Power equipment predictive maintenance method based on digital twinning

Abstract

The invention relates to a digital twinning-based power equipment predictive maintenance method, which comprises the following steps of: acquiring time sequence operation data of each power equipment in a power system in real time, and preprocessing the time sequence operation data; based on the preprocessed time sequence operation data, updating the constructed digital twinborn model integrated with the LSTM in real time to obtain a latest digital twinborn model synchronously operating with the physical entity of the power equipment; based on the time sequence operation data, an isolated forest algorithm and a newest digital twinborn model are adopted to carry out fault prediction, and a fault prediction result is obtained; and determining a plurality of maintenance schemes based on the fault prediction result, and evaluating each maintenance scheme to obtain an optimal maintenance scheme and complete a predictive maintenance process. Compared with the prior art, the method has the advantages of improving the reliability and safety of the power equipment and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment maintenance, and more particularly to a predictive maintenance method for power equipment based on digital twin. Background Art

[0002] With the continuous expansion of the scale of the power system and the continuous improvement of equipment complexity, the traditional power equipment maintenance method usually focuses on regular maintenance or maintenance after failure. However, this method is difficult to meet the requirements of modern power systems for high-reliability and economic operation and maintenance. Although regular maintenance can provide basic preventive maintenance, it lacks flexibility in response to the actual operating status of the equipment, which may lead to over-maintenance of healthy equipment or failure to detect problem equipment in a timely manner. Maintenance after failure, on the other hand, is likely to prolong the equipment outage time, resulting in greater economic losses and even potentially triggering serious systemic risks.

[0003] Traditional power equipment maintenance mainly relies on regular maintenance and maintenance after failure. This method has problems such as resource waste, insufficient real-time perception ability, and lack of intelligent maintenance strategies. Regular maintenance does not consider the actual operating status of the equipment, which may lead to over-maintenance of healthy equipment and waste of human and material resources. Maintenance after failure lacks monitoring of the real-time operating status of the equipment and cannot detect potential faults in a timely manner, which may lead to the expansion of faults. Maintenance strategies are usually based on experience and lack data-driven intelligent analysis, making it difficult to formulate the optimal maintenance plan in a timely and accurate manner.

[0004] In recent years, the rapid development of digital twin technology has provided a new solution for the intelligent maintenance of power equipment. By constructing a virtual model of a physical device and dynamically synchronizing it with real-time operation data, digital twin can achieve precise monitoring of the device operating status and prediction of potential faults. However, there are still many deficiencies in the actual application of existing technologies. First, in the construction of digital twin models, due to the complex operating environment and diverse fault modes of power equipment, existing models are difficult to fully cover key parameters and specific fault characteristics, resulting in limited prediction accuracy. Second, in the process of real-time data acquisition and synchronization, deficiencies in sensor layout, data transmission delay, and processing capacity limit the timeliness and reliability of digital twin models. In addition, there are still some deficiencies in fault prediction and maintenance plans. Moreover, most existing digital twin systems mainly focus on the device status monitoring function and have relatively weak capabilities in generating intelligent maintenance plans based on prediction results, making it difficult to provide comprehensive support for device operation and maintenance decisions.

[0005] Therefore, it is necessary to develop a predictive maintenance method for power equipment based on digital twin to improve the safety and reliability of power equipment operation, effectively reduce maintenance costs at the same time, and provide a new technical path for the intelligent operation and maintenance of power systems. Summary of the Invention

[0006] The object of the present invention is to provide a digital twin-based predictive maintenance method for power equipment to improve the safety and reliability of power equipment.

[0007] The object of the present invention can be achieved by the following technical solutions:

[0008] A digital twin-based predictive maintenance method for power equipment, comprising the following steps:

[0009] Collect the time-series operation data of each power equipment in the power system in real time and perform preprocessing;

[0010] Based on the preprocessed time-series operation data, the digital twin model integrated with LSTM is updated in real time to obtain the latest digital twin model that runs synchronously with the physical entity of the power equipment;

[0011] Based on the time-series operation data, the isolation forest algorithm and the latest digital twin model are used for fault prediction to obtain the fault prediction result;

[0012] Based on the fault prediction result, determine multiple maintenance plans, evaluate each maintenance plan, obtain the optimal maintenance plan, and complete the predictive maintenance process.

[0013] Further, the time-series operation data includes vibration frequency domain information, temperature change rate, and real-time power. Among them, vibration signals are collected by vibration sensors, and the vibration frequency domain information is extracted by Fourier transform. The calculation expression of the vibration frequency domain information is:

[0014]

[0015] In the formula, V(f) is the amplitude of the vibration signal at frequency f, v(t) is the time-domain signal, representing the time series of the vibration signal, e -j2πft is the complex exponential function, t is the time, and j is the imaginary unit;

[0016] Temperature data is collected by temperature sensors, and the temperature change situation within the time interval is solved as the temperature change rate. The expression of the temperature change rate is:

[0017]

[0018] In the formula, T(t) is the data sequence of the temperature data, and Δt is the time interval;

[0019] Current parameters and voltage parameters are collected by current sensors and voltage sensors, and the real-time power is obtained. The calculation expression of the real-time power is:

[0020] P(t) = U(t)·I(t)

[0021] Wherein, P(t) is the real-time power, U(t) is the voltage parameter, and I(t) is the current parameter.

[0022] Furthermore, the preprocessing step includes denoising, filtering, and feature extraction operations, where the weighted average method is used for denoising, expressed as:

[0023]

[0024] Wherein, x filtered (t) is the sampled value after denoising, N is the number of sampling points, w i is the weighting coefficient, and x i (t) is the sampled value.

[0025] Furthermore, the steps for constructing the digital twin model integrated with LSTM include:

[0026] Obtain the historical operation data and design parameters of the power equipment, and establish a digital twin model of the power equipment;

[0027] Extract the historical data with normal operation from the historical operation data, and train the LSTM network;

[0028] Introduce the trained LSTM network into the digital twin model to obtain the digital twin model integrated with LSTM.

[0029] Furthermore, the steps for obtaining the fault prediction result include:

[0030] Based on the time-series operation data, use the isolation forest algorithm to identify the abnormal data points and normal data points in the current operation state;

[0031] For all normal data points, use the digital twin model integrated with LSTM to perform time-series prediction of the health state, and output the fault prediction result of the normal data points;

[0032] Take the fault prediction results of the abnormal data points and normal data points as the final fault prediction result.

[0033] Furthermore, the steps for identifying the abnormal data points and normal data points in the current operation state include:

[0034] For each data point x in the time-series operation data, randomly select a dimension and the corresponding splitting threshold v, generate a random partitioning rule, and construct multiple independent random trees;

[0035] Calculate the path length h(x) of the data point x through each random tree, where the path length represents the number of splits required for the data point to be isolated;

[0036] Calculate the failure risk score of the data point x according to the expected value E(h(x)) of the path length h(x), where the calculation expression of the failure risk score is:

[0037]

[0038] In the formula, s(x) is the failure risk score of the data point x. The closer s(x) is to 1, the more abnormal the data point x is. On the contrary, it is a normal point. c(n) is the normalization factor, n is the number of samples, and γ is the Euler constant;

[0039] Obtain the abnormal data points and normal data points of the current operating state according to the failure risk score of the data point x.

[0040] Furthermore, the failure prediction result of the normal data point includes the remaining life prediction value, where the expression of the remaining life prediction value is:

[0041] RUL = T failure -T current

[0042] In the formula, RUL is the remaining life prediction value, T failure is the predicted failure occurrence time, and T current is the current time.

[0043] Furthermore, the steps for obtaining the optimal maintenance plan include:

[0044] Based on the failure prediction result, calculate the maintenance priority of the power equipment, where the calculation expression of the maintenance priority is:

[0045]

[0046] In the formula, p i is the maintenance priority of the power equipment i. The larger p i , the more urgent the maintenance requirement of the power equipment i is. w1 and w2 are weight coefficients, satisfying w1 + w2 = 1, s i is the failure risk score of the power equipment i, max(s) is the maximum value of the failure risk score, and RUL i is the remaining life prediction value of the power equipment i;

[0047] According to the maintenance priority, determine multiple maintenance plans in combination with the maintenance content, where the maintenance content includes the maintenance steps of the power equipment, required spare parts, maintenance time window, and maintenance cost;

[0048] Construct an evaluation objective function to evaluate the execution effect of each maintenance plan and obtain the optimal maintenance plan, where the evaluation objective function is expressed as:

[0049] Q = α·E reliability + β·C maintenance + γ·T downtime

[0050] In the formula, Q is the evaluation objective function, α, β, and γ are the weight coefficients of different objectives, which are adjusted according to actual requirements. E reliability is the improvement in the reliability of the power system after the maintenance of the power equipment, C maintenance is the maintenance cost, T downtime is the downtime caused by the maintenance;

[0051] Among them, the evaluation objective function is updated using a deep Q-network to evaluate using the updated evaluation objective function in the next predictive maintenance cycle to obtain the optimal maintenance plan in the next predictive maintenance cycle. The update expression is:

[0052] Q(s,a) ← Q(s,a) + α[r + γmax a' Q(s',a') - Q(s,a)]

[0053] In the formula, Q(s,a) is the value of performing action a in the current state s, α is the learning rate, r is the immediate reward, γ is the discount factor, s’ is the next state, and max a' Q(s',a') represents the optimal value of the next state.

[0054] Furthermore, it also includes iteratively optimizing the latest digital twin model to perform fault prediction on the optimized latest digital twin model in the next predictive maintenance cycle. The steps of the iterative optimization include:

[0055] 1) After the maintenance is completed, collect the operation data of the latest digital twin model when it enters the normal operation state, compare it with the operation data before the maintenance, and calculate the prediction error of the latest digital twin model. The calculation expression of the prediction error is:

[0056]

[0057] In the formula, ∈ is the prediction error, y i is the actual observation value, is the predicted value, and n is the number of samples;

[0058] 2) Determine whether the prediction error exceeds the set threshold. If not, do not update the parameters of the latest digital twin model. If so, update the parameters of the latest digital twin model by minimizing the error function. The parameter update expression is:

[0059]

[0060] where θ t+1 is the parameter at time t+1, i.e., the updated parameter, and θ t is the parameter at time t, i.e., the parameter before update. η is the learning rate, denotes the gradient of the error function L(θ) with respect to the parameter θ;

[0061] 3) During each predictive maintenance cycle, repeat steps 1)-2) for iterative calculation to continuously learn the optimal strategy, and continuously update the latest digital twin model until the end condition is reached.

[0062] Furthermore, it also includes:

[0063] When adding new power equipment to the power system, a cross-device knowledge transfer mechanism is adopted for knowledge transfer to transfer the LSTM, isolation forest algorithm, and reinforcement learning algorithm for the operation of the original power equipment in the power system to the new power equipment, and complete the predictive maintenance process of the new power equipment according to the knowledge transfer result, where the knowledge transfer process is expressed as:

[0064] ω target = λ·ω src +(1 - λ)·ω new

[0065] where ω target is the feature weight of the new power equipment, λ is the transfer factor, ω src is the feature weight of the original power equipment, and ω new is the new feature weight of the new power equipment.

[0066] Compared with the prior art, the present invention has the following beneficial effects:

[0067] (1) The present invention updates the digital twin model integrated with LSTM in real time, enabling the digital twin model to synchronize data with the physical entity of the power equipment, having good real-time performance, and predicting potential faults of the power equipment through the isolation forest algorithm and the latest digital twin model, and maintaining the potential faults by selecting the optimal maintenance plan, thereby improving the safety and reliability of the power equipment.

[0068] (2) The present invention synchronously updates the digital twin model by collecting real-time operation data, enabling the digital twin model to dynamically monitor and accurately evaluate the operation state of the power equipment, which helps to improve the accuracy and real-time performance of power equipment fault prediction.

[0069] (3) The present invention uses a reinforcement learning algorithm to continuously update the evaluation objective function and iteratively optimize the digital twin model, so that in the next predictive maintenance cycle, the optimized digital twin model can be used for fault prediction, thereby improving the accuracy of fault prediction in the next predictive maintenance cycle, and further obtaining more accurate maintenance priorities and optimal maintenance plans, significantly reducing the equipment maintenance cost and optimizing the equipment operation and maintenance efficiency.

[0070] (4) The present invention adopts a cross-device knowledge transfer mechanism, which can, when new power equipment is added to the power system, transfer the operation models of the original power equipment, including LSTM, isolation forest algorithm, reinforcement learning algorithm, etc., to the new power equipment, reducing the training time and data requirements of the new power equipment model, and having certain generality and expansion ability, being applicable to various types of power equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 It is a schematic flow chart of the method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0072] The present invention will be described in detail below with reference to the drawings and specific embodiments. This embodiment is implemented on the premise of the technical solution of the present invention, and gives detailed implementation manners and specific operation processes, but the protection scope of the present invention is not limited to the following embodiments.

[0073] This embodiment provides a power equipment predictive maintenance method based on digital twin, as Figure 1 shown, the method includes the following steps:

[0074] Step 1: Construct a digital twin model of the power equipment.

[0075] In the embodiment of the present invention, first, based on the historical operation data and design parameters of each target power equipment in the power system, its digital twin model is constructed. During the construction process, first, multi-dimensional historical data of the power equipment is collected, including but not limited to key operation parameters such as vibration, temperature, current, and voltage. By combining these historical data with the design parameters of the equipment (such as equipment specifications, working loads, maintenance records, etc.), a virtual model (digital twin model) of the power equipment is established. This virtual model not only includes the physical attributes of the equipment but also reflects the behavioral characteristics of the equipment under different operating conditions. The core part of this digital twin model uses a long short-term memory network (LSTM) to evaluate the health status of the equipment. LSTM is a special recurrent neural network (RNN) that can effectively capture long-term dependencies in time series data and is suitable for processing the time-series operation data of power equipment.

[0076] During the construction of the digital twin model, historical operation data of the device is first collected, including multi-dimensional parameters such as vibration, temperature, voltage, and current. After preprocessing, these data are used as the input of the LSTM. The calculation process of the LSTM unit is as follows:

[0077] (1) Forget Gate: Determines which information in the previous cell state C t-1 needs to be forgotten. The calculation formula is:

[0078] f t = σ(W f ·[h t-1 , x t +b f ) (1)

[0079] Where x t is the current input, h t-1 is the previous hidden state, W f and b f are the weight matrix and bias respectively, and σ represents the Sigmoid activation function.

[0080] (2) Input Gate: Determines which information needs to be written into the cell state at the current moment. It includes two steps:

[0081] ● Calculate the candidate cell state

[0082]

[0083] ● Calculate the activation value i of the input gate t :

[0084] i t = σ(W i ·[h t-1 , x t +b i ) (3)

[0085] Where W C , W i , b C , b i are the weight matrix and bias respectively, and tanh represents the hyperbolic tangent activation function.

[0086] (3) Update the cell state: Combine the outputs of the forget gate and the input gate to update the cell state. The formula is as follows:

[0087]

[0088] Where ⊙ represents element-wise multiplication.

[0089] (4) Output Gate: Determines the hidden state h at the current moment t , that is, the output of the model. The formula is as follows:

[0090] o t = σ(W o · [h t-1 , x t + b o ) (5)

[0091] h t = o t ⊙ tanh(C t ) (6)

[0092] Among them, W o and b o are the weight matrix and bias respectively.

[0093] Through the above steps, the digital twin model integrated with LSTM can effectively capture the temporal features in the operation data of power equipment using LSTM and evaluate the health status of the equipment. During the training process, the method of supervised learning is adopted, and using the known equipment health status labels, the error between the predicted value and the true value is minimized to optimize the model parameters. The constructed digital twin model can reflect the operation status of the equipment in real time, providing a basis for subsequent fault prediction and maintenance strategy formulation.

[0094] Step 2: Deploy a sensor network to collect the real-time operation data of power equipment.

[0095] To achieve real-time monitoring of the operation status of power equipment, this invention deploys a variety of sensors at key parts of the equipment to form a high-precision sensor network. The deployed sensors include vibration sensors, temperature sensors, and current / voltage sensors, etc., specifically as follows:

[0096] (1) Vibration sensor: Installed at key mechanical parts such as the bearings and rotors of the equipment, it monitors the vibration signal in real time to capture the early signs of mechanical faults. The collected vibration signal can be expressed as a time series v(t), and its features are used to extract frequency domain information through Fourier transform:

[0097]

[0098] Among them, V(f) represents the amplitude of the signal at frequency f, and v(t) is the time domain signal.

[0099] (2) Temperature sensor: Arranged at heat-sensitive parts such as the windings and fuel tanks of the equipment, it obtains temperature data in real time to prevent faults caused by overheating. The temperature data can be expressed as a time series T(t), and its rate of change dT / dt is used to monitor abnormal temperature rise:

[0100]

[0101] Among them, Δt represents the time interval, and T(t) is the current temperature.

[0102] (3) Current / voltage sensor: Connected to the input and output ports of the device, it measures electrical parameters in real time and monitors the stability of power transmission. If the current and voltage of the device are I(t) and U(t) respectively, the real-time power P(t) can be calculated by the following formula:

[0103] P(t) = U(t) · I(t) (9)

[0104] To ensure the high precision and real-time nature of the data, the data acquisition frequency of the sensor network is set not less than 100 Hz to meet the dynamic monitoring requirements of key operating parameters. The collected raw data is preprocessed by edge computing devices, including operations such as denoising, filtering, and feature extraction. The denoising process can be carried out by the weighted average method, and the specific formula is:

[0105]

[0106] Among them, x i (t) is the sampling value, w i is the weighting coefficient, and N is the number of sampling points.

[0107] The preprocessed data is transmitted to the digital twin model through the Industrial Internet of Things (IIoT) platform to achieve real-time synchronous update of the physical device and the virtual model. Through the deployment of the above sensor network and the design of the data acquisition process, the present invention realizes the comprehensive perception of the operating state of power equipment and provides reliable data support for subsequent fault prediction and maintenance strategies.

[0108] Step 3: Perform fault prediction based on the digital twin model.

[0109] Based on the constructed digital twin model, the present invention uses the real-time collected data to perform fault prediction on the operating state of the device through deep learning and anomaly detection algorithms.

[0110] First, the real-time data is input into the anomaly detection algorithm to identify the abnormal behaviors of the current operating state of the device. The anomaly detection algorithm adopts the Isolation Forest algorithm, and its core is to determine the anomaly degree of data points by constructing a random forest. The specific steps of the Isolation Forest are as follows:

[0111] (1) For each data point x, randomly select a dimension d and the corresponding splitting threshold v to generate a random partitioning rule and construct multiple independent random trees.

[0112] (2) Calculate the path length h(x) of the data point x in each tree of the random forest. The path length is defined as the number of splits required to isolate the data point.

[0113] (3) Calculate the anomaly score according to the expected value E(h(x)) of the path length:

[0114]

[0115] Among them, c(n) is the normalization factor, defined as:

[0116]

[0117] Among them, γ is the Euler constant and n is the number of samples. The closer the anomaly score s(x) is to 1, the more likely the data point x is to be an anomaly; otherwise, it is a normal point.

[0118] Next, use the long short-term memory network (LSTM) to perform time series prediction on the device health status for the normal operation data, and train and generate a model in combination with the historical operation data. The input data of the LSTM model is the time series data collected historically during the normal operation of the power equipment, and these data include but are not limited to key parameters such as the vibration, temperature, current, and voltage of the equipment. The output is the predicted device operation status. Since the formula of LSTM has been introduced in detail in the first part, it will not be elaborated here. Combining the anomaly detection and prediction results, the fault risk score and the predicted value of the remaining useful life (RUL) of the device are output, as shown in Equation (11) and Equation (13) respectively. The prediction of RUL uses a time series regression model, and the defined formula is as follows:

[0119] RUL = T failure -T current (13)

[0120] Among them, T failure is the predicted fault occurrence time, and T current is the current time.

[0121] Through the above steps, the present invention can accurately judge whether there are potential faults in the device, evaluate the health status of the device, and provide a basis for subsequent maintenance.

[0122] Step 4: Generate an optimal maintenance plan according to the fault prediction result.

[0123] Based on the fault prediction result of the device, the present invention generates intelligent maintenance suggestions, including specific maintenance steps, required spare parts, maintenance time window, and cost analysis. Through the digital twin platform, the influence of different maintenance plans on the device operation status is simulated, the maintenance effect is evaluated, and an optimized maintenance strategy is provided.

[0124] First, according to the fault risk score and the predicted value of the remaining useful life (RUL), determine the maintenance priority of the device. Let the fault risk score of the power equipment i be s i , and its maintenance priority pi It is defined as:

[0125]

[0126] where w1 and w2 are weight coefficients, satisfying w1 + w2 = 1. The maintenance priority p i The larger it is, the more urgent the maintenance requirement of device i is.

[0127] Next, clarify the content involved in the maintenance plan, which specifically includes the following points:

[0128] ● Maintenance steps: Identify the components that need to be repaired and the specific operation steps of the device.

[0129] ● Spare parts required: List the types and quantities of spare parts required for replacement or repair.

[0130] ● Maintenance time window: Combine the power grid operation plan to select the optimal maintenance timing to avoid affecting the normal operation of the system.

[0131] ● Maintenance cost analysis: Calculate the costs of different plans, including spare part costs, labor costs, and downtime losses.

[0132] Finally, optimize the maintenance strategy through a reinforcement learning algorithm. Use a deep Q-network (DQN) to evaluate the value of different maintenance plans. Let the objective function Q for simulation evaluation be:

[0133] Q = α·E reliability + β·C maintenance + γ·T downtime (15)

[0134] where: E reliability represents the improvement in the reliability of the power system after device maintenance; C maintenance represents the maintenance cost; T downtime represents the downtime caused by maintenance; α, β, γ are weight coefficients for different objectives, which are adjusted according to actual needs.

[0135] Its update formula is:

[0136] Q(s,a) ← Q(s,a) + α[r + γmax a' Q(s',a') - Q(s,a)] (16)

[0137] where Q(s,a) is the value of executing action a in the current state s; α is the learning rate, r is the immediate reward; γ is the discount factor, s’ is the next state, and max a' Q(s',a') represents the optimal value of the next state.

[0138] Through the above reinforcement learning framework, the system evaluates the value of different maintenance plans through the Deep Q-Network (DQN) algorithm, and optimizes by combining immediate rewards (such as equipment health recovery, risk reduction, etc.) and long-term benefits (such as extended equipment life, reduced downtime). The selection of maintenance plans takes into account the cost-benefit balance. The system simulates and evaluates the effects of different plans, and finally selects the optimal maintenance plan to achieve equipment health recovery, minimize failure risks, and optimize maintenance costs.

[0139] Finally, the best maintenance suggestions are pushed to the operation and maintenance personnel, and a visualization report is generated. The report includes the current state of the equipment, the results of risk assessment, the recommended maintenance time, and the expected maintenance effects, etc., providing scientific decision-making support for the operation and maintenance personnel. Through the above steps, the present invention realizes the generation of intelligent maintenance suggestions based on fault prediction, providing guarantee for the efficient maintenance of equipment and the reliable operation of the power grid.

[0140] 5. Iteratively optimize the digital twin model based on operation data.

[0141] The present invention continuously optimizes the performance of the digital twin model by continuously collecting and analyzing the real-time operation data of the equipment, making it adapt to the operation characteristics and change laws of different equipment, and improving the accuracy and generality of prediction. This process combines the reinforcement learning algorithm and the cross-device knowledge transfer mechanism to ensure that the model has dynamic adaptability.

[0142] (1) Data feedback and model optimization

[0143] After the maintenance is completed, the equipment enters the normal operation state. During this prediction period, new operation data is collected, including the equipment health status, vibration signals, temperature, etc. These data are compared with the historical data before maintenance, and the prediction error ∈ of the digital twin model is calculated, which is defined as:

[0144]

[0145] where y i is the actual observed value, is the model predicted value, and n is the number of samples. If the error ∈ exceeds the set threshold, the model parameters will be optimized.

[0146] The optimization adopts the gradient descent method, and updates the model parameters θ by minimizing the error function L(θ). The optimization process is as follows:

[0147]

[0148] where η is the learning rate, represents the gradient of the error function with respect to the parameter θ.

[0149] In each prediction cycle, the above steps are repeated to optimize and iterate the digital twin model, so as to predict faults for the optimized and latest digital twin model in the next predictive maintenance cycle.

[0150] (2) Knowledge transfer and model extension

[0151] In a multi-device scenario, the present invention utilizes a cross-device knowledge transfer mechanism to transfer the operation model of one device (i.e., the above-mentioned LSTM model, isolation forest algorithm model, etc.) to other similar devices, reducing the time and data requirements for training new device models. The knowledge transfer is completed through the following steps:

[0152] ● Extract the feature weight ω of the source device src ;

[0153] ● Adjust the transfer factor λ to achieve model adaptation to the target device:

[0154]

[0155] where ω new represents the new feature weight of the target device.

[0156] (3) Model evaluation and cyclic optimization

[0157] The optimized model is evaluated in a virtual environment to simulate the device state under different operating conditions. If the evaluation result meets the set performance requirements, the model is updated to the digital twin platform; otherwise, return to step 1 and continue iterative optimization.

[0158] Through the above steps, the present invention realizes the dynamic optimization of the digital twin model and the improvement of the cross-device adaptation ability, providing reliable technical support for the intelligent operation and maintenance of power equipment.

[0159] The embodiments of the present invention adopt the above method to perform actual operations on a certain power system, including the following steps:

[0160] 1) Obtain the operation data and design parameters of the target power equipment, including key data such as vibration, temperature, voltage, and current of the equipment, and at the same time collect the historical operation status and typical fault characteristics. Through a method combining data-driven and physical modeling, construct a digital twin model of the equipment. The core of the model is a health status evaluation module, which uses a long short-term memory network (LSTM) and is trained in combination with the historical operation data of the equipment to generate a time series prediction of the equipment health status. This model can reflect the operation status of the equipment in real time and provide support for subsequent analysis;

[0161] 2) Deploy a variety of high-precision sensors at key parts of the equipment, including vibration sensors, temperature sensors, and current / voltage sensors, etc. Collect the equipment operation data in real time, and set the data collection frequency not less than 100Hz to ensure high-precision dynamic monitoring of key operation parameters. The collected data is preprocessed by edge computing devices, including denoising, filtering, and feature extraction, and then transmitted to the digital twin model through the Industrial Internet of Things (IIoT) platform to achieve real-time synchronization between the physical device and the virtual model;

[0162] 3) Analyze the data collected in real time using the digital twin model, and identify potential risks in the current operation state of the equipment through anomaly detection algorithms. The anomaly detection algorithm adopts the Isolation Forest algorithm, and combines the historical operation state of the equipment to evaluate the health status score of the equipment. At the same time, the model outputs the fault risk score and the predicted value of the remaining useful life (RUL) of the equipment. Based on the fault prediction results, it can accurately judge whether there are potential faults in the equipment and provide a basis for status evaluation;

[0163] 4) Generate intelligent maintenance suggestions based on the fault prediction results, including specific maintenance steps, required spare parts, maintenance time window, and cost analysis. Combine the simulation evaluation of multiple maintenance plans, dynamically optimize the maintenance strategy, and select the optimal maintenance plan to reduce the maintenance cost and reduce the downtime. The maintenance suggestions are displayed through a visual report, including the current state of the equipment, risk score, recommended maintenance time, and expected maintenance effect, etc.;

[0164] 5) Continuously optimize the digital twin model through the equipment operation data collected in real time. The model optimization combines the reinforcement learning algorithm, and continuously adjusts the model parameters by minimizing the prediction error. Further, use the cross-device knowledge transfer mechanism to migrate the model of the existing equipment to the same type of equipment to quickly adapt to the operation characteristics of the new equipment. The optimized model is tested and evaluated in a virtual environment to ensure that its prediction accuracy meets the actual application requirements.

[0165] In summary, the present invention takes digital twin technology as the core, combines advanced technical means such as LSTM, Isolation Forest algorithm, and reinforcement learning, and proposes a predictive maintenance method for power equipment. By constructing a digital twin model of power equipment, deploying a high-precision sensor network to collect equipment operation data in real time, using deep learning and anomaly detection algorithms for fault prediction, and generating optimized maintenance suggestions based on the prediction results. At the same time, through the reinforcement learning and cross-device knowledge transfer mechanism, the digital twin model is iteratively optimized to continuously improve the generality and prediction accuracy of the model. The purpose of the present invention is to improve the perception ability of the operation state of power equipment, optimize the maintenance strategy, reduce the operation and maintenance cost, and improve the reliability and safety of the power system.

[0166] Although the preferred embodiments of the present invention have been described, additional changes and modifications can be made by those skilled in the art once they learn of the basic creative concept. Therefore, the appended claims are intended to be construed to include the preferred embodiments as well as all changes and modifications that fall within the scope of the present invention.

[0167] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalent technologies, the present invention is also intended to include these modifications and variations.

Claims

1. A predictive maintenance method for power equipment based on digital twin, characterized in that, It includes the following steps: Collect the time-series operation data of each power device in the power system in real time and perform preprocessing; Based on the preprocessed time-series operation data, the digital twin model integrated with LSTM is updated in real time to obtain the latest digital twin model that runs synchronously with the physical entity of the power device; Based on the time-series operation data, the isolation forest algorithm and the latest digital twin model are used for fault prediction to obtain the fault prediction result; Based on the fault prediction result, multiple maintenance plans are determined, and each maintenance plan is evaluated to obtain the optimal maintenance plan, completing the predictive maintenance process.

2. The predictive maintenance method for power equipment based on digital twin according to claim 1, characterized in that The time-series operation data includes vibration frequency domain information, temperature change rate, and real-time power. Among them, vibration signals are collected by vibration sensors, and the vibration frequency domain information is extracted by Fourier transform. The calculation expression of the vibration frequency domain information is: Wherein, V(f) is the amplitude of the vibration signal at frequency f, v(t) is the time-domain signal representing the time series of the vibration signal, e -j2πft is the complex exponential function, t is the time, and j is the imaginary unit; Temperature data is collected by temperature sensors, and the temperature change situation within the time interval is solved as the temperature change rate. The expression of the temperature change rate is: In the formula, T(t) is the data sequence of temperature data, and Δt is the time interval; Current parameters and voltage parameters are collected by current sensors and voltage sensors, and the real-time power is obtained. The calculation expression of the real-time power is: P(t) = U(t)·I(t) In the formula, P(t) is the real-time power, U(t) is the voltage parameter, and I(t) is the current parameter.

3. A predictive maintenance method for power equipment based on digital twin according to claim 1, characterized in that, The preprocessing steps include denoising, filtering, and feature extraction operations. Among them, the weighted average method is used for denoising, which is expressed as: where x filtered (t) is the sampled value after denoising, N is the number of sampling points, w i is the weighting coefficient, and x i (t) is the sampled value.

4. A predictive maintenance method for power equipment based on digital twin according to claim 1, characterized in that, The steps for constructing the digital twin model integrated with LSTM include: Obtain the historical operation data and design parameters of the power device, and establish the digital twin model of the power device; Extract the historical data with normal operation from the historical operation data and train the LSTM network; Introduce the trained LSTM network into the digital twin model to obtain the digital twin model integrated with LSTM.

5. A predictive maintenance method for power equipment based on digital twin according to claim 1, characterized in that The steps for obtaining the fault prediction result include: Based on the time-series operation data, the isolation forest algorithm is used to identify abnormal data points and normal data points in the current operation state; For all normal data points, the digital twin model integrated with LSTM is used for time-series prediction of the health state, and the fault prediction result of the normal data points is output; The fault prediction results of the abnormal data points and normal data points are used as the final fault prediction result.

6. A predictive maintenance method for power equipment based on digital twin according to claim 5, characterized in that, The steps for identifying abnormal data points and normal data points in the current operation state include: For each data point x in the time-series operation data, randomly select a dimension and the corresponding splitting threshold v, generate a random partitioning rule, and construct multiple independent random trees; Calculate the path length h(x) of the data point x through each random tree, where the path length represents the number of splits required for the data point to be isolated; According to the expected value E(h(x)) of the path length h(x), calculate the fault risk score of the data point x. The calculation expression of the fault risk score is: Where, s(x) is the fault risk score of data point x. The closer s(x) is to 1, the more abnormal the data point x is. On the contrary, it is a normal point. c(n) is the normalization factor, n is the number of samples, and γ is the Euler constant; Based on the fault risk score of the data point x, abnormal data points and normal data points in the current operating state are obtained.

7. A predictive maintenance method for power equipment based on digital twin according to claim 5, characterized in that, The fault prediction result of the normal data point includes the predicted remaining life value, and the expression of the predicted remaining life value is: RUL = T failure -T current Wherein, RUL is the predicted remaining useful life, T failure is the predicted fault occurrence time, and T current is the current time.

8. A predictive maintenance method for power equipment based on digital twin according to claim 1, characterized in that, The steps of obtaining the optimal maintenance plan include: Based on the fault prediction result, calculate the maintenance priority of the power equipment, and the calculation expression of the maintenance priority is: Where, p i is the maintenance priority of power equipment i. The larger p i is, the more urgent the maintenance requirement of power equipment i is. w1 and w2 are weight coefficients, satisfying w1 + w2 = 1. s i is the failure risk score of power equipment i, max(s) is the maximum value of the failure risk score, and RUL i is the predicted remaining life value of power equipment i; According to the maintenance priority, determine multiple maintenance plans in combination with the maintenance content, where the maintenance content includes the maintenance steps of the power equipment, required spare parts, maintenance time window, and maintenance cost; Construct an evaluation objective function to evaluate the execution effect of each maintenance plan, and obtain the optimal maintenance plan. The evaluation objective function is expressed as: Q = α·E reliability + β·C maintenance + γ·T downtime Where Q is the evaluation objective function, α, β, and γ are the weight coefficients of different objectives, which are adjusted according to actual requirements, and E reliability is the improvement in the reliability of the power system after the maintenance of power equipment, and C maintenance is the maintenance cost, and T downtime is the downtime caused by maintenance; Among them, the evaluation objective function is updated using a deep Q-network, and in the next predictive maintenance cycle, the updated evaluation objective function is used for evaluation to obtain the optimal maintenance plan in the next predictive maintenance cycle. The update expression is: Q(s,a) ← Q(s,a) + α[r + γ max a' Q(s',a') - Q(s,a)] Wherein, Q(s,a) is the value of executing action a in the current state s, α is the learning rate, r is the immediate reward, γ is the discount factor, s’ is the next state, and max a' Q(s',a') represents the optimal value of the next state.

9. A predictive maintenance method for power equipment based on digital twin according to claim 1, characterized in that, It also includes iterative optimization of the latest digital twin model to perform fault prediction on the optimized latest digital twin model in the next predictive maintenance cycle. The steps of the iterative optimization include: 1) After the maintenance is completed, collect the operation data of the latest digital twin model when it enters the normal operation state, compare it with the operation data before the maintenance, and calculate the prediction error of the latest digital twin model. The calculation expression of the prediction error is: where ∈ is the prediction error, y i is the actual observed value, is the predicted value, and n is the number of samples; 2) Determine whether the prediction error exceeds the set threshold. If not, do not update the parameters of the latest digital twin model. If so, update the parameters of the latest digital twin model by minimizing the error function. The parameter update expression is: θ t+1 = θ t -η·▽ θ L(θ) where θ t+1 is the parameter at t + 1, that is, the updated parameter, and θ t is the parameter at t, that is, the parameter before update. η is the learning rate, and ▽ θ L(θ) represents the gradient of the error function L(θ) with respect to the parameter θ; 3) In each predictive maintenance cycle, repeat steps 1)-2) for iterative calculation to continuously learn the optimal strategy, and the latest digital twin model is continuously updated until the end condition is reached.

10. A predictive maintenance method for power equipment based on digital twin according to claim 9, characterized in that, It also includes: When adding new power equipment to the power system, a cross-device knowledge transfer mechanism is used for knowledge transfer to transfer the LSTM, isolation forest algorithm, and reinforcement learning algorithm of the original power equipment operation in the power system to the new power equipment, and complete the predictive maintenance process of the new power equipment according to the knowledge transfer result. The knowledge transfer process is expressed as: ω target = λ·ω src +(1 - λ)·ω new where ω target is the feature weight of the newly added power equipment, λ is the migration factor, and ω src is the feature weight of the original power equipment, and ω new is the new feature weight of the newly added power equipment.

Images