Measurement board card fault prediction method and system based on multi-source data fusion and deep learning
By combining multi-source data fusion and deep learning methods with TimesNet models and knowledge graphs, the problem of insufficient accuracy and generalization ability in fault diagnosis of measurement boards in power systems has been solved, enabling early fault identification and warning, and improving operation and maintenance efficiency and system reliability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- UHV CONVERTER STATION BRANCH OF STATE GRID SHANGHAI ELECTRIC POWER CO
- Filing Date
- 2025-11-25
- Publication Date
- 2026-04-28
AI Technical Summary
In the fault diagnosis of key equipment in power systems, existing technologies rely on model matching methods, which depend on precise mathematical models, rule-based methods, which depend on human experience, and data-driven methods, which require manually designed features and have high requirements for sample data. This results in limited diagnostic accuracy, especially when fault samples are scarce and data is imbalanced, leading to insufficient generalization ability.
By employing a method based on multi-source data fusion and deep learning, multi-channel time-series monitoring data from measurement boards is collected. An improved TimesNet model is used for feature extraction and fault prediction. Combined with a multi-head attention mechanism and knowledge graph, disposal suggestions are generated to achieve early identification and warning of faults.
It enables early fault identification and warning of measurement boards, improves diagnostic accuracy and generalization ability, reduces human intervention, improves operation and maintenance efficiency and system reliability, and builds an adaptive fault handling solution recommendation system.
Smart Images
Figure CN121935510A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of power system automation technology, specifically relating to a method and system for predicting measurement board faults based on multi-source data fusion and deep learning. Background Technology
[0002] Currently, domestic and international approaches to fault diagnosis of critical equipment in power systems mainly employ model matching, rule-based methods, and data-driven methods. Model matching involves establishing a mathematical model of the equipment and comparing the actual operating data with the model output to determine the presence of faults. This method relies on accurate mathematical models and requires high precision in system modeling. Rule-based methods depend on expert experience and pre-defined rules for fault diagnosis, typically combining historical fault cases and operating procedures. This method heavily relies on human experience, suffers from lagging rule updates, struggles to handle novel or complex fault modes, and exhibits poor generalization ability. Data-driven methods utilize historical operating data, employing feature extraction and machine learning algorithms (such as SVM and decision trees) for fault classification. These methods require manual feature design and demand a high degree of sample data balance and quantity. However, in actual operation, fault samples are scarce, and data distribution is uneven. Traditional data-driven methods perform poorly on such data, resulting in insufficient model generalization ability. Furthermore, existing data-driven methods require manual feature design, a cumbersome process that struggles to comprehensively capture fault characteristics, leading to limited diagnostic accuracy. Summary of the Invention
[0003] In view of this, the purpose of this invention is to provide a method and system for predicting measurement board faults based on multi-source data fusion and deep learning, so as to realize early fault diagnosis and fault warning of measurement boards in UHVDC converter stations.
[0004] The objective of this invention is achieved through the following technical solution:
[0005] This invention provides a method for predicting faults in measurement boards based on multi-source data fusion and deep learning. The method includes:
[0006] S1: Collect multi-channel timing monitoring data from the measurement board. The multi-channel timing monitoring data is the status data generated by the measurement board during operation, including electrical quantity data, status quantity data, and timing waveform data of the measurement board.
[0007] S2: Perform data enhancement and cleaning on the multi-channel timing monitoring signal to obtain the first data;
[0008] S3: Input the first data into the improved TimesNet model, which includes multiple TimesBlock modules connected in sequence and multiple multi-head attention mechanism modules located after the multiple TimesBlock modules; the multiple TimesBlock modules perform layer-by-layer feature extraction on the first data through residual connections to obtain a preliminary feature vector; the multi-head attention mechanism module receives the preliminary feature vector and focuses on fault-sensitive features of different dimensions through multiple attention heads to generate enhanced feature vectors; wherein, different attention heads focus on capturing the temporal evolution pattern of single-channel signals and the coupling correlation pattern between multi-channel signals through their respective linear projections to enhance the model's ability to express complex dynamic features;
[0009] S4: Using the enhanced feature vector, a fault prediction result for the measurement board is generated.
[0010] Furthermore, the electrical quantity data of the measurement board corresponds to the internal circuit modules of the measurement board, including the instantaneous voltage value, instantaneous current value, effective voltage value, and effective current value of the power supply circuit of the measurement board; the status quantity data of the measurement board corresponds to the status monitoring module of the measurement board, including the operating temperature of the measurement board, the power supply operating status, the alarm signal level, and the temperature of the onboard sensor; the timing waveform data of the measurement board corresponds to the signal transmission and processing module of the measurement board, including high-frequency sampled internal specific signal waveforms, pulse signal waveforms, and photocurrent signal waveforms.
[0011] Further, feature extraction is performed on the first data to generate a preliminary feature vector, including:
[0012] The data from each channel in the first data is treated as a one-dimensional time series data. For each one-dimensional time series data, the features corresponding to the one-dimensional time series data are extracted through multiple residual-connected TimesBlock modules. The features corresponding to all channels are combined to form a preliminary feature vector. Each TimesBlock module performs the following operations:
[0013] The one-dimensional input is received and, through a fast Fourier transform, the frequencies corresponding to the first K peaks in the Fourier transform amplitude spectrum are selected. The frequencies are then converted into corresponding time periods to obtain the K target periods contained in the one-dimensional input.
[0014] Based on the K target periods, the one-dimensional input is reshaped into K two-dimensional time series tensors to characterize the changes within the period and the changes during the period.
[0015] Multi-scale convolution processing is performed on each two-dimensional temporal tensor to extract and fuse two-dimensional features at different scales, resulting in a two-dimensional feature tensor containing multi-scale spatiotemporal information.
[0016] All two-dimensional feature tensors are converted back to one-dimensional time-series feature vectors and then fused to obtain the one-dimensional output of the TimesBlock module.
[0017] Furthermore, the number of TimesBlock modules is 2 to 4.
[0018] Furthermore, the number of attention heads in the multi-head attention mechanism module is greater than the dimension of the initial feature vector.
[0019] Furthermore, after generating the fault prediction results for the measurement board, the method further includes: S5: Based on the RAG framework and knowledge graph, generating handling suggestions according to the fault prediction results, specifically including:
[0020] S51: Construct a knowledge graph of measurement board faults. The nodes of the knowledge graph include the fault phenomena, fault causes, handling measures, required spare parts, risks, and operating steps of the measurement board. The edges of the knowledge graph represent the causal relationships or logical connections between nodes.
[0021] S52: Using the fault prediction result as the query input, retrieve candidate knowledge fragments related to the fault prediction result in the knowledge graph through the RAG framework;
[0022] S53: A multi-index similarity evaluation algorithm is used to evaluate and filter candidate knowledge fragments to obtain the filtered knowledge fragments;
[0023] S54: Based on the filtered knowledge fragments, generate personalized handling suggestions for the fault prediction results.
[0024] Furthermore, step S51 specifically includes:
[0025] Collect text data such as historical maintenance records of measurement boards, expert experience manuals, equipment manuals, and accident reports, and perform formatting, noise reduction, and semantic segmentation on the text data;
[0026] We use a large language model to generate context vector embeddings for semantically segmented text slices and build a semantic index based on a vector database.
[0027] The RAG framework's generation module extracts key entities such as measurement card fault phenomena, fault causes, handling measures, required spare parts, risks, and operating procedures, and identifies the relationships between entities to generate standardized knowledge triples.
[0028] The knowledge triples are stored to realize a queryable, reasonable, and updatable knowledge graph of measurement board faults. Key entities are used as nodes in the knowledge graph, and the relationships between entities are used as edges in the knowledge graph.
[0029] Furthermore, in step S53, the multi-index similarity evaluation algorithm includes the TOPSIS algorithm and the GHA algorithm, and the evaluation indexes include at least the consistency of equipment type, the similarity of failure modes, the similarity of operating environment, and the severity of failure.
[0030] This invention also provides a measurement board fault prediction system based on multi-source data fusion and deep learning, the system comprising:
[0031] The data acquisition and storage module is used to acquire and store multi-channel timing monitoring signals from the measurement board. These multi-channel timing monitoring signals are status data generated by the measurement board during operation, including electrical quantity data, status quantity data, and timing waveform data.
[0032] The data enhancement and cleaning module is used to enhance and clean the multi-channel time-series monitoring signal to obtain the first data.
[0033] The intelligent diagnosis and prediction engine module extracts features from the first data using an improved TimesNet model and generates fault prediction results for the measurement board. The improved TimesNet model includes multiple TimesBlock modules connected in sequence, followed by multiple multi-head attention mechanism modules. The TimesBlock modules extract features from the first data layer by layer through residual connections to obtain preliminary feature vectors. The multi-head attention mechanism modules receive the preliminary feature vectors and use multiple attention heads to focus on fault-sensitive features of different dimensions, thereby generating enhanced feature vectors. Each attention head, through its own linear projection, focuses on capturing the temporal evolution pattern of a single-channel signal and the coupling correlation pattern between multi-channel signals, thereby enhancing the model's ability to express complex dynamic features.
[0034] Furthermore, the system also includes an operation and maintenance recommendation module, which generates handling suggestions based on the fault prediction results using the RAG framework and knowledge graph. The operation and maintenance recommendation module specifically executes the following process:
[0035] A knowledge graph of measurement board failures is constructed. The nodes of the knowledge graph include the failure phenomenon, failure cause, handling measures, required spare parts, risks, and operation steps of the measurement board. The edges of the knowledge graph represent the causal relationship or logical association between the nodes.
[0036] Using the fault prediction results as query input, the RAG framework is used to retrieve candidate knowledge fragments related to the fault prediction results from the knowledge graph.
[0037] A multi-index similarity evaluation algorithm is used to evaluate and filter candidate knowledge fragments to obtain the filtered knowledge fragments;
[0038] Based on the selected knowledge fragments, personalized handling suggestions are generated for the fault prediction results.
[0039] The beneficial effects of this invention are:
[0040] This invention provides a method for fault prediction based on real-time data streams, enabling early identification and warning of potential faults in measurement boards;
[0041] This invention combines generative artificial intelligence with the sample expansion and cleaning mechanism of mechanistic models to improve the model's fault identification ability under conditions of imbalanced and small sample data.
[0042] This invention employs deep learning structures such as the TimesNet model and multi-head attention mechanism to achieve automatic extraction and representation of fault features, reducing human intervention and improving diagnostic accuracy and generalization ability. Furthermore, by setting multiple attention heads, this invention adaptively focuses on the most fault-sensitive patterns in different time steps and signal channels, dynamically allocating attention weights to achieve enhanced expression of key features. Through this mechanism, the model can understand the coupling relationship and evolution law between features in the global time dimension, thereby achieving unsupervised deep feature extraction.
[0043] This invention constructs a fault handling solution recommendation system based on knowledge graphs, which realizes adaptive matching between fault types and handling measures, thereby improving operation and maintenance efficiency and system reliability.
[0044] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description
[0045] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will now be described in further detail with reference to the accompanying drawings, wherein:
[0046] Figure 1 This is a schematic flowchart of a measurement board fault prediction method based on multi-source data fusion and deep learning;
[0047] Figure 2This is a schematic diagram of the improved TimesNet model;
[0048] Figure 3 This is an illustration of how the TimesBlock module processes input data.
[0049] Figure 4 This is a schematic flowchart illustrating the intelligent recommendation method for operation and maintenance solutions;
[0050] Figure 5 This is a schematic block diagram of a measurement board fault prediction system based on multi-source data fusion and deep learning. Detailed Implementation
[0051] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be understood that the preferred embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.
[0052] Figure 1 This is a schematic flowchart of a measurement board fault prediction method based on multi-source data fusion and deep learning, such as... Figure 1 As shown, the method includes the following steps:
[0053] S1: Collect multi-channel timing monitoring data from the measurement board. The multi-channel timing monitoring data is the status data generated by the measurement board during operation, including electrical quantity data, status quantity data, and timing waveform data of the measurement board.
[0054] S2: Perform data enhancement and cleaning on the multi-channel timing monitoring signal to obtain the first data;
[0055] S3: Input the first data into the improved TimesNet model. For example... Figure 2As shown, the improved TimesNet model includes multiple TimesBlock modules connected in sequence and multiple multi-head attention mechanism modules following the multiple TimesBlock modules. The multiple TimesBlock modules perform layer-by-layer feature extraction on the first data through residual connections to obtain a preliminary feature vector. The multi-head attention mechanism module receives the preliminary feature vector and focuses on fault-sensitive features of different dimensions through multiple attention heads, thereby generating an enhanced feature vector. Among them, different attention heads focus on capturing the temporal evolution pattern of single-channel signals and the coupling correlation pattern between multi-channel signals through their respective linear projections, so as to enhance the model's ability to express complex dynamic features. This model integrates the time-frequency domain modeling capabilities of multi-head attention mechanism and Fourier transform, and completes deep feature extraction and intelligent prediction based on high-quality input data. The number of stacked TimesBlock modules is 2 to 4. The number of attention heads in the multi-head attention mechanism module is greater than the dimension of the preliminary feature vector.
[0056] S4: Using the enhanced feature vector, a fault prediction result for the measurement board is generated.
[0057] The electrical quantity data of the measurement board corresponds to the internal circuit modules of the measurement board, including the instantaneous voltage value, instantaneous current value, effective voltage value, and effective current value of the power supply circuit of the measurement board; the status quantity data of the measurement board corresponds to the status monitoring module of the measurement board, including the operating temperature of the measurement board, the power supply status, the alarm signal level, and the temperature of the onboard sensor; the timing waveform data of the measurement board corresponds to the signal transmission and processing module of the measurement board, including high-frequency sampled internal specific signal waveforms, pulse signal waveforms, and photocurrent signal waveforms.
[0058] The data acquisition system is based on industrial communication protocols such as OPC UA and MTConnect, and constructs a unified information model to achieve efficient integration of data from different sources and with different sampling frequencies. In some embodiments, correlation algorithms, variance screening, principal component analysis (PCA), and DBSCAN clustering algorithms are also used to reduce the dimensionality of high-dimensional data and extract key features, selecting the core feature data streams that best characterize the health status of the boards, and storing them in Redis (for high-speed caching of real-time data) and a NoSQL database (for storing historical and multi-granular data).
[0059] To address the issues of scarce fault samples and imbalanced data, data augmentation and cleaning are required after collecting a small amount of fault data.
[0060] First, generative adversarial networks (GANs) or generative algorithms based on large models can be used to learn the data distribution characteristics of a small number of fault samples and generate a large amount of synthetic data that conforms to the characteristics of real faults, effectively expanding the fault sample set.
[0061] Then, the generated synthetic data is input into the mechanistic model (a mathematical model based on circuit principles and board physical characteristics) for verification. This model determines whether the generated data conforms to known fault physics, eliminates "dirty data" that does not conform to the mechanism, and ensures the physical rationality and reliability of the enhanced dataset, thus obtaining the first data. After obtaining the first data, the data can be aligned, normalized, and processed for subsequent use.
[0062] Feature extraction is performed on the first data to generate a preliminary feature vector, which may specifically include:
[0063] Each channel's data in the first dataset is treated as a one-dimensional time-series data point. For each one-dimensional time-series data point, features corresponding to the data are extracted using a TimesBlock module with multiple residual connections. The features corresponding to all channels are combined to form a preliminary feature vector. The operations performed by each TimesBlock module can be found in [link to documentation]. Figure 2 The structure diagram of TimesBlock and Figure 3 And its related descriptions.
[0064] like Figure 3 As shown, each TimesBlock module receives a one-dimensional input, which is then processed by a Fast Fourier Transform (e.g., ...). Figure 2 The FFT in the Fourier transform selects the frequencies corresponding to the first K peaks in the amplitude spectrum of the Fourier transform, converts the frequencies into corresponding time periods, and obtains the K target periods contained in the one-dimensional input.
[0065] Based on the K target periods, the one-dimensional input is reshaped (e.g. Figure 2 In this context, "Reshape" refers to K two-dimensional time tensors to characterize intra-period and inter-period variations.
[0066] Perform multi-scale convolution processing on each of the two-dimensional temporal tensors (e.g.) Figure 2 The “Inception Block” extracts and fuses two-dimensional features at different scales to obtain a two-dimensional feature tensor containing multi-scale spatiotemporal information.
[0067] Transform all two-dimensional feature tensors back into one-dimensional temporal feature vectors (e.g.) Figure 2 The "Reshape Back" function in the TimeBlock module is used to perform feature fusion, thereby obtaining the one-dimensional output of the TimeBlock module.
[0068] The one-dimensional output of each TimesBlock module is added to its one-dimensional input and used as the input for the next TimesBlock module. If the TimesBlock module is the last one, the result of adding its one-dimensional output to its one-dimensional input is the initial feature vector.
[0069] The model employed in this invention, based on TimesNet, adaptively focuses on the most fault-sensitive patterns across different time steps and signal channels using a multi-head attention mechanism, dynamically allocating attention weights to enhance the representation of key features. Through this mechanism, the model can understand the coupling relationships and evolutionary patterns between features in a global temporal dimension, thereby achieving unsupervised deep feature extraction. Figure 2 As shown, several multi-head attention mechanism modules were added after TimesNet (i.e. Figure 2 (The "MutilAttention" module). While a single multi-head attention mechanism can focus on features from multiple subspaces, its effect is mainly concentrated on the data representation at the same level. For complex temporal-multichannel signal scenarios, a single multi-head attention mechanism module is insufficient to fully exploit these features. Different multi-head attention mechanism modules are responsible for modeling feature relationships at different levels, scales, and / or semantic spaces, thereby significantly improving the model's ability to express temporal patterns and cross-channel correlations in complex dynamic environments. The number of multi-head attention mechanism modules can be any suitable value, such as 2, 3, 4, etc.
[0070] After obtaining the enhanced features, the system generates fault prediction results for the measurement board and outputs the judgment results of the measurement board status, such as: board is normal, board data is abnormal (please check the board), board is damaged, etc.
[0071] , (1)
[0072] , (2)
[0073] Equation (1) represents autoregressive prediction, which uses historical observation data to predict the signal value at the next moment, where, Represents historical observations (historical observations of data in each channel). This represents the predicted value at time t+1. Representing the autoregressive coefficients, these can be obtained by fitting historical data. This represents the error term.
[0074] Equation (2) represents the state classification, that is, judging the state of the measurement board, where, Represents the probability distribution of states. Represents future state variables. The hidden state vector (a feature vector obtained by fusing the input data through the improved TimesNet model's feature extraction of time-domain features, frequency-domain features, and historical information from autoregressive predictions, and then through fully connected layers, residual connections, etc.) contains key information about the device's state. Represents the classification weight moments, Represents the bias vector. and It can also be obtained through training the model, and the training of the model can use the minimization of mean squared error (MSE) as the objective function.
[0075] Therefore, while performing real-time diagnostics, the system also utilizes TimesNet's multi-scale time modeling capabilities and trend extrapolation mechanism to predict the future health status of the equipment. The model, through time-frequency joint modeling of trend changes, periodic fluctuations, and abnormal patterns in historical feature sequences, can capture potential degradation trajectories. Combined with time series forecasting methods (such as autoregressive feature extrapolation and trend smoothing), the system can probabilistically predict the operating status of the board over several future time windows.
[0076] After generating the fault prediction results for the measurement board, the system automatically generates initial value suggestions. Therefore, the method of the present invention further includes: S5: generating disposal suggestions based on the fault prediction results using the Retrieval-Augmented Generation (RAG) framework and knowledge graph. Figure 4 This is a schematic flowchart illustrating the intelligent recommendation method for operation and maintenance solutions, such as... Figure 4 As shown, the process specifically includes:
[0077] S51: Construct a knowledge graph of measurement board faults. The nodes of the knowledge graph include the fault phenomena, fault causes, handling measures, required spare parts, risks, and operating steps of the measurement board. The edges of the knowledge graph represent the causal relationships or logical connections between nodes.
[0078] S52: Using the fault prediction result as query input, the RAG framework (i.e., the RAG retrieval phase) is used to retrieve candidate knowledge fragments related to the fault prediction result from the measurement board fault knowledge graph. Specifically, the previously generated fault diagnosis result or fault prediction result is used as natural language query input. A corresponding semantic representation vector is generated through a large language model, and a retrieval is performed in the semantic index of the knowledge graph to recall historical case nodes or knowledge fragments most similar to the current fault phenomenon. The retrieval process supports multimodal input (such as text descriptions, sensor data summaries, device models, etc.).
[0079] S53: A multi-index similarity evaluation algorithm is used to evaluate and filter candidate knowledge fragments to obtain the filtered knowledge fragments. The multi-index similarity evaluation algorithm includes the TOPSIS algorithm and the GHA algorithm. The evaluation indicators include at least the consistency of equipment type, the similarity of failure modes, the similarity of operating environment, and the severity of failure. By comprehensively calculating and ranking the weights of multiple evaluation indicators, the optimal or near-ideal initial value scheme is obtained.
[0080] S54: Based on the filtered knowledge fragments, generate personalized handling suggestions for the fault prediction results. For example, the filtered knowledge fragments obtained in S53 (or "highly relevant knowledge fragments") can be used as contextual prompts input into the large language model to generate specific personalized recommendation schemes. The model can output structured operation suggestions, such as: "It is recommended to replace the PS811-A board within 48 hours," "Check if there is any blockage in the air-cooling system."
[0081] In some embodiments, step S51 specifically includes:
[0082] First, collect text data such as historical maintenance records of measurement boards, expert experience manuals, equipment manuals, and accident reports, and then format, denoise, and semantically segment the text data. Sentence-by-sentence and paragraph-by-parameter semantic segmentation of the text content can ensure that the granularity of each knowledge unit is suitable for subsequent retrieval and generation tasks.
[0083] Next, the context vector embeddings of the semantically segmented text slices are generated using a large language model, and a semantic index is built based on a vector database (such as FAISS, Milvus, etc.). This index is used for semantic-level retrieval in large-scale text, so as to accurately find content fragments that are semantically related to the user's query, rather than relying solely on keyword matching.
[0084] Then, the RAG framework's generation module extracts key entities such as measurement card fault phenomena, fault causes, handling measures, required spare parts, risks, and operating procedures, and identifies the relationships between entities to generate standardized knowledge triples. Specifically, when receiving input text or a query task, the RAG framework first retrieves the most relevant context fragments from the vector index, and then combines the retrieval results with the generation capabilities of the large language model. Through Prompt Engineering, it controls the operation and maintenance model to perform the following tasks: extract key entities such as fault phenomena, fault causes, handling measures, required spare parts, and operational risks; identify causal, logical, or temporal relationships between entities (such as "A causes B", "B requires C for handling"); and output standardized knowledge triples (Entity–Relation–Entity) or attribute relationship structures.
[0085] Finally, the knowledge triples are stored to realize a queryable, reasonable, and updatable knowledge graph of measurement board faults. Key entities can be used as nodes in the knowledge graph, and the relationships between entities can be used as edges in the knowledge graph.
[0086] This invention also proposes a measurement board fault prediction system based on multi-source data fusion and deep learning, such as... Figure 5 As shown, the system includes: a data acquisition and storage module, a data enhancement and cleaning module, and an intelligent diagnosis and prediction engine module.
[0087] The data acquisition and storage module is used to acquire and store multi-channel timing monitoring signals from the measurement board. These signals are status data generated by the measurement board during operation, including electrical quantity data, status data, and timing waveform data. The electrical quantity data, status data, and timing waveform data are the same as described above and will not be repeated here. The data acquisition module includes various sensors (temperature, voltage, and current sensors), a data acquisition unit (DAQ), and a communication gateway supporting protocols such as OPC UA, deployed at the converter station site. The backend includes a Redis in-memory database and a NoSQL distributed database. Sensors and the data acquisition unit are connected via hardwired connections or fieldbus. The data acquisition unit connects to the industrial network through the communication gateway, pushing data to the Redis database in real time for streaming processing, while simultaneously storing historical data in batches into the database. This module provides data services to upstream modules.
[0088] The data augmentation and cleaning module is used to augment and clean the multi-channel time-series monitoring signals to obtain the first data. Specifically, this module may include a feature extraction unit (running algorithms such as PCA), a sample generation unit (running generative AI models such as GAN), and a data cleaning unit (this unit integrates the mechanism model). This module reads the raw data and multi-granularity feature data from the data acquisition and storage module. The feature extraction unit performs dimensionality reduction on the data; the sample generation unit and the data cleaning unit work together to generate a high-quality augmented dataset and output it to the downstream intelligent diagnosis and prediction engine module.
[0089] The intelligent diagnosis and prediction engine module may include a data processing unit, a time series modeling unit, and a prediction decision unit. The data processing unit is responsible for preprocessing, normalizing, and aligning the multidimensional signals from the acquisition layer to ensure the temporal consistency and quality of the input data. The time series modeling unit extracts features from the first data using an improved TimesNet model. This improved TimesNet model includes multiple TimesBlock modules and a multi-head attention mechanism module. The multiple TimesBlock modules are residually connected, and they extract features from the first data to generate a preliminary feature vector. The multi-head attention mechanism module receives the preliminary feature vector and uses multiple attention heads to focus on fault-sensitive features of different dimensions, thereby generating an enhanced feature vector. The attention heads include attention heads for capturing single-channel signal timing patterns and attention heads for capturing multi-channel signal coupling and correlation patterns. By combining the multi-head attention mechanism with Fourier transform feature extraction, it can capture the periodicity, abrupt changes, and multi-scale dependencies of equipment operating signals in the time and frequency domains, achieving adaptive extraction and dynamic modeling of key features.
[0090] The predictive decision unit is used to generate fault prediction results for the measurement board using the enhanced feature vector. This invention utilizes the multi-scale modeling capabilities of TimesNet to perform trend extrapolation and health assessment on the extracted features. Combining time series prediction and probabilistic regression methods, it calculates the health status change trend and remaining useful life (RUL) of the equipment in future operating stages, and outputs the current status diagnosis results and fault probability prediction information.
[0091] The entire intelligent diagnosis and prediction engine module is deployed on a high-performance computing framework, with end-to-end model self-learning and online update capabilities, and can automatically optimize parameters based on newly added operation and maintenance data or fault records.
[0092] The system may also include an operation and maintenance recommendation module, which generates handling suggestions based on the fault prediction results using the RAG framework and knowledge graph. This module includes a knowledge graph management unit, a similarity evaluation unit (running algorithms such as TOPSIS), and a web front-end application (for result visualization, alarm display, and report generation). The module specifically executes the following process:
[0093] A knowledge graph of measurement board failures is constructed. The nodes of the knowledge graph include the failure phenomenon, failure cause, handling measures, required spare parts, risks, and operation steps of the measurement board. The edges of the knowledge graph represent the causal relationship or logical association between the nodes.
[0094] Using the fault prediction results as query input, the RAG framework is used to retrieve candidate knowledge fragments related to the fault prediction results from the knowledge graph.
[0095] A multi-index similarity evaluation algorithm is used to evaluate and filter candidate knowledge fragments to obtain the filtered knowledge fragments;
[0096] Based on the selected knowledge fragments, personalized handling suggestions are generated for the fault prediction results.
[0097] The operations and maintenance recommendation module receives fault information from the intelligent diagnosis module, performs matching and retrieval in the knowledge graph through the similarity evaluation unit, and generates recommended solutions. Finally, the warning information, diagnostic reports, and operations and maintenance suggestions are pushed to users through the web frontend. User feedback and new operations and maintenance experiences can also flow back into the knowledge graph, enabling the system to self-optimize.
[0098] The intelligent diagnosis and prediction engine module receives multi-dimensional time-series input data from the data cleaning and enhancement module for model training and online inference. After the model is loaded by the model service engine, it performs feature extraction, health assessment, and trend prediction on the real-time data stream. The diagnosis and prediction results are sent to the operation and maintenance recommendation module to generate corresponding handling plans and operation and maintenance decisions. Simultaneously, they are written to the knowledge graph and database modules to support historical case association, model retraining, and knowledge updates. Through dynamic interaction with the knowledge graph, the system can semantically associate diagnostic results with historical fault experience, thereby improving the accuracy and interpretability of predictions and realizing an integrated closed-loop intelligent analysis process from status monitoring to health prediction.
[0099] Therefore, the method and system proposed in this invention realize the transformation of the operation and maintenance mode from "post-event maintenance" to "pre-event early warning". By integrating multi-source real-time data with the TimesNet deep time series model, the system can accurately capture early abnormal characteristics of equipment operation status, realize early identification and early warning of potential faults, significantly reduce the risk of unplanned outages caused by sudden faults, and improve the reliability and safety of power grid operation.
[0100] The method of this invention effectively solves the model training problem under small sample and imbalanced data. By introducing a data augmentation and cleaning mechanism that combines generative adversarial networks (GANs) and mechanistic models, high-quality fault samples are greatly expanded while preserving physical authenticity. This significantly improves the generalization ability and diagnostic accuracy of deep learning models in real industrial data scenarios with scarce samples and class imbalance.
[0101] This invention also achieves automated and intelligent extraction of fault features. It adopts a deep time series modeling method that combines multi-head attention mechanism with Fourier transform. The system can adaptively extract the deep features most relevant to the fault from complex time series data, avoiding the limitations of relying on manually designed features in traditional methods, and improving the comprehensiveness of feature expression and the self-learning ability of the model.
[0102] A smart operation and maintenance knowledge brain with continuous evolution capabilities has been built.
[0103] By using the RAG framework based on the Large Language Model (LLM), the system can automatically extract, construct, and update the knowledge graph of power equipment faults from massive unstructured documents, realizing intelligent matching and recommendation of fault diagnosis results and handling solutions. This not only improves the efficiency of operation and maintenance response, but also enables the system to continuously accumulate and reuse expert experience.
[0104] It provides an end-to-end intelligent closed-loop operation and maintenance solution.
[0105] The system organically integrates functional modules such as data acquisition, intelligent diagnosis, health prediction, and operation and maintenance recommendation, forming a complete closed loop from status perception to decision support. This greatly reduces manual intervention and improves the automation, intelligence, and overall efficiency of UHV converter station operation and maintenance.
[0106] The following uses the PS811 board dataset as an example to explain in detail the implementation process of this algorithm:
[0107] 1. Environment Configuration
[0108] (1) Software: PyCharm 2021.1.0
[0109] (2) Hardware: 14th Gen Intel(R) Core(TM) i7-14700kf @ 5.6GHz, NVIDIA GeForce3090Ti (24 GB)
[0110] (3) Operating system: Ubuntu 22.04
[0111] 2. Data Preparation
[0112] Data from January to September 2025, collected using a PS811 board, totaling 470,000 time steps, was used. 80% of the data was used as the training set, and 20% as the test set. A sliding window approach was employed to construct sequence data of length 100.
[0113] 3. Parameter Settings
[0114] (1) dropout = 0.1
[0115] (2) batch_size = 64
[0116] (3) learning_rate = 1e-4
[0117] (4) top_k = 5 (meaning select the top 5 frequency components with the highest energy (amplitude) in the frequency domain.)
[0118] (5) Optimizer Adam
[0119] 4. Algorithm Execution Steps
[0120] a. Initialize model parameters using a random seed
[0121] b. Feed the data into the improved TimesNet model.
[0122] c. The model output is trained using the MSE loss function.
[0123] 5. Results Output and Evaluation
[0124] Output the model's reconstruction error (MSE) metric. The model's training level is determined by calculating its reconstruction error.
[0125] During the testing phase, after inputting the test data into the model, the reconstruction error is used to determine whether the PS811 board is malfunctioning or needs to be replaced.
[0126] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A method for predicting faults in measurement boards based on multi-source data fusion and deep learning, characterized in that, include: S1: Collect multi-channel timing monitoring data from the measurement board. The multi-channel timing monitoring data is the status data generated by the measurement board during operation, including electrical quantity data, status quantity data, and timing waveform data of the measurement board. S2: Perform data enhancement and cleaning on the multi-channel timing monitoring signal to obtain the first data; S3: Input the first data into the improved TimesNet model, which includes multiple TimesBlock modules connected in sequence and multiple multi-head attention mechanism modules located after the multiple TimesBlock modules; the multiple TimesBlock modules perform layer-by-layer feature extraction on the first data through residual connections to obtain a preliminary feature vector; the multi-head attention mechanism module receives the preliminary feature vector and focuses on fault-sensitive features of different dimensions through multiple attention heads to generate enhanced feature vectors; wherein, different attention heads focus on capturing the temporal evolution pattern of single-channel signals and the coupling correlation pattern between multi-channel signals through their respective linear projections to enhance the model's ability to express complex dynamic features; S4: Using the enhanced feature vector, a fault prediction result for the measurement board is generated.
2. The measurement board fault prediction method based on multi-source data fusion and deep learning according to claim 1, characterized in that, The electrical quantity data of the measurement board corresponds to the internal circuit modules of the measurement board, including the instantaneous voltage value, instantaneous current value, effective voltage value, and effective current value of the power supply circuit of the measurement board; the status quantity data of the measurement board corresponds to the status monitoring module of the measurement board, including the operating temperature of the measurement board, the power supply operating status, the alarm signal level, and the temperature of the onboard sensor; the timing waveform data of the measurement board corresponds to the signal transmission and processing module of the measurement board, including high-frequency sampled internal specific signal waveforms, pulse signal waveforms, and photocurrent signal waveforms.
3. The measurement board fault prediction method based on multi-source data fusion and deep learning according to claim 1, characterized in that, Feature extraction is performed on the first data to generate a preliminary feature vector, including: Each channel of the first data is treated as a one-dimensional time series data. For each one-dimensional time series data, the features corresponding to the one-dimensional time series data are extracted through multiple residual-connected TimesBlock modules. The features corresponding to all channels are combined to form a preliminary feature vector. Each TimesBlock module performs the following operations: The one-dimensional input is received and, through a fast Fourier transform, the frequencies corresponding to the first K peaks in the Fourier transform amplitude spectrum are selected. The frequencies are then converted into corresponding time periods to obtain the K target periods contained in the one-dimensional input. Based on the K target periods, the one-dimensional input is reshaped into K two-dimensional time series tensors to characterize the changes within the period and the changes during the period. Multi-scale convolution processing is performed on each two-dimensional temporal tensor to extract and fuse two-dimensional features at different scales, resulting in a two-dimensional feature tensor containing multi-scale spatiotemporal information. All two-dimensional feature tensors are converted back to one-dimensional time-series feature vectors and then fused to obtain the one-dimensional output of the TimesBlock module.
4. The measurement board fault prediction method based on multi-source data fusion and deep learning according to claim 1, characterized in that, The number of TimesBlock modules is 2 to 4.
5. The measurement board fault prediction method based on multi-source data fusion and deep learning according to claim 1, characterized in that, The number of attention heads in the multi-head attention mechanism module is greater than the dimension of the initial feature vector.
6. The measurement board fault prediction method based on multi-source data fusion and deep learning according to claim 1, characterized in that, After generating the fault prediction results for the measurement board, the method further includes: S5: Based on the RAG framework and knowledge graph, generating handling suggestions according to the fault prediction results, specifically including: S51: Construct a knowledge graph of measurement board faults. The nodes of the knowledge graph include the fault phenomena, fault causes, handling measures, required spare parts, risks, and operating steps of the measurement board. The edges of the knowledge graph represent the causal relationships or logical connections between nodes. S52: Using the fault prediction result as the query input, retrieve candidate knowledge fragments related to the fault prediction result in the knowledge graph through the RAG framework; S53: A multi-index similarity evaluation algorithm is used to evaluate and filter candidate knowledge fragments to obtain the filtered knowledge fragments; S54: Based on the filtered knowledge fragments, generate personalized handling suggestions for the fault prediction results.
7. The measurement board fault prediction method based on multi-source data fusion and deep learning according to claim 6, characterized in that, Step S51 specifically includes: Collect text data such as historical maintenance records of measurement boards, expert experience manuals, equipment manuals, and accident reports, and perform formatting, noise reduction, and semantic segmentation on the text data; We use a large language model to generate context vector embeddings for semantically segmented text slices and build a semantic index based on a vector database. The RAG framework's generation module extracts key entities such as measurement card fault phenomena, fault causes, handling measures, required spare parts, risks, and operating procedures, and identifies the relationships between entities to generate standardized knowledge triples. The knowledge triples are stored to realize a queryable, reasonable, and updatable knowledge graph of measurement board faults, wherein key entities are used as nodes of the knowledge graph and the relationships between entities are used as edges of the knowledge graph.
8. The measurement board fault prediction method based on multi-source data fusion and deep learning according to claim 6, characterized in that, In step S53, the multi-index similarity evaluation algorithm includes the TOPSIS algorithm and the GHA algorithm. The evaluation indexes include at least the consistency of equipment type, the similarity of failure modes, the similarity of operating environment, and the severity of failure.
9. A measurement board fault prediction system based on multi-source data fusion and deep learning, characterized in that, include: The data acquisition and storage module is used to acquire and store multi-channel timing monitoring signals from the measurement board. These multi-channel timing monitoring signals are status data generated by the measurement board during operation, including electrical quantity data, status quantity data, and timing waveform data. The data enhancement and cleaning module is used to enhance and clean the multi-channel time-series monitoring signal to obtain the first data. The intelligent diagnosis and prediction engine module extracts features from the first data using an improved TimesNet model and generates fault prediction results for the measurement board. The improved TimesNet model includes multiple TimesBlock modules connected in sequence, followed by multiple multi-head attention mechanism modules. The TimesBlock modules extract features from the first data layer by layer through residual connections to obtain preliminary feature vectors. The multi-head attention mechanism modules receive the preliminary feature vectors and use multiple attention heads to focus on fault-sensitive features of different dimensions, thereby generating enhanced feature vectors. Each attention head, through its own linear projection, focuses on capturing the temporal evolution pattern of a single-channel signal and the coupling correlation pattern between multi-channel signals, thereby enhancing the model's ability to express complex dynamic features.
10. The measurement board fault prediction system based on multi-source data fusion and deep learning according to claim 9, characterized in that, The system also includes an operation and maintenance recommendation module, which generates handling suggestions based on the fault prediction results using the RAG framework and knowledge graph. The operation and maintenance recommendation module specifically performs the following process: A knowledge graph of measurement board failures is constructed. The nodes of the knowledge graph include the failure phenomenon, failure cause, handling measures, required spare parts, risks, and operation steps of the measurement board. The edges of the knowledge graph represent the causal relationship or logical association between the nodes. Using the fault prediction results as query input, the RAG framework is used to retrieve candidate knowledge fragments related to the fault prediction results from the knowledge graph. A multi-index similarity evaluation algorithm is used to evaluate and filter candidate knowledge fragments to obtain the filtered knowledge fragments; Based on the selected knowledge fragments, personalized handling suggestions are generated for the fault prediction results.