Intelligent equipment management method and system based on digital twinning

Abstract

The invention discloses an intelligent equipment management method and system based on digital twinning. The method comprises the following steps: constructing an initial digital twinning model based on a multi-source data fusion result; monitoring the operation state of the initial model in real time and collecting feedback data; based on difference analysis of feedback data and a fusion result, evaluating a model deviation degree; and iteratively correcting the parameters of the digital twinborn model according to the deviation degree to optimize the precision. The method can solve the problem of how to carry out iterative correction on the model precision according to the multi-source data fusion result so as to solve the problem of overlarge deviation of the digital twinborn model.

Description

Technical Field

[0001] The present application relates to the field of digital twin technology, and specifically to a smart device management method and system based on digital twins. Background Art

[0002] Digital twin-based intelligent device management methods construct virtual models of physical entities, collect and fuse multi-source sensor data in real time, and achieve accurate monitoring of device status, fault prediction, and efficient management, thereby improving device operating efficiency and lifespan. However, this method faces a key issue: how to effectively iteratively correct the accuracy of the digital twin model based on the results of multi-source data fusion. Due to complex interference factors in the actual operating environment or errors in the initial modeling, there may be excessive deviations between the digital twin model and the physical entity, which in turn affects decision-making accuracy. Therefore, solving the model correction problem is crucial to improving the reliability and adaptability of the system. Summary of the Invention

[0003] In view of this, the present invention discloses a smart device management method and system based on digital twins, which at least partially solve the problems existing in the prior art.

[0004] The smart device management method based on digital twins includes:

[0005] Build an initial digital twin model based on the fusion results of multi-source data;

[0006] Performing real-time monitoring on the operating status of the initial model and collecting feedback data;

[0007] Evaluate the degree of model deviation based on the difference analysis between feedback data and fusion results;

[0008] The digital twin model parameters are iteratively corrected according to the degree of deviation to optimize accuracy.

[0009] Preferably, the multi-source data fusion result further includes:

[0010] Obtaining a first amount of data M1 from multiple data sources;

[0011] Calculate the initial weight W0;

[0012] Adjust the weight according to the following formula: W = W0*exp(α*|F-F_model|), where F represents the real-time data fusion eigenvalue, F_model represents the model expected eigenvalue, and α is the weight adjustment parameter;

[0013] If the deviation is greater than the preset threshold θ (i.e., |F-F_model|>θ), the weights of each data source are redistributed.

[0014] Preferably, the difference analysis based on the feedback data and the fusion result further includes:

[0015] The average variance D_avg of the statistical feedback data;

[0016] Set the deviation sensitivity coefficient k as the evaluation parameter;

[0017] According to the formula J = k*log(D_avg / σ), where σ represents the standard deviation of the fusion result, the comprehensive deviation evaluation index J is obtained;

[0018] When J exceeds the threshold J_max, it is determined to be a high deviation state and enters the model parameter correction step.

[0019] Preferably, the method further comprises the following steps:

[0020] Extract the core feature vector C_t of the real-time running status;

[0021] Define the model prediction core feature vector P_t;

[0022] Use the formula ΔE=||C_t-P_t||2 to measure the current state error;

[0023] If ΔE is greater than the given error bound E_bound, the deep learning optimization strategy is executed.

[0024] Preferably, further defined:

[0025] Collect N sets of multi-source synchronous data;

[0026] Establish a fusion matrix X∈R^(m×N) to standardize each set of data;

[0027] The maximum main characteristic direction contribution coefficient λ is determined by the calculation formula λ=max(eigen(X^T*X));

[0028] If the λ value is abnormal (λ<λ_min or λ>λ_max), the fusion logic path is replanned.

[0029] Preferably, the iterative correction model further includes:

[0030] Define the initial value of the parameter group θ_k to be adjusted;

[0031] Introduce the loss function L(θ_k), expressed as:

[0032] L=(y_{text{actual}}-y_{text{predict}})^2;

[0033] Apply the formula δL / δθ_k to derive and update the optimal value of θ_k;

[0034] The updated θ_k is used for the next moment prediction to reduce the overall prediction deviation.

[0035] Preferably, the following contents are also included when optimizing the accuracy:

[0036] Record the historical error set {e_t} and calculate the average value e_avg;

[0037] Set the compensation coefficient ρ to adjust the dynamic compensation intensity;

[0038] Calculate the real-time error compensation value δ_e = ρ*(e_t-e_avg);

[0039] If δ_e does not fall within the reasonable range (Δ, Δ), trigger an alarm or re-evaluate the model configuration.

[0040] Preferably, a finer-grained calibration operation is performed in the case of larger deviations:

[0041] Construct a short-term sliding window S(t) with a time span of τ;

[0042] Calculate the dynamic average value A_t=mean(S(t)) based on S(t);

[0043] Use the formula B(t)=|A_t-A(t1)| to calculate the change range index B(t) between adjacent time periods;

[0044] If B(t)>β_max (exceeds the tolerance limit β_max), the fast correction mechanism is activated.

[0045] Preferably, the following processing rules are also included to refine the utilization efficiency of the feedback data:

[0046] Group the monitoring data into G_i (i=1 to m) category intervals;

[0047] Estimate the weight factor w_i of each interval separately, where w_i is proportional to the importance of the data;

[0048] Calculate the overall normalized weight sum Σ_w = sum(w_i), and use the weight balance formula H(i) = √[Σ(G_j^2) / Σ_j] to recheck each group of influencing factors H(i);

[0049] Finally, based on H(i), it is judged whether the model deviation repair has completed the convergence standard H(i)≤ε.

[0050] Preferably, the following constraints are added when evaluating the validity of multi-source data:

[0051] Define the quality score Q(x)∈[0,1] for each data point and derive the basic reliability score F_q based on its confidence interval;

[0052] The quality summary score of all input data is weighted sum Z = Σ(Q(x_i)*α_i), where α_i represents the relative importance of a single input;

[0053] If Z>=threshold (preset quality boundary value) is satisfied, the input data is considered to participate in the validity determination; otherwise, it is eliminated and the input sequence is reconstructed.

[0054] Preferably, the following advanced features are added to further improve the model iteration effect:

[0055] The global performance index PSI and the corresponding parameter set are stored at the end of each calibration iteration;

[0056] Select a candidate reference model R based on the latest PSI value and use the formula Fitness = γ * (PSI_new / PSI_old) to determine its optimization potential (γ represents a scaling factor used to unify scalar level relationships).

[0057] If the candidate Fitness value continues to increase and reaches a stable period, some parameters are frozen to avoid overshoot;

[0058] Finally, a complete optimized trajectory record is generated to guide the formulation of automatic decision-making plans for similar tasks in the future.

[0059] Intelligent device management system based on digital twins, including:

[0060] Model building module, used to build the initial digital twin model based on the multi-source data fusion results;

[0061] A data acquisition and monitoring module is used to monitor the operating status of the initial model in real time and collect feedback data;

[0062] Model evaluation module, used to evaluate the degree of model deviation based on the difference analysis between feedback data and fusion results;

[0063] The model optimization module is used to iteratively correct the parameters of the digital twin model according to the degree of deviation to optimize the accuracy.

[0064] Beneficial effects of the present invention:

[0065] The present invention can solve the problem of how to iteratively correct the model accuracy based on the multi-source data fusion results to solve the problem of excessive deviation of the digital twin model. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 is a flow chart of the smart device management method based on digital twins according to an embodiment of the present invention;

[0067] Figure 2 It is a functional module structure diagram of the digital twin-based intelligent device management system described in an embodiment of the present invention. DETAILED DESCRIPTION

[0068] The embodiments of the present disclosure are described in detail below with reference to the accompanying drawings.

[0069] The following describes the embodiments of the present disclosure through specific examples, and those skilled in the art can easily understand other advantages and effects of the present disclosure from the contents disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present disclosure, rather than all of the embodiments. The present disclosure can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present disclosure. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present disclosure.

[0070] It should be noted that various aspects of the embodiments within the scope of the appended claims are described below. It should be apparent that the aspects described herein can be embodied in a wide variety of forms, and any specific structure and / or function described herein is merely illustrative. Based on this disclosure, it should be understood by those skilled in the art that an aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects described herein can be used to implement an apparatus and / or practice a method. In addition, other structures and / or functionalities other than one or more of the aspects described herein can be used to implement this apparatus and / or practice this method.

[0071] It should also be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present disclosure. The illustrations only show components related to the present disclosure and are not drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component can be changed at will, and the component layout type may also be more complicated.

[0072] Additionally, in the following description, specific details are provided to provide a thorough understanding of the examples. However, one skilled in the art will appreciate that the aspects described can be practiced without these specific details.

[0073] like Figure 1As shown, an embodiment of the present invention discloses a method for intelligent device management based on digital twins. This method includes the following four main steps: constructing an initial digital twin model, real-time monitoring and feedback data collection, assessing model deviations, and iteratively correcting deviations through analysis. Specifically, these steps work together to ensure that digital twin technology can provide high-precision dynamic simulation and decision-making support in the field of device management.

[0074] First, the present invention constructs an initial digital twin model by fusing multi-source data. In this process, it is necessary to integrate information streams from physical device sensor data (such as operating temperature, pressure, vibration intensity, etc.), historical operation and maintenance records, external environmental conditions, and other relevant parameters. In order to effectively convert multi-source information into a usable digital representation, in one embodiment, advanced algorithms such as deep learning or Bayesian networks are used to process and map conflicts or overlapping areas between different data formats to form a preliminary model architecture with predictive capabilities. Such a model design can accurately reflect the behavioral characteristics of actual equipment and its potential performance change trajectory.

[0075] After creating the initial model, the next step is to continuously monitor this virtual replica and regularly collect various on-site feedback data from the real world. This phase primarily utilizes advanced IoT perception layer technology and cloud computing platforms to automate large-scale data acquisition. For example, when an industrial robot is performing an assembly line task, all displacement sensors installed in its joints upload hundreds of values ​​per second to a central server for back-end calculations. Simultaneously, high-level business indicators such as order completion progress are obtained from the production process control system as auxiliary evidence to form a comprehensive status snapshot that further verifies and enriches the digital simulation.

[0076] Then, the third step, the difference analysis phase, is carried out to determine whether there are any anomalies between the data from the two sources mentioned above and to calculate the corresponding error range level. It is particularly emphasized here that for those major problems that may cause distortion of the final forecast, specific mathematical statistical theories such as root mean square error calculation, Z-score transformation or other more complex regression analysis methods must be used to analyze the details of the cause one by one. In an example, if the expected cooling effect of the air-conditioning system is set too high due to seasonal temperature fluctuations, but the actual power consumption is significantly lower than normal, potential improvement breakthroughs can be found by carefully examining the accuracy of the relevant weather forecast API access and the probability of internal heat dissipation pipe blockage.

[0077] The last step is also the key link - make corresponding adjustments to the core framework structure of the digital twin currently in use based on the various quantitative measurement standards obtained previously, thereby continuously improving the effectiveness of the overall solution. The core concept of this methodology lies in the cyclical and progressive optimization strategy. By introducing an adaptive mechanism, the system is allowed to explore the possibility of better parameter combinations on its own until the preset quality goals are achieved. For example, a thermal power plant hopes to extend the service life of the boiler while maintaining a high power output efficiency by deploying this management model. After several rounds of trial runs to adjust its steam pressure curve generation rules and fuel mixture ratio weight distribution, it has indeed observed ideal performance improvements. Therefore, we can see that the entire operation sequence has always been centered around how to make full use of existing resources to achieve efficient governance goals.

[0078] The present invention further includes four main steps based on the multi-source data fusion results: first, obtain the first data volume M1 of multiple data sources; second, calculate the initial weight W0; then use the formula to adjust the weight W; finally, determine whether the deviation is greater than the threshold θ, and if so, reallocate the data source weights.

[0079] The first step is to obtain the first amount of data, M1, from various data sources. These data sources may include sensors, environmental monitoring equipment, historical data records, and other multi-channel data sets. M1 refers to the total amount of data or the number of feature dimensions in each specific time slice.

[0080] The second step is to determine the initial weights W0 based on the importance of each data source or other relevant indicators. The weight initialization process can be set through prior knowledge or uniform distribution to ensure that the initial values ​​roughly reflect the importance distribution in the actual scenario.

[0081] The formula used in the third step, W = W0*exp(α*|F - F_model|), adjusts the weights based on the difference between the real-time data fusion result and the expected value of the estimated model. α is the weight adjustment parameter, which determines the magnitude of the change. The default range is from -2 to 2 to optimize balance. The optimal choice requires analysis based on the specific situation. For example, in a stable system, a value close to zero can be used to reduce drastic changes. This expression allows for dynamic weight updates, improving the quality of subsequent data fusion and avoiding the mismatch that can occur with static configurations.

[0082] In the fourth step, if the deviation exceeds a certain value, a re-matching process is automatically triggered: if the condition |F - F_model| > θ holds, the original weights are deemed no longer suitable for the actual situation and need to be revised. Setting a clear limit here is to prevent frequent changes from interfering with the normal operation of the entire system.

[0083] In one example, consider monitoring the operating temperature of a critical piece of industrial machinery. Three primary collection points are initially assigned equal weights: the outer shell, the bearings, and the hydraulic oil reservoir. Initially, these three are weighted equally, with W0 = 1 / 3. During operation, a leak in the cooling system is detected, causing an abnormal increase in the inner wall temperature. Using the aforementioned method, this deviation from the model's expected value exceeds a set threshold θ (e.g., ±5%). Therefore, the weights of the corresponding locations monitored are increased while the weights of the remaining two locations are appropriately reduced, resulting in more accurate and reliable decision-making.

[0084] Finally, after all adjustments are completed, the new ratio is integrated to generate information that is more in line with the current real-world status for use in smart device management.

[0085] The present invention's discrepancy analysis between feedback data and fusion results further includes four main steps: the first step is to calculate the average variance D_avg of the feedback data; the second step is to set the bias sensitivity coefficient k as an evaluation parameter; the third step is to calculate the comprehensive bias evaluation index J according to the formula J = k * log (D_avg / σ), where σ represents the standard deviation of the fusion result; and the fourth step is to determine the high bias state and decide whether to proceed to the model parameter correction step. The following details each step and its application.

[0086] The first step involves collecting feedback data and using statistical methods to calculate the average variance (D_avg) of this data. Feedback data is typically actual data collected by smart devices during operation, and the average variance measures the degree of fluctuation in this data. A higher D_avg indicates greater dispersion of data points, potentially reflecting instability in actual operating conditions.

[0087] The second step defines the deviation sensitivity coefficient k. This is a manually set parameter used to adjust the sensitivity of the deviation assessment to feedback fluctuations. Generally, the value of k ranges from 0 to 1, depending on the specific needs of device management. The optimal k value is typically determined through experimentation to balance sensitivity and false alarm rate. For example, k = 0.5 is a good reference value for many common scenarios.

[0088] The third step is to use the formula J = k*log(D_avg / σ) for calculation. In this formula, D_avg reflects the stability of the real-time data, while σ represents the standard error of the fusion result (usually the predicted or expected result). The formula uses logarithmic operations to quantify the relative gap between the two distribution characteristics. Due to the characteristics of the log function, when σ is large, it can avoid the problem of extreme deviations causing excessive changes in the J value. The purpose of setting this formula is to provide a standardized way to uniformly measure the overall deviation between feedback data and expected results.

[0089] The fourth step compares J with the threshold J_max. If the calculated J exceeds a predefined maximum value, the system is considered to be in a high-deviation state and a correction procedure must be initiated. In one example, suppose an industrial robot exhibits significant inconsistency in its operating speed. Continuous measurement reveals that D_avg reaches a certain value while σ remains relatively constant. If the calculated J > J_max, the next step is to re-optimize the sensor model and parameters. This dynamic mechanism effectively ensures the continued accuracy of intelligent management methods within the digital twin platform.

[0090] The present invention also includes the following steps: first, extracting the core feature vector C_t of the real-time operating state; second, defining the model prediction core feature vector P_t; then calculating the current state error through the formula ΔE=||C_t-P_t||2; finally, if the error ΔE exceeds a given error limit E_bound, executing a deep learning optimization strategy.

[0091] In the first step, extracting the core feature vector C_t of the real-time operating status means selecting the most representative part from the actual operating parameters of the smart device. For example, these parameters may include temperature, vibration frequency or energy consumption values. The dimension of the core feature vector C_t can be set to N dimensions according to the specific scenario, and the range is usually 1 to 50 to ensure that it is sensitive enough to key performance indicators without being overly complicated. The model prediction core feature vector P_t defined in the second step is a data point generated by the trained digital twin prediction model, which reflects the predicted value of the ideal state of the smart device at a certain moment in the future. The dimension of P_t is the same as C_t and depends on the quality of the training data and the depth of the neural network architecture.

[0092] In the third step, ΔE = ||C_t - P_t||2 represents the Euclidean distance measurement of the error between the actual state and the expected state. The core meaning of this formula is to numerically identify the deviation of the current state from the expected state, helping to determine whether the system has potential problems or needs adjustment. ΔE is a unitless scalar value typically set in the range [0,∞). Its optimal value is close to zero, indicating that the device state fully matches the model prediction. This formula was established because of its computational simplicity and sensitivity to deviations between variables.

[0093] When ΔE exceeds a preset threshold, E_bound (which can be calculated based on historical data fluctuations, such as three times the standard deviation), the deep learning optimization strategy execution phase begins. Specifically, in one embodiment, suppose the equipment management system used for wind farm blade inspection detects that ΔE suddenly exceeds the normal fluctuation range during a certain period of time. This triggers the optimization logic, inputs the abnormal situation into the pre-trained deep learning framework for retraining, and updates the equipment behavior model. Ultimately, the predictive model can more accurately capture equipment operating trends under new operating conditions, thereby improving the reliability and long-term applicability of the digital twin model.

[0094] The present invention is further defined as comprising four steps: acquiring synchronous data from multiple sources, establishing a fusion matrix to standardize the data, calculating the maximum principal feature direction contribution coefficient using a specific formula, and replanning the fusion logic path based on the coefficient value. Each step has a clear definition and scope.

[0095] The first step is to collect N sets of synchronized data from multiple sources. This means obtaining a set of time-consistent data from multiple sources simultaneously. The key here is synchronization, ensuring that data from different sources is collected at the same time. For example, in a digital twin-based intelligent device management approach, it may be necessary to collect data from multiple devices, such as sensors, environmental monitoring equipment, and operating status records, and ensure strict temporal consistency of the data.

[0096] Subsequently, a fusion matrix X∈R^(m×N) is established to standardize each set of data, where m is the number of features in a single set of data and N is the number of data samples. The standardization process can adjust data of different units and magnitudes to a unified range, facilitating subsequent processing. In one embodiment, if a sensor outputs voltage values ​​and temperature and humidity data for a smart device management system, by constructing a corresponding fusion matrix, both can be converted to a similar numerical range, such as the range [0,1], for unified comparison and analysis.

[0097] The calculation formula λ = max(eigen(X^T*X)) is defined to determine the contribution coefficient of the maximum main characteristic direction. Specifically, X^T*X calculates the covariance property matrix of the data, and eigen represents the process of obtaining the eigenvalue. By extracting the maximum eigenvalue of this matrix as the λ value, the most influential direction or data combination pattern in the system can be effectively reflected. In this scenario, the optimal parameter range must meet the accuracy and rationality standards in the actual application scenario. For example, the settings of λ_min and λ_max can be configured according to the statistical distribution of the historical data of the device to ensure that abnormal situations do not interfere with the smoothness of the overall process.

[0098] If the calculated λ value exceeds the set range (λ < λ_min or λ > λ_max), it indicates that the existing fusion mechanism has problems or cannot accurately reflect the actual state changes of the equipment. Replanning the fusion logic path is necessary. For example, if λ is significantly low while monitoring the operating status of a large piece of machinery, considering adding more real-time sensor feedback channels or modifying and upgrading the original algorithm to improve the reliability of the results should be considered.

[0099] The iterative correction model of the present invention further includes: defining the initial value of the parameter group θ_k to be adjusted, which means specifying an initial value for θ_k according to the design requirements of the model and the actual application scenario. The parameter usually includes a set of various variables associated with the device status, sensor data or other external inputs. In the intelligent device management method based on digital twins, θ_k may cover a variety of attribute values ​​such as temperature coefficient, vibration frequency, energy consumption ratio, etc. Its range is limited by physical laws or empirical values, and the initial value is usually set to an estimated value close to the actual situation or a zero value. When the loss function L(θ_k) is introduced, it is expressed as:

[0100] The core purpose of this formula, L = (y_{text{actual}} - y_{text{predict}})^2, is to assess the degree of deviation between model predictions and actual results by quantifying the error. In the formula, (y_{text{actual}}) is the actual measured output value of the device, while (y_{text{predict}}) is the theoretical predicted value generated by the model. Both are observable data types of the system. This form of squared loss is used because it converts the deviation into a continuous non-negative value, which is easy to calculate and intuitively reflects the room for accuracy improvement.

[0101] Applying the formula (delta L / deltatheta_k) to determine the optimal value of θ_k, this method minimizes the loss function by calculating the current parameter adjustment amplitude. In this step, the gradient information guides parameter changes in a direction that minimizes error, thereby achieving the optimization goal. For example, in one specific embodiment, when managing the power system performance of an industrial robot, if the initial prediction fails to fully reflect its energy consumption behavior, the gradient calculation mechanism can gradually approach the optimal parameter configuration that meets the actual situation.

[0102] The updated θ_k is then applied to the next-moment forecast to reduce overall forecast deviation. This cyclical operation helps continuously improve the accuracy and reliability of the system. In one example, considering the time efficiency of an intelligent robot performing a task on an assembly line in a factory automation environment, by iteratively revising the model based on an initial estimate of θ_k, the model not only more closely tracks actual operational performance but also reduces the accumulated deviation over the long term. This approach ultimately supports a more efficient and stable digital twin mapping process and management strategy development.

[0103] The present invention also includes the following steps when optimizing accuracy: During the accuracy optimization process, higher accuracy and stability are achieved through the following steps: The first step is to record a set of historical errors and calculate their mean; the second step is to set a compensation coefficient to adjust the dynamic compensation strength; the third step is to calculate a real-time error compensation value according to a formula; and the fourth step is to determine whether the compensation value is within a reasonable range. If not, an alarm is triggered or a reassessment is performed.

[0104] The first step is to compile statistics on all historical errors generated during equipment operation, generating a set of errors {e_t}, and then calculating the mean of this set, e_avg. Here, e_t represents the error value at different time points t, which can be understood as the degree of deviation between the model's predictions and actual operating data. The significance of this step is to use the long-term accumulated error information to establish a stable reference baseline.

[0105] The second step is to set the compensation coefficient ρ. This parameter controls the strength of the dynamic compensation mechanism and is typically a decimal greater than 0 and less than or equal to 1. When it is close to 0, the impact of dynamic adjustment is very weak; when it is close to 1, the impact of real-time errors is fully accounted for. The optimal value of ρ depends on the specific use case and analysis of error characteristics. The introduction of ρ ensures that the system can flexibly adjust its sensitivity to error fluctuations based on demand.

[0106] In the third step, the real-time error compensation value corresponding to the current moment is derived based on the formula δ_e = ρ*(e_t - e_avg). δ_e represents the error compensation amount. The above formula focuses on the portion of the error value that varies around the average error, and combines this with the adjustment factor ρ to determine the direction and magnitude of the compensation. This helps mitigate the impact of extreme errors, thereby improving overall model accuracy. The logic behind this formula is to find an optimal path that both smooths past performance and quickly adapts to future deviations.

[0107] Finally, in one embodiment, the error range needs to be determined. If δ_e is not within the reasonable bounds [Δ, Δ], appropriate action should be taken, such as triggering an alarm or re-examining the configuration. For example, when monitoring the movement of an industrial robot using digital twin technology, if the real-time position deviation exceeds a predefined safety threshold, engineers will be prompted to verify the physical hardware or correct the initial conditions of the virtual simulation. This ensures the continued accuracy and safety of the system.

[0108] The present invention performs a more fine-grained calibration operation in the case of large deviations:

[0109] First, list the steps as follows:

[0110] 1. Construct a short-term sliding window S(t) with a time span of τ.

[0111] 2. Calculate the dynamic average value A_t=mean(S(t)) based on S(t).

[0112] 3. Use the formula B(t)=|A_t-A(t1)| to calculate the change amplitude index B(t) between adjacent time periods.

[0113] 4. If B(t)>β_max (exceeds the tolerance limit β_max), the fast correction mechanism is activated.

[0114] Then, describe the specific meaning of each step one by one:

[0115] In the first step, a short-term sliding window S(t) is constructed, integrating historical data from a specific time period into a dynamic interval. The time span τ of the sliding window is set to reflect fluctuation trends within a relatively short period while avoiding excessive noise interference. The time span τ typically ranges from seconds to minutes, for example, from 1 minute to 10 minutes. The optimal value depends on the data collection frequency and stability requirements of the target device in the specific scenario. The sliding window provides a more accurate quantitative basis for local data features.

[0116] In the second step, a dynamic average value, A_t = mean(S(t)), is calculated based on S(t). A_t here represents a concentrated summary of the real-time data distribution within the window, reflecting the central position of the state within the window. This average calculation captures the trend characteristics of device behavior based on smoothed data, rather than relying solely on specific values ​​at a specific moment. This average calculation helps mitigate the misleading effects of individual extreme data values.

[0117] In the third step, the formula B(t) = |A_t - A(t1)| is used to calculate the amplitude of change indicator B(t) between adjacent time periods. The parameter A(t1) is the dynamic average calculated for the previous window. The degree of state transition is characterized by comparing the absolute difference between the current and previous dynamic averages. The purpose of this formula is to detect whether there are significant periodic changes. A higher B(t) indicates an increased risk of potential error, making it a key factor in determining whether to initiate additional adjustment strategies. Generally speaking, the sensitivity of the amplitude of change indicator is determined by the maximum acceptable deviation range in the application environment. For example, a reasonable range of β_max can be pre-calibrated experimentally to [0.1, 1.0]. The optimal choice is related to the actual stability tolerance of the device.

[0118] In step 4, if the result of B(t) exceeds the preset threshold β_max, the system proceeds to step 4 to activate the rapid correction mechanism. This action ensures a prompt response when actual performance deviates and approaches loss of control, while minimizing the waste of resources caused by false alarms. The core role of this step is to improve overall system responsiveness through detailed management, especially ensuring precise control under high load or multi-variable coupling conditions.

[0119] In one embodiment, assume that a key sensor for temperature control on a factory production line is equipped with a digital twin-based intelligent device management method. The method processes a set of temperature readings recorded by the sensor as the basic data stream. Assuming that new measurement data is obtained every second, it is found in a specific analysis phase that the current short-term sliding window is 5 minutes long (τ=5). After a series of calculations, the dynamic average value A_t of this period is close to the target set value. Combined with the previous historical data evaluation B(t)=0.7, which is higher than the predetermined limit β_max=0.6, the hot backup or remote resynchronization logic is immediately triggered to restore to the optimal operating state. This example fully demonstrates the importance of the above four key steps in the interconnected operation process for realizing an efficient closed-loop monitoring system.

[0120] The present invention also includes the following processing rules to refine the utilization efficiency of feedback data: first, list the steps in order, then explain the specific meaning of each step, and finally give a relevant example for further explanation.

[0121] The first step is to group the monitoring data into (G_i) ((i=1) to (m)) category intervals. The significance of this step is to classify the feedback data according to its characteristics and source, forming multiple groups with inherent logic, such as classification according to dimensions such as temperature, humidity, vibration frequency or pressure. These groups can provide a basic framework for subsequent refined calculations. The value of the number of groups (m) is determined by the actual operating requirements of the specific equipment, usually ranging from 5 to 15 groups, so that the characteristics of each category can be studied and processed separately.

[0122] The second step is to estimate the weight factor (w_i) for each interval, where (w_i) is proportional to the importance of the data type. Estimating weights relies on prior knowledge or expert judgment. For example, in the management of a smart device, temperature may be more important than voltage fluctuations, so the weight (w_i) for the corresponding temperature interval should be set larger. The weight range is generally ([0,1]), and it is necessary to ensure that (sum(w_i) ≤ 1) to maintain normalization consistency. The purpose of this setting is to quantify the importance of each category and thus rationally allocate computing resources.

[0123] The third step is to calculate the overall normalized weight sum (Σ_w = sum(w_i)) and recheck each group of influencing factors (H(i)) using the weight balancing formula (H(i) = √[Σ(G_j^2) / Σ_j]). Here, (Σ(G_j^2)) represents the sum of the squares of each data set, reflecting the intensity of fluctuations, while the denominator (Σ_j) is the sum of the elements within the group for averaging. This formula comprehensively corrects the contribution of each type of feedback and introduces a quadratic term to increase attention to significant outliers, ensuring a more accurate and reliable evaluation system. The ideal convergence state in the formula requires that (H(i)) is sufficiently small.

[0124] The fourth step ultimately determines whether the model deviation correction has met the convergence criteria (H(i) ≤ ε) based on (H(i)). The core of this process is to compare the actual calculated impact factor with a given threshold ε. ε, as a tolerance, defines the maximum acceptable error for the system. In the context of digital twin management, it may be set to (ε = 0.001) or lower to ensure highly stable results. This standard test confirms the effectiveness of the optimized and adjusted intelligent management strategy.

[0125] In one specific example, a digital twin is used to monitor the status of an industrial robot in a factory environment. If the sensor data contains multiple variables, such as joint temperature, rotational speed, and load weight, it is first classified into three main categories (G_1) (joint temperature), (G_2) (rotational speed), and (G_3) (load) using the aforementioned method. Then, the highest estimate (w_1) is used, followed by the others. The aforementioned complex function is then used to iterate until all (H(i)) values ​​are less than the target value ε, indicating that the robot control algorithm is becoming more sophisticated and adapting to the new working conditions.

[0126] The present invention adds the following constraint rules when evaluating the validity of multi-source data:

[0127] First, this process involves three main steps. The first step is to define the quality score Q(x)∈[0,1] for each data point; the second step is to calculate the quality summary score of all input data using the weighted comprehensive formula Z=Σ(Q(x_i)*α_i), and determine whether to participate in the judgment based on the preset boundary value; the third step is to eliminate data that does not meet the conditions and reconstruct the input sequence to optimize the multi-source data analysis process.

[0128] For the first step, Q(x) is a measure of the validity of a single data point, ranging from [0,1], where higher values ​​indicate more reliable data points, and vice versa. This is a quality score generated by performing statistical analysis on the data points and evaluating the confidence interval to obtain the basic reliability score F_q, which is used to ensure accurate judgment of each independent data. This parameter setting helps to distinguish which data may be affected by noise or errors, thereby improving the overall accuracy of subsequent analysis. For example, in a digital twin-based device management scenario, if the data collected by the sensor has a significant offset, the corresponding quality score will drop significantly.

[0129] Then, in the formula Z=Σ(Q(x_i)*α_i) in the second step, x_i represents a single piece of input data, and α_i is the relative importance of the corresponding data in the overall task. It is a weight parameter pre-configured according to the actual task needs, usually in the range of [0,1]. The specific optimal value depends on the distribution of each task requirement. The core goal of this weighted sum is to combine the importance of data from different sources with their own effectiveness to form a unified evaluation system. At the same time, this result is compared with the threshold threshold (the setting of the threshold needs to refer to the historical data experience value and the system stability requirements). Only data above the threshold will be recognized as available to enter the next step of processing. This setting makes the entire process flexible and highly targeted, and can adjust the threshold size and weight value distribution strategy according to different application scenarios.

[0130] In the third step, if one or more sets of data fail to meet the aforementioned conditions (i.e., Z>=threshold does not hold), these data are deleted and the remaining data sets that meet the requirements are rebuilt as new input sequences to continue the calculation and verification operations. This is done to minimize the interference effect caused by invalid data and make the final conclusion more accurate and scientific. In one embodiment, if three of the indicators for monitoring the operating status of several smart devices in an industrial environment are a set of values ​​such as vibration frequency, energy consumption change trend, and operating temperature fluctuation amplitude record, it is found that some sensors have feedback abnormalities that deviate from the normal value range due to failures. The above mechanism can quickly screen and retain truly reliable signal data for key decision-making links such as predicting the probability of failure. Specifically, for the monitoring of a key machine tool, if the sensor feedback within a certain specific time window is affected by a large amount of noise components due to short-term power jitter, the above method can be applied to ensure that the subsequent state update of the digital mapping model of this machine still stably and efficiently reflects the actual working condition information.

[0131] This invention adds the following advanced features to further enhance the model iteration effect:

[0132] At the end of each calibration iteration, the global performance index (PSI) and the corresponding parameter set are stored. The core purpose of this operation is to establish a historical data accumulation system for the iterative optimization process. By recording the PSI value and its associated parameters generated after each iteration, it lays the foundation for subsequent analysis and model tuning. The parameter set in this step can include the device's state variables, control parameters, and their adjustment coefficients.

[0133] Next, candidate reference models R are selected based on the latest PSI value, and their optimization potential is determined using the formula: Fitness = γ * (PSI_new / PSI_old). In this formula, γ represents a scaling factor, typically ranging from [0.5 to 2]. The preferred value is adjusted based on the actual scenario, and a default value of 1 is recommended. The structure of the formula embodies the principle of evaluating the relative degree of performance improvement, calculating Fitness by comparing the PSI values ​​of the current cycle with those of the previous cycle. Because direct numerical values ​​may have absolute magnitude differences, the inclusion of γ ensures greater comparability and universality in the comparison process.

[0134] If the candidate model's Fitness value continues to increase over multiple consecutive evaluations and reaches a stable state, the system triggers the freezing of some overly sensitive or saturated parameters to prevent overfitting or overshoot. This logic aims to protect areas close to the optimal solution, preventing the algorithm from wasting computing resources or straying from the target path due to unnecessary fluctuations.

[0135] Finally, the entire optimization process data trace is fully integrated and exported to a historical archive, which serves as an important reference for future solutions to similar problems. This step not only speeds up the solution of similar tasks but also helps improve the model's generalization capabilities.

[0136] In one embodiment, the above characteristics are applied to an industrial cooling equipment management platform based on digital twins. Specifically, in a case of operating status adjustment of a cooling tower, real-time data such as temperature and flow are collected through the digital twin system and dynamically fed back to the model training mechanism. Each time an iteration is completed, the corresponding PSI and related adjustment strategy set are recorded. Assuming that γ=1 is determined after preliminary testing, and then a set of candidate Fitness values ​​are found to gradually increase and have no significant changes for three consecutive rounds, several non-core but key variable parameters that highly affect stability factors (such as pumping frequency) are automatically locked to maintain the system in a more ideal equilibrium state until the end of the project cycle. At the same time, all intermediate processes are carefully archived to facilitate direct reference to these mature experience practices when reviewing and optimizing the configuration of other related projects.

[0137] The digital twin-based intelligent device management method of the present invention includes: constructing an initial digital twin model based on the results of multi-source data fusion. In this step, data from different sources (such as sensor data, system operation logs, and environmental parameters) are comprehensively analyzed and integrated and converted into digital representations that can reflect the operating characteristics of real physical devices, thereby constructing an initial digital twin model with strong biomimetic and mapping capabilities. Subsequently, the method further implements real-time monitoring of the operating status of the initial model and feedback data collection, that is, the monitoring system continuously obtains new dynamic data from the device end. At this stage, the collected real-time data can provide a comparison basis for the subsequent process.

[0138] In order to solve the problem of excessive deviation of the digital twin model by iteratively correcting the model accuracy based on the results of multi-source data fusion, the invention proposes to conduct a detailed analysis based on the difference between the feedback data and the original fused data, and evaluate the degree of deviation of the digital twin model. This degree of deviation reflects the deviation between the current model and the actual operating conditions. If the deviation exceeds the set threshold range, it means that the existing model can no longer accurately reproduce the behavior or characteristics of the device, and the key parameters of the digital twin model need to be adjusted. This adjustment process involves complex algorithm processing, such as using advanced methods such as machine learning to identify and quantify which parameters cause large errors, and then design targeted correction strategies to improve model accuracy. This method ensures continuous optimization and adaptive adjustment of the model through a closed-loop mechanism, so that the final model always accurately fits the state of the real-world smart device, significantly improving overall efficiency and prediction accuracy.

[0139] like Figure 2As shown, the present invention also discloses an intelligent device management system based on digital twins. The above-mentioned functional modules of the intelligent device management system based on digital twins described in the embodiments of the present invention respectively correspond to the various operating steps of the intelligent device management method based on digital twins of the present invention, which will not be repeated here.

[0140] The specific implementation methods described above further illustrate the purpose, technical solutions and beneficial effects of the embodiments of the present disclosure. It should be understood that the above description is only a specific implementation method of the embodiments of the present disclosure and is not intended to limit the scope of protection of the embodiments of the present disclosure. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the embodiments of the present disclosure should be included in the scope of protection of the embodiments of the present disclosure.

Claims

1. The intelligent device management method based on digital twins is characterized by: include: Build an initial digital twin model based on the fusion results of multi-source data; Performing real-time monitoring on the operating status of the initial model and collecting feedback data; Evaluate the degree of model deviation based on the difference analysis between feedback data and fusion results; The digital twin model parameters are iteratively corrected according to the degree of deviation to optimize accuracy.

2. The method for managing intelligent devices based on digital twins according to claim 1, characterized in that: The multi-source data fusion result further includes: Obtaining a first amount of data M1 from multiple data sources; Calculate the initial weight W0; Adjust the weight according to the following formula: W = W0*exp(α*|F-F_model|), where F represents the real-time data fusion eigenvalue, F_model represents the model expected eigenvalue, and α is the weight adjustment parameter; If the deviation is greater than the preset threshold θ, the weights of the data sources are redistributed, where the deviation is |F-F_model|.

3. The method for managing intelligent devices based on digital twins according to claim 2, characterized in that: The difference analysis based on the feedback data and the fusion result further includes: The average variance D_avg of the statistical feedback data; Set the deviation sensitivity coefficient k as the evaluation parameter; According to the formula J = k*log(D_avg / σ), where σ represents the standard deviation of the fusion result, the comprehensive deviation evaluation index J is obtained; When J exceeds the threshold J_max, it is determined to be a high deviation state and enters the model parameter correction step.

4. The method for managing intelligent devices based on digital twins according to claim 3, characterized in that: The following steps are also included: Extract the core feature vector C_t of the real-time running status; Define the model prediction core feature vector P_t; Use the formula ΔE=||C_t-P_t||2 to measure the current state error; If ΔE is greater than the given error bound E_bound, the deep learning optimization strategy is executed.

5. The method for managing intelligent devices based on digital twins according to claim 4, characterized in that: Further qualification: Collect N sets of multi-source synchronous data; Establish a fusion matrix X∈R^(m×N) to standardize each set of data; The maximum main characteristic direction contribution coefficient λ is determined by the calculation formula λ=max(eigen(X^T*X)); If the λ value is abnormal (λ<λ_min or λ>λ_max), the fusion logic path is replanned.

6. The method for managing intelligent devices based on digital twins according to claim 5, characterized in that: The iterative correction model further includes: Define the initial value of the parameter group θ_k to be adjusted; Introduce the loss function L(θ_k), expressed as: L=(y_{text{actual}}-y_{text{predict}})^2; Apply the formula δL / δθ_k to derive and update the optimal value of θ_k; The updated θ_k is used for the next moment prediction to reduce the overall prediction deviation.

7. The method for managing smart devices based on digital twins according to claim 6, characterized in that: Optimizing for accuracy also includes the following: Record the historical error set {e_t} and calculate the average value e_avg; Set the compensation coefficient ρ to adjust the dynamic compensation intensity; Calculate the real-time error compensation value δ_e = ρ*(e_t-e_avg); If δ_e does not fall within the reasonable range (Δ, Δ), trigger an alarm or re-evaluate the model configuration.

8. The method for managing intelligent devices based on digital twins according to claim 7, characterized in that: Perform finer-grained calibration in cases of larger deviations: Construct a short-term sliding window S(t) with a time span of τ; Calculate the dynamic average value A_t=mean(S(t)) based on S(t); Use the formula B(t)=|A_t-A(t1)| to calculate the change range index B(t) between adjacent time periods; If B(t)>β_max (exceeds the tolerance limit β_max), the fast correction mechanism is activated.

9. The method for managing intelligent devices based on digital twins according to claim 8, characterized in that: The following processing rules are also included to refine the efficiency of feedback data utilization: Group the monitoring data into G_i (i=1 to m) category intervals; Estimate the weight factor w_i of each interval separately, where w_i is proportional to the importance of the data; Calculate the overall normalized weight sum Σ_w = sum(w_i), and use the weight balance formula H(i) = √[Σ(G_j^2) / Σ_j] to recheck each group of influencing factors H(i); Finally, based on H(i), it is judged whether the model deviation repair has completed the convergence standard H(i)≤ε.

10. The intelligent device management system based on digital twins is characterized by: include: Model building module, used to build the initial digital twin model based on the multi-source data fusion results; A data acquisition and monitoring module is used to monitor the operating status of the initial model in real time and collect feedback data; Model evaluation module, used to evaluate the degree of model deviation based on the difference analysis between feedback data and fusion results; The model optimization module is used to iteratively correct the parameters of the digital twin model according to the degree of deviation to optimize the accuracy.

Images