Intelligent electric appliance cabinet full life cycle management system based on digital twinning

Abstract

The invention belongs to the technical field of data processing, and particularly relates to an intelligent electric appliance cabinet full life cycle management system based on digital twinning, which comprises a pre-training and parameter division module, an online performance monitoring trigger module and a dynamic parameter adjustment synchronization module, the method comprises the following steps: pre-training based on historical data to obtain a basic model, constructing a comprehensive evaluation index by calculating an average gradient amplitude and gradient direction information entropy of each parameter, and dividing a steady-state parameter set and a dynamic parameter set according to the comprehensive evaluation index; calculating an absolute error value at each moment by acquiring a real-time predicted value and a real value, constructing an absolute error sequence for calculating a performance degradation index, and generating an update trigger signal of a dynamic parameter set by comparing the performance degradation index with a threshold value; in response to the update trigger signal, a total loss function is constructed and only the dynamic parameter set is iteratively updated. According to the invention, the reliability of intelligent electric appliance cabinet full-life-cycle management is improved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology. More specifically, this invention relates to a full lifecycle management system for intelligent appliance cabinets based on digital twins. Background Technology

[0002] With the in-depth development of the Industrial Internet, digital twin technology provides core support for realizing the digital management of the entire life cycle of intelligent switchgear from design and manufacturing to operation and maintenance by constructing a real-time mapping between physical equipment and virtual models. Among existing technologies, long short-term memory networks, with their unique gating mechanism and cell state, can effectively learn long-term dependencies in historical operating data, thereby achieving accurate prediction of equipment degradation trends. They have been applied to fault prediction and health management of intelligent electrical cabinets.

[0003] However, in the real-time synchronization and dynamic evolution scenarios required by digital twins, traditional long short-term memory networks have fundamental limitations: First, their network parameters are fixed after pre-training and cannot be adjusted online, resulting in lag and inaccurate predictions when the virtual model's operating conditions change; second, their cell state memory update mechanism lacks the ability to dynamically assess the importance of information and cannot autonomously decide what to remember and what to forget in non-stationary data streams; this defect causes the system to retain outdated historical memories, seriously interfering with the accurate judgment of the current state, and making it difficult to quickly learn new operating modes and fault characteristics.

[0004] The aforementioned problems make it difficult for traditional long short-term memory networks to achieve intelligent memory management, ultimately leading to a disconnect between the virtual model and the dynamic evolution of the appliance cabinet. This severely restricts the continuous cognitive synchronization capability and decision reliability that the digital twin system should have in the whole life cycle management, and ultimately significantly reduces the accuracy of the digital twin system's assessment of the health status of the smart appliance cabinet. Summary of the Invention

[0005] To address the technical problems of reduced accuracy in assessing the health status of smart appliance cabinets in digital twin scenarios due to the fixed network parameters that cannot be adjusted online and the lack of dynamic assessment of the importance of cell states, this invention provides a digital twin-based full lifecycle management system for smart appliance cabinets. The system includes the following modules: a pre-training and parameter partitioning module, used to pre-train a basic model predicting the remaining lifespan based on historical operating data of the smart appliance cabinet; constructing a second validation set and calculating the average gradient magnitude and gradient direction information entropy of each parameter in the basic model; and constructing a comprehensive evaluation index for each parameter using a weighted geometric average, used to partition the parameters of the basic model into a steady-state parameter set and a dynamic parameter set; and an online performance monitoring triggering module. The system is used to obtain the predicted and actual values ​​of the remaining lifespan of the basic model at each time point; it forms an absolute error sequence by combining the absolute errors of the predicted and actual values ​​at all time points in the sliding window, and calculates the model performance degradation index of the absolute error sequence using an exponentially weighted moving average method, compares it with a preset performance degradation threshold, and generates an update trigger signal for the dynamic parameter set based on the comparison result; the dynamic parameter adjustment synchronization module is used to construct a total loss function using recent running data as training samples in response to the update trigger signal, the total loss function including a prediction error loss term and a logarithmic constraint term on the relative update magnitude of the dynamic parameters; the optimizer is used to iteratively update only the dynamic parameter set to obtain the updated dynamic parameter set, thereby realizing the full life cycle management of the smart appliance cabinet.

[0006] This invention utilizes a pre-training and parameter partitioning module to construct a comprehensive evaluation index using average gradient magnitude and gradient direction information entropy. This scientifically divides model parameters into steady-state and dynamic parameter sets, structurally decoupling core principle solidification from dynamic adaptability, laying the foundation for stable and efficient online learning. Through an online performance monitoring and triggering module, the system calculates the model performance degradation index based on the absolute error value sequence within a sliding window and automatically generates an update trigger signal when it exceeds a performance degradation threshold. This achieves sensitive, automatic monitoring and precise triggering of synchronization deviations in the digital twin virtual model. Through a dynamic parameter adjustment and synchronization module, the system constructs a total loss function with logarithmic constraints during updates and iteratively updates only the dynamic parameter set. This method rapidly reduces prediction errors while effectively constraining the relative update magnitude of parameters, ensuring the stability of the learning process and avoiding model oscillations. The synergy of these three modules forms a closed-loop management system encompassing parameter partitioning, monitoring and triggering, and stable updates. This significantly improves the long-term accuracy, environmental robustness, and automation level of health status prediction, achieving efficient and reliable management of the entire lifecycle of intelligent electrical cabinets.

[0007] Preferably, the basic model for obtaining the predicted remaining service life includes: the input data of the basic model is multi-source time-series sensor data collected during the historical operation of the smart appliance cabinet in the past month. The multi-source time-series sensor data is stored in sequence with timestamps as indexes, specifically including current, voltage, temperature, and vibration signals of the core parts. The output data is the remaining service life of the smart appliance cabinet in a future preset time window.

[0008] Preferably, constructing the second verification set includes: selecting a portion of samples from historical normal operating condition data as a base, and generating a series of adversarial samples as the second verification set by injecting noise that conforms to actual physical laws to simulate sensor drift.

[0009] Preferably, the calculation of the average gradient magnitude of each parameter in the basic model includes: for each sample in the second validation set ,in, This refers to the sample index in the second validation set; For sensor data sequences, For the corresponding true remaining service life label; sample the sensor data sequence. Input the base model to obtain the predicted value. And calculate the mean squared error loss value. ; The sample is calculated using the backpropagation algorithm. Mean squared error loss value Compared to the first Calculate the gradient magnitude of each parameter; calculate the mean squared error loss of all samples in the second validation set relative to the first parameter. The average gradient magnitude of the nth parameter is used as the 1st parameter. The average gradient magnitude of each parameter.

[0010] This method calculates the mean squared error loss based on the predicted values ​​and actual remaining useful life labels of the second validation set samples. It then uses the backpropagation algorithm to solve for the gradient magnitude of the loss of a single sample with respect to each parameter, and finally takes the average of all sample gradient magnitudes. This allows for precise quantification of the average impact of each parameter on the model's predicted loss. This approach avoids the randomness interference of single-sample gradients, ensuring the stability and reliability of gradient magnitude evaluation, and providing an objective and quantifiable core basis for subsequent parameter classification and selection.

[0011] Preferably, the comprehensive evaluation index for constructing each parameter includes: In the formula, It is the first A comprehensive evaluation index for each parameter; It is the first The average gradient magnitude of each parameter; It is the first The gradient direction information entropy of each parameter; It is an adjustment weighting coefficient; It is a very small positive number.

[0012] This method constructs a comprehensive evaluation index that integrates the average gradient magnitude and gradient direction entropy, achieving a dual-dimensional consideration of parameter characteristics: the average gradient magnitude reflects the sensitivity of the parameter to the model loss, while the gradient direction entropy reflects the stability of the parameter gradient change; the introduction of an adjustable weight coefficient allows for flexible adjustment of the importance ratio of the two dimensions, and the extremely small positive number avoids the abnormal situation of zero or infinity during the calculation process; this comprehensive evaluation index breaks through the limitations of single-dimensional evaluation, and can more comprehensively and accurately characterize the dynamic characteristics of parameters in the model, laying the foundation for the scientific classification of parameters.

[0013] Preferably, the step of dividing the parameters of the basic model into a steady-state parameter set and a dynamic parameter set includes: sorting the comprehensive evaluation index of all parameters in ascending order to obtain a sorting sequence; and using the third and fourth quantiles of the sorting sequence as quantile thresholds to perform parameter division: parameters whose comprehensive evaluation index is less than or equal to the quantile threshold are assigned to the steady-state parameter set; while parameters whose comprehensive evaluation index is greater than the quantile threshold are assigned to the dynamic parameter set.

[0014] This method uses the third and fourth quantiles as quantile thresholds to divide the ranking sequence of comprehensive evaluation indicators, enabling objective classification of parameters based on statistical laws and avoiding biases caused by subjective threshold settings. The steady-state parameter set contains parameters that have a relatively stable impact on model performance and low sensitivity, which can be kept relatively fixed in subsequent training. The dynamic parameter set contains parameters that have a significant impact on model performance and fluctuate greatly, requiring key adjustments during training. This classification method enables differentiated management of parameters, improving the targeting and efficiency of model training.

[0015] Preferably, the step of calculating the model performance degradation index of the absolute error sequence using the exponentially weighted moving average method includes: ;middle, It is the first The model performance degradation index at time step; It is the first The absolute error value at time , where For the backtracking step size; It is the length of the sliding window; To represent an increasing weighting factor, the following conditions must be met: .

[0016] This method sets an increasing weighting factor to give higher weight to the absolute error value at recent times within the sliding window. At the same time, the sliding window limits the calculation range, which can accurately capture the recent degradation trend of model performance. Compared with the simple arithmetic mean method, this exponentially weighted moving average method highlights the dominant role of the latest error data in performance evaluation, and filters out the interference of long-term errors through the window length. It realizes real-time, sensitive and stable monitoring of the model performance degradation state, and provides timely decision-making basis for the dynamic adjustment of subsequent model parameters.

[0017] Preferably, constructing the total loss function includes: In the formula, It is the total loss function; It is the standard prediction error loss; It is the constraint strength coefficient; It is a dynamic parameter set; It is a single dynamic parameter in the dynamic parameter set; It is a dynamic parameter The amount of updates during the adjustment process; It is the scaling factor; It is a very small positive number; It is the absolute value symbol; It is a logarithmic function.

[0018] This method introduces a constraint term for the dynamic parameter set on top of the standard prediction error loss, which enables fine-grained control over the amount of dynamic parameter updates while ensuring the model's prediction accuracy. The constraint term penalizes the relative change ratio of parameter updates through a logarithmic function and a scaling factor, avoiding excessive fluctuations in dynamic parameters during adjustment and preventing the model from falling into an unstable state or overfitting. The constraint strength coefficient can flexibly balance the weights of prediction loss and parameter constraints, and a very small positive number avoids the calculation anomaly of zero denominator, ultimately achieving a balance between improving model performance and parameter stability.

[0019] Preferably, the optimizer iteratively updates only the dynamic parameter set, including: in the iterative optimization, only the gradients of all dynamic parameters in the dynamic parameter set are calculated and updated; steady-state parameters in the steady-state parameter set are not calculated and do not participate in the update; when the total loss function When the maximum number of iterations is reached, the adjustment process terminates; finally, the system outputs the updated dynamic parameter set.

[0020] Preferably, the method for achieving full lifecycle management of the intelligent electrical cabinet includes: the digital twin virtual model continues to operate online by combining a locked steady-state parameter set with an updated dynamic parameter set; during the operation phase, the digital twin virtual model continuously maps key operating parameters such as current, voltage, and temperature of the intelligent electrical cabinet and its remaining service life, providing maintenance personnel with 24-hour uninterrupted visual monitoring, covering the status perception needs of the operation phase; during the maintenance phase, when the remaining service life predicted by the digital twin virtual model shows an abnormal trend, a fault warning will be automatically triggered; finally, based on the remaining service life assessment results, a maintenance cycle and maintenance content adapted to the intelligent electrical cabinet are automatically generated.

[0021] The beneficial effects of this invention are as follows: the collaborative operation of the pre-training and parameter partitioning module, the online performance monitoring triggering module, and the dynamic parameter adjustment synchronization module realizes a full-chain technological innovation from model preprocessing and online monitoring to adaptive updates, bringing significant beneficial effects; the pre-training and parameter partitioning module innovatively introduces a second validation set containing slightly adversarial samples, calculates the average gradient magnitude and gradient direction information entropy of each parameter, and uses weighted geometric average to construct a comprehensive evaluation index, thereby scientifically dividing the parameters of the basic model into a steady-state parameter set and a dynamic parameter set; this process essentially solidifies the learned stable physical laws of the intelligent appliance cabinet into the steady-state parameter set, while endowing the dynamic parameter set with the ability to adapt to dynamic environmental changes, achieving decoupling of stability and adaptability from the source of model structure, laying a core foundation for subsequent safe and efficient online learning; the online performance monitoring triggering module constructs a sensitive synchronization deviation quantification mechanism, which obtains reliable true values ​​of health status through multi-source cross-validation, calculates the absolute error value, and then uses a sliding window and exponentially weighted moving average method to dynamically calculate the model performance degradation index. This index, by assigning higher weight to recent errors, can keenly capture early performance degradation trends. By comparing with a preset performance degradation threshold, it automatically and accurately generates update trigger signals for the dynamic parameter set, achieving intelligent control from performance degradation perception to maintenance decision triggering, overcoming the lag of traditional manual threshold setting. The dynamic parameter adjustment synchronization module is crucial for ensuring the long-term stable operation of the system. Upon responding to the update signal, it constructs a total loss function including a prediction error loss term and a logarithmic constraint term, and iteratively optimizes only the dynamic parameter set. The innovative logarithmic constraint term, by smoothing the relative update magnitude, ensures the smoothness of the parameter update process, effectively preventing model oscillations caused by parameter mutations. While quickly adapting to new operating conditions, it strictly protects the core knowledge maintained by the steady-state parameter set, eliminating catastrophic forgetting, and thus significantly improving the prediction accuracy and long-term operational reliability of the digital twin model in the full lifecycle management of intelligent electrical cabinets. Attached Figure Description

[0022] Figure 1 This is a schematic diagram illustrating a system block diagram of a digital twin-based intelligent appliance cabinet lifecycle management system according to the present invention; Figure 2 This is a schematic diagram illustrating the comparison of prediction effects before and after updating the LSTM model of a digital twin-based intelligent appliance cabinet lifecycle management system in this invention. Figure 3 This diagram schematically illustrates a comparison of the prediction performance of the updated LSTM method and the traditional method in a digital twin-based intelligent appliance cabinet lifecycle management system according to the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0025] This invention discloses a full lifecycle management system for intelligent appliance cabinets based on digital twins, referring to... Figure 1 It includes a pre-training and parameter partitioning module 100 and a dynamic parameter adjustment synchronization module 300.

[0026] The pre-training and parameter set partitioning module 100 is used to pre-train based on the historical operating data of the smart appliance cabinet to obtain a basic model for predicting the remaining service life; construct a second validation set and calculate the average gradient magnitude and gradient direction information entropy of each parameter in the basic model, and construct a comprehensive evaluation index for each parameter through weighted geometric average, which is used to partition the parameters of the basic model into a steady-state parameter set and a dynamic parameter set.

[0027] It should be noted that this step is the basic preprocessing stage, and its core logic is as follows: the gradient magnitude of a parameter reflects its influence on the model output; the larger the average gradient magnitude, the more sensitive the model prediction is to changes in that parameter; adversarial samples are introduced in this module to simulate possible interference noise in the online environment, so that the parameter sensitivity assessment is closer to the real scenario, rather than being limited to the performance under stable conditions, and provides a more generalizable data foundation for subsequent fine division of parameters.

[0028] Specifically, a traditional Long Short-Term Memory (LSTM) network is pre-trained based on the historical operating data of the smart appliance cabinet over the past month. The network is divided into an 80% training set and a 20% first validation set, according to a preset ratio of 8:2, to complete model training and hyperparameter tuning. This yields a basic model capable of predicting the remaining lifespan of the smart appliance cabinet. The input data for this basic model consists of multi-source time-series sensor data collected at a sampling frequency of 1 Hz during the smart appliance cabinet's historical operation over the past month. This multi-source time-series sensor data is stored sequentially, indexed by timestamps, and includes data such as current, voltage, temperature, and vibration signals from core components. The output data is the remaining lifespan of the smart appliance cabinet within a preset future time window, expressed in hours. For example, the preset future time window can be set to 24 hours. The basic model preferentially uses mean squared error as its loss function to minimize the difference between the predicted and actual values.

[0029] Furthermore, to evaluate the performance and robustness of the basic model under realistic and complex operating conditions, and to provide data support for subsequent online sensitivity calibration, a second validation set needs to be constructed. The construction method of the second validation set includes: selecting a portion of samples from historical normal operating condition data as a basis, and simulating sensor drift by injecting noise that conforms to actual physical laws to generate a series of adversarial samples. This second validation set is not used to modify the parameters of the basic model, but is used to quantitatively evaluate the predictive stability of the model under non-ideal operating conditions.

[0030] Specifically, the network weights and biases, iteratively updated during pre-training using backpropagation and optimizers such as Adam, are used as parameters; based on the second validation set, each parameter in the base model is analyzed to obtain the average gradient magnitude and gradient direction information entropy of each parameter; taking the first... Taking one parameter as an example for analysis: First, for each sample in the second validation set... ;in, This refers to the sample index in the second validation set; The input is a sequence of sensor data. For the corresponding actual remaining useful life label; the sample Input the base model to obtain the predicted value. And calculate the mean squared error loss value. The mean squared error loss value is calculated using the backpropagation algorithm. Compared to the first The gradient magnitude of the nth parameter is calculated by repeating the above process for all samples in the second validation set. The average gradient magnitude of the parameters is denoted as . Among them, the first The average gradient magnitude of each parameter Characterized the first The average influence of each parameter on the model prediction error: The average gradient magnitude of each parameter The larger the value, the more sensitive the model prediction is to changes in this parameter. The more active a parameter is in the model, the better; repeat this operation to obtain the average gradient magnitude of each parameter.

[0031] After obtaining the average gradient magnitude of each parameter, it is necessary to further evaluate the consistency of the gradient direction of each parameter across different samples in the second validation set, in order to characterize the stability of the learning behavior of each parameter; with the first Taking the parameter as an example for analysis: to obtain the first parameter... The gradient values ​​calculated for each parameter on each sample of the second validation set are the signed raw gradients, not the gradient magnitudes. The gradient magnitudes are the absolute values ​​of the gradient values, thus forming a gradient value sequence. To analyze the directional distribution of the gradient values, the numerical range of this sequence is divided into... A continuous interval, for example The threshold for dividing the continuous interval is set as the third quartile of the gradient value sequence; the statistical... The frequency with which the gradient values ​​of the parameters fall into each continuous interval is obtained using the well-known entropy calculation formula. The gradient direction information entropy of each parameter, denoted as . Among them, the first Gradient direction information entropy of each parameter Characterized the first Consistency of gradient direction of the parameter among different samples in the second validation set: when the... Gradient direction information entropy of each parameter The smaller the number, the more significant the difference. The gradient directions of the parameters are highly consistent, and the learning behavior is stable; when the gradient direction of the parameters is highly consistent, the learning behavior is stable; Gradient direction information entropy of each parameter The larger the value, the greater the maximum value of 1, indicating that the gradient direction changes randomly and the learning behavior is unstable; repeat this operation to obtain the gradient direction information entropy of each parameter.

[0032] It should be noted that since the average gradient magnitude and gradient direction information entropy have different dimensions and numerical ranges, a direct linear combination may affect the fairness of the evaluation due to the difference in dimensions. To solve this problem, a weighted geometric mean is used to construct a dimensionless comprehensive evaluation index.

[0033] Based on this, a comprehensive evaluation index is constructed, which is the weighted geometric mean of the average gradient magnitude and the gradient direction information entropy; taking the first... Taking the parameter as an example for analysis, we obtain the first parameter... The comprehensive evaluation index for each parameter is expressed by the following formula: ; In the formula, It is the first A comprehensive evaluation index for each parameter; It is the first The average gradient magnitude of each parameter; It is the first The gradient direction information entropy of each parameter; It is an adjustment weighting coefficient used to control the relative importance of average gradient magnitude and gradient direction information entropy in the comprehensive evaluation index; It is a very small positive number used to prevent calculation anomalies when the average gradient magnitude or gradient direction entropy is zero; when the... The smaller the average gradient magnitude of the nth parameter, and the more significant the average gradient magnitude of the nth parameter, the more significant the average The smaller the gradient direction information entropy of the first parameter, the better. The smaller the comprehensive evaluation index of the first parameter, the better. The stronger the stability of the first parameter, the weaker the stability of the second parameter. The weaker the stability of each parameter.

[0034] Among them, setting the adjustment weight coefficient The logic is: As a hyperparameter, it needs to be adjusted according to the actual application scenario of the smart appliance cabinet, taking into account the emphasis on parameter sensitivity and operational stability: if the operating environment of the smart appliance cabinet is complex and variable, and more attention needs to be paid to the sensitivity of the parameter to changes, then the hyperparameter can be appropriately increased. If the system requires extremely high operational stability and needs to focus more on the consistency of parameter learning behavior, then the size can be appropriately reduced. ; The value can be set according to actual needs. For example, the present invention will... Set it to 0.7.

[0035] It should be noted that the construction logic of the comprehensive evaluation index can be further explained: For a parameter to be classified as a stable steady-state parameter, it must simultaneously satisfy two conditions: a small average gradient magnitude (i.e., low sensitivity to disturbances) and a small gradient direction entropy (i.e., stable learning behavior). The weighted geometric mean product form used strengthens the logical relationship of "sum" between these two conditions. If a parameter has a large average gradient magnitude and a large gradient direction entropy, then this parameter is more likely to be fitting noise and should be classified as a dynamic parameter. If a parameter has a small average gradient magnitude and a small gradient direction entropy, then this parameter is likely to have learned the underlying stable laws of the system and should be classified as a steady-state parameter.

[0036] Furthermore, a comprehensive evaluation index for all parameters is obtained, and all parameters are sorted in ascending order based on the comprehensive evaluation index to obtain a sorted sequence; the third quartile of the sorted sequence is used as a quartile threshold to classify parameter types.

[0037] Specifically, based on quantile thresholds, parameter partitioning is performed: parameters with comprehensive evaluation indicators less than or equal to the quantile thresholds are deemed to have stable learning behavior and are crucial to the core capabilities of the model, and are assigned to the steady-state parameter set; while parameters with comprehensive evaluation indicators greater than the quantile thresholds are deemed to have unstable learning behavior and are sensitive to changes in the learning environment, and are assigned to the dynamic parameter set for online learning. This allows for rapid adaptation to changes in the operating status of the intelligent appliance cabinet, laying a solid foundation for subsequent safe, efficient, and stable online learning, and ensuring that the model's core predictive capabilities do not degrade when adapting to new operating conditions.

[0038] The online performance monitoring trigger module 200 is used to: obtain the predicted and actual values ​​of the remaining lifetime of the basic model at each time point; form an absolute error sequence by combining the absolute error values ​​of the predicted and actual values ​​at all times in the sliding window, and calculate the model performance degradation index of the absolute error sequence using the exponential weighted moving average method, compare it with the preset performance degradation threshold, and generate an update trigger signal for the dynamic parameter set based on the comparison result.

[0039] It should be noted that the core purpose of this module is to establish a sensitive and reliable mechanism for quantifying the synchronization deviation between the digital twin virtual model and the smart appliance cabinet. The steady-state parameter set of the online running short-term memory basic model has been locked. The steady-state parameter set is used to retain the learned stable physical laws of the smart appliance cabinet. Therefore, the core cause of the significant degradation of model performance can be located as the dynamic parameter set being unable to adapt to the new operating conditions. Based on this, monitoring the overall performance degradation of the model is essentially to evaluate the effectiveness of the dynamic parameter set, thereby providing a triggering basis for subsequent accurate and efficient parameter updates.

[0040] Specifically, during the online synchronization and operation phase of the digital twin system for the smart appliance cabinet, the system continuously receives real-time sensor data streams from the smart appliance cabinet via an IoT interface. These real-time sensor data streams are from the same source and have the same format as the data from the pre-training phase, both collected at a sampling frequency of 1 Hz, and include current, voltage, temperature, and vibration signals from core components. Forward inference is performed by the basic model of the long short-term memory network after dividing the steady-state parameter set and the dynamic parameter set, generating the predicted values ​​of the basic model for that batch of time periods. These predicted values ​​are obtained through forward computation of the basic model: the real-time sensor data streams collected at each moment are input into the basic model, and through linear transformations and nonlinear activation functions such as the Sigmoid function within the basic model, the predicted values ​​of the basic model at each moment are finally output. For example, the predicted value of the basic model at each moment is... The predicted value of the basic model at time is denoted as .

[0041] Furthermore, the true value of the basic model is obtained through a combination of multi-path data interfaces. The specific acquisition process is as follows: monitoring data from the sensors deployed in the smart appliance cabinet is obtained to ensure the reliability of core parameter measurements; the equipment logs associated with the smart appliance cabinet are accessed to extract quantitative data such as fault records; based on this, the true value needs to be verified to confirm its accuracy. The verification rule is: first, through… The criteria are used to remove outliers, then the deviation between the main sensor data and the redundant sensor data is calculated. The deviation must be less than or equal to 5%; otherwise, data re-acquisition is triggered, and finally, the true value of the base model at that moment is generated. For example, the first... The true value of the underlying model at time is denoted as ; will the first The predicted value of the base model at this time Compared with the true value The difference, as the first The absolute error value at time t is denoted as . The larger the absolute error value, the lower the instantaneous prediction accuracy of the model at the current moment; conversely, the smaller the absolute error value, the higher the instantaneous prediction accuracy of the model at the current moment.

[0042] Specifically, to further capture the evolution trend of model performance, a fixed length is set. The sliding window with a fixed length The setting needs to be based on the typical working cycle of the smart appliance cabinet. For example, if the typical working cycle of the smart appliance cabinet is 8 hours, then the fixed length... It can be set to 8 to ensure that the sliding window can cover the entire working cycle of the smart appliance cabinet.

[0043] Furthermore, using a fixed length of The sliding window dynamically stores the most recent The absolute error sequence at each time step; the update rule for the absolute error sequence is: after calculating the absolute error value at the current time step, add it to the sliding window; if the window is full, then... If the first absolute error value is removed, then the earliest absolute error value is removed simultaneously, i.e., the first absolute error value is removed. absolute error value at time This forms a dynamically updated absolute error sequence, denoted as... .

[0044] Furthermore, the model performance degradation index is calculated using an exponentially weighted moving average function. Its core advantage is that it assigns higher weight to recent errors, making the index more sensitive to initial changes in performance degradation and avoiding the lag problem of simple moving averages. Based on this, the first... The model performance degradation exponent corresponding to time point is expressed by the following formula: ; In the formula, It is the first The model performance degradation index at time t is a dimensionless scalar. The larger the value, the more significant the recent overall predictive performance degradation of the model and the greater the synchronization deviation with the smart appliance cabinet. It is the first The absolute error value at time , where To backtrack the step size, ;when hour, It is the first The absolute error value at time t; It is the length of the sliding window; To represent an increasing weighting factor, the following conditions must be met: Exemplary ; This constitutes an exponentially increasing weight sequence; this means that as... Decrease, meaning the closer to the current moment. hour, The larger the value, the greater the weight given to the most recent absolute error value.

[0045] It should be noted that the first Model performance degradation index at time step The physical meaning of is: the exponentially weighted average of recent prediction errors. Compared with the simple moving average, its logical advantage lies in: significantly amplifying the contribution of recent errors through exponential weighting; when the model performance begins to show a continuous but slight downward trend, even if the absolute error of a single time batch does not exceed the alarm threshold, the exponent will accumulate and rise rapidly due to the high weighting of recent errors, thus enabling timely detection of the signs of model performance degradation and providing an accurate basis for subsequent online adaptive updates of the dynamic parameter set.

[0046] Furthermore, the calculated model performance degradation index is compared with a preset performance degradation threshold. The performance degradation threshold needs to be determined through offline testing. Preferably, the 95th percentile of all model performance degradation indices is taken as the performance degradation threshold to ensure that the false trigger rate is controlled below 5%. If the model performance degradation index is greater than or equal to the performance degradation threshold, it is determined that the synchronization deviation between the digital twin virtual model and the smart appliance cabinet is significant, and an update trigger signal for the dynamic parameter set is generated. If the model performance degradation index is less than the performance degradation threshold, it is determined that the current performance of the model is within an acceptable range, and monitoring will continue without triggering any update operation. The update trigger signal for the dynamic parameter set will initiate subsequent steps to perform online incremental learning on the divided dynamic parameter set so that the digital twin virtual model can resynchronize with the smart appliance cabinet.

[0047] The dynamic parameter adjustment synchronization module 300 is used to respond to the update trigger signal, construct a total loss function using recent operating data as training samples, the total loss function includes a prediction error loss term and a logarithmic constraint term on the relative update magnitude of the dynamic parameters; the optimizer is used to iteratively update only the dynamic parameter set to obtain the updated dynamic parameter set, thereby realizing the full life cycle management of the smart appliance cabinet.

[0048] It should be noted that this step is the core adaptive link for the digital twin virtual model to achieve dynamic synchronization with the smart appliance cabinet; its design is based on the following core logic: when the performance degradation index exceeds the performance degradation threshold, it indicates that the dynamic parameter set of the basic model after locking the steady-state parameter set can no longer effectively represent the current operating conditions and must be updated; in order to ensure the stability of the update process and avoid model instability due to parameter mutations, the dynamic parameter set needs to be adjusted.

[0049] Specifically, when the performance degradation index exceeds the performance degradation threshold, the dynamic parameter set needs to be adjusted. This adjustment process uses data within a sliding window as training samples to ensure that the training data fully matches the data source and time dimension of the performance evaluation. Based on this, the total loss function is constructed. The relationship is as follows.

[0050] ; In the formula, It is the total loss function, used to guide the optimization direction of dynamic parameter updates; It is the standard prediction error loss, consistent with the loss function type in the pre-training stage. The mean squared error is preferred. Its calculation method is: calculate the average of the predicted values ​​and effective true values ​​at all times in the current batch according to the mean squared error formula, and use it as the standard prediction error loss. This is the constraint strength coefficient, used to adjust the update magnitude of dynamic parameters. An example value is 0.1, which can be adjusted appropriately according to the stability requirements of the actual working conditions. It is a dynamic parameter set, containing the weights and biases in the base model that need to be adaptively updated; It is a single dynamic parameter in the dynamic parameter set; It is a dynamic parameter The amount of updates during the adjustment process; This is the scaling factor, with an example value of 1.0, used to ensure the relative update magnitude. The calculation is not affected by the magnitude of the absolute value of the dynamic parameter; It is a very small positive number, used to avoid calculation errors when the denominator is zero; It is the absolute value symbol; It is a logarithmic function.

[0051] The core logic of the total loss function lies in: the prediction error term This ensures that prediction errors are reduced after model updates, guarantees prediction accuracy, and enables the model to quickly adapt to the current operating conditions of the smart appliance cabinet; the constraint terms are in logarithmic function form. Used for relative update magnitude Applying a smoothing penalty: When the relative update magnitude is too large, the value of the constraint term will increase significantly, thus strongly suppressing drastic changes in the dynamic parameter and forcing the optimizer to choose a smaller update amount; this will make... This naturally became the step size for its current update. Adaptive regulator: The larger the absolute value of the dynamic parameter, the larger the allowable update amount. The larger the value, the greater the allowable update amount; the smaller the absolute value of the dynamic parameter, the smaller the allowable update amount. The smaller the value, the more stable the update will be, ultimately achieving a scale that matches the dynamic parameters themselves.

[0052] Furthermore, the Adam optimizer in the pre-training stage is used for iterative optimization; in the iterative optimization, only the gradients of all dynamic parameters in the dynamic parameter set are calculated and updated; the steady-state parameters in the steady-state parameter set are not calculated and do not participate in the update; when the total loss function The adjustment process terminates when the maximum number of iterations is reached; for example, the maximum number of iterations is set to 50, which can be adjusted appropriately according to actual needs.

[0053] It should be noted that the core innovation of this module lies in solving the technical problem of model instability caused by parameter mutations in traditional long short-term memory networks, and realizing stable and smooth updates of dynamic parameter sets. Its logical advantages are: it ensures that the model can quickly adapt to new working conditions and improve prediction accuracy through standard prediction error loss, and it limits the update amount of dynamic parameters through logarithmic function constraints to avoid model oscillations caused by drastic changes in individual dynamic parameters. Ultimately, it achieves a balance between adapting to changes and maintaining stability in the digital twin virtual model. This adjustment method is a key technical feature to ensure the long-term reliable operation of the intelligent electrical cabinet full life cycle management system.

[0054] Finally, the system outputs an updated set of dynamic parameters; the digital twin virtual model will continue to run online, combining the locked steady-state parameter set with the updated dynamic parameter set, thereby restoring the synchronization state with the smart appliance cabinet without disrupting the core laws already learned by the smart appliance cabinet.

[0055] Based on the synchronized digital twin virtual model, the remaining service life of the smart appliance cabinet is output, thereby realizing full lifecycle management of the smart appliance cabinet. The system refers to a management system centered on digital twin technology and oriented towards the full lifecycle of the smart appliance cabinet: During the operation phase: the digital twin virtual model continuously maps key operating parameters such as current, voltage, and temperature of the smart appliance cabinet and its remaining service life, providing maintenance personnel with 24-hour uninterrupted visual monitoring, covering the status awareness needs of the operation phase; During the maintenance phase: when the remaining service life of the smart appliance cabinet predicted by the digital twin virtual model shows an abnormal trend, such as a rapid shortening of the remaining service life, the system automatically triggers a fault warning; Finally, the system combines the assessment results of the remaining service life to automatically generate a maintenance cycle and maintenance content suitable for the smart appliance cabinet, thereby improving the management efficiency of the smart appliance cabinet throughout its entire lifecycle.

[0056] Following the steps above, the updated LSTM model can be obtained, leading to a comparison chart of the prediction performance before and after the update of the LSTM model for a digital twin-based intelligent appliance cabinet lifecycle management system. (Refer to...) Figure 2 And a comparison chart of the prediction performance of the updated LSTM and the traditional method, see reference. Figure 3 .

[0057] in, Figure 2This reflects the synergistic effect of the pre-training and parameter set partitioning module, the online performance monitoring triggering module, and the dynamic parameter adjustment synchronization module; the deviation between the prediction curve and the actual lifetime curve of the unupdated Long Short-Term Memory (LSTM) network continues to widen, which is consistent with the actual situation that the performance of the basic model degrades with changes in operating conditions; the updated LSTM prediction curve fits the actual lifetime curve, which is the result of the steady-state parameter set retaining the core laws and the dynamic parameter set adapting to new operating conditions, thus achieving a balance between adaptability and stability.

[0058] in, Figure 3 This reflects that the remaining service life of the smart appliance cabinet is a result of time-series correlation. Traditional methods, which use static modeling, cannot capture the temporal correlation of data, and the predicted curve deviates significantly from the actual service life curve. In contrast, the Long Short-Term Memory (LSTM) network captures long- and short-term temporal dependencies through its own structure and combines dynamic parameter updates to adapt to changes in operating conditions. Its prediction accuracy far exceeds that of traditional methods, and it can still accurately predict the remaining service life of the smart appliance cabinet in the later stages of its operation when it is more sensitive.

[0059] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A digital-twin-based intelligent electrical cabinet whole life cycle management system, characterized in that, The application relates to a method for implementing full-life-cycle management of an intelligent electric appliance cabinet, which comprises the following steps: a pre-training and parameter division module is used to pre-train based on historical operation data of the intelligent electric appliance cabinet, to obtain a basic model for predicting residual service life; a second verification set is constructed, and average gradient amplitude and gradient direction information entropy of each parameter in the basic model are calculated; a comprehensive evaluation index of each parameter is constructed by weighted geometric mean, and the parameters of the basic model are divided into a steady-state parameter set and a dynamic parameter set; an online performance monitoring trigger module is used to obtain predicted values and real values of residual service life of the basic model at each moment; absolute error values of the predicted values and the real values at all moments in a sliding window are combined to form an absolute error sequence, and a model performance degradation index of the absolute error sequence is calculated by using an exponential weighted moving average method; the model performance degradation index is compared with a preset performance degradation threshold value, and an update trigger signal of the dynamic parameter set is generated according to a comparison result; a dynamic parameter adjustment synchronization module is used to respond to the update trigger signal, to construct a total loss function by taking recent operation data as training samples, wherein the total loss function comprises a prediction error loss term and a logarithmic constraint term for a relative update amplitude of the dynamic parameters; an optimizer is used to iteratively update only the dynamic parameter set, to obtain an updated dynamic parameter set, and to realize full-life-cycle management of the intelligent electric appliance cabinet.

2. The intelligent electric cabinet whole life cycle management system based on digital twinning according to claim 1, characterized in that, The basic model for predicting residual service life comprises the following steps: input data of the basic model are multi-source time-series sensor data collected in a historical operation process of the intelligent electric appliance cabinet in the past one month, the multi-source time-series sensor data are stored in sequence form with time stamps as indexes, and the input data specifically comprise current, voltage, temperature and vibration signals of core parts; and output data are residual service life of the intelligent electric appliance cabinet in a future preset time window.

3. The intelligent electric cabinet whole life cycle management system based on digital twinning according to claim 1, characterized in that, The second verification set is constructed by the following steps: a part of samples are selected as a base from historical normal working condition data, and a series of adversarial samples are generated by injecting noise conforming to actual physical laws to simulate sensor drift, to serve as the second verification set.

4. The intelligent electric cabinet whole life cycle management system based on digital twinning according to claim 1, characterized in that, The average gradient amplitude of each parameter in the basic model is calculated by the following steps: For each sample in the second validation set ,in, This refers to the sample index in the second validation set; For sensor data sequences, For the corresponding true remaining service life label; sample the sensor data sequence. Input the base model to obtain the predicted value. And calculate the mean squared error loss value. The sample is calculated using the backpropagation algorithm. Mean squared error loss value Compared to the first Calculate the gradient magnitude of each parameter; calculate the mean squared error loss of all samples in the second validation set relative to the first parameter. The average gradient magnitude of the nth parameter is used as the 1st parameter. The average gradient magnitude of each parameter.

5. The intelligent electric cabinet whole life cycle management system based on digital twinning according to claim 1, characterized in that, The comprehensive evaluation index of each parameter is constructed by the following steps: ； In the formula, is the comprehensive evaluation index of the first parameter; is the average gradient amplitude of the first parameter; is the gradient direction information entropy of the first parameter; is the adjustment weight coefficient; is a very small positive number.

6. The intelligent electric cabinet whole life cycle management system based on digital twinning according to claim 1, characterized in that, The parameters of the basic model are divided into the steady-state parameter set and the dynamic parameter set by the following steps: comprehensive evaluation indexes of all parameters are sorted in ascending order to obtain a sorting sequence; third and fourth quantile thresholds of the sorting sequence are obtained; and parameter division is performed: parameters with a comprehensive evaluation index less than or equal to the quantile threshold are divided into the steady-state parameter set; and parameters with a comprehensive evaluation index greater than the quantile threshold are divided into the dynamic parameter set.

7. The intelligent electric cabinet whole life cycle management system based on digital twinning according to claim 1, characterized in that, The model performance degradation index of the absolute error sequence is calculated by the following steps: ； wherein is the model performance degradation index at the moment; is the absolute error value at the moment, wherein is the backtracking step size; is the length of the sliding window; denotes the weight increasing factor, which needs to satisfy the condition: . 8.The intelligent electric cabinet whole life cycle management system based on digital twinning of claim 1, wherein, The total loss function is constructed by the following steps: ； wherein is the total loss function; is the standard prediction error loss; is the constraint strength coefficient; is the dynamic parameter set; is an individual dynamic parameter in the dynamic parameter set; is the dynamic parameter is the update amount in the adjustment process; is the scale scaling coefficient; is a very small positive number; is the absolute value symbol; is the logarithmic function. 9.The intelligent electric cabinet full life cycle management system based on digital twinning of claim 1, wherein, The dynamic parameter set is iteratively updated by the following steps: In the iterative optimization, only the gradients of all dynamic parameters in the dynamic parameter set are calculated and updated; the steady-state parameters in the steady-state parameter set are not calculated and do not participate in the update; when the total loss function reaches the maximum number of iterations, the adjustment process is terminated; finally, the system outputs the updated dynamic parameter set.

10. The intelligent electric cabinet whole life cycle management system based on digital twinning according to claim 1, characterized in that, The method for implementing full-life-cycle management of the intelligent electric appliance cabinet comprises the following steps: The digital twin virtual model will continue to run online in combination with the locked steady-state parameter set and the updated dynamic parameter set. In the running stage: the digital twin virtual model continuously maps the key operating parameters such as current, voltage, temperature and remaining service life of the intelligent electric cabinet, providing 24-hour uninterrupted visual monitoring for operation and maintenance personnel, covering the state perception needs in the running stage; in the maintenance stage: when the remaining service life predicted by the digital twin virtual model shows an abnormal trend, a fault warning will be automatically triggered; finally, combined with the remaining service life evaluation result, the maintenance cycle and maintenance content suitable for the intelligent electric cabinet are automatically generated.

Images