Device operation and maintenance strategy optimization method based on digital twinning

Abstract

The application provides a device operation and maintenance strategy optimization method based on digital twinning, which comprises the following steps: collecting multi-dimensional working condition data of a physical device within a set time window, inputting the preprocessed data into a time sequence feature extraction model to generate a state vector representing the current operation state of the device; inputting the state vector into a digital twinning model to obtain a predicted behavior vector, comparing the predicted behavior vector with a reference behavior vector, and calculating a deviation vector and a deviation aggregation scalar reflecting the deviation degree of each dimension; based on the state vector, the deviation vector and the deviation aggregation scalar, a calibration strategy vector is dynamically generated through a strategy decision model, and the strategy vector contains the activation probability of multiple strategy paths; and according to the calibration strategy vector, the parameters of the digital twinning model are selectively fine-tuned to complete the online update of the model. The application has strong universality and engineering feasibility.

Description

Technical Field

[0001] This invention belongs to the field of digital twin technology and intelligent equipment operation and maintenance, and particularly relates to a method for optimizing equipment operation and maintenance strategies based on digital twins. Background Technology

[0002] With the deepening of industrial digitalization and intelligent manufacturing, digital twin systems, as a key technology for achieving collaborative operation between physical equipment and virtual models, have been widely applied in fields such as equipment health monitoring, operational status prediction, and production process optimization. Digital twins collect real-time equipment operation data to drive dynamic updates of the virtual model, achieving virtual-real fusion and predictive control, providing support for equipment operation assessment and maintenance decisions. However, in actual operation, influenced by factors such as ambient temperature, load fluctuations, equipment wear, and sensor accuracy degradation, the operating conditions of equipment continuously evolve. This causes the digital twin model to gradually deviate from the actual equipment behavior over long-term operation, thus affecting the accuracy and reliability of maintenance strategies and adjustment decisions based on the twin model.

[0003] Existing model calibration or operation and maintenance (O&M) adjustment methods typically rely on periodic training, manual intervention, or static correction triggered by fixed thresholds. While these methods can alleviate model bias to some extent, they struggle to continuously perceive and comprehensively assess equipment operating status and model behavior deviations under complex operating conditions, nonlinear coupling effects, and multidimensional dynamic disturbances. This often leads to problems such as response lag, over-calibration, or inappropriate O&M strategy selection. Traditional methods often separate model calibration from O&M decision-making, lacking a mechanism to systematically correlate operating conditions, model behavior deviations, and the O&M strategy generation process. This makes it difficult for the system to comprehensively determine when O&M intervention is needed, what adjustment strategy should be adopted, and how much adjustment should be set. Consequently, it is difficult to achieve continuous optimization of equipment operating performance and O&M efficiency while ensuring system stability. Therefore, how to construct a method based on digital twin models, combined with equipment operating conditions and model behavior deviations, capable of dynamically generating and optimizing equipment O&M strategies, and supporting O&M decision implementation through reasonable model parameter updates and strategy execution mechanisms, has become a critical technical problem that urgently needs to be solved. Summary of the Invention

[0004] The purpose of this invention is to propose a device operation and maintenance strategy optimization method based on digital twins to address the above shortcomings.

[0005] To achieve the above objectives, this invention provides a method for optimizing equipment operation and maintenance strategies based on digital twins, the method comprising:

[0006] S1. Collect multi-dimensional operating condition data of physical equipment within a set time window, preprocess the data, input the data into the time series feature extraction model, and generate a state vector representing the current operating state of the equipment.

[0007] S2. Input the state vector into the digital twin model to obtain the predicted behavior vector, and compare it with the reference behavior vector to calculate the deviation vector and the deviation aggregation scalar that reflect the degree of deviation in each dimension.

[0008] S3. Based on the state vector, the deviation vector, and the deviation aggregation scalar, a calibration policy vector is dynamically generated through a pre-constructed policy decision model. The policy vector contains the activation probabilities of multiple policy paths.

[0009] S4. Based on the calibration strategy vector, selectively fine-tune the parameters of the digital twin model to complete the online update of the model.

[0010] Furthermore, the preprocessing involves performing channel compression on the multidimensional operating condition data and mapping it to an intermediate representation as a hidden state tensor.

[0011] Furthermore, the temporal feature extraction model employs a single lightweight Transformer encoder for temporal modeling;

[0012] The Transformer encoder structure includes a multi-head attention mechanism, a feedforward neural network layer, and a residual connection and layer normalization module.

[0013] After encoding, average pooling is performed on the representation vector within the time window to generate the final state vector representing the current operating state of the device.

[0014] Furthermore, the digital twin model consists of a two-layer feedforward neural network, with the feedforward neural network having an input layer dimension, two hidden layers of 128 and 64 units respectively, and the activation function being ReLU.

[0015] Furthermore, the comparison with the reference behavior vector to calculate the deviation vector and the deviation aggregation scalar reflecting the degree of deviation in each dimension specifically includes:

[0016] The reference behavior vector is obtained; the reference behavior vector originates from data collected by the actual device's sensors.

[0017] By combining the reference behavior vector and the predicted behavior vector, a bias vector is constructed using a weighted relative bias form to characterize the degree of deviation of each behavior variable;

[0018] Based on the aforementioned deviation vector, a confidence-weighted deviation aggregation is introduced to generate a deviation aggregation scalar.

[0019] Furthermore, S3 specifically includes:

[0020] The deviation distribution is calculated by combining the mean and variance of the deviation vector.

[0021] The calibration initiation score of the current system is calculated by weighted summation of the deviation distribution and the deviation aggregate scalar; wherein, the higher the calibration initiation score, the more inclined the system is to perform calibration; the weights of the weighted summation correspond to the weights of the overall deviation intensity and the unevenness of the deviation distribution, respectively;

[0022] Based on the calibration start score of the current system, a device state influence factor is introduced. The activation level of each strategy path is adjusted by mapping the state vector to control weights to generate a calibration strategy vector. Each component of the calibration strategy vector represents the activation probability of the corresponding strategy path.

[0023] Furthermore, the mapping structure of the calibration strategy vector is a linear combination of matrices;

[0024] The strategy path includes high probability of performing minor adjustments, not recommending module replacement, and moderately recommending switching redundant paths.

[0025] Furthermore, the selective fine-tuning is based on the current Siamese model parameter set, and is updated by combining the mask matrix generated based on the calibration policy vector, the gradient of the loss function corresponding to the model behavior error, and the sign function, thereby generating an updated Siamese model parameter set.

[0026] Furthermore, the mask matrix generated based on the calibration strategy vector is used to control the parameter update region and combined with a lightweight regularization term to stabilize the update direction, so that the system modeling capability can be maintained even when the device state changes drastically.

[0027] Furthermore, S4 also includes:

[0028] If current computing resources are limited or multiple consecutive calibrations do not significantly improve the deviation, the system can store the calibration strategy vector and deviation vector in a buffer queue, and perform the update process in batches when resources allow or the accumulated deviation exceeds a preset threshold, so as to avoid frequent fine-tuning from affecting system stability.

[0029] The beneficial technical effects of the present invention are at least as follows:

[0030] This invention addresses the problems of existing digital twin models, such as behavioral deviations and lack of dynamic adaptive calibration capabilities, by proposing a digital twin-based method and system for optimizing equipment operation and maintenance strategies. The method constructs a multi-level closed-loop structure centered on operational condition modeling, model deviation estimation, dynamic strategy decision-making, and model parameter updates. This structure continuously identifies the differences between equipment operating conditions and model outputs during actual operation and adaptively generates calibration strategies based on deviation distribution characteristics and operational condition stability, achieving real-time model correction. By introducing a Transformer-based state representation model, the system can extract implicit features from time-series operational conditions, accurately describing the current operating state of the equipment. Combining weighted deviation estimation and a dynamic confidence mechanism, it quantifies model prediction errors and improves the accuracy of identifying complex disturbances. Furthermore, a strategy generation model coupled with deviation concentration and operational conditions is designed, enabling dynamic decision-making for multi-strategy parallel and hierarchical control. Finally, a strategy mask-driven parameter update mechanism achieves lightweight online calibration without altering the model structure, significantly reducing computational costs and improving system stability. This invention enables digital twin models to maintain a high degree of fit and operational consistency with physical entities in complex, nonlinear, and frequently changing environments through a continuous dynamic calibration process of "state perception - deviation quantification - strategy generation - model adjustment". It has strong versatility and engineering feasibility. Attached Figure Description

[0031] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0032] Figure 1 This is a flowchart of the device operation and maintenance strategy optimization method based on digital twins according to the present invention. Detailed Implementation

[0033] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0034] like Figure 1 As shown in the embodiment of the present invention, the equipment operation and maintenance strategy optimization method based on digital twins includes:

[0035] S1. Collect multi-dimensional operating condition data of physical equipment within a set time window, preprocess the data, input the data into the time series feature extraction model, and generate a state vector representing the current operating state of the equipment.

[0036] Specifically, this step aims to construct a vector that can dynamically represent the current operating state of the physical equipment based on key operating condition data during operation. This vector will serve as the foundational input for subsequent judgments of behavioral biases in the digital twin model and the generation of dynamic calibration strategies. Since the device's operating state changes continuously over time, and the data collected by the sensors has temporal characteristics, this step employs a temporal modeling approach and introduces a lightweight Transformer structure to achieve compressed representation of states, correlation modeling, and representation aggregation. This step avoids introducing redundant data categories and uses only core physical quantities directly related to the twin model's behavior, ensuring that the modeling process is both efficient and practical for engineering applications.

[0037] Among them, input data This represents a multidimensional operating condition data sequence within a fixed time window, where... Indicates the length of the time window (e.g., 50 frames are captured within 5 seconds, at 10 frames per second). This indicates the number and types of sensors. Data originates from a standard industrial-grade sensor system integrated with the modeled device, acquired through conventional communication interfaces. For example, temperature is acquired by an embedded thermocouple module via an RS485 interface; current is obtained by a Hall current sensor via an analog acquisition card; rotational speed is transmitted by an encoder via a CAN bus; load torque is provided by a strain gauge sensor connected to a Modbus interface; and device power is provided by a smart meter module (transmitted via OPCUA). The system uniformly samples and caches the data in matrix format through a centralized acquisition platform (such as a PLC or edge computing gateway), forming tensor inputs.

[0038] Furthermore, considering that different physical quantities have different dimensions and significantly different value ranges, the data first enters the normalization preprocessing module after collection. The linear mapping function constructed by historical statistical values ​​converts all original values ​​into a unit-independent normalized space. The normalization operation is implemented in the front-end data platform, relying only on historical extreme values ​​and moving averages, without the need for model intervention, thus ensuring the stability of the input data.

[0039] Furthermore, in order to reduce the input dimensionality and fuse semantic information from different sensor channels before the data enters the deep model, the first step is to... Perform channel compression and map it to an intermediate representation. The calculation method is as follows:

[0040] ;

[0041] in, It is the weight matrix for channel compression, used to map the original sensor channels to a low-dimensional semantic space; It is the corresponding bias vector. This is the compressed tensor, representing the fused state vector at each time step. This operation is performed in the device-side inference module via floating-point matrix multiplication. and The model is determined through training on historical data before deployment.

[0042] Obtain the compressed hidden state tensor Subsequently, a lightweight Transformer encoder is used for temporal modeling. This Transformer architecture includes a multi-head attention mechanism (4 attention heads, each with a dimension of [missing information]). ), feedforward neural network layer (hidden layer size is The system uses GELU as the activation function, and includes residual connections and layer normalization modules. After encoding, average pooling is performed on the representation vectors within the time window to generate the final device state representation vector. The calculation method is as follows:

[0043] ;

[0044] in, This refers to the Transformer encoder module. For the first The compressed state vector at each time step, where T is the total time length. This is an aggregated representation of the operating conditions over the entire time period. It reflects the overall operating trend within the time window, such as high temperature, continuous overload, and intermittent current fluctuations.

[0045] Output state vector This will serve as the sole input for subsequent steps (behavioral bias estimation), and its physical meaning is an overall representation of the device's operating state in the current time period. The vector itself does not directly represent specific temperature or current values, but through processes such as compression, modeling, and pooling, each dimension implicitly contains temporal information related to the device's physical operating characteristics.

[0046] S2. Input the state vector into the digital twin model to obtain the predicted behavior vector, and compare it with the reference behavior vector to calculate the deviation vector and the deviation aggregate scalar that reflect the degree of deviation in each dimension.

[0047] Specifically, the purpose of this step is to improve the current digital twin model. Under actual working conditions The deviation between the predicted behavior and the reference behavior is quantitatively estimated, and the deviation vector is output. and weighted bias scalar This provides a basis for decision-making in the next step of calibration strategy selection. Unlike traditional error calculation methods, this step takes into account the special problems existing in the digital twin scenario, such as partial observation missing, large differences in behavioral output dimensions, and non-unique reference behavior, and proposes a behavioral bias estimation method with interpretability and strong resistance to uncertainty.

[0048] Among them, the model in the digital twin system It is a behavior prediction module consisting of two layers of feedforward neural networks, with its structure being based on the input layer dimension. (and (Same), the two hidden layers are respectively and Unit, activation function is ReLU, output dimension is This represents the key behavioral parameters of the system (e.g., robotic arm position, gripper speed, system energy consumption, local stress, etc.). Model The model is trained using historical operational data from the device before deployment and undergoes no structural changes after deployment. It accepts state input during inference. And generate predictive behavior This refers to the predicted output of the device in its current state.

[0049] Reference behavior vector The acquisition method is dynamically selected based on the deployment environment of the twin system. In scenarios with real-time feedback, Data is derived from actual device sensors (e.g., the end effector position of a robotic arm is obtained using a laser rangefinder); in the absence of real-time feedback, the system generates data based on historical state behavior pairs or simplified physical models. The estimated value. To account for the differences in reliability of reference behaviors from different sources, the system assigns a reliability weight to each dimension of the reference value. , forming a weight vector .

[0050] Furthermore, to accommodate the inconsistency in the output dimensions and unit differences of twin behaviors (e.g., position in mm, power in W), this step employs a weighted relative deviation form to construct a deviation vector. :

[0051] ;

[0052] in, The first prediction of the current model Individual behavior quantity For reference only. To prevent division by zero constants (such as...) ), As the velocity regularization factor, This represents the model prediction value at the previous time step. The first term is the time interval (in seconds). The second term is the behavior change rate, a penalty term used to supplement the model's own change trend when the reference behavior is missing or unreliable. This structure is designed for real-time operation of digital twins, enhancing sensitivity to abnormal fluctuations in behavior prediction. For example, if the robotic arm's position changes by 5cm in the past 0.1 seconds, this regularization term can indicate an anomaly in the current prediction even if the reference value is missing.

[0053] Furthermore, after constructing the bias vector, to provide the bias strength scalar required for policy decision-making, a confidence-weighted bias aggregation formula is introduced as follows:

[0054] ;

[0055] in, It is the first The reference confidence weight for each behavior item, where k is the total number of behavior items. This is the information enhancement factor, used to emphasize the non-negligibility of regions with moderate bias. It adds to the bias term. The structure has two engineering implications: first, it avoids excessive concentration of linear deviations in large error terms, improving the identifiability of small and medium deviations; second, it amplifies rapidly fluctuating but small-absolute-value behavioral terms, making them perceptible to the strategy model. This design is particularly important in industrial scenarios, such as when the behavior of a device system is physically in a "critically unstable" state, traditional deviation quantities may not have reached the threshold but may actually be indicative of a danger signal.

[0056] The final output includes two variables: one is the deviation vector. One is used to characterize the degree of bias of each behavioral variable, and the other is the bias aggregation scalar. These serve as crucial inputs for selecting calibration strategies in subsequent steps. Both outputs are entirely based on the state vector from the previous step. and the current model It does not depend on any uncontrollable external variables.

[0057] S3. Based on the state vector, the deviation vector, and the deviation aggregation scalar, a calibration policy vector is dynamically generated through a pre-constructed policy decision model. The policy vector contains the activation probabilities of multiple policy paths.

[0058] Specifically, this step is used to analyze the model behavior deviation information (i.e., the deviation vector) output from the previous step. and deviation aggregate scalar ), combined with the current device operating state vector Dynamically generate a set of policy vectors to drive adjustments to the digital twin model. ,in This indicates the number of predefined calibration strategies in the system (such as weight fine-tuning, submodule replacement, control layer parameter adjustment, etc.). This step is the key core control module in the entire patent solution, determining "whether to calibrate, when to calibrate, and how to calibrate." Its role is to transform the "deviation phenomenon" of the digital twin model into "executable calibration operations," serving as the decision interface for achieving "dynamics." Compared to traditional systems that rely on static threshold triggers or manual rule judgments, this step, through dynamic logic modeling that integrates multiple factors, achieves for the first time a mapping expression between deviation, state, and control actions. It also introduces regularization and constraint mechanisms specifically designed for the operating characteristics of digital twin systems, improving the accuracy and stability of strategy selection.

[0059] Furthermore, to ensure the reliability and stability of strategy selection, this step employs a structured two-stage strategy generation mechanism: the first stage aggregates scalars based on deviations. and the state vector of the current running state The combination of factors determines whether calibration should be performed; the second stage is based on the deviation vector. The distribution characteristics and bias concentration determine the specific strategy to be adopted and the execution intensity. To this end, a logical expression model based on structure-sensitive weights is introduced, the core innovation of which lies in encoding the strategy probability through the following dual feature functions:

[0060] ;

[0061] in, This indicates the current system's calibration initiation score (the higher the score, the more likely it is to perform calibration). and These are adjustment coefficients, corresponding to the weights of overall deviation intensity and uneven deviation distribution, respectively; This represents the average value of the deviation vector. The variance of the deviation vector is used to reflect the fluctuation of deviations across different behavioral dimensions. This is a coefficient that enhances the impact of fluctuations. The first term of the formula reflects whether the system's deviation is large, and the second term reflects whether the deviation is concentrated. The multiplication of these two terms forms a scoring gating value used to trigger the strategy selection module. This structural design is a key innovation in the patent scenario, enabling dynamic adaptation to the behavioral accuracy requirements of different working conditions. For example, in robot trajectory control scenarios, Lower but Local calibration should be triggered when there are sharp fluctuations on a specific axis, rather than silent processing of the entire model.

[0062] Furthermore, after obtaining Next, the strategy vector generation stage begins. This stage introduces device state influence factors, by... The mapping is used to control weights to adjust the activation level of each policy path. The specific structure is as follows:

[0063] ;

[0064] in, For calibration strategy vectors;

[0065] This structure is a linear combination of matrices, where... The state mapping weight matrix, The influence matrix of deviation behavior, For dimension unit vector, To calibrate the startup gain, The Sigmoid activation function makes the output... Controlled Interval. The key innovation of this structure is that it combines the three types of variables—state, bias, and rating—into a single policy weight expression, improving the sensitivity and interpretability of policy selection. For example, when High value but state vector When the display device is in a high-risk operating condition, the system will suppress high-risk strategy items (such as structural reconfiguration) and select low-risk lightweight strategies (such as controller gain fine-tuning), reflecting consideration for the overall system security.

[0066] Final output Each component in the equation represents the activation probability of the corresponding policy path. For example:

[0067] This indicates a high probability of performing minor adjustments. This indicates that module replacement is not recommended. The suggestion is to switch to a redundant path.

[0068] These values ​​will be read and parsed into specific model parameter operations in the next step, the "Model Calibration Execution Module".

[0069] S4. Based on the calibration strategy vector, selectively fine-tune the parameters of the digital twin model to complete the online update of the model.

[0070] Specifically, this step is based on the calibration policy vector output by the policy module. Regarding the current twin model Perform parameter fine-tuning and selective module updates within the structure to generate the updated model. This step is used for state prediction in the next operating cycle. It is the core "taking action" link in the entire dynamic calibration chain, and its rationality, timeliness, and stability directly determine the calibration effect and the sustainable operation capability of the entire system. Digital twin model It is a deep neural network deployed on the edge or local server based on prior training. Its structure has been defined in the previous steps. It consists of two fully connected layers with 128 and 64 hidden units, and the input is the current state vector of the device. The output is the behavior prediction value. The model is deployed on a processing platform with inference capabilities, such as an industrial control terminal equipped with NVIDIA Jetson or Intel Movidius.

[0071] Among them, the policy vector The output from the previous step consists of multiple policy channels, each component corresponding to the activation level of a calibration technique, such as weight fine-tuning, path reweighting, or light quantum model activation. These policy channels have been mapped to their corresponding parameter regions during model initialization, for example... Control the fine-tuning amplitude of the first-layer weight tensor Controls whether to enable pre-trained low-dimensional path submodules. This includes controlling whether to replace the output layer's bias terms, etc. In actual deployment, the system will configure each component through a configuration file. With parameter tensor Each segment of weight or structural position in the data is bound to a specific element to form a controllable update interface.

[0072] Furthermore, to achieve precise and controllable online updates, this step employs "residual fine-tuning" to selectively adjust the target model parameters. Unlike traditional large-scale retraining based on full-model backpropagation, residual fine-tuning focuses only on the model's prediction error for the current state and performs weighted updates based on existing parameters. The core update formula for this process is as follows:

[0073] ;

[0074] in, For updating the set of parameters for the twin model, This is the current set of parameters for the twin model. The learning rate is a constant (the system default value is 0). ), It is by The generated mask matrix has the following dimensions: Correspondingly, this is used to specify which regions are allowed to be updated; The gradient of the loss function corresponding to the model behavior error can be obtained through error backpropagation. The strength parameter of the regularization term. The sign function is represented to suppress excessive parameter oscillations. The core innovation of this structure lies in the introduction of... By controlling the parameter update region and combining it with lightweight regularization terms to stabilize the update direction, the system's modeling capabilities can be maintained even when the device state changes drastically. For example, in a certain operating cycle, if the strategy... If the value reaches 0.9, the corresponding mask weight will be close to 1, allowing for a significant update of the first layer weights. When the value is 0.1, the corresponding region will only accept weak disturbances or even freeze.

[0075] Furthermore, the data input for this update process depends on the current state. and model output With behavioral deviation vector The system first inputs the state vector into the model to generate the predicted behavior. And then with reference behavior The deviation is obtained through comparison, and the target region is then updated based on the direction of the residual. For example, in an industrial welding arm controlled by a digital twin, if the deviation is concentrated in the end-effector pose error (such as...), the deviation is further analyzed. If this deviation is mapped through a policy, it will primarily affect the output layer parameters; the system will then automatically construct the output layer parameters based on this mapping. And restrict the update operation to the relevant weight region of the end output.

[0076] Furthermore, to accommodate the computational limitations at the edge, especially in embedded deployments on PLC or low-power MCU platforms, this step supports a "delayed calibration buffer mechanism." That is, if current computing resources are limited or multiple consecutive calibrations fail to significantly improve the deviation, the system can... and Store the data in a buffer queue and execute the update process in batches when resources allow or the accumulated deviation exceeds a threshold, thus avoiding frequent fine-tuning from impacting system stability.

[0077] Finally, the output of this step is the updated twin model. Its structure remains unchanged, but its parameters are updated within a controllable region, giving it new behavior mapping capabilities, which will be used for the state prediction task in the next cycle. The model update process is entirely dependent on the policy value output from the previous step. It deviates from the current model's predictions without introducing external training or online backpropagation mechanisms, and is reproducible and has low deployment costs.

[0078] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0079] In the embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection of apparatuses or units may be electrical, mechanical, or other forms.

[0080] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0081] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A method for optimizing equipment operation and maintenance strategies based on digital twins, characterized in that, include: S1. Collect multi-dimensional operating condition data of physical equipment within a set time window, preprocess the data, input the data into the time series feature extraction model, and generate a state vector representing the current operating state of the equipment. S2. Input the state vector into the digital twin model to obtain the predicted behavior vector, and compare it with the reference behavior vector to calculate the deviation vector and the deviation aggregation scalar that reflect the degree of deviation in each dimension. S3. Based on the state vector, the deviation vector, and the deviation aggregation scalar, a calibration policy vector is dynamically generated through a pre-constructed policy decision model. The policy vector contains the activation probabilities of multiple policy paths. S4. Based on the calibration strategy vector, selectively fine-tune the parameters of the digital twin model to complete the online update of the model; S3 specifically includes: The deviation distribution is calculated by combining the mean and variance of the deviation vector. The calibration initiation score of the current system is calculated by weighted summation of the deviation distribution and the deviation aggregate scalar; wherein, the higher the calibration initiation score, the more inclined the system is to perform calibration; the weights of the weighted summation correspond to the weights of the overall deviation intensity and the unevenness of the deviation distribution, respectively; Based on the calibration start score of the current system, a device state influence factor is introduced. The activation level of each strategy path is adjusted by mapping the state vector to control weights, generating a calibration strategy vector. Each component of the calibration strategy vector represents the activation probability of the corresponding strategy path. Specifically: ； in, For calibration strategy vectors; in The state mapping weight matrix is... The influence matrix of deviation behavior, For dimension unit vector, To calibrate the startup gain, The Sigmoid activation function makes the output... Controlled interval; It is a state vector; The deviation vector; To calibrate the startup score; The mapping structure of the calibration strategy vector is a linear combination of matrices; The strategy path includes high probability of performing minor adjustments, not recommending module replacement, and moderately recommending switching redundant paths. The comparison with the reference behavior vector yields a deviation vector and a deviation aggregation scalar reflecting the degree of deviation in each dimension, specifically including: The reference behavior vector is obtained; the reference behavior vector originates from data collected by the actual device's sensors. By combining the reference behavior vector and the predicted behavior vector, a bias vector is constructed using a weighted relative bias form to characterize the degree of deviation of each behavior variable; Based on the aforementioned deviation vector, a confidence-weighted deviation aggregation is introduced to generate a deviation aggregation scalar.

2. The equipment operation and maintenance strategy optimization method based on digital twin as described in claim 1, characterized in that, The preprocessing involves performing channel compression on the multidimensional operating condition data and mapping it to an intermediate representation, which serves as a hidden state tensor.

3. The equipment operation and maintenance strategy optimization method based on digital twin according to claim 1, characterized in that, The temporal feature extraction model uses a lightweight Transformer encoder for temporal modeling. The Transformer encoder structure includes a multi-head attention mechanism, a feedforward neural network layer, and a residual connection and layer normalization module. After encoding, average pooling is performed on the representation vector within the time window to generate the final state vector representing the current operating state of the device.

4. The equipment operation and maintenance strategy optimization method based on digital twin according to claim 1, characterized in that, The digital twin model consists of a two-layer feedforward neural network. The feedforward neural network has an input layer dimension, two hidden layers with 128 and 64 units respectively, and the activation function is ReLU.

5. The equipment operation and maintenance strategy optimization method based on digital twin according to claim 1, characterized in that, The selective fine-tuning is based on the current set of Siamese model parameters, and is updated by combining the mask matrix generated based on the calibration policy vector, the gradient of the loss function corresponding to the model behavior error, and the sign function, thereby generating an updated set of Siamese model parameters.

6. The equipment operation and maintenance strategy optimization method based on digital twin according to claim 5, characterized in that, The mask matrix generated based on the calibration strategy vector is used to control the parameter update region and combined with a lightweight regularization term to stabilize the update direction, so that the system modeling capability can be maintained even when the device state changes drastically.

7. The equipment operation and maintenance strategy optimization method based on digital twin according to claim 5, characterized in that, The S4 further includes: If current computing resources are limited or multiple consecutive calibrations do not significantly improve the deviation, the calibration strategy vector and deviation vector are stored in a buffer queue, and the update process is executed in batches when resources allow or the accumulated deviation exceeds a preset threshold, so as to avoid frequent fine-tuning from affecting system stability.

Images