Power equipment fault prediction method based on multi-source heterogeneous data and twinborn model

Abstract

The invention relates to a power equipment fault prediction method based on multi-source heterogeneous data and a twinborn model, and the method comprises the following steps: S1, obtaining multi-modal data related to the operation of power equipment, constructing a joint feature encoder through an attention mechanism in deep learning, and carrying out the joint modeling of multi-modal information, and obtaining a joint feature representation; s2, based on joint feature representation, adopting a dynamic twin model of self-supervised learning and fusion reinforcement learning, and dynamically updating the equipment health state; s3, constructing a complete-cycle prediction framework based on Transform, capturing long-term dependency and dynamic change characteristics in equipment operation, and identifying trigger factors of faults in combination with causal relationship analysis; and S4, based on the equipment health state and the fault prediction result, optimizing equipment operation parameters by adopting an intelligent operation and maintenance optimization engine, and dynamically generating an optimal operation and maintenance plan in combination with a genetic algorithm. The reliability and the operation efficiency of the power equipment are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of fault prediction, and in particular to a method for predicting faults of electric power equipment based on multi-source heterogeneous data and a twin model. Background Art

[0002] With the acceleration of global energy transformation and smart grid construction, large-scale access to new energy has significantly changed the operating characteristics of the power system. The intermittent and random nature of new energy generation not only exacerbates power load fluctuations, but also poses a great challenge to the operational stability of power equipment. At the same time, traditional power equipment mainly relies on fixed-cycle maintenance and empirical judgment operation management methods, which can no longer meet the needs of modern power systems for efficient, reliable and economical operation of equipment.

[0003] During long-term operation, power equipment will deteriorate due to environmental impacts, load fluctuations, material aging and other reasons, thus causing failures. The occurrence of failures often has complex triggering mechanisms and time-space correlation characteristics, and the operating status and health of equipment are subject to dynamic, nonlinear and multi-factor constraints. Therefore, how to comprehensively and accurately evaluate the health status of equipment in real time and predict potential failures in advance is of great significance to ensure the continuous and stable operation of the power system. Summary of the invention

[0004] In order to solve the above problems, the purpose of the present invention is to provide a method for predicting power equipment faults based on multi-source heterogeneous data and twin models, which realizes the full-cycle closed-loop optimization of power equipment fault prediction and operation and maintenance management, and significantly improves the reliability and operational efficiency of power equipment.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] A method for predicting power equipment faults based on multi-source heterogeneous data and twin models, comprising the following steps:

[0007] S1: Obtain multimodal data related to the operation of power equipment, including sensor data, maintenance records and spatial data, use the attention mechanism in deep learning to build a joint feature encoder, and jointly model the multimodal information of sensor data, maintenance records and spatial data to obtain a joint feature representation;

[0008] S2: Based on joint feature representation, a dynamic twin model that combines self-supervised learning and reinforcement learning is adopted to enable the twin to adapt to the device operating status and environmental changes in real time, and dynamically update the device health status;

[0009] S3: Build a full-cycle prediction framework based on Transformer to capture the long-term dependencies and dynamic change characteristics of equipment operation, combine causal relationship analysis, identify the triggering factors of faults, and enhance the accuracy and explainability of predictions;

[0010] S4: Based on the equipment health status and fault prediction results, an intelligent operation and maintenance optimization engine is used to optimize equipment operating parameters, and combined with genetic algorithms, the optimal operation and maintenance plan is dynamically generated.

[0011] Furthermore, S1 is specifically:

[0012] Define the time series collected by the sensor as a matrix:

[0013] X ts ∈R N×T×F ;

[0014] Where N is the number of samples; T is the time step; F is the number of sensor features;

[0015] Define the maintenance record as the embedding sequence after word segmentation:

[0016] X text ={w1,w2,...,w a ,...,w L};

[0017] Among them, L is the length of the word segmentation of the text; w a Indicates the ath word;

[0018] Define the GIS spatial information associated with each device:

[0019] X gis =(P,A);

[0020] P is the geographical coordinates of the device; A is the adjacency matrix between devices, indicating the connection relationship between devices;

[0021] For each type of data, different neural network models are used to extract unimodal features, and based on the extraction of separate features of each modality, cross-modal attention mechanism is used for joint modeling.

[0022] Furthermore, for each type of data, different neural network models are used to extract unimodal features, and on the basis of extracting individual features of each modality, a cross-modal attention mechanism is used for joint modeling, as follows:

[0023] The time series collected by the sensor is modeled based on the Transformer time series model to capture the global dependency of the time series. The position code P is superimposed on the input sequence to form the input:

[0024] Z ts =Xts +P;

[0025] The attention mechanism is used to model the temporal dependency. The attention formula is:

[0026]

[0027] Where Q = Z ts W Q ,K=Z ts W K ,V=Z ts W V Represents query, key and value respectively; W Q , W K and W V , is the trainable weight matrix;

[0028] After merging multiple outputs, we get the time sequence features:

[0029]

[0030] Where H is the number of attention mechanism heads; W o is the weight coefficient;

[0031] The BERT model is used to process the embedded sequence after word segmentation and extract semantic information:

[0032] H text =BERT(X text );

[0033] Add a self-attention layer to enhance key semantic extraction and obtain text features:

[0034] F text =Attention(H text , H text , H text );

[0035] Using graph neural network to extract spatial dependency features between devices gis ;

[0036] Concatenate time series, text, and spatially dependent features:

[0037] F cat =Concat(F ts , F text , F gis );

[0038] Get the joint feature representation:

[0039] Z multi =F cat .

[0040] Furthermore, based on the data after joint modeling of multimodal information, a dynamic twin model that integrates self-supervised learning and reinforcement learning is used to enable the twin to adapt to the equipment operating status and environmental changes in real time, and dynamically update the equipment health status, as follows:

[0041] Through self-supervised learning, a reference benchmark for equipment operating status is constructed to detect and quantify health status deviations; time series reconstruction based on joint feature representation:

[0042]

[0043] in, is the predicted value at time t+1; Z multi,t is the joint feature representation at time t; f SSL is the self-supervised learning model function

[0044] The reconstruction loss is calculated using the mean squared error:

[0045]

[0046] Among them, X ts,t+1 is the actual value at time t+1;

[0047] Construct positive and negative sample pairs between normal and abnormal states, and optimize through contrast loss:

[0048]

[0049] sim() is cosine similarity; Z i is the feature vector representation of the current sample i; Z j is the feature vector representation of the positive sample of the current sample i; Z k For comparison samples;

[0050] The ultimate goal of self-supervised learning is:

[0051]

[0052] Among them, λ1,λ2 are weight coefficients;

[0053] Reinforcement learning is used to enhance the dynamic adaptability of the twin, so that it can be updated in real time according to environmental changes. t Current device operation status

[0054] S t =(Z multi,t ,HS t ,RUL t );

[0055] Among them, HS t For health status prediction, RULt for the remaining life;

[0056] Action A t Update the model parameters of the twin:

[0057] A t =Δθt;

[0058] Reward R t , rewards based on equipment health, failure risk, and performance:

[0059] R t =-α(1-HS t ) 2 -βP fail,t +γη t ;

[0060] Among them, α, β, and γ are weight coefficients; P fail,t is the failure risk; t For operational efficiency;

[0061] Deep reinforcement learning DDPG is used to optimize the twin, with the goal of maximizing the cumulative reward:

[0062]

[0063] Through the policy network π θ (St) Output action, the goal is:

[0064]

[0065] Health status prediction:

[0066] HS t =σ(f HS (Z multi,t ));

[0067] Combined with the adjustment action output by reinforcement learning, f HS It is the health status prediction model function.

[0068] Furthermore, a full-cycle prediction framework based on Transformer is constructed to capture the long-term dependencies and dynamic change characteristics of equipment operation. Specifically:

[0069] The joint feature is represented as Z multi Input to the Transformer model and mapped to the input dimension of the Transformer through a linear layer:

[0070] Z emb =Z multi W emb +b emb ;

[0071] Among them, W emb and b emb are weight and bias respectively;

[0072] Add position encoding P to time series features E , capturing the time sequence information, the final input is:

[0073] Z input =Z emb +P E ;

[0074] Using stacked Transformer encoders to model long-term dependencies in time series, the output of the Transformer is:

[0075] H t = TransformerEnc(Z input，t );

[0076] Through the fully connected layer and the Sigmoid activation function, the failure probability of the future time step is output:

[0077]

[0078] Where Ht is the hidden state vector at time step t; W fail ,b fail Weights and biases for failure probability prediction;

[0079] Predict potential fault categories through classification networks:

[0080]

[0081] Among them, W class ,b class Weights and bias terms predicted for potential fault classes;

[0082] Remaining life estimation, regression prediction through a fully connected layer:

[0083]

[0084] Among them, W RUL ,b RUL Weights and biases for remaining lifetime prediction.

[0085] Furthermore, combined with causal analysis, the triggering factors of the fault are identified to enhance the accuracy and explainability of the prediction, as follows:

[0086] Based on domain knowledge and data-driven methods, we build a causal relationship graph between multimodal features:

[0087] Node: input feature; Edge: causal relationship between features;

[0088] The causal effect of key features on the probability of failure is calculated using causal inference methods:

[0089]

[0090] Among them, Causal Effect (x i →P fail ) describes the variable x i The failure probability P fail causal effect;

[0091] Through causal analysis, the weights of input features are optimized to enhance the prediction accuracy and interpretability of the model.

[0092] Furthermore, based on the equipment health status and fault prediction results, an intelligent operation and maintenance optimization engine is used to optimize equipment operating parameters, as follows:

[0093] Dynamically monitor the health status of equipment, update the prediction model in real time, and obtain the latest prediction data, including

[0094] Integrate dynamic data to build the global state of the system, which is used to optimize operation and maintenance strategies and predict equipment status representation

[0095]

[0096] Determine the optimal operating parameters A t ′, so as to optimize the following objective function:

[0097]

[0098] Among them, Cost(A t ′) The operating cost of the equipment under the current parameters; a is the weight coefficient;

[0099] Design goal-related reward function:

[0100] R′=R HS +R fail +R cost ;

[0101] Among them, R HS ,R fail ,R cost They are equipment life bonus, failure rate bonus and cost bonus;

[0102] Use the DDPG algorithm based on deep reinforcement learning to obtain the optimal equipment operating parameters.

[0103] Furthermore, combined with genetic algorithms, the optimal operation and maintenance plan is dynamically generated, as follows:

[0104] Dynamically generate maintenance scheduling plans based on equipment health status and predicted remaining life, optimizing the following objectives:

[0105]

[0106] Among them, C maint (i) is the cost of the ii-th maintenance; P fail (i) is the residual probability of failure during maintenance ii; C fail (i) is the cost after the failure occurs; λ4 is the weight, which controls the balance between maintenance and failure costs;

[0107] Each individual represents the maintenance time and maintenance type of all equipment, and the quality of the maintenance plan is estimated by the fitness function: Fitness = -J maint , optimize maintenance costs and failure risks

[0108] Select excellent individuals according to the fitness function;

[0109] Generate children by exchanging time or maintenance type;

[0110] Randomly change the maintenance time or adjust the maintenance type to explore the solution space; continuously iterate and finally output the optimal maintenance scheduling plan.

[0111] A power equipment fault prediction system based on multi-source heterogeneous data and twin models, characterized in that it includes a processor, a memory and a computer program stored in the memory. When the processor executes the computer program, it specifically executes the steps in the power equipment fault prediction method based on multi-source heterogeneous data and twin models as described above.

[0112] The present invention has the following beneficial effects:

[0113] 1. The present invention realizes the full-cycle closed-loop optimization of power equipment fault prediction and operation and maintenance management, significantly improving the reliability and operation efficiency of power equipment;

[0114] 2. The self-supervised learning of the present invention allows the twin model to autonomously learn key feature representations through the historical status and environmental information of the equipment operation without a large amount of manually labeled data, avoiding the dependence of traditional supervised models on large-scale high-quality fault labeling data, reducing the model training cost, and improving the generalization ability. Through reinforcement learning, the twin model is given the ability to dynamically optimize parameters, so that it can adaptively adjust the model weights when the equipment operating environment changes, and maintain prediction accuracy;

[0115] 3. The present invention forms an active optimization closed loop through a multi-level feedback mechanism to achieve continuous iteration of "prediction-optimization-feedback". Short-term feedback: dynamically adjust equipment operating parameters (such as load, voltage, etc.) through prediction results; long-term feedback: combine predicted health status and historical maintenance records, and use genetic algorithms (GA) to dynamically generate maintenance priorities and time points to achieve a global optimal solution for maintenance costs and failure risks, generate long-term operation and maintenance optimization plans, and dynamically adjust maintenance cycles. BRIEF DESCRIPTION OF THE DRAWINGS

[0116] Figure 1 The figure is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0117] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0118] refer to Figure 1 In this embodiment, a method for predicting power equipment faults based on multi-source heterogeneous data and twin models is provided, comprising the following steps:

[0119] S1: Obtain multimodal data related to the operation of power equipment, including sensor data, maintenance records and spatial data (GIS geographic information), use the attention mechanism in deep learning to build a joint feature encoder, and jointly model the multimodal information of sensor data, maintenance records and spatial data to obtain a joint feature representation;

[0120] S2: Based on joint feature representation, a dynamic twin model that combines self-supervised learning and reinforcement learning is adopted to enable the twin to adapt to the device operating status and environmental changes in real time, and dynamically update the device health status;

[0121] S3: Build a full-cycle prediction framework based on Transformer to capture the long-term dependencies and dynamic change characteristics of equipment operation, combine causal relationship analysis, identify the triggering factors of faults, and enhance the accuracy and explainability of predictions;

[0122] S4: Based on the equipment health status and fault prediction results, an intelligent operation and maintenance optimization engine is used to optimize equipment operating parameters, and combined with genetic algorithms, the optimal operation and maintenance plan is dynamically generated.

[0123] In this embodiment, S1 is specifically:

[0124] Define the time series collected by the sensor as a matrix:

[0125] X ts ∈R N×T×F ;

[0126] Where N is the number of samples; T is the time step; F is the number of sensor features (such as temperature, current, voltage, etc.);

[0127] Define the maintenance record as the embedding sequence after word segmentation:

[0128] X text ={w1,w2,...,w a ,...,w L};

[0129] Among them, L is the length of the word segmentation of the text; w a Indicates the ath word;

[0130] Define the GIS spatial information associated with each device:

[0131] X gis =(P,A);

[0132] P is the geographical coordinates (latitude and longitude) of the device; A is the adjacency matrix between devices, indicating the connection relationship between devices;

[0133] For each type of data, different neural network models are used to extract unimodal features, and based on the extraction of separate features of each modality, cross-modal attention mechanism is used for joint modeling.

[0134] In this embodiment, for each type of data, different neural network models are used to extract single-modal features, and on the basis of extracting separate features of each modality, a cross-modal attention mechanism is used for joint modeling, as follows:

[0135] The time series collected by the sensor is modeled based on the Transformer time series model to capture the global dependency of the time series. The position code P is superimposed on the input sequence to form the input:

[0136] Z ts =X ts +P;

[0137] The attention mechanism is used to model the temporal dependency. The attention formula is:

[0138]

[0139] Where Q = Z ts W Q ,K=Z ts W K ,V=Z ts W V Represents query, key and value respectively; W Q , W K and W V , is the trainable weight matrix;

[0140] After merging multiple outputs, we get the time sequence features:

[0141]

[0142] Where H is the number of attention mechanism heads; W o is the weight coefficient;

[0143] The BERT model is used to process the embedded sequence after word segmentation and extract semantic information:

[0144] H text =BERT(X text );

[0145] Add a self-attention layer to enhance key semantic extraction and obtain text features:

[0146] F text =Attention(H text , H text , H text );

[0147] Using graph neural network to extract spatial dependency features between devices gis ;

[0148] Concatenate time series, text, and spatially dependent features:

[0149] F cat =Concat(F ts , F text , F gis );

[0150] Get the joint feature representation:

[0151] Z multi =F cat .

[0152] In this embodiment, based on the data after joint modeling of multimodal information, a dynamic twin model of self-supervised learning and fusion reinforcement learning is adopted to enable the twin to adapt to the equipment operation status and environmental changes in real time, and dynamically update the equipment health status, as follows:

[0153] Through self-supervised learning, a reference benchmark for equipment operating status is constructed to detect and quantify health status deviations; time series reconstruction based on joint feature representation:

[0154]

[0155] in, is the predicted value at time t+1; Z multi,t is the joint feature representation at time t; f SSL is the self-supervised learning model function

[0156] The reconstruction loss is calculated using the mean squared error:

[0157]

[0158] Among them, X ts,t+1 is the actual value at time t+1;

[0159] Construct positive and negative sample pairs between normal and abnormal states, and optimize through contrast loss:

[0160]

[0161] sim() is cosine similarity; Z i is the feature vector representation of the current sample i; Z j is the feature vector representation of the positive sample of the current sample i; Z k For comparison samples;

[0162] The ultimate goal of self-supervised learning is:

[0163]

[0164] Among them, λ1,λ2 are weight coefficients;

[0165] Reinforcement learning is used to enhance the dynamic adaptability of the twin, so that it can be updated in real time according to environmental changes. t Current device operation status

[0166] S t =(Z multi,t ,HS t ,RUL t );

[0167] Among them, HS t For health status prediction, RUL t for the remaining life;

[0168] Action A t Update the model parameters of the twin:

[0169] A t =Δθt;

[0170] Reward R t , rewards based on equipment health, failure risk, and performance:

[0171] R t =-α(1-HS t ) 2 -βP fail，t +γη t ;

[0172] Among them, α, β, and γ are weight coefficients; Pfail,t is the failure risk; t For operational efficiency;

[0173] Deep reinforcement learning DDPG is used to optimize the twin, with the goal of maximizing the cumulative reward:

[0174]

[0175] Through the policy network π θ (S t ) Output action (update parameters), the goal is:

[0176]

[0177] Health status prediction:

[0178] HS t =σ(f HS (Z multi,t ));

[0179] Combined with the adjustment action output by reinforcement learning, f HS It is the health status prediction model function.

[0180] In this embodiment, a full-cycle prediction framework based on Transformer is constructed to capture the long-term dependencies and dynamic change characteristics of equipment operation, specifically:

[0181] The joint feature is represented as Z multi Input to the Transformer model and mapped to the input dimension of the Transformer through a linear layer:

[0182] Z emb =Z multi W emb +b emb ;

[0183] Among them, W emb and b emb are weight and bias respectively;

[0184] Add position encoding P to time series features E , capturing the time sequence information, the final input is:

[0185] Z input =Z emb +P E ;

[0186] Using stacked Transformer encoders to model long-term dependencies in time series, the output of the Transformer is:

[0187] H t= TransformerEnc(Z input，t );

[0188] Through the fully connected layer and the Sigmoid activation function, the failure probability of the future time step is output:

[0189]

[0190] Where Ht is the hidden state vector at time step t; W fail ,b fail Weights and biases for failure probability prediction;

[0191] Predict potential fault categories through classification networks:

[0192]

[0193] Among them, W class ,b class Weights and bias terms predicted for potential fault classes;

[0194] Remaining life estimation, regression prediction through a fully connected layer:

[0195]

[0196] Among them, W RUL ,b RUL Weights and biases for remaining lifetime prediction.

[0197] In this embodiment, causal relationship analysis is combined to identify the triggering factors of the fault and enhance the accuracy and explainability of the prediction, as follows:

[0198] Based on domain knowledge and data-driven methods, we build a causal relationship graph between multimodal features:

[0199] Node: input features (such as temperature, current, environmental data, etc.); Edge: causal relationship between features;

[0200] The causal effect of key features on the probability of failure is calculated using causal inference methods:

[0201]

[0202] Among them, Causal Effect (x i →P fail ) describes the variable x i The failure probability P fail causal effect;

[0203] Through causal analysis, the weights of input features are optimized to enhance the prediction accuracy and interpretability of the model.

[0204] In this embodiment, based on the equipment health status and fault prediction results, an intelligent operation and maintenance optimization engine is used to optimize equipment operating parameters, as follows:

[0205] Dynamically monitor the health status of equipment, update the prediction model in real time, and obtain the latest prediction data, including

[0206] Integrate dynamic data to build the global state of the system, which is used to optimize operation and maintenance strategies and predict equipment status representation

[0207]

[0208] Determine the optimal operating parameters A t ′ (such as load, speed, voltage, etc.), to optimize the following objective function:

[0209]

[0210] Among them, Cost(A t ′) The operating cost of the equipment under the current parameters; a is the weight coefficient;

[0211] Design goal-related reward function:

[0212] R′=R HS +R fail +R cost ;

[0213] Among them, R HS ,R fail ,R cost They are equipment life bonus, failure rate bonus and cost bonus;

[0214] Use the DDPG algorithm based on deep reinforcement learning to obtain the optimal equipment operating parameters.

[0215] In this embodiment, the optimal operation and maintenance plan is dynamically generated by combining the genetic algorithm, as follows:

[0216] Dynamically generate maintenance scheduling plans based on equipment health status and predicted remaining life, optimizing the following objectives:

[0217]

[0218] Among them, C maint (i) is the cost of the ii-th maintenance; P fail (i) is the residual probability of failure during maintenance ii; C fail (i) is the cost after the failure occurs; λ4 is the weight, which controls the balance between maintenance and failure costs;

[0219] Each individual represents the maintenance time and maintenance type of all equipment, and the quality of the maintenance plan is estimated by the fitness function: Fitness = -J maint , optimize maintenance costs and failure risks

[0220] Select excellent individuals according to the fitness function;

[0221] Generate children by exchanging time or maintenance type;

[0222] Randomly change the maintenance time or adjust the maintenance type to explore the solution space; continuously iterate and finally output the optimal maintenance scheduling plan.

[0223] A power equipment fault prediction system based on multi-source heterogeneous data and twin models, characterized in that it includes a processor, a memory and a computer program stored in the memory. When the processor executes the computer program, it specifically executes the steps in the power equipment fault prediction method based on multi-source heterogeneous data and twin models as described above.

[0224] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0225] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0226] These computer program instructions may also be stored in a computer readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1A function specified in one or more boxes.

[0227] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0228] The above is only a preferred embodiment of the present invention, and does not limit the present invention in other forms. Any technician familiar with the profession may use the above disclosed technical content to change or modify it into an equivalent embodiment with equivalent changes. However, any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present invention without departing from the technical solution of the present invention still belongs to the protection scope of the technical solution of the present invention.

Claims

1. A method for predicting power equipment faults based on multi-source heterogeneous data and twin models, characterized in that: The following steps are involved: S1: Obtain multimodal data related to the operation of power equipment, including sensor data, maintenance records and spatial data, use the attention mechanism in deep learning to build a joint feature encoder, and jointly model the multimodal information of sensor data, maintenance records and spatial data to obtain a joint feature representation; S2: Based on joint feature representation, a dynamic twin model that combines self-supervised learning and reinforcement learning is adopted to enable the twin to adapt to the device operating status and environmental changes in real time, and dynamically update the device health status; S3: Build a full-cycle prediction framework based on Transformer to capture the long-term dependencies and dynamic change characteristics of equipment operation, combine causal relationship analysis, identify the triggering factors of faults, and enhance the accuracy and explainability of predictions; S4: Based on the equipment health status and fault prediction results, an intelligent operation and maintenance optimization engine is used to optimize equipment operating parameters, and combined with genetic algorithms, the optimal operation and maintenance plan is dynamically generated.

2. The method for predicting power equipment faults based on multi-source heterogeneous data and twin models according to claim 1 is characterized in that: The S1 is specifically: Define the time series collected by the sensor as a matrix: X ts ∈R N×T×F ; Where N is the number of samples; T is the time step; F is the number of sensor features; Define the maintenance record as the embedding sequence after word segmentation: X text ={w1,w2,...,w a ,...,w L }; Among them, L is the length of the word segmentation of the text; w a Indicates the ath word; Define the GIS spatial information associated with each device: X gis =(P,A); P is the geographical coordinates of the device; A is the adjacency matrix between devices, indicating the connection relationship between devices; For each type of data, different neural network models are used to extract unimodal features, and based on the extraction of separate features of each modality, cross-modal attention mechanism is used for joint modeling.

3. The method for predicting power equipment faults based on multi-source heterogeneous data and twin models according to claim 2 is characterized in that: For each type of data, different neural network models are used to extract single-modal features, and based on the extraction of individual features of each modality, a cross-modal attention mechanism is used for joint modeling, as follows: The time series collected by the sensor is modeled based on the Transformer time series model to capture the global dependency of the time series. The position code P is superimposed on the input sequence to form the input: Z ts =X ts +P; The attention mechanism is used to model the temporal dependency. The attention formula is: Where Q = Z ts W Q ,K=Z ts W K ,V=Z ts W V Represents query, key and value respectively; W Q , W K and W V , is the trainable weight matrix; After merging multiple outputs, we get the time sequence features: Where H is the number of attention mechanism heads; W o is the weight coefficient; The BERT model is used to process the embedded sequence after word segmentation and extract semantic information: H text =BERT(X text );; Add a self-attention layer to enhance key semantic extraction and obtain text features: F text =Attention(H text ,H text ,H text ); Using graph neural network to extract spatial dependency features between devices gis ; Concatenate time series, text, and spatially dependent features: F cat =Concat(F ts ,F text ,F gis ); Get the joint feature representation: Z multi =F cat 。 4. The method for predicting power equipment faults based on multi-source heterogeneous data and twin models according to claim 1 is characterized in that: The data after joint modeling based on multimodal information adopts a dynamic twin model that combines self-supervised learning and reinforcement learning to enable the twin to adapt to the device operating status and environmental changes in real time and dynamically update the device health status, as follows: Through self-supervised learning, a reference benchmark for equipment operating status is constructed to detect and quantify health status deviations; time series reconstruction based on joint feature representation: in, is the predicted value at time t+1; Z multi,t is the joint feature representation at time t; f SSL is the self-supervised learning model function The reconstruction loss is calculated using the mean squared error: Among them, X ts,t+1 is the actual value at time t+1; Construct positive and negative sample pairs between normal and abnormal states, and optimize through contrast loss: sim(·) is the cosine similarity; Zi is the feature vector representation of the current sample i; Z j is the feature vector representation of the positive sample of the current sample i; Z k For comparison samples; The ultimate goal of self-supervised learning is: Among them, λ1,λ2 are weight coefficients; Reinforcement learning is used to enhance the dynamic adaptability of the twin, so that it can be updated in real time according to environmental changes. t Current device operation status S t =(Z multi,t ,HS t ,ROLL t ); Among them, HS t For health status prediction, RUL t for the remaining life; Action A t Update the model parameters of the twin: A t =Δθt; Reward R t , rewards based on equipment health, failure risk, and performance: R t =-α(1-HS t ) 2 -βP fail，g +earth t ; Among them, α, β, and γ are weight coefficients; P fail,t is the failure risk; t For operational efficiency; Deep reinforcement learning DDPG is used to optimize the twin, with the goal of maximizing the cumulative reward: Through the policy network π θ (S t ) output action, the target is: Health status prediction: HS t =σ(f HS (Z multi,t )); Combined with the adjustment action output by reinforcement learning, f HS It is the health status prediction model function.

5. The method for predicting power equipment faults based on multi-source heterogeneous data and twin models according to claim 1 is characterized in that: The Transformer-based full-cycle prediction framework is constructed to capture the long-term dependencies and dynamic change characteristics of equipment operation, specifically: The joint feature is represented as Z multi Input to the Transformer model and mapped to the input dimension of the Transformer through a linear layer: WITH emb =Z multi IN emb +b emb ; Among them, W emb and b emb are weight and bias respectively; Add position encoding P to time series features E , capturing the time sequence information, the final input is: WITH input =Z emb +P E ; Using stacked Transformer encoders to model long-term dependencies in time series, the output of the Transformer is: H t =TransformerEnc(Z input，t ); Through the fully connected layer and the Sigmoid activation function, the failure probability of the future time step is output: Where Ht is the hidden state vector at time step t; W fail ,b fail Weights and biases for failure probability prediction; Predict potential fault categories through classification network: Among them, W class ,b class Weights and bias terms predicted for potential fault classes; Remaining life estimation, regression prediction through a fully connected layer: Among them, W RUL ,b RUL Weights and biases for remaining lifetime prediction.

6. The method for predicting power equipment faults based on multi-source heterogeneous data and twin models according to claim 5 is characterized in that: The above-mentioned causal relationship analysis is combined to identify the triggering factors of the fault and enhance the accuracy and explainability of the prediction, as follows: Based on domain knowledge and data-driven methods, we build a causal relationship graph between multimodal features: Node: input features; Edge: causal relationship between features; The causal effect of key features on the probability of failure is calculated using causal inference methods: Among them, Causal Effect (x i →P fail ) describes the variable x i The failure probability P fail causal effect; Through causal analysis, the weights of input features are optimized to enhance the prediction accuracy and interpretability of the model.

7. The method for predicting power equipment faults based on multi-source heterogeneous data and twin models according to claim 6 is characterized in that: Based on the equipment health status and fault prediction results, the intelligent operation and maintenance optimization engine is used to optimize the equipment operating parameters, as follows: Dynamically monitor the health status of equipment, update the prediction model in real time, and obtain the latest prediction data, including HS t , Integrate dynamic data to build the global state of the system, which is used to optimize operation and maintenance strategies and predict equipment status representation Determine the optimal operating parameters A′ t , so that the following objective function is optimized: Among them, Cost(A′ t ) The operating cost of the equipment under the current parameters; a is the weight coefficient; Design goal-related reward function: R′=R HS +R fail +R cost ; Among them, R HS ,R fail ,R cost They are equipment life bonus, failure rate bonus and cost bonus; Use the DDPG algorithm based on deep reinforcement learning to obtain the optimal equipment operating parameters.

8. The method for predicting power equipment faults based on multi-source heterogeneous data and twin models according to claim 6 is characterized in that: The genetic algorithm is combined to dynamically generate the optimal operation and maintenance plan, as follows: Dynamically generate maintenance scheduling plans based on equipment health status and predicted remaining life, optimizing the following objectives: Among them, C maint (i) is the cost of the ii-th maintenance; P fail (i) is the residual probability of failure during maintenance ii; C fail (i) is the cost after the failure occurs; λ4 is the weight, which controls the balance between maintenance and failure costs; Each individual represents the maintenance time and maintenance type of all equipment, and the quality of the maintenance plan is estimated by the fitness function: Fitness = -J maint , optimize maintenance costs and failure risks Select excellent individuals according to the fitness function; Generate children by exchanging time or maintenance type; Randomly change the maintenance time or adjust the maintenance type to explore the solution space; continuously iterate and finally output the optimal maintenance scheduling plan.

9. A power equipment fault prediction system based on multi-source heterogeneous data and twin models, characterized in that: It includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically executes the steps in the power equipment fault prediction method based on multi-source heterogeneous data and twin models as described in any one of claims 1 to 8.

Images