observation noise covariance in real time. Combined with the frequency sweep fingerprint benchmark established during factory or overhaul, when the deviation between the real-time frequency sweep response and the frequency sweep fingerprint benchmark exceeds a preset threshold, parameter recalibration or fault alarm is triggered. Specifically, the online calibration unit uses key physical parameters from the physical mechanism model as state variables. ; The system state equation is a parametric slow random walk, specifically: ; Observation equation: Utilizing sensors to measure actuator force Computational power of mechanism model Difference: ; in: The force measured by the sensor; Calculation force for the mechanism model; This is the sensitivity matrix; The Sage-Husa adaptive algorithm is used to update the process noise covariance Q and observation noise covariance R in real time. Specifically: Adaptive Noise Covariance Update (Sage-Husa): ; ; ; in: Forgetting factor; For residuals; Kalman gain; It is the covariance matrix; Initialization and Divergence Suppression: Initial State Extract the manufacturer's nominal value and the initial covariance. ; Divergence suppression: Monitor the autocorrelation function of the residual sequence; if it exceeds a threshold (e.g., 0.3) for 10 consecutive steps, temporarily increase the threshold. Or reset ; After the equipment leaves the factory or undergoes a major overhaul, a frequency sweep excitation test is performed (0.1Hz~10Hz, sinusoidal sweep, amplitude 0.1g~0.5g), and the amplitude-frequency response at each frequency point is recorded. Phase frequency response Harmonic distortion Force-displacement transfer function of each actuator, motor current-output force transfer function; The data is stored as a fingerprint database in JSON format, containing frequency vectors and their corresponding complex values. During online calibration, the current frequency scan test result is compared with the fingerprint database; if the deviation exceeds 5%, parameter recalibration or a fault alarm is triggered. The state assessment and diagnosis layer is configured to: perform dimensionality reduction mapping on the spatial pose residuals, thereby obtaining the independent length unidirectional residuals and residual accumulation energy of each actuator; and decouple the residual accumulation energy by combining the real-time operating condition characteristics of the actuators to obtain the physical performance degradation index of each actuator. In this embodiment, the spatial pose residual is subjected to dimensionality reduction mapping to obtain the independent length unidirectional residual of each actuator and the residual accumulation energy, specifically: Multiplying the spatial pose residual containing six degrees of freedom components by the inverse Jacobian matrix yields independent length unidirectional residuals corresponding to the absolute deviations of the extension and retraction strokes of the six actuators. Discrete square integrals are then performed on the independent length unidirectional residuals of each actuator within a preset time sliding window, and normalized using the maximum stroke of each actuator to obtain the values ​​for each actuator. Residual accumulated energy; Specifically, the spatial pose residual, comprising six degrees of freedom, is: ; After filtering with a first-order low-pass filter (cutoff frequency approximately 5Hz), y(n) = 0.95y(n-1) + 0.05e pose (n); Using the inverse Jacobian matrix Convert the spatial pose residuals into unidirectional length residuals for each actuator: ; Within a preset time sliding window Tw=0.5s, the discrete square integral is performed on the independent length unidirectional residuals of each actuator, and the summation is based on the maximum stroke of each actuator. After normalization, the residual accumulated energy of each actuator is obtained: ; Among them, the maximum stroke of the actuator (e.g., 500mm); Furthermore, by combining the real-time operating characteristics of the actuators, the residual accumulated energy is decoupled and calculated to obtain the physical performance degradation index of each actuator, specifically: The real-time operating condition characteristics of the actuator are obtained, including the operating speed and push-pull load factors. Take the absolute values ​​of the operating speeds of each actuator and the absolute values ​​of the push-pull load factors (specifically, the absolute values ​​of the real-time speeds). and the absolute value of real-time load capacity ), and then multiply them by the offline calibrated speed correction coefficient and load force correction coefficient respectively, and then add the basic constant (1) to construct the decoupling correction coefficient; multiply the ratio of the residual accumulated energy to the baseline residual energy in the healthy state by the decoupling correction coefficient, and subtract the product from 1 to obtain the physical performance degradation index of the actuator; The formula for the decoupling correction coefficient is as follows: ; in: The speed of the actuator; For load capacity; These are the speed correction factor and the load force correction factor, obtained through offline calibration. Preferably, s / mm, N - ¹; The formula for the physical performance degradation index is as follows: ; in, This represents the baseline residual energy under healthy conditions. ∈ [0,1], where 1 represents complete health and 0 represents failure; The fault location mapping rules are as follows: When When the length residual decreases and deviates positively from the current, it is determined to be a decrease in external leakage; when When the temperature decreases and the length residual increases with increasing temperature, it is determined that internal leakage is increasing; when When the vibration decreases and the high-frequency components increase, it is determined to be bearing / guide wear; Pressure pulsation signal and vibration acceleration signal Perform 4-level wavelet packet decomposition (db4 wavelet) to obtain 16 frequency band nodes, and extract the energy of each node. and kurtosis Feature vector dimension: 16 frequency bands × 2 features × 2 signals = 64; LuGre friction model parameter identification was performed using a low-speed triangular wave velocity excitation (±5 mm / s, 0.1 Hz) and offline identification using the Levenberg-Marquardt method. During online operation, parameters were updated every 10 minutes using recursive least squares (RLS). The LuGre friction model formula is as follows: ; ; ; in: The average deformation of the bristles; This is the stiffness coefficient; The damping coefficient; It is the coefficient of viscous friction; Coulomb friction; It is static friction; Stribeck speed; A deep neural network classifier is used to identify fault modes based on the fused multi-source features. The classifier structure consists of an input layer, a first hidden layer, a second hidden layer, and an output layer. The layers are connected in series with full connectivity, and the output tensor of the previous layer is used as the input tensor of the next layer. The workflow is as follows: The feature vector received by the input layer has a dimension of 106. This feature vector is composed of three parts: First, 64-dimensional frequency band energy and kurtosis features extracted by 4-layer wavelet packet decomposition, which reflects the energy distribution and impact characteristics of pressure pulsation signals and vibration acceleration signals in different frequency bands; Second, 6-dimensional friction parameters, namely σ0 (stiffness coefficient), obtained by online identification through the LuGre model. σ1 (damping coefficient), σ2 (viscous friction coefficient), F c (Coulomb friction), F s (Static friction) and v s (Stribeck velocity), reflecting the dynamic changes in the frictional characteristics of the actuator; thirdly, 36-dimensional time-domain statistical features, including the mean, variance, peak value, skewness, kurtosis, and peak factor of each actuator's displacement, velocity, and current signals; the first hidden layer contains 128 neurons, which perform a linear transformation h1=ReLU(W1x+b1) on the input 106-dimensional feature vector, where W1 is a 128×106 weight matrix, b1 is a 128-dimensional bias term, and the 106-dimensional input is transformed into a 128-dimensional feature vector through the ReLU activation function f(x)=max(0,x); this layer also sets a dropout rate of 0.3, that is, during training, a portion of the neuron outputs are randomly dropped with a 30% probability. To prevent overfitting and enhance the model's generalization ability, the second hidden layer contains 64 neurons. It takes the 128-dimensional feature vector output from the first hidden layer as input and performs a linear transformation h2 = ReLU(W2h1 + b2), where W2 is a 64×128 weight matrix and b2 is a 64-dimensional bias term. After ReLU activation, it outputs a 64-dimensional feature vector, further abstracting high-level semantic features related to faults. This layer also has a dropout rate of 0.3 to further suppress the risk of overfitting. The output layer is a Softmax classification layer, taking the 64-dimensional feature vector output from the second hidden layer as input and performing a linear transformation o = W3h2 + b3, where W3 is an 8×64 weight matrix and b3 is an 8-dimensional bias term. It then undergoes a Softmax activation function. The 8-dimensional output vector is mapped to a probability distribution [p1,p2,...,p8], and each component corresponds to the posterior probability of the input sample belonging to one of the 8 fault modes. Finally, the fault category corresponding to the maximum probability is taken as the classification result output. The 8 fault modes include: normal, internal leakage, external leakage, bearing wear, seal wear, sensor failure, lead screw wear, and motor demagnetization. Fault label correspondence: Significant increase in energy in frequency band 8-12 (approximately 1-2kHz) → Decrease in external leakage; Increased kurtosis in low frequency band (<100Hz) → Wear of seals; In addition, it also includes remaining life prediction, the specific process of which is as follows: Degradation index is defined as Alternatively, friction parameters and leakage coefficients can be used directly; the RUL prediction model uses an LSTM network, with the input being the historical sequence of degradation indicators (100 points, time resolution of 1 hour), and the output being the remaining service life (hours). The RUL prediction model uses an LSTM network to perform time-series modeling and life prediction of equipment degradation trends; the input of the LSTM network is the historical sequence of degradation indicators of each key component X=[D(t T+1),D(t [T+2),...,D(t)], where T=100 is the time window length, the time resolution is 1 hour, that is, the input contains degradation index data of the past 100 hours, used to predict the remaining service life (hours) after the current moment; for each actuator, the degradation index D(t) is defined as D(t=1 DI i (t) or directly use physical degradation parameters such as friction parameter σ2(t) and screw transmission efficiency η(t), which monotonically increase with the equipment running time; The LSTM network structure consists of a first LSTM layer, a second LSTM layer, and a fully connected layer, with each layer connected in series. The output tensor of the previous layer serves as the input tensor of the next layer. The first LSTM layer contains 64 LSTM units, receiving a sequence of degradation indices X as input. Its input tensor has a dimension of 100×1 (100 time steps, each step corresponding to one degradation indice feature). At each time step t, the LSTM units pass through a forget gate. Determine the memory unit C of the previous moment t The forgetting rate of 1, through the input gate and candidate memories Determine the amount of new information to be written at the current moment, through the output gate. Generate the hidden state h at the current moment. t =o t ⊙tanh(C tThe first LSTM layer enables long-range memory and retrieval of temporal dependencies. This layer iteratively processes 100 time steps along the time axis, transforming the output dimension into a 100×64 hidden state sequence. The second LSTM layer also contains 64 LSTM units, using the 100×64 hidden state sequence output from the first LSTM layer as input. It employs the same gated recurrent mechanism for secondary temporal feature extraction, further encoding the 64-dimensional input from the 100 time steps, maintaining the output dimension at 100×64. Only the 64-dimensional hidden state hT from the last time step (encoding all historical degradation information) is input to the subsequent fully connected layer. The fully connected layer contains one neuron, receiving the 64-dimensional feature vector hT∈R64 output from the last moment of the second LSTM layer. It performs a linear transformation on the weight matrix to 1×64, where b is a scalar bias, compressing the 64-dimensional features into a 1-dimensional scalar output, representing the predicted remaining lifespan (in hours) at the current moment.

[0024] LSTM network training uses a loss function that includes a monotonicity constraint penalty term: ,in, The mean squared error term ensures that the deviation between the predicted value and the true value is minimized. As a monotonicity penalty term, the predicted remaining lifetime curve should exhibit a monotonically decreasing trend over time (i.e., the equipment will not "self-heal" with increasing operating time), in accordance with the physical laws of equipment degradation; λ is a balance coefficient used to adjust the relative weights of the two losses; to quantify the uncertainty of the prediction results, the network uses the MC Dropout method to perform 10 random forward propagations during the inference phase (i.e., in each forward propagation, each LSTM layer still randomly discards some neurons at a dropout rate of 0.2 during the inference phase), obtaining 10 prediction values. Calculate its mean and standard deviation Finally, μ is used as the expected predicted value of RUL, and μ±1.96σ is used as the 95% confidence interval output, providing a probabilistic lifetime estimation basis for maintenance decisions; In addition, it also includes multi-source feature fusion diagnosis, the specific implementation process of which is as follows: Pressure pulsation signal and vibration acceleration signal Perform 4-level wavelet packet decomposition (db4 wavelet) to obtain 16 frequency band nodes, and extract the energy of each node. and kurtosis Feature vector dimension: 16 frequency bands × 2 features × 2 signals = 64; LuGre friction model parameter identification was performed using a low-speed triangular wave velocity excitation (±5 mm / s, 0.1 Hz) and offline identification using the Levenberg-Marquardt method. During online operation, parameters were updated every 10 minutes using recursive least squares (RLS). The LuGre friction model formula is as follows: ; ; ; in: The average deformation of the bristles; This is the stiffness coefficient; The damping coefficient; It is the coefficient of viscous friction; Coulomb friction; It is static friction; Stribeck speed; A classifier is used to output fault modes. The classifier structure is as follows: input layer, with a dimension of 106 (64-dimensional wavelet packet frequency domain features + 6-dimensional LuGre friction parameters + 36-dimensional time domain statistical features); first hidden layer, containing 128 neurons, with ReLU activation function and dropout rate of 0.3; second hidden layer, containing 64 neurons, with ReLU activation function and dropout rate of 0.3; output layer, with Softmax activation function, outputs 8 fault modes: normal, internal leakage, external leakage, bearing wear, seal wear, sensor failure, leadscrew wear, and motor demagnetization. Fault label correspondence: Significant increase in energy in frequency band 8-12 (approximately 1-2kHz) → Decrease in external leakage; Increased kurtosis in low frequency band (<100Hz) → Wear of seals; In addition, it also includes remaining life prediction, the specific process of which is as follows: Degradation index is defined as Alternatively, friction parameters and leakage coefficients can be used directly. The RUL prediction model employs an LSTM network, with the input being a historical sequence of degradation indicators (100 points, time resolution of 1 hour), and the output being the remaining lifetime (hours). The LSTM network structure is as follows: a first LSTM layer containing 64 LSTM units; a second LSTM layer containing 64 LSTM units; and a fully connected layer containing 1 neuron. The loss function is the mean squared error plus a monotonicity constraint penalty term. The MCDropout method is used for 10 random forward propagations, with the confidence interval taken as mean ± 1.96σ (95% confidence level). Decision optimization layer, see Figure 3configured to: set a dynamic early warning threshold based on the characteristics of the current flight training subject; when the physical performance degradation index triggers the dynamic early warning threshold, perform boundary optimization by combining with a multi-dimensional safety constraint set, and solve to generate the maximum available degraded operation envelope under the current health state; In this embodiment, the specific process of setting the dynamic early warning threshold based on the characteristics of the current flight training subject is: acquire the identifier of the current flight training subject from the instructor station in real time, and obtain the corresponding amplitude factor by looking up the pre-stored subject parameter table , frequency factor and load intensity , and the subject parameter table is shown in Table 3 below; Table 3

[0025] calculate the dynamic threshold based on the weighted combination of the basic threshold and the amplitude factor, frequency factor and load intensity; the calculation formula of the dynamic threshold is: ; wherein, mm is the basic threshold; is a weight coefficient; the dynamic threshold is limited between a preset minimum threshold and a preset maximum threshold (preferably ), and the dynamic threshold is subjected to smoothing filtering treatment (time constant is 0.5 seconds), and the threshold linearly transitions within 0.5 seconds when the subject is switched, so as to obtain the dynamic early warning threshold; Risk level calculation: health degree , subject intensity , risk value ; Grade division: Risk<0.3: low risk, normal use; 0.3<Risk<0.7: medium risk, prompt the instructor to pay attention and recommend avoiding high-intensity subjects; Risk>0.7: high risk, the subject is prohibited and an alarm is triggered; when the physical performance degradation index triggers the dynamic early warning threshold, performing boundary optimization by combining with a multi-dimensional safety constraint set, and solving to generate the maximum available degraded operation envelope under the current health state is specifically: comparing the physical performance degradation index with the dynamic early warning threshold, and triggering degraded envelope optimization when the physical performance degradation index of any actuator is lower than the dynamic early warning threshold; during the 0.5-second linear transition period of subject switching, the dynamic threshold monotonically transitions from an old value to a new value; if the new threshold is lower than the current value during the transition period, the transition curve keeps monotonically decreasing; if the new threshold is higher than the current value, the transition curve keeps monotonically increasing, which ensures that the threshold changes monotonically throughout the transition and avoids false triggering caused by threshold jump; An optimization model is constructed with the objective function of maintaining the maximum available envelope volume at the platform's end execution point. A multi-dimensional safety constraint set is superimposed, which is constrained by the positive contraction of the physical performance degradation exponent. This multi-dimensional safety constraint set includes at least travel constraints, velocity constraints, force constraints, and current constraints, wherein the upper limits of the velocity constraints, force constraints, and current constraints contract proportionally with the decrease of the physical performance degradation exponent. It is a dimensionless linear scaling factor, with a value range of [0,1]; when When =1, all constraints return to their rated maximum values; when When the value is 0.5, the upper limit of each constraint is reduced to 50% of the nominal value; when multiple constraints shrink in parallel, the constraint that first reaches the boundary value is used as the dominant constraint condition of the current envelope, and the remaining constraints are automatically satisfied; the optimization model is solved in the feasible region of the multidimensional safety constraint set to obtain the optimal motion envelope parameters under the current healthy state and generate the optimal envelope boundary accordingly; the optimal envelope boundary constitutes the maximum available degraded operating envelope; The parameter optimization solution process of the optimization model is as follows: The objective function for maintaining the maximum available envelope volume is specifically expressed as maximizing the product of the travel distances of the platform's three core degrees of freedom: pitch, roll, and heel. ; in, The envelope volume is (pitch angle range × roll angle range × lift range). Let i be the amount of violation of the i-th constraint. The penalty weights related to health (the lower the health, the greater the weight) are as follows: Penalty weight w i (DI i The design principle is as follows: when the health of an actuator decreases, the cost of violating its corresponding constraints should increase significantly, thereby actively limiting the working range of the actuator during optimization; the specific determination method is as follows: ① Basic weights: Define the basic penalty weight w for the i-th actuator. i The base is determined based on the load percentage of the actuator in the current kinematic configuration. ; in, The average output amplitude of the i-th actuator under standard operating conditions (obtained through historical data statistics or static balance calculation) is used to ensure that actuators with larger loads have higher constraint priority in optimization. ② Health moderating factors Define the health penalty amplification factor k w (DI iUsing a piecewise or exponential function, the weights increase non-linearly as health decreases: Method: Exponential amplification factor ; Where λ is the magnification intensity coefficient (ranging from 2.0 to 5.0, and in this embodiment, λ = 3.0). When DI i When k = 1.0 w =1.0; when DI i When k = 0.5 w =e 1.5 ≈4.48; when DI i When k = 0.2 w =e 2.4 ≈11.0; The upper limits of the allowed velocity constraints and force constraints in the optimization model are respectively narrowed down to the product of the system's rated maximum speed and the current physical performance degradation index of each actuator, and the product of the rated maximum force and the current physical performance degradation index of each actuator; that is, the multidimensional safety constraint set is shown in Table 4 below: Table 4

[0026] The upper limit of the velocity constraint is narrowed to the rated maximum velocity of the system. Current physical performance degradation index of each actuator The product of the forces and the upper limit of the force constraint are contracted to the rated maximum force of the system. Current physical performance degradation index of each actuator The product; The Sequential Quadratic Programming (SQP) algorithm is used to iteratively search within the feasible region of the multidimensional safety constraint set, outputting the optimal motion envelope parameters under the current decay state. Specifically, the calculation cycle does not exceed 100ms, the iteration limit is 50 times, and a warm start is performed using the previous optimal solution as the initial value to obtain the optimal motion envelope parameters under the current healthy state and generate the optimal envelope boundary, forming the maximum available degraded running envelope. When there is no feasible solution to the optimization problem, a step-by-step degradation strategy is executed: the first level relaxes the travel constraint by a preset percentage (preferably 10%), the second level reduces the target envelope volume by a preset percentage (preferably 20%), and the third level triggers an emergency stop command. The human-machine interface layer is configured to: convert the maximum available degraded operating envelope into a limiting control command and send it to the underlying motion controller; and render a 3D visualization interface of the device and generate a maintenance work order based on the physical performance degradation index. (See the 3D visualization interface for details.) Figure 4 ; The human-computer interaction layer receives the degraded operating envelope parameter X from the decision optimization layer. optIt is converted into amplitude limiting control commands that can be executed by the motion controller and sent out through the real-time communication bus to realize dynamic adaptive constraints on the platform's operating boundaries; Degraded operating envelope X opt The constraints include multi-dimensional parameters such as platform pose, motion velocity, and motion acceleration. The specific parameter definitions are shown in Table 11 below. Table 11

[0027] The converted limiting control command format is a structured limiting table, which is sent to the motion control computer via shared memory or reflected memory (delay <50μs); When generating motion commands (the motion commands output after the shuffling algorithm, i.e., platform pose commands), the motion control computer performs triple constraint processing on commands that exceed the limit table range, including saturation limiting, second-order low-pass filtering smoothing, and rate limiting. Specifically: ① Saturation limit: ; Where, x cmd To extract the motion commands generated by the algorithm, x min and x max To degrade the lower and upper limits of the operating package; ② Second-order low-pass filtering for smooth transition: ; Wherein, the natural frequency ωn = 2π × 10 rad / s (cutoff frequency 10Hz), and the damping ratio ζ = 0.707; the filtered output To avoid high-frequency mutations introduced by amplitude limiting; ③ Rate limiter: ; Prevent the rate of change of instructions from exceeding physical capabilities; The communication fault tolerance protection logic is configured so that when the communication interruption between the human-machine interface layer and the motion control computer exceeds 10ms, the motion controller automatically reverts to the fixed threshold protection mode and uses the factory-preset safety envelope (x). fallback ), to ensure system security; In this embodiment, the human-computer interaction layer extracts the extreme value boundary array of the maximum usable degraded operating envelope, configures a limiting filter component in the pose washing algorithm signal transmission channel input to the bottom motion control computer, writes the extreme value boundary array as the saturation cutoff parameter of the limiting filter component into the instruction and sends it down for execution, and the degraded envelope transitions with limiting surface sliding. Unlike existing technologies that use fixed envelope limiting and only allow selection of either "sending the command to the motion control computer, which then saturates and limits commands exceeding the envelope and uses a second-order low-pass filter to continue full-range operation" or "complete shutdown," the degraded envelope of this invention has the following substantial improvements: the constraint boundary is dynamically adjusted according to the health status, for example, velocity constraints. ,when When the speed is reduced from 1 to 0.5, the maximum speed is allowed to decrease by 50%; multi-constraint joint optimization considers six types of constraints, including stroke, speed, force, pressure difference, current and temperature rise, and solves the maximum usable envelope through sequential quadratic programming; a step-by-step degradation strategy is adopted, with three levels of degradation (the first level relaxes the stroke constraint by 10%, the second level reduces the target envelope volume by 20%, and the third level triggers an emergency stop command) to avoid the harsh treatment of direct shutdown; The human-computer interaction layer builds a digital twin 3D visualization system for the equipment, which renders the equipment's operating and health status in real time based on the physical performance degradation index of each actuator. See the visualization interface for details. Figure 4 ; The visualization system uses Unity 3D or Unreal Engine as the 3D rendering engine, running on the GPU (NVIDIA RTX 4000) of the central health management server, with a rendering frame rate of ≥60fps, and the model update cycle is synchronized with the health management loop (100Hz); the core components of the rendering engine are shown in Table 21 below: Table 21

[0028] The visualization system defines a color mapping function C(DI). i ), and the performance degradation index DI i ∈[0,1] is mapped to RGB color values: C(DI) i )=HSL(h(DI i ), 0.85, 0.50); ​​The adaptive warm and cool color spectrum is used to refresh the model material colors in real time (green indicates health, yellow indicates attention, and red indicates alarm), so as to realize the intuitive visualization of the health status of each actuator; Based on the constructed digital twin 3D model, the system supports human-computer interaction and refined status query: clicking on any actuator model will display the actuator's real-time operating parameters, remaining service life prediction results, failure risk analysis and corresponding maintenance suggestions in a pop-up window; it also supports functions such as backtracking analysis of historical operating data of equipment and fault simulation injection testing, realizing visualized management and control of the equipment's status throughout its entire life cycle. The human-machine interaction layer integrates multiple indicators such as the degree of actuator performance degradation, remaining service life, and fault identification confidence, establishes a hierarchical alarm mechanism, and automatically matches fault modes based on the fault tree database to generate standardized maintenance work orders, thereby realizing proactive operation and maintenance management of equipment faults and degradation status. The specific alarm judgment rules are as follows: When the physical performance degradation index of any actuator reaches the first threshold, an "Attention" level alarm is triggered; when the physical performance degradation index of any actuator reaches the second threshold, a "Warning" level alarm is triggered; when the equipment predicts that the remaining service life is less than 100 flight hours, an "Information" level alarm is triggered; when the fault classifier outputs a confidence level greater than 0.8 and determines that a fault exists, a "Warning" level alarm is triggered. The visualization system inputs the physical performance degradation index and fault characteristic parameters that trigger alarms into a preset fault tree database for matching and retrieval, and automatically generates standardized maintenance work orders. The work order includes complete fields such as equipment ID, fault description, fault mode, severity level, suggested maintenance time window, required spare parts model and quantity, standardized maintenance steps, estimated maintenance time, and safety precautions.

[0029] The generated maintenance work orders are connected to the enterprise asset management system (CMMS) and spare parts management system via REST API in JSON data format and HTTPS protocol, so as to realize the automatic issuance, circulation and archiving of maintenance tasks and intelligent linkage management of spare parts inventory, and complete the closed loop of the entire process of equipment health management and operation and maintenance business. Thus, the data perception layer synchronously acquires data from multiple types of sensors via the EtherCAT bus, and obtains operational data after denoising, anomaly removal, missing data completion, and coordinate unification transformation. Based on this operational data, the digital twin model layer calculates the theoretical driving force of each actuator using the Newton-Euler dynamics equations and the LuGre friction model. A two-layer LSTM compensation network generates a corrected compensation force, which is then superimposed after being subject to triple physical safety constraints of amplitude limiting, energy, and power to form a hybrid driving force. This force is substituted into the forward dynamics equations to deduce the theoretical desired pose and generate spatial pose residuals. Adaptive Kalman filtering is then used for online calibration based on frequency scanning fingerprints. The state assessment and diagnosis layer uses the inverse Jacobian matrix to reduce the dimensionality of each actuator... The independent length residuals of the device are accumulated and normalized, and then combined with the decoupled correction of the operating rate and load force to calculate the physical performance degradation index. At the same time, wavelet packet decomposition is used to extract frequency band energy and kurtosis features, and LuGre friction parameter identification is fused into the input deep neural network to realize multi-type fault classification. The remaining service life is predicted by the degradation index time series. The decision optimization layer dynamically sets the warning threshold according to the flight training subjects. With the goal of maximizing the envelope volume, multi-dimensional safety constraints of velocity, force and current constraints shrink linearly with the degradation index are superimposed to solve the degradation envelope and execute a three-level progressive degradation strategy. The human-machine interaction layer converts the degradation envelope into a limiting command and issues it for execution. The three-dimensional visualization renders the health status and generates a standardized maintenance work order.

[0030] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process and related explanations of the methods described above can be found in the corresponding processes in the foregoing system embodiments, and will not be repeated here.

[0031] A second embodiment of the present invention proposes a health management method for a flight simulator motion platform based on digital twins. Based on the aforementioned health management system for a flight simulator motion platform based on digital twins, the method includes the following steps: S1 collects real-time multi-source sensor data from the six-degree-of-freedom motion platform of the full-motion flight simulator, and obtains operational data after preprocessing; S2, based on the running data, the theoretical driving force and the correction compensation force are mixed and calculated and forward evolution is performed to calculate the spatial pose residual between the actual feedback pose and the theoretical expected pose. S3, perform dimensionality reduction mapping on the spatial pose residual, and then obtain the independent length unidirectional residual and residual accumulation energy of each actuator; and combine the real-time operating condition characteristics of the actuator to decouple the residual accumulation energy to obtain the physical performance degradation index of each actuator. S4. Set a dynamic early warning threshold based on the characteristics of the current flight training subject; when the physical performance degradation index triggers the dynamic early warning threshold, perform boundary optimization by combining the multi-dimensional safety constraint set, and solve to generate the maximum available degraded operation envelope under the current health state. S5, convert the maximum available degraded operating envelope into a limiting control command and send it to the underlying motion controller; and render the device's three-dimensional visualization interface and generate a maintenance work order based on the physical performance degradation index.

[0032] It should be noted that the health management method for a flight simulator motion platform based on digital twins provided in the above embodiments is only an example of the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the modules or steps in the embodiments of the present invention can be further decomposed or combined. For example, the modules in the above embodiments can be merged into one module, or further divided into multiple sub-modules to complete all or part of the functions described above. The names of the modules and steps involved in the embodiments of the present invention are only for distinguishing the various modules or steps and are not considered as an improper limitation of the present invention.

[0033] Those skilled in the art will recognize that the modules and method steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. The programs corresponding to the software modules and method steps can be placed in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disks, removable disks, CD-ROMs, or any other form of storage medium known in the art. To clearly illustrate the interchangeability of electronic hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in electronic hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the invention.

[0034] The terms “first”, “second”, etc., are used to distinguish similar objects, not to describe or indicate a specific order or sequence.

[0035] The term "comprising" or any other similar term is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus / device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent in such process, method, article, or apparatus / device.

[0036] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A health management system for a flight simulator motion platform based on digital twins, characterized in that, The system includes: The data perception layer is configured to: collect real-time multi-source sensor data from the six-degree-of-freedom motion platform of the full-motion flight simulator, and obtain operational data through preprocessing; The digital twin model layer is configured to: based on the running data, calculate the theoretical driving force and the correction compensation force by mixing and performing forward evolution, and measure the spatial pose residual between the actual feedback pose and the theoretical expected pose; The state assessment and diagnosis layer is configured to: perform dimensionality reduction mapping on the spatial pose residuals, thereby obtaining the independent length unidirectional residuals and residual accumulation energy of each actuator; and decouple the residual accumulation energy by combining the real-time operating condition characteristics of the actuators to obtain the physical performance degradation index of each actuator. The decision optimization layer is configured as follows: a dynamic early warning threshold is set based on the characteristics of the current flight training subject; when the physical performance degradation index triggers the dynamic early warning threshold, boundary optimization is performed in combination with the multi-dimensional safety constraint set to solve and generate the maximum available degraded operation envelope under the current health state. The human-computer interaction layer is configured to: convert the maximum available degraded operating envelope into a limiting control command and send it to the underlying motion controller; and render the device's 3D visualization interface and generate maintenance work orders based on the physical performance degradation index.

2. The health management system for a flight simulator motion platform based on digital twins according to claim 1, characterized in that, The spatial pose residuals are dimensionality-reduced and mapped to obtain the independent length unidirectional residuals of each actuator and the residual accumulation energy, specifically as follows: Multiply the spatial pose residual containing six degrees of freedom components by the inverse Jacobian matrix to obtain independent length unidirectional residuals corresponding to the absolute deviations of the extension strokes of the six actuators; perform discrete square integral summation on the independent length unidirectional residuals of each actuator within a preset time sliding window, and normalize them according to the maximum stroke of each actuator to obtain the residual accumulated energy of each actuator.

3. The health management system for a flight simulator motion platform based on digital twins according to claim 1, characterized in that, Furthermore, by combining the real-time operating characteristics of the actuators, the residual accumulated energy is decoupled and calculated to obtain the physical performance degradation index of each actuator, specifically: The real-time operating condition characteristics of the actuator are obtained, including the operating speed and push-pull load factors. Take the absolute value of the operating speed of each actuator and the absolute value of the push-pull load factor, multiply them by the offline calibrated speed correction coefficient and load force correction coefficient respectively, and then add the basic constant to construct the decoupling correction coefficient; The physical performance degradation index of the actuator is obtained by multiplying the ratio of the accumulated residual energy to the baseline residual energy in the healthy state by the decoupling correction coefficient and subtracting the product from 1.

4. The health management system for a flight simulator motion platform based on digital twins according to claim 1, characterized in that, The spatial pose residual between the actual feedback pose and the theoretical expected pose was calculated, specifically as follows: The operational data is input into the physical mechanism model and the data-driven compensation network, respectively. The physical mechanism model calculates the theoretical driving force of each actuator based on the current platform motion state, and the data-driven compensation network generates the corrected compensation force of each actuator based on the current and historical time-series sensing characteristics. The corrected compensation force is nonlinearly constrained to obtain the constrained corrected compensation force; the nonlinear constraint is specifically a multi-dimensional physical safety constraint, including amplitude constraint, energy constraint and power constraint. The constrained corrective compensation force is superimposed component by component onto the theoretical driving force to form a hybrid driving force; The hybrid driving force is substituted into the platform's forward dynamic differential equation for numerical integration to obtain the theoretical desired pose of the platform under the action of the hybrid driving force. The difference between the theoretical desired pose and the actual feedback pose is used to generate a spatial pose residual containing six degrees of freedom.

5. The health management system for a flight simulator motion platform based on digital twins according to claim 4, characterized in that, The amplitude limiting constraint specifically refers to limiting the amplitude of the correction compensation force to no more than a preset percentage of the theoretical driving force amplitude. The energy constraint specifically involves: performing discrete integration of the square of the correction compensation force amplitude over time; when the accumulated integral value exceeds the maximum compensation energy limit set by the system, the correction compensation force is subjected to amplitude attenuation processing. The power constraint specifically refers to: real-time estimation of the additional power expected to be generated by each actuator due to the superimposed correction compensation force, ensuring that the expected additional power does not exceed a preset percentage threshold of the motor's rated power.

6. The health management system for a flight simulator motion platform based on digital twins according to claim 1, characterized in that, The process of setting dynamic early warning thresholds based on the characteristics of the current flight training subjects is as follows: The current flight training subject identifier is obtained in real time from the instructor's station, and the corresponding amplitude factor, frequency factor and load intensity are retrieved through the pre-stored subject parameter table; The dynamic threshold is calculated based on a weighted combination of the basic threshold and the amplitude factor, frequency factor, and load intensity. The dynamic threshold is limited to a preset minimum threshold and a maximum threshold, and the dynamic threshold is smoothed and filtered to obtain a dynamic warning threshold.

7. The health management system for a flight simulator motion platform based on digital twins according to claim 1, characterized in that, When the physical performance degradation index triggers the dynamic early warning threshold, boundary optimization is performed using a multi-dimensional safety constraint set to solve for and generate the maximum available degraded operating envelope under the current healthy state. Specifically: The physical performance degradation index is compared with the dynamic warning threshold. When the physical performance degradation index of any actuator is lower than the dynamic warning threshold, a degradation envelope optimization is triggered. An optimization model is constructed with the objective function of maintaining the maximum available envelope volume at the platform's end execution point. A multidimensional safety constraint set is superimposed, which is subject to a positive contraction in the physical performance degradation exponent. The multidimensional safety constraint set includes at least travel constraints, velocity constraints, force constraints, and current constraints, wherein the upper limits of the velocity constraints, force constraints, and current constraints contract proportionally as the physical performance degradation exponent decreases. The optimization model is solved within the feasible region of the multidimensional safety constraint set to obtain the optimal motion envelope parameters under the current healthy state and generate the optimal envelope boundary accordingly. The optimal envelope boundary constitutes the maximum available degraded operating envelope.

8. The health management system for a flight simulator motion platform based on digital twins according to claim 7, characterized in that, The parameter optimization solution process of the optimization model is as follows: The objective function for maintaining the maximum available envelope volume is specifically expressed as maximizing the product of the travel distances of the platform's three core degrees of freedom: pitch, roll, and heel. The upper limits of the allowed velocity constraints and the upper limits of the allowed force constraints in the optimization model are respectively narrowed and limited to the product of the rated maximum speed of the system and the current physical performance degradation index of each actuator, and the product of the rated maximum force and the current physical performance degradation index of each actuator; a sequential quadratic programming algorithm is used to iteratively search within the feasible region of the multidimensional safety constraint set to output the optimal motion envelope parameters under the current degradation state.

9. The health management system for a flight simulator motion platform based on digital twins according to claim 1, characterized in that, The digital twin model layer also includes an online calibration unit. The online calibration unit uses key physical parameters in the mechanism model as state variables, and uses an adaptive Kalman filter to dynamically update the state variables online. It also uses an adaptive algorithm to update the process noise covariance and observation noise covariance in real time. Combined with the frequency sweep fingerprint benchmark established during factory or overhaul, when the deviation between the real-time frequency sweep response and the frequency sweep fingerprint benchmark exceeds a preset threshold, parameter recalibration or fault alarm is triggered.

10. A health management method for a flight simulator motion platform based on digital twins, based on the health management system for a flight simulator motion platform based on digital twins as described in any one of claims 1-9, characterized in that, The method includes the following steps: S1 collects real-time multi-source sensor data from the six-degree-of-freedom motion platform of the full-motion flight simulator, and obtains operational data after preprocessing; S2, based on the running data, the theoretical driving force and the correction compensation force are mixed and calculated and forward evolution is performed to calculate the spatial pose residual between the actual feedback pose and the theoretical expected pose. S3, perform dimensionality reduction mapping on the spatial pose residual, and then obtain the independent length unidirectional residual and residual accumulation energy of each actuator; and combine the real-time operating condition characteristics of the actuator to decouple the residual accumulation energy to obtain the physical performance degradation index of each actuator. S4. Set a dynamic early warning threshold based on the characteristics of the current flight training subject; when the physical performance degradation index triggers the dynamic early warning threshold, perform boundary optimization by combining the multi-dimensional safety constraint set, and solve to generate the maximum available degraded operation envelope under the current health state. S5, convert the maximum available degraded operating envelope into a limiting control command and send it to the underlying motion controller; and render the device's three-dimensional visualization interface and generate a maintenance work order based on the physical performance degradation index.

Images