A working condition parameter dynamic loading method and system suitable for a digital twin model
By combining LSTM prediction models and event-driven triggering logic, the resource efficiency of data loading and the reliability of fault diagnosis in digital twin models are solved, achieving efficient and accurate data synchronization and fault capture.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING UNIV OF CHEM TECH
- Filing Date
- 2026-02-10
- Publication Date
- 2026-06-12
AI Technical Summary
In existing technologies, the data loading methods for digital twin models cannot effectively distinguish between redundant data under stable operating conditions and high-value information under dynamic changes. This leads to excessive burden on communication bandwidth and processing capacity or missing key fault characteristic signals, resulting in a disconnect between the model and the physical entity's state, which affects the reliability of fault diagnosis.
An intelligent data synchronization decision mechanism based on an LSTM prediction model is adopted. By defining the trigger logic for data loading, data synchronization is only performed when the physical entity state deviates from the expectation. Combined with event-driven trigger functions, efficient and accurate data loading is achieved.
This approach improves sensitivity to critical information and fault detection capabilities while reducing data transmission volume, ensuring timely response of the model during rapid changes in critical operating conditions and enhancing the reliability and predictability of fault diagnosis.
Smart Images
Figure CN122195524A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of digital twin technology, specifically to a method and system for intelligent data synchronization, state perception and high-confidence parameter fusion between large and complex physical mechanical equipment and its high-fidelity virtual model. It is particularly suitable for the efficient and intelligent loading of operating environment parameters (such as vibration, stress, temperature, pressure, speed, voltage, current, etc.) of key mechanical equipment such as aero-engines, marine power systems, and gas turbines. Background Technology
[0002] The core value of digital twin technology lies in its fidelity—the extent to which a virtual model can accurately and in real-time mirror the state, behavior, and attributes of its physical counterpart. For large mechanical equipment such as aircraft and ships, high-fidelity digital twins are fundamental for predictive maintenance, performance optimization, and remaining service life assessment of these complex machines. However, this high fidelity heavily relies on continuous and accurate data streams collected by numerous sensors deployed in key parts of the equipment, such as vibration, eddy current, temperature, voltage, and current sensors.
[0003] In existing technologies, methods for loading massive amounts of sensor data into digital twin models generally employ a periodic synchronization (fixed-frequency polling) strategy. This strategy has an inherent and irreconcilable flaw: to capture critical transient conditions (such as engine surge precursors or early bearing damage), an extremely high update frequency is required, which generates enormous data traffic, placing a heavy burden on the bandwidth of airborne / shipborne communication links and the processing power of edge computing units. Conversely, if the update frequency is reduced to save resources, critical fault characteristic signals are likely to be missed, causing the twin model to become disconnected from the physical entity's state, rendering it unreliable and even misleading in fault diagnosis and safety decision-making. This method is essentially "blind," unable to distinguish between redundant data under stable operating conditions and high-value information under dynamic changes. Summary of the Invention
[0004] This invention aims to solve the problem of how to achieve continuous and accurate performance prediction and fault diagnosis using discontinuously loaded parameters in practical engineering applications of digital twin models. The technical approach of this invention lies not only in providing an intelligent data synchronization decision-making mechanism based on an LSTM prediction model, but more importantly, in defining fusion parameters. The coupling mapping relationship between the digital twin physics simulation engine and the physical simulation engine. That is: It's not just a numerical result; it's the real-time boundary condition driving the evolution of the physical partial differential equations in the digital twin model. The model... Driven by this, it can update its internal "damage accumulation status" or "energy dissipation index" in real time, thereby achieving accurate extrapolation of future states without relying on the full data of physical equipment.
[0005] To address the aforementioned technical problems, this invention proposes a method for dynamically loading operating parameters suitable for digital twin models. The core of this invention lies in an intelligent triggering logic driven by a neural network, defined by the following mathematical model, which determines the timing of data loading:
[0006] Define a data loading trigger decision function T(t) at time t, which executes a data loading once if and only if T(t) = 1.
[0007]
[0008] in, This is the logical OR operator. The two sub-functions are defined as follows:
[0009] Predictive state depends on trigger function :
[0010] The core idea of this function is to synchronize data only when the actual state of a physical entity deviates from its "expected" normal state. This expected state is predicted in real time by a pre-trained recurrent neural network, specifically a Long Short-Term Memory (LSTM) network model.
[0011]
[0012] in:
[0013] The total number of monitored operating condition features;
[0014] For the first Individual operating condition characteristics at time Real-time measurement values;
[0015] It is the LSTM neural network at time 1000. For the The predicted value for each feature. This prediction is based on past... Historical feature sequence at each time step (This sequence specifically includes vibration amplitude, bearing fault characteristic frequency energy, major axis dimension of the shaft center trajectory, minor axis dimension of the shaft center trajectory, and ellipticity) and the current external operating condition vector. (This vector is generated by including flight altitude, flight speed, and operator-commanded thrust.)
[0016] The calculated value is called the residual, which quantifies the degree of deviation between the actual behavior and the expected behavior;
[0017] It is the first The dynamic threshold set for each feature is not fixed, but rather adaptively adjusted according to the current operating conditions and equipment health status.
[0018] Event-driven trigger function :
[0019]
[0020] in, It is the set of all events that occur at time t; This is a predefined set of critical events (which includes, and only includes, motor start commands, motor stop commands, ESC over-temperature alarm signals, and Level 1 fault codes broadcast by the flight control system). This function ensures that the model state can be forcibly and with high priority synchronized when critical operations or system alarms occur.
[0021] Based on the above model, the technical solution of the present invention specifically includes the following steps:
[0022] Step 1: Multi-source data acquisition and real-time preprocessing. Through the sensor interface module, raw time-series signals such as vibration, temperature, pressure, rotational speed, voltage, and current are acquired. The real-time preprocessing engine processes the signals, for example, performing a Fast Fourier Transform (FFT) on the vibration signal to extract the energy or amplitude of key frequency bands as operating condition features, and using moving average filtering to eliminate noise from slowly varying signals such as temperature and pressure.
[0023] Step 2: Adaptive Synchronization Triggering. The adaptive triggering module continuously monitors the preprocessed operating condition characteristic values. And according to the decision function defined above Determine whether data loading has been triggered.
[0024] Step 3: Data fusion based on consistency and variance awareness triggers a signal activation, initiating the core data fusion process. This does not directly transmit the raw data, but rather performs deep cleaning and fusion of multiple redundant sensor data for the same physical quantity. This step includes three key sub-processes: consistency matrix construction, support-based outlier sensor removal, and adaptive weighting based on inverse variance. This is the core element of this invention to ensure data quality.
[0025] Step 4: Loading Holistic Model Parameters: After merging and packaging all parameters that need updating, update the digital twin model holistically via the model loading interface. In mechanical engineering, this holistic approach is a "complete commit or complete undo" protection mechanism. It ensures that a set of related operating parameters (such as the vibration, temperature, and lubricating oil pressure of a bearing) are updated as an indivisible unit. This prevents issues such as communication interruptions from causing only partial parameter updates, thus importing a physically non-existent and misaligned model state, guaranteeing data consistency and reliability during simulation and analysis of the digital twin model.
[0026] Compared with the prior art, the present invention has the following significant advantages:
[0027] Extremely high resource efficiency and enhanced fault detection capabilities: When equipment is operating smoothly, the actual operating conditions closely match the predictions of the LSTM model, effectively suppressing state-dependent triggering and reducing data transmission volume by more than an order of magnitude. Simultaneously, due to the model's deep understanding of "normal" behavior, any minute abnormal deviations (i.e., the generation of residuals) that foreshadow early faults can be sensitively captured, achieving a reduction in total data volume while simultaneously enhancing sensitivity to critical information.
[0028] Exceptional model fidelity and predictability: Event-driven triggers ensure sub-second model response during critical operational transitions. More importantly, the predictive state-dependent triggering mechanism ensures data loading always occurs when the "information value" is highest—that is, the moment physical reality begins to deviate from the expected trajectory. This provides the most critical data input for the digital twin model, enabling it to be used more reliably for real-time fault diagnosis and performance prediction.
[0029] Enhanced system robustness and intelligence: The data fusion step effectively mitigates random noise and measurement errors from individual sensors, improving the data quality of the input model. Simultaneously, replacing the fixed heartbeat mechanism with LSTM-based dynamic triggering frees the system from the constraints of static rules, enabling it to autonomously learn and adjust its monitoring strategies based on individual equipment differences and degradation processes, achieving a higher level of intelligent operation. Attached Figure Description
[0030] Figure 1 This is a schematic diagram of the system architecture of the method described in this invention, illustrating the relationship between the physical equipment layer, the edge data processing layer, and the digital twin model layer.
[0031] Figure 2 This is a detailed flowchart of the dynamic loading method described in this invention, showing the complete steps from data acquisition to model loading;
[0032] Figure 3This is a schematic diagram of the state machine of the hybrid triggering logic described in this invention, which explains the internal working logic of the adaptive triggering module; Detailed Implementation
[0033] Example: An adaptive loading method for environmental parameters of a large rotor motor based on a digital twin model.
[0034] This embodiment uses a rotor motor applied to a drone as an example to illustrate in detail the specific implementation of the dynamic loading method and system for operating parameters described in this invention. The system is deployed on an airborne edge computing unit and aims to resolve the contradiction between limited bandwidth for high-frequency sampling data transmission and insufficient real-time performance in fault feature capture.
[0035] System Setup and Architecture – Handling Massive Amounts of Real-Time Data
[0036] Reference Figure 1 The system architecture shown in this embodiment strictly follows the four-step closed-loop logic of "acquisition and preprocessing - intelligent triggering and discrimination - multi-source data fusion - overall loading".
[0037] Detailed execution flow of the loading method:
[0038] This implementation method strictly follows Figure 2 The flowchart shown illustrates the dynamic loading of operating parameters. For long-duration hovering tasks, the process is as follows:
[0039] Step 1: Data Acquisition and Real-time Preprocessing
[0040] Multi-source heterogeneous data is acquired through the sensor interface module and standardized using a real-time preprocessing engine.
[0041] Multi-source sensing network deployment: Each rotor motor unit is equipped with a complete sensor suite to achieve comprehensive monitoring of the motor's electromagnetic, thermal, and dynamic characteristics. Specific configuration is as follows:
[0042] Triaxial accelerometer: An inertial sensor used to monitor high-frequency vibrations, it is securely mounted on the base of each motor unit and is key to accurately capturing the dynamic excitation force of the rotor system.
[0043] Thermistor: A resistive sensor used to measure core temperature. It is embedded in the slots of the motor stator winding. This is the core guarantee to prevent the motor from aging or burning out due to overheating.
[0044] Hall effect sensor: A magnetic field sensor used to monitor electromagnetic properties, integrated into the power supply phase line of each motor controller, which is the basis for evaluating motor operating efficiency and diagnosing electrical faults.
[0045] Eddy current displacement sensor: This is a pair of orthogonally arranged non-contact sensors used to accurately measure the axis trajectory of the motor rotor spindle within the bearing. This is a core indicator for judging the dynamic stability of the rotor, bearing wear, and shaft alignment.
[0046] Real-time preprocessing execution:
[0047] The onboard edge computing unit (ECU)'s real-time preprocessing engine continuously processes the raw data stream from the entire aircraft's sensor network and extracts key operational features:
[0048] Vibration signal: The vibration amplitude of each motor's rotational frequency (1X) is extracted through FFT transformation. and bearing fault characteristic frequency energy .
[0049] Eddy current signal: The original voltage signal is solved into the real-time displacement of the rotor in the XY plane, and further processed to obtain key parameters of the shaft center trajectory, such as the major and minor axis dimensions and ellipticity of the trajectory, denoted as the characteristic. .
[0050] Temperature signal: Electromagnetic interference noise is eliminated by using a moving average filter.
[0051] Step Two: Adaptive Synchronization Triggering Based on Hybrid Model – Demonstrating Data Cost Advantages
[0052] The adaptive triggering module determines when data is loaded. This module runs a set of intelligent triggering logic driven by a neural network, defined at time... Triggering decision function for:
[0053]
[0054] If and only if At that time, data loading is performed. The specific implementation logic of the two sub-functions is as follows:
[0055] Predictive state depends on trigger function
[0056] This function is used to determine whether the actual state of a physical entity deviates from its "expected" normal state.
[0057] LSTM Neural Network Prediction: A pre-trained Long Short-Term Memory (LSTM) network is deployed within the system. The model receives past... Historical feature sequence at each time step and the current external operating condition vector .
[0058] Historical feature sequence Specifically, this includes vibration amplitude, bearing fault characteristic frequency energy, major axis dimension of the shaft center trajectory, minor axis dimension of the shaft center trajectory, and ellipticity.
[0059] External operating condition vector Specifically, this includes the PWM speed command (i.e., operator-instructed thrust) issued by the flight controller, flight altitude, and flight speed.
[0060] Residual calculation: The neural network outputs a predicted value of the current operating condition characteristics. Calculate real-time measurements Residuals between the predicted and actual values:
[0061]
[0062] The residual quantifies the degree of deviation between the actual motor behavior and the theoretical healthy behavior.
[0063] Explanation of variable correspondence: In the formula This is a general representation, referring to the i-th type of feature. In this embodiment, the specific correspondence is as follows:
[0064] when hour, Corresponding to the motor frequency vibration amplitude extracted in step one;
[0065] when hour, Corresponding to the bearing fault characteristic frequency energy extracted in step one;
[0066] when hour, The key parameters of the axis trajectory (such as ellipticity) extracted in step one.
[0067] This refers to the corresponding predicted values calculated by the neural network for each of the aforementioned specific physical quantities.
[0068] Dynamic threshold Quantitative discrimination logic:
[0069]
[0070] in This is a dynamic threshold. In this embodiment, its specific determination method is as follows:
[0071] Set the window length for calculating the variance of the operation instructions to be [length]. (For example (each time step) calculates the current flight control thrust command. Variance within the window Set a baseline variance limit for determining a stationary state. (For example, the value is 0.05).
[0072] Stable cruise period ( ):
[0073] At this point, the equipment is in a stable operating state, and the system needs to detect minor early signs of failure (such as slight bearing wear vibration).
[0074] The threshold is set as follows:
[0075] in This represents the baseline noise standard deviation of the sensor under ideal conditions. This threshold is quite stringent; any minute abnormal fluctuation will trigger data loading.
[0076] Intense maneuvering period ( ):
[0077] At this point, the equipment is in a dynamic response state (such as rapid acceleration), and the prediction model may produce reasonable dynamic errors. To avoid false triggers, the threshold needs to be adaptively relaxed according to the intensity of the maneuver.
[0078] The threshold is set as follows:
[0079] in This is the adjustment coefficient (e.g., 0.5). In this case, the threshold increases linearly with the variance of the operating command, thus ensuring the capture of major faults while filtering out prediction residuals caused by violent maneuvers.
[0080] Event-driven triggering function
[0081] This function handles a set of critical events. .
[0082]
[0083] Predefined set of key events Includes and only includes the following signals:
[0084] Motor start command;
[0085] Motor stop command;
[0086] ESC over-temperature alarm signal;
[0087] The flight control system broadcasts a Level 1 fault code.
[0088] Once the above set Any signal in the event set at the current time appears In the process, the system will ignore residual judgment and force the highest priority to trigger data synchronization to ensure that the twin model can respond to critical situations in a timely manner.
[0089] Step 3: Data Fusion Based on Consistency and Variance Awareness
[0090] When the adaptive trigger module determines At that time, the data fusion core was activated.
[0091] Variable definition clarification: In this step, variables... and The meaning here is different from steps one and two. Each of these represents a sensor number used for redundant measurement of a specific physical quantity (e.g., stator temperature). This represents the total number of redundant sensors. and They represent sensors respectively. and sensors At any moment The reading.
[0092] Execute the following high-confidence fusion algorithm:
[0093] Constructing a consistency support matrix: Calculate the support of any two sensors ( and The compatibility relationship between them.
[0094] Confidence distance calculation: Calculate the statistical distance between the conditional probability density functions of two sensor observations. (i.e., confidence distance):
[0095]
[0096] in, Indicates at time sensor Observed values In the case of physical quantities The posterior probability density function. In this embodiment, the measurement noise follows a normal distribution. .
[0097] Support mapping: Mapping distance to support using a decay function. Construct the support matrix:
[0098]
[0099] in This is the consistency tolerance parameter.
[0100] Abnormal sensor elimination: Calculate the overall support for each sensor. :
[0101]
[0102] Calculate the anomaly detection threshold in This is a preset consistency coefficient (e.g., a value between 0.7 and 0.8). If... If the sensor is found to be abnormal (e.g., it has drifted or been damaged), it will be removed from the fusion calculation.
[0103] Weighted fusion generates comprehensive parameters: for the selected set of effective sensors Perform a weighted summation.
[0104] Normalized weight calculation: The sliding window algorithm is used to calculate the weights of each effective sensor in real time at time 1. Measurement variance .
[0105] Where L is the time series length of the sliding window used to calculate the variance (i.e., the number of sampling points within the window). The mean of the data within this window:
[0106]
[0107] Calculate weights based on the minimum variance criterion Note: To ensure consistency of signs, the variance within the summation term is specified in the formula. Same label , indicating that it is a time Real-time variance:
[0108]
[0109] Final fusion: Obtaining high-confidence comprehensive parameter values :
[0110]
[0111] Step 4: Loading the overall model
[0112] The vibration, temperature, and electrical parameters of the rotor motor, after being fused, are packaged and loaded into the digital twin model in a "holistic manner" through the model loading interface.
[0113] Specific implementation method:
[0114] The system employs a "data packet + double buffer mutual exclusion" mechanism to achieve overall loading:
[0115] Data Packaging: A structure object TwinUpdatePacket is allocated in memory. This structure contains all the fused physical quantity fields (such as vibration_vector, stator_temp, bus_current) at time t. The packet is marked as "ready" only after all fields have been calculated and written to the structure.
[0116] Atomic update: Utilizes a mutex lock mechanism to lock the input buffer of the digital twin model. During the locking period, the contents of TwinUpdatePacket are copied to the input buffer once and uninterrupted using memcpy or a similar command, and then the lock is released immediately.
[0117] Double buffering mechanism: The digital twin model reads from the "front buffer" for simulation rendering, while the data loading system writes to the "back buffer". Once the back buffer has been written and verified, the front and back buffers are switched instantly through an atomic pointer swap operation.
[0118] Step 5: Based on high confidence parameters Closed-loop decision-making applications.
[0119] The system has entered the "closed-loop decision-making application stage." Taking aero-engine bearing monitoring as an example, This represents the high-confidence vibrational energy value after fusion. The digital twin model receives... Then, perform the following operations:
[0120] Physics-driven: The dynamic excitation source, mapped to a rotating mechanical structure, drives the finite element simulation engine within the model to calculate the stress distribution at key nodes. High-confidence parameter inputs ensure that the virtual strain field generated by the simulation engine closely matches the actual state of the physical entity.
[0121] Condition assessment: Using real-time stress data output by the simulation engine and combined with material fatigue characteristic curves, the residual fatigue life under the current working condition is calculated. Through calculation... The statistical distance (e.g., Euclidean distance) between the current simulation result and the preset historical health baseline sequence. ), and compare it with a preset security threshold. Perform real-time comparison.
[0122] Decision Output: Based on the magnitude by which the Euclidean distance exceeds a threshold, the system automatically executes closed-loop feedback logic. Specifically, this manifests as: real-time pop-up of tiered early warning information on the control interface (such as "Inspection Recommendation," "Reduced Load Operation," or "Immediate Shutdown" commands), and simultaneous automatic generation of an electronic maintenance work order. This work order includes the current fault characteristics. The model predicts the performance evolution trend curve and provides preliminary diagnostic conclusions based on simulation analysis.
[0123] Through the above mechanism, it is ensured that the multi-physics parameters (such as the temperature rise caused by a sudden increase in current) are strictly and synchronously updated at the same time, which fundamentally eliminates the model state misalignment caused by inconsistent arrival times of different parameters due to network latency or multi-thread competition (e.g., the current data has been updated, but the temperature data is still stuck in the previous moment), thereby significantly improving the accuracy and reliability of digital twins in fault diagnosis and predictive maintenance.
[0124] This embodiment clearly demonstrates the invention's superior processing capabilities when faced with massive amounts of real-time data. Instead of passively processing all data, it uses intelligent prediction to focus monitoring on unexpected anomalies, thereby achieving precise detection of early, subtle faults in large, complex equipment at extremely low data costs.
Claims
1. A method for dynamically loading operating parameters suitable for digital twin models, characterized in that, Includes the following steps: The sensor interface module collects multi-source heterogeneous raw operating condition data of the equipment, and the raw data is processed by a real-time preprocessing engine to extract operating condition features. The adaptive triggering module uses a hybrid model that combines predictive state-dependent triggering and event-driven triggering to determine the data loading method. When the adaptive triggering module determines that data loading needs to be performed, it fuses the operating features from one or more related sensors through the data fusion core to generate a high-confidence comprehensive parameter value. The fused integrated parameter values are packaged and loaded into the digital twin model in a holistic manner.
2. The method according to claim 1, characterized in that, The real-time preprocessing includes performing Fourier transform on the vibration signal to extract frequency domain features, and filtering and reducing noise for temperature and pressure.
3. The method according to claim 1, characterized in that, The predictive state-dependent triggering includes: Using a pre-trained Long Short-Term Memory (LSTM) network model, the expected value of the operating condition characteristics at the next moment is predicted in real time based on the historical operating condition feature sequence and the current operating parameters. Calculate the residual between the real-time measured value of the operating condition characteristic and the expected value; When the residual exceeds the dynamic adjustment threshold, data loading is triggered; the dynamic adjustment threshold is determined based on the operation command variance of the equipment and the standard deviation of the sensor reference noise; the operation command variance is used to quantify whether the current operating state is in a stable period or a period of drastic change.
4. The method according to claim 1, characterized in that, The event-driven triggering refers to the forced data loading when a critical event signal is received from the equipment control system; the critical event signal includes only: motor start command, motor stop command, ESC over-temperature alarm signal, and level one fault code broadcast by the flight control system.
5. The method according to claim 1, characterized in that, The steps for generating high-confidence comprehensive parameter values through the data fusion core specifically include: By constructing a consistency support matrix; for multiple redundant sensors measuring the same physical quantity, the support of any two sensors is calculated. With sensors The confidence distance between the sensors is calculated, and the confidence distance is mapped to mutual support using a decay function, thereby constructing a support matrix that reflects the compatibility relationship between the sensors. The overall support of each sensor is calculated based on the support matrix. Sensors with an overall support lower than a preset consistency threshold are identified as abnormal sensors and their observation data is removed from the current fusion cycle. The sliding window algorithm is used to calculate the measurement variance of each effective sensor after screening in real time. Based on the minimum variance criterion, the reciprocal of the sensor measurement variance is used as the benchmark to calculate the normalized weight. The high-confidence comprehensive parameter value is obtained by weighting and summing the operating characteristics of each effective sensor using the normalized weights.
6. The method according to claim 5, characterized in that, The support matrix reflecting the compatibility relationship between sensors for: in, This is a consistency tolerance parameter used to control the rate of support decay; The confidence distance, in its physical sense, is the statistical distance between the conditional probability density functions of two sensor observations, and is defined as follows: in, These are all sensor serial numbers. This represents the total number of sensors; To indicate at time sensor The observed value is In the case of a physical quantity x, the posterior probability density function is given; it is usually assumed that the measurement noise follows a normal distribution, i.e. 。 7. The method according to claim 5, characterized in that, Calculate overall support and anomaly detection threshold : in, The total number of sensors participating in the fusion. The preset consistency coefficient is 0.7-0.8; if Then determine the sensor This is an abnormal sensor.
8. The method according to claim 5, characterized in that, Calculate normalized weights And the comprehensive parameter value X with high confidence is: in A set of selected, effective sensors; To calculate the first in real time using the sliding window algorithm Each sensor at time Measurement variance: in The length of the time series of the sliding window used to calculate the variance (i.e., the number of sampling points within the window). This represents the mean of the data within the window. in, For effective sensors The measurement value at the current moment.
9. The method according to claim 1, characterized in that, The digital twin model utilizes The steps for fault diagnosis include: Input the data into the preset physical degradation model, compare the Euclidean distance between the current simulation state and the healthy baseline state, and locate the fault location when the distance exceeds the preset deviation range.
10. A dynamic loading system for operating parameters suitable for digital twin models, characterized in that, include: The sensor interface module is used to collect multi-source heterogeneous raw operating condition data of the equipment; A real-time preprocessing engine, connected to the sensor interface module, is used to perform real-time preprocessing of the raw data and extract operating condition features. An adaptive triggering module, connected to the real-time preprocessing engine, has a Long Short-Term Memory (LSTM) network model deployed within it and is configured to use a hybrid model that combines predictive state-dependent triggering and event-driven triggering to determine whether data loading needs to be performed. The core of the data fusion is used to fuse operating condition features when the adaptive triggering module determines that data loading needs to be performed. The model loading interface is used to load the fused data into the digital twin model in a holistic manner.