Digital twin system and construction method thereof

Abstract

The application discloses a kind of digital twinborn systems and its construction method, it is related to digital twinborn technical field, including: the multiple-source heterogeneous data of digital twinborn entity is classified and collected, through adaptive wavelet threshold filtering denoising, data is mapped as multidimensional feature vector and filters key features, and key feature data is output;Physical mechanism layer is established based on general physical law, defines core physical parameter and constraint equation, initializes particle swarm to build data-driven layer, constructs LSTM time series prediction module and executes particle filtering state calibration, selects configuration communication protocol to build virtual-actual interaction layer, realizes the state mapping and instruction feedback of digital twinborn entity and virtual model;Real-time calculation virtual-actual state deviation and carry out attribution diagnosis, according to the deviation source adjustment model parameter or key feature, based on long time series prediction result generates optimization parameter combination and simulates and verifies in virtual environment, adjusts entity parameter through instruction feedback channel control, realizes predictive optimization.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of digital twinning, in particular to a digital twinning system and a construction method thereof. BACKGROUND

[0002] As an important bridge connecting physical entities and virtual models, digital twinning technology has been widely applied in industrial manufacturing, smart cities, energy management and other fields in recent years. The existing technology usually adopts multi-source data acquisition, dynamic modeling and real-time interaction to construct a digital twinning system. For example, entity state data is collected through a sensor network, a physical model or a data-driven algorithm is combined to realize dynamic updating of the virtual model, and a communication protocol is used to complete the instruction transmission between the virtual and real. Some technologies also introduce noise filtering, feature screening and spatio-temporal alignment methods to improve the quality of data and the accuracy of the model.

[0003] However, the existing technology has obvious deficiencies in the data preprocessing stage. The noise filtering of multi-source heterogeneous data often uses a fixed threshold filtering method, which cannot dynamically adapt to the distribution changes of noise and effective signals in the data, leading to problems such as loss of sudden signals or residual noise. In addition, the feature screening process usually relies on artificial experience or simple statistical methods, lacks dynamic evaluation of the correlation between features and entity core states, and is prone to introduce redundant features or ignore key features, affecting the subsequent modeling accuracy. In terms of model construction and optimization, the existing technology also faces challenges. The coupling between the physical mechanism layer and the data-driven layer is weak, which may cause the state output of the virtual model to deviate from the physical rationality boundary. At the same time, the real-time performance and reliability of virtual-real interaction are insufficient. For example, improper selection of communication protocols or mismatched transmission rates may cause data synchronization delays or instruction loss problems. These deficiencies limit the prediction accuracy and optimization capability of the digital twinning system in complex scenarios. SUMMARY

[0004] (I) Technical problems solved

[0005] In view of the deficiencies of the existing technology, the present application provides a digital twinning system and a construction method thereof. Through adaptive wavelet threshold filtering denoising, dynamic feature screening, deep coupling of physical mechanism and data-driven, and high-reliability virtual-real interaction mechanism, the problems of incomplete multi-source data noise filtering, insufficient feature correlation evaluation, lack of model physical rationality, and low real-time synchronization efficiency are solved, and the prediction accuracy and optimization capability of the digital twinning system are significantly improved.

[0006] (II) Technical solutions

[0007] To achieve the above purpose, the present application is implemented by the following technical solutions: a construction method of a digital twinning system, comprising:

[0008] The multi-source heterogeneous data of the digital twin entity is classified and collected, the self-adaptive wavelet threshold filtering denoising is performed, the filtering threshold is dynamically adjusted, the denoised data is reconstructed, the data is mapped into a multi-dimensional feature vector and the key features are screened, the time stamp alignment and the spatial coordinate alignment are performed, and the key feature data aligned in time and space is output;

[0009] A physical mechanism layer is established based on general physical laws, core physical parameters and constraint equations are defined, a data driven layer is built by initializing a particle swarm, an LSTM time series prediction module is constructed and particle filtering state calibration is performed, a virtual-real interaction layer is built by selecting and configuring a communication protocol, and state mapping and instruction feedback of the digital twin entity and the virtual model are realized;

[0010] The virtual-real state deviation is calculated in real time and attributed diagnosis is performed, the model parameters or key features are adjusted according to the deviation source, the optimized parameter combination is generated based on the long time series prediction result and is simulated and verified in the virtual environment, the entity parameter is adjusted through the instruction feedback channel, and predictive optimization is realized.

[0011] Further, the specific steps of self-adaptive wavelet threshold filtering denoising include:

[0012] A db4 wavelet basis is selected for 3-layer wavelet decomposition to separate the noise high-frequency components and the effective signal low-frequency components;

[0013] The energy proportion of each layer of high-frequency coefficients is calculated, and the energy proportion is the proportion of the sum of the absolute values of the high-frequency coefficients of the layer to the total energy of all high-frequency layers;

[0014] A soft threshold function is used to dynamically adjust the filtering threshold, if the high-frequency energy proportion of each layer is less than 15%, the filtering threshold is increased to 1.2 times the original threshold, and if the high-frequency energy proportion of any layer is more than 30%, the filtering threshold of the layer is reduced to 50%-60% of the stable scene;

[0015] The adjusted wavelet coefficients are subjected to 3-layer wavelet inverse transformation, and the denoised data sequence is reconstructed.

[0016] Further, the specific steps of the key feature screening include:

[0017] The multi-source heterogeneous data is mapped into a multi-dimensional feature vector, a multi-head attention mechanism is used to calculate the feature correlation weight, the multi-head attention mechanism is configured with 8 attention heads, and the weight is output through the dot product attention formula;

[0018] The weights output by all attention heads are averaged and pooled to obtain the final correlation weight of each feature, and the weight value range is [0, 1];

[0019] The features ranked in the top 70%-80% of the weight are selected as the key features, and the redundant features with a correlation weight lower than a preset correlation threshold are removed.

[0020] Further, the specific steps of timestamp alignment and spatial coordinate alignment processing include:

[0021] Timestamp alignment is performed based on the timestamp of the entity core state acquisition device, the sampling frequency of all key feature data is unified to 2-3 times the entity dynamic response frequency, and linear interpolation or mean down-sampling method is used for timestamp calibration;

[0022] A three-dimensional rectangular coordinate system with the entity geometric center as the origin is constructed, spatial coordinate alignment is performed, the installation position coordinates of each acquisition device are measured, and the spatial coordinate field is marked in the key feature data to verify the timestamp consistency of the same spatial position data.

[0023] Further, the construction steps of the physical mechanism layer include:

[0024] Select a general physical law as the modeling basis; define entity core physical parameters, including inertia coefficient, damping coefficient, and energy loss rate, and mark the physical meaning and value boundary; establish a physical constraint equation to limit the physical rationality boundary of virtual model state output.

[0025] Further, the construction steps of the data-driven layer include:

[0026] Initialize the particle swarm, generate 500-1000 virtual particles according to the entity state complexity, and each particle contains core state parameters and feature association weights;

[0027] Build an LSTM time series prediction module, configure the number of hidden layer nodes to be 32-128, input historical key feature data, and output state prediction results for the next 5-10 time steps;

[0028] Perform particle filtering state calibration, output the optimal state estimation through resampling and weight updating, and check whether it meets the physical constraint equation.

[0029] Further, the construction steps of the virtual-real interaction layer include:

[0030] Select and configure the MQTT-SN protocol, enable QoS2 level, and adjust the transmission rate to 2 times the entity dynamic response frequency;

[0031] Build a state mapping channel from entity to virtual, achieve millisecond-level data synchronization, and set up a CRC check mechanism;

[0032] Build a command feedback channel from virtual to entity, convert the optimization parameter command to JSON format, and ensure the effective transmission of the command through the confirmation mechanism.

[0033] Further, the specific steps of real-time calculation of virtual-real state deviation and attribution diagnosis include:

[0034] The absolute error of the virtual optimal state estimation value and the actual state value of the entity is calculated, and the time stamp and the associated features corresponding to the deviation are recorded; the multi-modal error diagnosis model is used to locate the source of the deviation, including data noise, model parameter drift or change of entity dynamic characteristics; the filter threshold, model parameter or key feature set is adjusted according to the source of the deviation, and the model is corrected.

[0035] Further, the specific steps of predictive optimization include:

[0036] Based on the long-time sequence prediction result of LSTM, a genetic algorithm is called to generate multiple sets of optimization parameter combinations;

[0037] The parameter combinations are simulated and verified in a virtual environment, and the optimal combination meeting the requirements of running efficiency and energy consumption is selected;

[0038] The entity parameter adjustment is controlled through the instruction feedback channel to realize predictive optimization.

[0039] A digital twin system comprises:

[0040] A data acquisition and processing module classifies and acquires multi-source heterogeneous data of the digital twin entity, removes noise through adaptive wavelet threshold filtering, dynamically adjusts the filtering threshold and reconstructs the denoised data, maps the data into a multi-dimensional feature vector and selects key features, performs timestamp alignment and spatial coordinate alignment processing, and outputs the spatio-temporally aligned key feature data;

[0041] A model construction module establishes a physical mechanism layer based on general physical laws, defines core physical parameters and constraint equations, initializes a particle swarm to build a data-driven layer, constructs an LSTM time sequence prediction module and performs particle filtering state calibration, selects and configures a communication protocol to build a virtual-real interaction layer, and realizes state mapping and instruction feedback of the digital twin entity and the virtual model;

[0042] An optimization control module calculates the virtual-real state deviation in real time and performs attribution diagnosis, adjusts the model parameters or key features according to the source of the deviation, generates optimization parameter combinations based on long-time sequence prediction results and simulates and verifies them in a virtual environment, controls the entity parameter adjustment through the instruction feedback channel, and realizes predictive optimization.

[0043] (Three) beneficial effects

[0044] The present application provides a digital twin system and a construction method thereof, which has the following beneficial effects:

[0045] (1) Through adaptive wavelet threshold filtering denoising, dynamic feature screening and space-time alignment processing, the quality and efficiency of data preprocessing are significantly improved. The adaptive filtering technology can dynamically adjust the threshold according to the noise distribution, effectively retain the mutation signal and filter out the noise. The multi-head attention mechanism screens the key features, reduces the interference of redundant data on the model, and the space-time alignment ensures the synchronization and consistency of multi-source data, providing high-precision input for subsequent modeling, enhancing the data reliability and modeling accuracy of the digital twin system.

[0046] (2) By coupling the physical mechanism layer and the data driven layer, the accuracy and reliability of the digital twin model are significantly improved. The physical mechanism layer defines the core parameters and constraint equations based on general physical laws, ensuring that the model output meets the physical rationality boundary. The data driven layer uses particle swarm initialization and LSTM time series prediction module, combined with particle filtering dynamic calibration, to enhance the accuracy of state estimation. At the same time, the high-reliability virtual-real interaction layer is built by selecting the MQTT-SN protocol, realizing millisecond-level synchronization and instruction feedback, solving the problems of model deviation from physical logic and real-time deficiency in traditional methods, and providing a high-precision foundation for predictive optimization.

[0047] (3) By calculating the virtual-real state deviation in real time and performing attribution diagnosis, the dynamic correction and predictive optimization ability of the digital twin system is significantly improved. The multi-modal error diagnosis model accurately locates the deviation source and adjusts the filtering threshold, model parameters or key feature set accordingly, ensuring that the model continuously fits the entity state. Combined with LSTM long-time series prediction and genetic algorithm to generate optimized parameter combination, after simulation and verification in the virtual environment, the entity operating parameters are adjusted in advance through the high-reliability instruction feedback channel, realizing the transition from passive response to predictive optimization, effectively preventing potential failures and improving operating efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0048] Fig. 1 The figure is a schematic diagram of the construction method steps of the digital twin system of the present application.

[0049] Fig. 2 The figure is a schematic diagram of the optimization control process of the digital twin system of the present application.

[0050] Fig. 3 The figure is a schematic diagram of the structure of a digital twin system of the present application. DETAILED DESCRIPTION

[0051] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0052] Referring to Figs. 1-2 The application provides a construction method of a digital twin system, comprising the following steps:

[0053] Step one: classified collection of multi-source heterogeneous data of the digital twin entity, denoising through adaptive wavelet threshold filtering, dynamic adjustment of the filtering threshold and reconstruction of the denoised data, mapping of the data into a multi-dimensional feature vector and screening of key features, time stamp alignment and spatial coordinate alignment processing, and output of the key feature data aligned in time and space;

[0054] The step one comprises the following contents:

[0055] Step 101: classified collection of multi-source heterogeneous data of the digital twin entity in the whole life cycle according to three dimensions of state data, environment data and operation and maintenance data, wherein the state data is collected through a collection device such as a sensor or an upper computer, which is self-provided or externally connected, to collect parameters reflecting the core state of the entity, and covers key state indicators of the entity in different running stages; the environment data is collected through a special collection device deployed in the environment of the entity to collect external environmental factor data that may affect the running state of the entity, and covers various environmental variables faced by the entity in the running process; and the operation and maintenance data is collected through manual input or connection with an entity operation and maintenance management system to collect operation and maintenance related record data in the whole life cycle of the entity, including but not limited to maintenance operation, fault handling, spare part replacement and other information;

[0056] Step 102: denoising through adaptive wavelet threshold filtering, random noise contained in the multi-source heterogeneous data such as inherent noise of a sensor and transmission interference, selection of an adaptive wavelet basis function such as a db4 wavelet basis, consideration of denoising smoothness and signal detail preservation, determination of 3-layer wavelet decomposition layers through experimental verification, effective separation of noise high-frequency components and effective signal low-frequency components, and avoidance of excessive decomposition redundancy; 3-layer wavelet decomposition of the data sequence, calculation of the energy of each layer of high-frequency coefficients, i.e. the sum of the squares of the absolute values of all high-frequency coefficients in the layer, statistics of the proportion of the energy of each layer in the total energy of all high-frequency layers, determination of the presence of an entity state mutation signal when the proportion is more than 30%, and determination of the presence of noise when the proportion is less than 10%;

[0057] Step 103: dynamic adjustment of the filtering threshold and reconstruction of the denoised data, adoption of a soft threshold function, increase of the filtering threshold in the soft threshold function to strengthen noise removal when in a data stable scenario, i.e. the proportion of the energy of each layer of high-frequency coefficients is less than 15%, for example, setting the filtering threshold to 1.2 times the original value, and reduction of the filtering threshold to 50% to 60% of the value in the stable scenario to avoid loss of a mutation signal when in a data mutation scenario, i.e. the proportion of the energy of any layer of high-frequency coefficients is more than 30%; 3-layer wavelet inverse transformation of the wavelet coefficients after adjustment of the filtering threshold, and reconstruction of a clean data sequence after denoising;

[0058] Step 104: Map various types of data into a multi-dimensional feature vector, use a preset number of attention heads, such as 8, to balance the diversity and efficiency of weight calculation, use dot product attention formula to output weight, dot product attention formula is 64, core state such as running stability, energy consumption, etc., average pooling is performed on the weights output by all attention heads to obtain the final correlation weight of each feature, the value range is [0, 1]; filter the top 70%-80% features by weight to ensure that the retained features can fully reflect the entity core state, and eliminate features with correlation weight lower than the preset correlation threshold, such as 0.1, to avoid increasing the complexity of subsequent modeling by redundant features;

[0059] Step 105: Perform timestamp alignment operation to determine the reference timestamp, select the timestamp of the entity core state acquisition device, such as the dedicated sensor reflecting running stability and energy efficiency, as the reference, and require the sampling frequency of the device to be ≥ the dynamic response frequency of the entity, the minimum time interval of the change of the entity state; unify the sampling frequency of all key feature data to 2-3 times the dynamic response frequency of the entity: when the original sampling frequency is lower than the target frequency, use linear interpolation method to complete the missing time point data; when it is higher than the target frequency, use mean down-sampling method to take the average; calibrate all data timestamps to integer multiples of the reference timestamp to ensure that the same timestamp contains all key feature data.

[0060] Step 106: Implement spatial coordinate alignment processing, construct a three-dimensional right-handed coordinate system with the geometric center of the entity as the origin, the X-axis along the main motion direction of the entity, the Y-axis perpendicular to the X-axis in the horizontal direction, and the Z-axis perpendicular to the X-Y plane in the vertical direction, to ensure suitability for various entities; measure the three-dimensional coordinates of the installation positions of the key feature acquisition devices relative to the geometric center of the entity using a laser range finder; add a spaceCoordinate field to the key feature data to mark the corresponding three-dimensional coordinates; verify the timestamp consistency of multiple types of data collected at the same spatial location to ensure spatio-temporal dimension matching; output high-quality key feature data aligned in space and time.

[0061] In use, the contents of steps 101 to 106 are combined:

[0062] Through adaptive wavelet threshold filtering, dynamic feature selection, and spatio-temporal alignment processing, the quality and efficiency of data preprocessing are significantly improved. The adaptive filtering technology can dynamically adjust the threshold according to the noise distribution, effectively retaining sudden signals and filtering out noise. The multi-head attention mechanism selects key features, reducing the interference of redundant data on the model. Spatio-temporal alignment ensures the synchronization and consistency of multi-source data, providing high-precision input for subsequent modeling, enhancing the data reliability and modeling accuracy of the digital twin system.

[0063] Step two: establish a physical mechanism layer based on general physical laws, define core physical parameters and constraint equations, initialize particle swarm to build a data-driven layer, construct an LSTM time series prediction module and perform particle filtering state calibration, select and configure a communication protocol to build a virtual-real interaction layer, and realize state mapping and instruction feedback between digital twin entities and virtual models;

[0064] The step two includes the following contents:

[0065] Step 201: Select general physical laws as the basis for modeling the physical mechanism layer. Based on the common operating characteristics of digital twin entities, select universal physical laws as the basis for modeling. The physical laws include the law of conservation of mass, the law of conservation of energy, and Newton's law of motion. The selected laws should cover common state change scenarios such as entity composition, energy conversion, and motion under force, and should not depend on specific entity types to ensure universal adaptability and provide a bottom layer basis for subsequent physical reasonableness of the limited model;

[0066] Step 202: Define the core physical parameters of the entity. According to the selected physical laws, determine the key influence parameters of the entity state change. If motion characteristics are involved, define inertia coefficient, damping coefficient, and friction coefficient. If energy characteristics are involved, define energy loss rate and energy conversion efficiency coefficient. All parameters should be clearly labeled with physical meaning and value boundaries to avoid invalidation of subsequent constraint equations due to ambiguous parameters;

[0067] Step 203: Establish the physical constraint equation of the entity state change. Based on the defined core physical parameters and selected general physical laws, construct the constraint equation that limits the output boundary of the virtual model state. The equation should be related to the core state and physical parameters of the entity, such as running stability and energy consumption. For example, for energy characteristics, a constraint equation can be established as entity energy consumption ≤ input energy × energy conversion efficiency coefficient. By inputting key feature data such as entity input energy data, it can be verified whether the output results of the subsequent data-driven layer conform to the physical logic. If the prediction results of the data-driven layer exceed the equation limit range, it is determined to be physically infeasible and needs to be calibrated subsequently;

[0068] Step 204: Initialize the particle swarm to build the data-driven layer. With key feature data as input, generate 500-1000 virtual particles according to the complexity of the entity state. When the entity state dimension ≤10, select 500 particles, and when the dimension >10, select 800-1000 particles to balance estimation accuracy and computational efficiency. Each particle contains entity core state parameters and feature correlation weights, which completely represent an estimate of the current state of the entity. Core state parameters include indicators reflecting running stability and energy consumption indicators. Assign initial weights based on the cosine similarity of particles and key feature data. The higher the similarity, the greater the initial weight, and the sum of all particle weights is normalized to 1;

[0069] Step 205: Construct and run the LSTM time series prediction module, configure the LSTM network parameters, set the number of hidden layer nodes to 32-128, select 64-128 nodes when the historical data volume > 100,000, select 32-64 nodes when the data volume ≤ 100,000, and set the training period to 12-24 hours; input the key feature historical data of the past 3-5 time windows into the trained LSTM network, and output the state change trend of the entity in the next 5-10 time steps, such as the energy consumption change curve in the next 5 time steps, which is used as the prior state reference for subsequent particle filtering to avoid random prediction caused by the absence of prior guidance;

[0070] The length of the time window is defined as the dynamic response period of the entity, and the dynamic response frequency of the entity is determined through vibration testing or historical operation data analysis, i.e., the minimum time interval for the state of the entity to change significantly. For example, the dynamic response frequency of a high-speed rotating machine can be measured by an acceleration sensor, and the reciprocal of the main frequency of the vibration signal is taken as the response period. The response period of static equipment can be determined by the minimum time interval of state parameter mutation in historical fault records;

[0071] Step 206: Perform particle filtering state calibration to output the optimal state estimation, obtain the real-time state data of the entity, i.e., the latest key feature data, calculate the similarity between each particle and the real-time state data of the entity and update the particle weight, the higher the similarity, the greater the weight increase; perform resampling operation, retain the top 30%-50% high-weight particles, such as the first 300 particles in 800 particles, and supplement 200-300 random particles to avoid estimation unification caused by particle degradation; weight average the resampled particles by weight, output the current optimal state estimation result of the entity, and simultaneously input the established physical constraint equation for rationality check, if it exceeds the constraint range, re-execute weight update;

[0072] Step 207: Select and configure a general communication protocol to build a virtual-real interaction layer foundation, select the MQTT-SN protocol that adapts to multiple device communication, configure the protocol parameters: enable QoS2 level to ensure that messages are received only once, avoiding data duplication or loss; adjust the transmission rate according to the dynamic response frequency of the entity to ensure that the transmission rate ≥ 2 times the dynamic response frequency of the entity, such as setting the transmission rate to 2Hz when the dynamic response frequency of the entity is 1Hz, to avoid synchronization lag caused by data accumulation;

[0073] Step 208: Build a virtual-to-physical state mapping channel to transmit the latest key feature data in real time to the data-driven layer through the configured MQTT-SN protocol, update the particle swarm state of the virtual model; shorten the heartbeat packet interval to 10-20 ms to control the data transmission delay within 50 ms, achieve millisecond-level synchronization of the entity state to the virtual model; set CRC data integrity check at the channel receiving end, if data is missing or incorrect, trigger the retransmission mechanism immediately to ensure data validity;

[0074] Step 209: Build a virtual-to-physical instruction feedback channel to convert the entity optimal state estimation results and long-term prediction results output in step 205, such as the state deviation and optimal operating parameters that the entity may have in the future, into a general instruction format that the entity control unit can recognize, such as a JSON format parameter adjustment instruction, according to the preset rules; when the prediction result shows that the future state of the entity may exceed the safety range, such as energy consumption exceeding the energy consumption threshold, the channel automatically triggers instruction transmission and sends the optimal parameter instruction to the entity control unit, such as the PLC module; set up an instruction receiving confirmation mechanism, if the entity control unit does not feedback the successful reception within 100 ms, the channel immediately re-sends the instruction to ensure effective communication of the instruction; output a dynamic twin model coupled with the physical mechanism layer, data-driven layer and virtual-physical interaction layer.

[0075] In use, in combination with the contents of steps 201 to 209:

[0076] By coupling the physical mechanism layer and the data-driven layer, the accuracy and reliability of the digital twin model are significantly improved, the physical mechanism layer defines the core parameters and constraint equations based on general physical laws to ensure that the model output meets the physical rationality boundary, the data-driven layer uses particle swarm initialization and LSTM time series prediction module combined with particle filter dynamic calibration to enhance the accuracy of state estimation, at the same time, the high-reliability virtual-physical interaction layer is built by selecting MQTT-SN protocol to realize millisecond-level synchronization and instruction feedback, solving the problems of model deviation from physical logic and insufficient real-time in traditional methods, providing a high-precision foundation for predictive optimization.

[0077] Step three: Real-time calculation of virtual-physical state deviation and attribution diagnosis, adjust model parameters or key features according to the source of deviation, generate optimal parameter combination based on long-term prediction results and simulate and verify in virtual environment, control entity parameter adjustment through instruction feedback channel, realize predictive optimization.

[0078] The step three includes the following contents:

[0079] Step 301: Real-time calculation of virtual-real state deviation, i.e. virtual-real residual, takes the output of the data-driven layer of the constructed dynamic twin model as the virtual end data source, and takes the latest collected actual state data of the entity as the entity end data source, and uses the absolute error formula: virtual-real residual = |virtual optimal state estimate value-actual state data value| to calculate the deviation between the two, and record the time stamp corresponding to the deviation, and the associated features, such as the entity state features and environmental features related to the deviation, to provide basic data support for subsequent attribution diagnosis;

[0080] Step 302: Adopting a multi-modal error diagnosis model to attribute diagnose the virtual-real residual and locate the source of the deviation, including: if the correlation degree of the residual and any noise feature is ≥0.8, the correlation degree uses Pearson correlation coefficient, and the noise feature is, for example, environmental interference feature, then it is determined that the deviation is caused by data noise; according to the constraint equation established by the physical mechanism layer, if the residual exceeds the physical reasonable range defined by the constraint equation, such as the virtual estimated entity rotating speed exceeding the mechanical structure tolerance value defined by the physical mechanism layer, then it is determined that the deviation is caused by model parameter drift; according to the entity state prediction trend output by the LSTM network of the data-driven layer, if the residual increases with time and is opposite to the prediction trend, such as the LSTM predicts energy consumption to decrease but the residual shows actual energy consumption to increase, then it is determined that the deviation is caused by the change of entity dynamic characteristics;

[0081] Step 303: If it is determined that the deviation is caused by data noise, immediately return to step one and adjust the filtering threshold: for the wavelet coefficient corresponding to the noise feature that causes the deviation, increase the original filtering threshold by 10%-20%, such as the original threshold is 0.6, adjust it to 0.66-0.72, to strengthen the filtering effect of this kind of noise; re-execute the process of step one on the latest collected original data of the entity, generate new key feature data, and input it into the data-driven layer of step two, update the particle swarm state and optimal state estimation result of the virtual model, and complete the noise-oriented correction of the model;

[0082] Step 304: If the deviation is determined to be caused by model parameter drift, adjust the parameters in three layers: first, adjust the number of hidden layer nodes of the LSTM network in the data-driven layer of step two. When the deviation continues to increase, such as when the residual error increases by more than 5% for three consecutive time steps, increase the number of hidden layer nodes by 5-10. For example, if the original number of nodes is 64, adjust it to 69-74. Second, adjust the number of particles in the particle filter of the data-driven layer. When the particle diversity is insufficient, such as when the weight of high-weight particles accounts for more than 80%, increase the number of virtual particles by 200. For example, if the original number of particles is 800, adjust it to 1000. Third, correct the core parameters of the physical mechanism layer in step two. According to the running time and historical operation data of the entity, such as when the entity has been running for 1000 hours and the aging degree has increased by 15%, increase the energy loss rate and other parameters by the aging proportion. For example, if the original energy loss rate is 0.1, increase it to 0.115. Adjust the data-driven layer modeling process in step two to output a new virtual optimal state estimation, and realize the drift-oriented correction of model parameters.

[0083] Step 305: If the deviation is determined to be caused by changes in the dynamic characteristics of the entity, immediately trigger the key feature extraction guided by the attention mechanism in step one, and recalculate the feature correlation weight: include the newly appeared state features related to the dynamic characteristic changes of the entity in the feature set, such as the newly added vibration frequency features during the operation of the entity. Use the multi-head attention mechanism to recalculate the correlation weight of all features and the core state of the entity, select a new Top 70%-80% key feature set, and eliminate the original redundant features with a correlation degree less than 0.1. Input the new key feature set into the data-driven layer of step two, update the training data of the LSTM network and the particle information of the particle filter, and complete the dynamic-oriented correction of the model features.

[0084] Step 306: Based on the long-term prediction results of the LSTM network in the data-driven layer of step two, select the prediction data of the next three time steps, which is consistent with the dynamic response period of the entity, such as 3s when the response period is 1s. When the prediction result shows that the core indicator will exceed the preset threshold, such as energy consumption and failure rate, for example, the energy consumption threshold is 500W, and the predicted energy consumption will rise to 550W after 3 time steps, call the general optimization algorithm library of the entity, select the parameter optimization algorithm based on genetic algorithm, and generate multiple sets of optimization parameter combinations, such as adjusting the running speed, load, and working voltage of the entity.

[0085] Step 307: The generated multiple sets of optimized parameter combinations are input into the constructed dynamic twin model one by one to simulate the running state of the entity under each parameter combination in the virtual environment: the simulation duration is consistent with the time step of the LSTM prediction, such as 3s, and the core indicator change data corresponding to each parameter combination is recorded; by comparing the simulation results of each parameter combination, the optimal parameter combination that simultaneously meets the running efficiency standard and the energy consumption threshold is selected, such as the combination of speed adjustment of 1500rpm and load adjustment of 80%, with a simulation energy consumption of 480W and an efficiency of 92%;

[0086] Step 308: The selected optimal parameter combination is converted into a general instruction format recognizable by the entity control unit such as PLC module, consistent with the instruction format of the virtual-real interaction layer, such as JSON format, to generate entity optimization control instructions; the instructions are transmitted to the entity control unit in real time through the constructed virtual-to-physical instruction feedback channel, using the MQTT-SN protocol and QoS2 level, to control the entity to adjust the running parameters in advance, realizing predictive optimization of the entity, which is different from the existing passive operation and maintenance mode of correcting after the problem occurs.

[0087] In use, the contents of steps 301 to 308 are combined:

[0088] By calculating the virtual-real state deviation in real time and performing attribution diagnosis, the dynamic correction and predictive optimization capabilities of the digital twin system are significantly improved, the multi-modal error diagnosis model accurately locates the deviation source, and adjusts the filtering threshold, model parameters or key feature set accordingly to ensure that the model continuously fits the entity state, generates optimized parameter combinations using LSTM long-time sequence prediction and genetic algorithm, and verifies them in a virtual environment, and then adjusts the entity running parameters in advance through a high-reliability instruction feedback channel, realizing the transition from passive response to predictive optimization, effectively preventing potential failures and improving running efficiency.

[0089] Please refer to Fig. 3 The application also provides a digital twin system, comprising,

[0090] The data acquisition and processing module classifies and collects the multi-source heterogeneous data of the digital twin entity, denoises through adaptive wavelet threshold filtering, dynamically adjusts the filtering threshold and reconstructs the denoised data, maps the data into a multi-dimensional feature vector and selects key features, performs timestamp alignment and spatial coordinate alignment processing, and outputs the spatio-temporally aligned key feature data;

[0091] The model construction module establishes a physical mechanism layer based on general physical laws, defines core physical parameters and constraint equations, initializes a particle swarm to build a data-driven layer, constructs an LSTM time series prediction module and performs particle filtering state calibration, selects and configures a communication protocol to build a virtual-real interaction layer, and realizes state mapping and instruction feedback of the digital twin entity and the virtual model;

[0092] The optimization control module calculates the virtual-real state deviation in real time and performs attribution diagnosis, adjusts model parameters or key features according to the deviation source, generates an optimized parameter combination based on long-time sequence prediction results and simulates and verifies in the virtual environment, controls entity parameter adjustment through the instruction feedback channel, and realizes predictive optimization.

[0093] In the application, the several formulas involved are dimensionless values for numerical calculation, and the formula is obtained by software simulation of a large amount of data to obtain a formula of the most recent real situation, and the coefficients in the formula are set by a person skilled in the art according to the actual situation.

[0094] The above embodiments can be realized wholly or partially by software, hardware, firmware or any other combination. When realized by software, the above embodiments can be realized wholly or partially in the form of a computer program product. Those skilled in the art can realize that the units and algorithm steps of the examples described in connection with the embodiments disclosed herein can be realized by electronic hardware, computer software and a combination of electronic hardware and computer software. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solutions.

[0095] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, which can be located in one place or distributed on multiple network units. Part or all of the units can be selected to achieve the purpose of the embodiments according to actual needs.

[0096] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical range disclosed in the present application, which should be covered within the protection scope of the present application.

Claims

1. A method for constructing a digital twin system, characterized in that: The application relates to a digital twin entity state estimation method and device. The method comprises the following steps: collecting multi-source heterogeneous data of a digital twin entity, performing self-adaptive wavelet threshold filtering denoising, dynamically adjusting a filtering threshold and reconstructing denoised data, mapping the data into a multi-dimensional feature vector and screening key features, performing timestamp alignment and spatial coordinate alignment processing, and outputting key feature data aligned in time and space; establishing a physical mechanism layer based on general physical laws, defining core physical parameters and constraint equations, initializing a particle swarm to build a data-driven layer, generating 500-1000 virtual particles according to the complexity of the entity state, and each particle containing core state parameters and feature correlation weights; constructing an LSTM time series prediction module, configuring the number of hidden layer nodes to be 32-128, inputting historical key feature data, and outputting state prediction results of 5-10 time steps in the future, which are used as prior state references for subsequent particle filtering; performing particle filtering state calibration, outputting optimal state estimation through resampling and weight updating, and checking whether the optimal state estimation meets the physical constraint equation; selecting and configuring a communication protocol to build a virtual-real interaction layer, and realizing state mapping and instruction feedback of the digital twin entity and the virtual model; calculating the absolute error between the virtual optimal state estimation value and the actual state value of the entity, recording the timestamp and associated features corresponding to the deviation, positioning the deviation source by using a multi-modal error diagnosis model, including data noise, model parameter drift or entity dynamic characteristic change, adjusting the filtering threshold, model parameters or key feature set according to the deviation source, and completing model correction; based on the LSTM long time series prediction result, a genetic algorithm is used to generate multiple sets of optimized parameter combinations; the parameter combinations are simulated and verified in a virtual environment, and the optimal combination meeting the requirements of running efficiency and energy consumption is selected; the entity parameter is adjusted through the instruction feedback channel to realize predictive optimization. The specific steps of the self-adaptive wavelet threshold filtering denoising include the following steps: selecting a db4 wavelet basis for 3-layer wavelet decomposition to separate noise high-frequency components and effective signal low-frequency components; calculating the energy proportion of each layer high-frequency coefficient, wherein the energy proportion is the ratio of the sum of the absolute values of the high-frequency coefficients of the layer to the total energy of all high-frequency layers; dynamically adjusting the filtering threshold by using a soft threshold function, wherein if the energy proportion of each layer high-frequency is lower than 15%, the filtering threshold is increased to 1.2 times of the original threshold; if the energy proportion of any layer high-frequency is higher than 30%, the filtering threshold of the layer is reduced to 50%-60% of the stable scene; and performing 3-layer wavelet inverse transformation on the adjusted wavelet coefficients to reconstruct the denoised data sequence. The specific steps of the key feature screening include the following steps: mapping the multi-source heterogeneous data into a multi-dimensional feature vector, calculating the feature correlation weight by using a multi-head attention mechanism, configuring the multi-head attention mechanism to have 8 attention heads, and outputting the weight by using a dot product attention formula; performing average pooling on the weights output by all the attention heads to obtain the final correlation weight of each feature, wherein the weight value ranges from 0 to 1; and screening the features with a weight ranking of 70%-80% as key features, and eliminating redundant features with a correlation weight lower than a preset correlation threshold. The specific steps of the timestamp alignment and spatial coordinate alignment processing include the following steps:

2. The method of claim 1, wherein: ​ ​ ​ ​ ​ 3. The method of claim 2, wherein: ​ ​ ​ ​ 4. The method of claim 3, wherein: ​ The timestamp alignment is performed based on the timestamp of the entity core state acquisition device, the sampling frequency of all key feature data is unified to 2-3 times of the entity dynamic response frequency, and the timestamp is calibrated by using a linear interpolation method or a mean down-sampling method; A three-dimensional rectangular coordinate system with the geometric center of the entity as the origin is constructed, the spatial coordinate alignment is performed, the installation position coordinates of each acquisition device are measured, and the spatial coordinate field is marked in the key feature data to verify the timestamp consistency of the same spatial position data.

5. The method of claim 1, wherein: The construction steps of the physical mechanism layer include: A general physical law is selected as the modeling basis, the entity core physical parameters are defined, the core physical parameters include the inertia coefficient, the damping coefficient, and the energy loss rate, and the physical meaning and value boundary are marked, and the physical constraint equation is established to limit the physical rationality boundary of the virtual model state output.

6. The method of claim 5, wherein: The construction steps of the virtual-real interaction layer include: MQTT-SN protocol is selected and configured, QoS2 level is enabled, and the transmission rate is adjusted to 2 times of the entity dynamic response frequency; The entity-to-virtual state mapping channel is built to realize data synchronization, and a CRC check mechanism is set; The virtual-to-physical instruction feedback channel is built to convert the optimization parameter instruction into JSON format, and the instruction is ensured to be effectively conveyed through the confirmation mechanism.

7. A digital twin system for implementing the method of any one of claims 1 to 6, characterized by: It includes: The data acquisition and processing module classifies and collects the multi-source heterogeneous data of the digital twin entity, removes noise through adaptive wavelet threshold filtering, dynamically adjusts the filtering threshold, reconstructs the denoised data, maps the data to a multi-dimensional feature vector, selects key features, performs timestamp alignment and spatial coordinate alignment processing, and outputs the time and space aligned key feature data; The model construction module establishes the physical mechanism layer based on the general physical law, defines the core physical parameters and constraint equation, initializes the particle swarm to build the data driven layer, generates 500-1000 virtual particles according to the entity state complexity, each particle contains core state parameters and feature correlation weights; An LSTM time series prediction module is constructed, the number of hidden layer nodes is configured to be 32-128, historical key feature data is input, and future state prediction results of 5-10 time steps are output, which are used as prior state reference for subsequent particle filtering; particle filtering state calibration is performed, the optimal state estimation is output through resampling and weight updating, and whether it meets the physical constraint equation is verified; a communication protocol is selected and configured to build a virtual-real interaction layer, and state mapping and instruction feedback of the digital twin entity and the virtual model are realized; The optimization control module calculates the absolute error between the virtual optimal state estimation value and the entity actual state value, records the timestamp and associated features corresponding to the deviation, locates the deviation source by using a multi-modal error diagnosis model, including data noise, model parameter drift or entity dynamic characteristic change, adjusts the filtering threshold, model parameters or key feature set according to the deviation source, and completes model correction; based on the LSTM long time series prediction result, a genetic algorithm is called to generate multiple sets of optimization parameter combinations; the parameter combinations are simulated and verified in a virtual environment, and the optimal combination meeting the operation efficiency and energy consumption requirements is selected; the entity parameter adjustment is controlled through the instruction feedback channel to realize predictive optimization.

Images