Power equipment fault prediction system based on big data analysis

Abstract

The invention discloses a power equipment fault prediction system based on big data analysis. The method comprises the following steps: acquiring initial equipment multi-dimensional data; constructing a dynamic topology network of the power equipment, including a dependency relationship between the equipment and a fault propagation path, performing embedded learning on the dynamic topology network by using a GNN graph neural network, and extracting equipment collaboration features in the initial equipment multi-dimensional data; a multi-task learning framework is constructed in combination with the equipment cooperation features to predict the equipment fault probability and the remaining service life, and an equipment health index is obtained; and acquiring environmental parameters, dynamically adjusting a fault judgment threshold based on the equipment health index and the environmental parameters, generating a prediction result, integrating the prediction result with an SCADA system, and triggering graded early warning. And the influence of environmental factors on the operation state of the equipment is fully considered. Under different environmental conditions, the equipment fault risk can be judged timely and accurately.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of power equipment monitoring technology, and in particular to a power equipment fault prediction system based on big data analysis. BACKGROUND

[0002] During long-term operation, power equipment is affected by multiple factors such as electrical stress, mechanical stress and environmental factors, and fault problems occur frequently. Traditional power equipment fault prediction methods mainly rely on manual inspection and experience, and have problems such as low efficiency, strong subjectivity and difficulty in finding potential fault hidden dangers. The state monitoring method based on sensor technology and monitoring system is gradually applied, but is limited to the analysis of single equipment or single parameter, cannot comprehensively consider the mutual relationship and fault propagation characteristics between equipment, and is difficult to accurately predict equipment faults in complex power systems, so that the accuracy and efficiency of power equipment fault analysis are not high. SUMMARY

[0003] The application aims to solve the above problems, and designs a power equipment fault prediction system based on big data analysis.

[0004] To achieve the above purpose, the technical scheme of the application is that, further, in the power equipment fault prediction system based on big data analysis, the power equipment fault prediction system comprises: An equipment data acquisition module is configured to acquire equipment multidimensional data of power equipment, and obtain initial equipment multidimensional data by sequentially performing noise reduction, missing value filling, normalization processing and time-frequency domain conversion on the equipment multidimensional data. A collaborative feature acquisition module is configured to construct a dynamic topology network of power equipment, including a dependency relationship between equipment and a fault propagation path, and to extract equipment collaborative features in the initial equipment multidimensional data by using GNN graph neural network embedding learning on the dynamic topology network. An equipment fault prediction module is configured to process time series data using a Transformer encoder to extract long-term dependency features, and to construct a multi-task learning framework to predict equipment fault probability and remaining useful life in combination with the equipment collaborative features to obtain an equipment health index. A fault grading early warning module is configured to acquire environmental parameters, dynamically adjust a fault discrimination threshold based on the equipment health index and the environmental parameters, generate a prediction result, integrate the prediction result with a SCADA system, and trigger a grading early warning.

[0005] Further, in the power equipment fault prediction system based on big data analysis, the equipment data acquisition module comprises the following units: A wavelet packet decomposition unit is configured to perform wavelet packet decomposition on the vibration signal and the voltage signal, reconstruct the signal after separating the noise frequency band, filter the real-time temperature and pressure data, and suppress random disturbances; A feature compression unit is configured to perform feature compression on the high-dimensional data by using a self-encoder, separate noise and effective features, and obtain the denoised equipment data. A missing value filling unit is configured to fill missing values in the sensor fault data, communication interruption data, and data storage abnormal data in the denoised equipment data, normalize the filled data, and obtain complete equipment data.

[0006] Further, in the power equipment fault prediction system based on big data analysis, the equipment data acquisition module comprises the following units: A spectrum feature extraction unit is configured to extract spectrum features of stable periodic signals in the complete equipment data by using FFT Fourier transform. A time-frequency analysis unit is configured to perform time-frequency analysis on non-stationary signals by wavelet transform, capture transient impact events, extract IMF intrinsic mode functions of nonlinear and non-stationary signals by HHT Hilbert-Huang transform, and perform marginal spectrum analysis to obtain initial equipment multidimensional data.

[0007] Further, in the power equipment fault prediction system based on big data analysis, the collaborative feature acquisition module comprises the following units: A quantitative analysis unit is configured to construct the static physical connection between devices based on the power grid design drawings and SCADA system data, and quantify the direct dependence strength between devices by topological analysis. An association mining unit is configured to extract fault propagation patterns from fault logs by using an Apriori association rule mining algorithm, generate propagation paths, combine the current device state and physical topology, and predict potential fault propagation paths by Bayesian network. A feature labeling unit is configured to detect the mutation characteristics of abnormal events by using a sliding window analysis method, and label them as high-priority propagation nodes.

[0008] Further, in the power equipment fault prediction system based on big data analysis, the collaborative feature acquisition module further comprises the following units: A weight adjustment unit is configured to introduce an attention mechanism based on a GAT graph attention network, and dynamically adjust the weights of the dependence relationships between different devices. A vector calculation unit is configured to take the initial equipment multidimensional data as a node feature vector, iteratively aggregate neighbor node information through a GNN layer, and calculate a node embedding vector. A feature splicing unit is configured to splice the node embedding vector with the device's own features to obtain device collaborative features.

[0009] Furthermore, in the aforementioned power equipment fault prediction system based on big data analysis, the equipment fault prediction module includes the following units: The data segmentation unit is used to segment the initial device's multi-dimensional data into a time series according to a time window, and add position information to the time series using sinusoidal position coding to obtain the target time series. The weight calculation unit is used to calculate the query, key, and value matrix of the target time series, and calculate the weights by scaling dot product attention to obtain the time series features; The feature splicing unit is used to splice the time-series features and the device collaborative features, and output the failure probability and the estimated remaining life of the device at the current moment through a fully connected network.

[0010] Furthermore, in the aforementioned power equipment fault prediction system based on big data analysis, the fault classification and early warning module includes the following units: The relation mapping unit is used to acquire environmental parameters through sensors, establish a mapping relationship model between the equipment health index and environmental parameters using random forest, and dynamically adjust the fault discrimination threshold according to changes in environmental parameters. The fault judgment unit is used to judge the health index of the equipment based on the adjusted fault judgment threshold, generate the fault prediction result of the equipment, and design a graded early warning strategy according to the fault risk level of the equipment.

[0011] Furthermore, in the method for implementing the above-mentioned power equipment fault prediction system based on big data analysis, the method includes the following steps: The device multi-dimensional data of the power equipment is obtained. The device multi-dimensional data is then subjected to noise reduction, missing value filling, normalization and time-frequency domain transformation in sequence to obtain the initial device multi-dimensional data. A dynamic topology network of power equipment is constructed, including inter-equipment dependencies and fault propagation paths. A graph neural network (GNN) is used to embed the dynamic topology network and extract the equipment collaboration features from the initial multi-dimensional equipment data. A Transformer encoder is used to process time-series data, extract long-term dependency features, and combine the device collaborative features to construct a multi-task learning framework to predict the probability of device failure and remaining service life, thereby obtaining the device health index. The system acquires environmental parameters, dynamically adjusts the fault discrimination threshold based on the equipment health index and environmental parameters, generates prediction results, integrates the prediction results with the SCADA system, and triggers tiered early warnings.

[0012] Furthermore, in the method for implementing the above-mentioned power equipment fault prediction system based on big data analysis, the method includes the following steps: Based on power grid design drawings and SCADA system data, static physical connections between devices are constructed, and the direct dependence strength between devices is quantified through topology analysis. The Apriori association rule mining algorithm is used to extract fault propagation patterns from fault logs and generate propagation paths; combined with the current device status and physical topology, potential fault propagation paths are predicted using a Bayesian network. The sliding window analysis method is used to detect the mutation characteristics of abnormal events and mark them as high-priority propagation nodes.

[0013] Furthermore, in the method for implementing the above-mentioned power equipment fault prediction system based on big data analysis, the method includes the following steps: Based on the GAT graph attention network, an attention mechanism is introduced to dynamically adjust the weights of dependencies between different devices; The initial device's multi-dimensional data is used as node feature vectors, and neighbor node information is iteratively aggregated through a GNN layer to calculate the node embedding vector. The node embedding vector is concatenated with the device's own features to obtain the device collaborative features.

[0014] Its beneficial effects are as follows: 1. Maintenance personnel can intuitively understand the overall status of equipment through health indices. Combined with failure probability and remaining service life, they can formulate scientific and reasonable maintenance plans in advance, changing the previous passive maintenance mode, reducing the risk of sudden equipment failures, and improving the reliability and stability of power system operation. 2. It fully considers the impact of environmental factors on equipment operating status. Under different environmental conditions, it can timely and accurately judge the risk of equipment failure. Compared with the fault identification method with fixed thresholds, it has stronger environmental adaptability and flexibility, reducing the occurrence of false alarms and missed alarms. 3. Maintenance personnel can intuitively obtain equipment fault information through the SCADA system, and quickly take corresponding measures according to the warning level, improving fault handling efficiency, shortening fault handling time, reducing power outage losses caused by equipment failures, and providing strong support for the optimized scheduling and operation management of the power system. Attached Figure Description

[0015] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0016] Figure 1 This is a schematic diagram of the first embodiment of the power equipment fault prediction system based on big data analysis in this invention. Figure 2 This is a schematic diagram of the second embodiment of the power equipment fault prediction system based on big data analysis in this invention. Figure 3 This is a schematic diagram of the third embodiment of the power equipment fault prediction system based on big data analysis in this invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0018] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0019] The present invention will now be described in detail with reference to the accompanying drawings, such as... Figure 1 As shown, a power equipment fault prediction system based on big data analysis includes: The equipment data acquisition module is used to acquire multi-dimensional data of power equipment. By performing noise reduction, missing value filling, normalization and time-frequency domain transformation on the multi-dimensional data of the equipment in sequence, the initial multi-dimensional data of the equipment is obtained. Specifically, this embodiment also includes a wavelet packet decomposition unit, which is used to perform wavelet packet decomposition on vibration signals and voltage signals, reconstruct signals after separating noise frequency bands, and perform real-time filtering on dynamically changing temperature and pressure data to suppress random disturbances; The feature compression unit is used to compress the features of high-dimensional data using an autoencoder, separate noise from effective features, and obtain noise reduction device data. The missing value filling unit is used to fill in missing values ​​for sensor fault data, communication interruption data, and data storage anomaly data in the noise reduction equipment data. The filled data is then normalized to obtain complete equipment data.

[0020] The spectral feature extraction unit is used to extract spectral features from stable periodic signals in complete device data using FFT Fourier transform. The time-frequency analysis unit is used to perform time-frequency analysis on non-stationary signals through wavelet transform, capture transient impact events, extract IMF intrinsic mode functions from nonlinear and non-stationary signals based on HHT Hilbert-Huang transform, and perform marginal spectrum analysis to obtain multi-dimensional data of the initial equipment.

[0021] Specific; 1.1 Multi-dimensional data collection; Data source: Electrical parameters: Data such as voltage, current, power, and frequency are collected in real time through the SCADA system; Physical state data: Data is collected through sensors such as equipment vibration signals (accelerometer), temperature (infrared thermal imager), pressure (pressure sensor), and sound (microphone array); Environmental data: Humidity, air pressure, corrosive gas concentration (SF6 leak detection), ambient temperature, etc. are collected through IoT sensors; Operation log data: Extract unstructured text data such as historical maintenance records, load change records, and downtime events from the Equipment Management System (EMS).

[0022] Data format: Time-series data (voltage curves, vibration signals) are stored in a structured table format indexed by timestamps; Environmental parameters are stored in a gridded format with spatial location as the dimension; The operation logs extract key event labels ("overload", "insulation degradation") through natural language processing (NLP).

[0023] 1.2 Data noise reduction processing; Noise reduction objective: To eliminate abnormal fluctuations such as sensor noise and electromagnetic interference, while preserving the true characteristics of the device.

[0024] Implementation method: Wavelet transform denoising: Wavelet packet decomposition is performed on vibration and voltage signals, and the noise frequency band is separated and the signal is reconstructed (using db4 wavelet basis function, threshold selection is based on UniversalThreshold criterion). Kalman filtering: Real-time filtering of dynamically changing temperature and pressure data to suppress random disturbances; Machine learning noise reduction: Using an autoencoder to compress features from high-dimensional data (multi-sensor joint signals) and separate noise from effective features.

[0025] Results: The signal-to-noise ratio (SNR) is improved by more than 15%, ensuring the accuracy of subsequent feature extraction.

[0026] 1.3 Missing value imputation; Analysis of reasons for data loss: Sensor failure, communication interruption, and abnormal data storage led to partial data loss.

[0027] Filling strategy: Numerical data: Linear interpolation: suitable for short time intervals (<5 seconds), filling the gaps by linear fitting of neighboring data points; Time series forecasting models: For data with long-term missing values ​​(>1 minute), ARIMA or LSTM models are used to predict missing values; Mean / Median Filling: For periodically fluctuating data (daily average load curve), statistical values ​​are calculated and filled by periodic grouping.

[0028] Categorical data (device status labels): KNN padding: Nearest neighbor matching is performed based on the historical state labels of similar devices; Rule filling: Infer missing states based on the device's operating logic ("overload → high temperature").

[0029] Results: Data integrity improved to over 98%, avoiding model bias caused by missing data.

[0030] Normalization objective: To eliminate the dimensional differences between data of different dimensions (voltage is in kV, temperature is in °C) and make features comparable on a uniform scale.

[0031] Segmented normalization: For non-stationary data (load fluctuations), the normalization parameters are dynamically adjusted according to a time window (1 hour).

[0032] Effects: Accelerates the convergence speed of machine learning models and improves the rationality of feature weight allocation.

[0033] 1.5 Time-frequency domain conversion; Transformation objective: To transform time-domain signals to frequency domain or time-frequency domain, and extract hidden periodic features and transient anomaly patterns.

[0034] Implementation method: Fourier Transform (FFT): Extracts spectral features (amplitude and frequency distribution) from stable periodic signals (voltage harmonics); Wavelet transform: performs time-frequency analysis on non-stationary signals (vibration during equipment startup) to capture transient impact events (high-frequency components caused by bearing failure). Hilbert-Huang Transform (HHT): Extracts intrinsic mode functions (IMFs) from nonlinear, non-stationary signals (complex mechanical vibrations) and performs marginal spectrum analysis.

[0035] Feature extraction: Frequency domain characteristics: dominant frequency, harmonic distortion rate, and bandwidth energy ratio; Time-frequency domain features: time-frequency plot texture features (entropy, kurtosis), duration of transient events.

[0036] Effects: Enhances sensitivity to early faults (partial discharge, insulation degradation), providing richer feature inputs for subsequent models.

[0037] The collaborative feature acquisition module is used to construct a dynamic topology network of power equipment, including inter-equipment dependencies and fault propagation paths. It uses a GNN graph neural network to perform embedding learning on the dynamic topology network and extracts the collaborative features of equipment from the initial multi-dimensional data of the equipment. Specifically, this embodiment also includes a quantitative analysis unit, which is used to construct static physical connections between devices based on power grid design drawings and SCADA system data, and to quantify the direct dependency strength between devices through topology analysis; The association mining unit is used to extract fault propagation patterns from fault logs using the Apriori association rule mining algorithm and generate propagation paths; combined with the current device status and physical topology, it predicts potential fault propagation paths through a Bayesian network. The feature labeling unit is used to detect the mutation features of abnormal events using the sliding window analysis method and mark them as high-priority propagation nodes.

[0038] The weight adjustment unit is used to dynamically adjust the weights of dependencies between different devices by introducing an attention mechanism based on the GAT graph attention network. The vector computation unit is used to take the initial device multi-dimensional data as node feature vectors, iteratively aggregate neighbor node information through GNN layers, and calculate the node embedding vector. The feature concatenation unit is used to concatenate the node embedding vector with the device's own features to obtain the device collaborative features.

[0039] Specific; (a) Construction of dynamic topology networks; Device dependency modeling: Based on the physical wiring diagram of the power system and the connection relationships of the equipment, determine the electrical connection dependencies between the equipment, such as the connection relationships between generators and transformers, transformers and transmission lines, and transmission lines and loads.

[0040] Consider the functional dependencies between devices, such as the relationship between protection devices and protected equipment, and the relationship between measuring devices and measured equipment.

[0041] Correlation analysis is introduced between equipment operating status parameters. By calculating correlation coefficients (Pearson correlation coefficient, Spearman correlation coefficient, etc.), the operating status dependencies between equipment are determined, and a dynamic equipment dependency matrix is ​​constructed.

[0042] Fault propagation path analysis: Collect historical fault data, analyze the propagation patterns and paths of different types of faults in the power system, and identify the key equipment and critical paths for fault propagation.

[0043] A fault propagation model is established, taking into account the type and severity of equipment failures and the operating mode of the power system (normal operation, overload operation, post-fault operation, etc.) on the impact on fault propagation.

[0044] The results of fault propagation path analysis are verified and optimized through simulation experiments (power system fault simulation software) to ensure that the dynamic topology network can accurately reflect the fault propagation characteristics between power equipment.

[0045] Dynamic update mechanism: Real-time monitoring of equipment operating status and connection changes, such as equipment commissioning, shutdown, and switching operations, and timely updating of the structure and parameters of the dynamic topology network.

[0046] Regularly (daily, weekly) reassess and update inter-device dependencies and fault propagation paths to adapt to changes in the power system's operating status.

[0047] (ii) GNN (Graph Neural Network) Embedding Learning; Graph Neural Network Model Selection: Graph Neural Networks (GCN), Graph Attention Networks (GAT), or Graph Recurrent Neural Networks (GRNN), which are suitable for processing graph-structured data, are adopted. A suitable model architecture is selected based on the characteristics of dynamic topology networks and the requirements for device collaborative feature extraction.

[0048] Node feature definition: The preprocessed initial equipment multi-dimensional data (including time domain features, frequency domain features, time-frequency domain joint features, etc.) are used as node features of the graph neural network, while the basic attributes of the equipment (model, manufacturer, commissioning time, etc.) are considered as static features of the nodes.

[0049] Edge feature definition: The strength of the inter-device dependency (correlation coefficient value, connection impedance value, etc.) and the probability of fault propagation (historical fault propagation probability) are used as edge features of the graph neural network.

[0050] Model training: Unsupervised or semi-supervised learning methods are used to train the GNN model to make full use of the large amount of unlabeled device data.

[0051] Define an appropriate loss function, such as a node classification loss function, a link prediction loss function, or a graph reconstruction loss function, and optimize the model parameters according to the training objective.

[0052] During training, regularization techniques (dropout, L2 regularization, etc.) are introduced to prevent model overfitting and improve the model's generalization ability.

[0053] Device collaboration feature extraction: Through the forward propagation process of the GNN model, the embedding vector of each device node is obtained. This embedding vector contains the device's own features and the collaborative relationship features with other devices, which serve as one of the inputs for subsequent multi-task learning.

[0054] The equipment failure prediction module is used to process time-series data with a Transformer encoder, extract long-term dependency features, and combine equipment collaborative features to build a multi-task learning framework to predict the probability of equipment failure and remaining service life, thereby obtaining the equipment health index. Specifically, this embodiment also includes a data segmentation unit, which is used to segment the initial device multi-dimensional data into time series according to the time window, and add position information to the time series using sinusoidal position coding to obtain the target time series; The weight calculation unit is used to calculate the query, key, and value matrix of the target time series, and calculate the weights through scaling dot product attention to obtain the time series features; The feature splicing unit is used to splice time-series features and device collaborative features, and outputs the failure probability and device remaining lifetime estimate at the current moment through a fully connected network.

[0055] Specific; (a) Temporal feature extraction from the Transformer encoder; Time series data processing: The preprocessed time series data is divided into time windows, each containing equipment operation data of a certain duration. The size of the time window is reasonably selected based on the time scale of equipment failure and the length of historical data. For example, for short-term failure prediction, the time window can be set to 1 hour; for long-term failure prediction, the time window can be set to 1 day or longer.

[0056] Transformer encoder structure: A multi-layer Transformer encoder is used, with each encoder layer containing a multi-head self-attention mechanism and a feed-forward neural network. The multi-head self-attention mechanism can capture long-term dependencies between different positions in time series data, while the feed-forward neural network performs further nonlinear transformations on the output of the self-attention mechanism.

[0057] Positional encoding: Since the Transformer encoder does not contain recursive or convolutional structures and cannot automatically capture the sequential information of the time series, positional encoding of the time series data is required. Common positional encoding methods include sine / cosine positional encoding and learnable positional encoding. The positional encoding is added to the input time series data and used as the input to the Transformer encoder.

[0058] Long-term dependency feature extraction: The Transformer encoder processes the time-series data within the time window to extract long-term dependency features, such as the trend changes and periodic fluctuations of equipment operating parameters over a long period of time.

[0059] (II) Design of a multi-task learning framework; Task definition: Failure probability prediction task: The probability of equipment failure within a certain period of time is used as the output. The model is based on a binary classification problem (failure or no failure) or a multi-classification problem (probability of different types of failure).

[0060] Remaining useful life (RUL) prediction task: Predicting the remaining time of a device from the current moment until it fails or needs to be replaced, which is a regression problem.

[0061] Equipment Health Index (HI) Calculation Task: Based on the results of comprehensive failure probability prediction and remaining service life prediction, combined with the equipment's operating status and historical data, calculate the equipment's health index. The health index range can be set from 0 to 100, with higher values ​​indicating better equipment health.

[0062] Model fusion: The long-term dependency features extracted by the Transformer encoder and the device-coordinated features extracted by the GNN graph neural network are fused. Fusion methods can include concatenation, weighted summation, or gating mechanisms. For example, the two feature vectors can be concatenated by dimension and then input into a subsequent fully connected layer for processing.

[0063] Multi-task loss function: The loss function for multi-task learning is defined as a weighted sum of the failure probability prediction loss, remaining useful life prediction loss, and equipment health index calculation loss. The weighting coefficients for each task's loss are set appropriately based on the importance of different tasks and the data distribution. For the failure probability prediction task, the cross-entropy loss function is used; for the remaining useful life prediction task, the mean squared error (MSE) loss function is used; the equipment health index calculation loss can be designed according to the definition of the health index and actual needs.

[0064] Model Training and Optimization: The multi-task learning model is trained using optimization algorithms such as Stochastic Gradient Descent (SGD) and Adam. During training, the weight coefficients of each task are dynamically adjusted to balance the learning effects of different tasks. Simultaneously, techniques such as early stopping and model checkpointing are employed to prevent overfitting and improve the model's training efficiency and generalization ability.

[0065] The fault classification and early warning module is used to acquire environmental parameters, dynamically adjust the fault discrimination threshold based on the equipment health index and environmental parameters, generate prediction results, integrate the prediction results with the SCADA system, and trigger graded early warnings.

[0066] Specifically, this embodiment also includes a relationship mapping unit, which is used to acquire environmental parameters through sensors, establish a mapping relationship model between the device health index and environmental parameters using random forest, and dynamically adjust the fault discrimination threshold according to the changes in environmental parameters; The fault judgment unit is used to judge the health index of the equipment based on the adjusted fault judgment threshold, generate the fault prediction result of the equipment, and design a graded early warning strategy according to the fault risk level of the equipment.

[0067] Specific; (a) Obtaining environmental parameters; Environmental parameter range: Identify environmental parameters that affect the operation of power equipment, including temperature, humidity, air pressure, wind speed, rainfall, pollution level (salt density, dust density), electromagnetic interference intensity, etc.

[0068] Environmental parameter acquisition: Environmental parameters are collected in real time using environmental monitoring sensors (temperature sensors, humidity sensors, weather stations, etc.) installed near power equipment to ensure the accuracy and timeliness of these parameters. For environmental parameters that cannot be directly collected (such as the degree of regional pollution), estimates can be made using meteorological forecast data, Geographic Information System (GIS) data, or historical statistical data.

[0069] (ii) Dynamic threshold adjustment; Threshold Adjustment Model: A mapping model between equipment health indices and environmental parameters is established. Regression analysis and machine learning methods (random forest, support vector machine) are used to analyze the impact of environmental parameters on equipment health status. The fault detection threshold is dynamically adjusted based on changes in environmental parameters. For example, when the ambient temperature rises, the equipment's heat dissipation conditions worsen, and the fault detection threshold can be appropriately lowered to improve the sensitivity of fault prediction.

[0070] Threshold adjustment strategy: Real-time dynamic adjustment: Based on real-time collected environmental parameters, the fault identification threshold is calculated and updated in real time to ensure that the threshold can reflect the fault risk of the equipment under the current environmental conditions in a timely manner.

[0071] Periodic adjustments: Based on the equipment's operating cycle and the periodic changes (seasonal changes) of environmental parameters, the fault discrimination thresholds are periodically adjusted in batches to adapt to the impact of long-term environmental changes on the equipment's health status.

[0072] (III) Generation and integration of prediction results; Prediction result generation: Based on the adjusted fault discrimination threshold, the health index of the equipment is judged, and the fault prediction result of the equipment is generated, including the probability of fault occurrence, remaining service life, fault type, and the health status level of the equipment (normal, attention, warning, serious fault, etc.).

[0073] Integration with SCADA systems: Data interface design: Establish communication protocols and data interface specifications for the SCADA system to ensure accurate and real-time transmission of prediction results. Commonly used communication protocols include Modbus, DNP3, and IEC61850.

[0074] Functional Integration: A fault prediction and early warning module is added to the SCADA system, displaying the predicted results of the equipment in the form of charts, lists, alarm signals, etc., to achieve real-time monitoring and early warning of the equipment's operating status. Simultaneously, the prediction results are correlated with historical and real-time data from the SCADA system to improve the system's overall decision-making capabilities.

[0075] Tiered early warning mechanism: A tiered early warning strategy is designed based on the equipment's fault risk level (minor fault, moderate fault, severe fault). Different early warning methods (audible alarm, SMS notification, email reminder, etc.) and processing procedures are adopted for different levels of early warning signals to ensure that maintenance personnel can obtain equipment fault information in a timely and accurate manner and take corresponding maintenance measures.

[0076] V. System Implementation and Verification Module; (a) System architecture design; Hardware architecture: Determine the hardware devices required for the system, including data acquisition servers, data storage servers, model training servers, application servers, etc. Based on the data volume and computing requirements, select appropriate hardware configurations, such as the server's CPU, memory, hard drive capacity, and graphics card performance.

[0077] Software architecture: A layered architecture is adopted, including a data acquisition layer, a data preprocessing layer, a model training layer, a predictive analysis layer, and an application presentation layer. The layers communicate with each other through interfaces, achieving modularity and scalability of the system.

[0078] Development tools and technology stack: Select appropriate development tools and technology stacks. For example, data acquisition uses Python libraries such as Pandas and NumPy for data processing, GNN models use graph neural network frameworks such as PyTorchGeometric and DGL, Transformer models use deep learning frameworks such as TensorFlow and PyTorch, and system integration uses Java and Web frameworks (SpringBoot, Flask).

[0079] Its beneficial effects are as follows: 1. Maintenance personnel can intuitively understand the overall status of equipment through health indices. Combined with failure probability and remaining service life, they can formulate scientific and reasonable maintenance plans in advance, changing the previous passive maintenance mode, reducing the risk of sudden equipment failures, and improving the reliability and stability of power system operation. 2. It fully considers the impact of environmental factors on equipment operating status. Under different environmental conditions, it can timely and accurately judge the risk of equipment failure. Compared with the fault identification method with fixed thresholds, it has stronger environmental adaptability and flexibility, reducing the occurrence of false alarms and missed alarms. 3. Maintenance personnel can intuitively obtain equipment fault information through the SCADA system, and quickly take corresponding measures according to the warning level, improving fault handling efficiency, shortening fault handling time, reducing power outage losses caused by equipment failures, and providing strong support for the optimized scheduling and operation management of the power system.

[0080] Please see Figure 2 In a power equipment fault prediction system based on big data analytics, the equipment data acquisition module includes: The wavelet packet decomposition unit is used to perform wavelet packet decomposition on vibration and voltage signals, reconstruct the signal after separating noise frequency bands, and perform real-time filtering on dynamically changing temperature and pressure data to suppress random disturbances. The feature compression unit is used to compress the features of high-dimensional data using an autoencoder, separate noise from effective features, and obtain noise reduction device data. The missing value filling unit is used to fill in missing values ​​for sensor fault data, communication interruption data, and data storage anomaly data in the noise reduction equipment data. The filled data is then normalized to obtain complete equipment data.

[0081] Please see Figure 3 In a power equipment fault prediction system based on big data analytics, the collaborative feature acquisition module includes: The quantitative analysis unit is used to construct static physical connections between devices based on power grid design drawings and SCADA system data, and to quantify the direct dependency strength between devices through topology analysis. The association mining unit is used to extract fault propagation patterns from fault logs using the Apriori association rule mining algorithm and generate propagation paths; combined with the current device status and physical topology, it predicts potential fault propagation paths through a Bayesian network. The feature labeling unit is used to detect the mutation features of abnormal events using the sliding window analysis method and mark them as high-priority propagation nodes.

[0082] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A power equipment fault prediction system based on big data analysis, characterized in that, The power equipment fault prediction system includes the following modules: The equipment data acquisition module is used to acquire multi-dimensional data of power equipment. By performing noise reduction, missing value filling, normalization and time-frequency domain transformation on the multi-dimensional data of the equipment in sequence, the initial multi-dimensional data of the equipment is obtained. The collaborative feature acquisition module is used to construct a dynamic topology network of power equipment, including inter-equipment dependencies and fault propagation paths. It uses a GNN graph neural network to perform embedding learning on the dynamic topology network and extracts the collaborative features of the equipment from the initial multi-dimensional data of the equipment. The equipment failure prediction module is used to process time-series data with a Transformer encoder, extract long-term dependency features, and combine the equipment collaborative features to construct a multi-task learning framework to predict the probability of equipment failure and the remaining service life, thereby obtaining the equipment health index. The fault classification and early warning module is used to acquire environmental parameters, dynamically adjust the fault discrimination threshold based on the equipment health index and environmental parameters, generate prediction results, integrate the prediction results with the SCADA system, and trigger graded early warnings.

2. The power equipment fault prediction system based on big data analysis as described in claim 1, characterized in that, The device data acquisition module includes the following units: The wavelet packet decomposition unit is used to perform wavelet packet decomposition on vibration and voltage signals, reconstruct the signal after separating noise frequency bands, and perform real-time filtering on dynamically changing temperature and pressure data to suppress random disturbances. The feature compression unit is used to compress the features of high-dimensional data using an autoencoder, separate noise from effective features, and obtain noise reduction device data. The missing value filling unit is used to fill in the missing values ​​of sensor fault data, communication interruption data and data storage abnormal data in the noise reduction device data, and then normalize the filled data to obtain complete device data.

3. The power equipment fault prediction system based on big data analysis as described in claim 2, characterized in that, The device data acquisition module includes the following units: The spectral feature extraction unit is used to extract spectral features from stable periodic signals in complete device data using FFT Fourier transform. The time-frequency analysis unit is used to perform time-frequency analysis on non-stationary signals through wavelet transform, capture transient impact events, extract IMF intrinsic mode functions from nonlinear and non-stationary signals based on HHT Hilbert-Huang transform, and perform marginal spectrum analysis to obtain multi-dimensional data of the initial equipment.

4. The power equipment fault prediction system based on big data analysis as described in claim 1, characterized in that, The collaborative feature acquisition module includes the following units: The quantitative analysis unit is used to construct static physical connections between devices based on power grid design drawings and SCADA system data, and to quantify the direct dependency strength between devices through topology analysis. The association mining unit is used to extract fault propagation patterns from fault logs using the Apriori association rule mining algorithm and generate propagation paths; combined with the current device status and physical topology, it predicts potential fault propagation paths through a Bayesian network. The feature labeling unit is used to detect the mutation features of abnormal events using the sliding window analysis method and mark them as high-priority propagation nodes.

5. The power equipment fault prediction system based on big data analysis as described in claim 1, characterized in that, The collaborative feature acquisition module further includes the following units: The weight adjustment unit is used to dynamically adjust the weights of dependencies between different devices by introducing an attention mechanism based on the GAT graph attention network. The vector calculation unit is used to take the multi-dimensional data of the initial device as node feature vectors, and iteratively aggregate neighbor node information through the GNN layer to calculate the node embedding vector. The feature splicing unit is used to splice the node embedding vector with the device's own features to obtain the device collaborative features.

6. The power equipment fault prediction system based on big data analysis as described in claim 1, characterized in that, The equipment fault prediction module includes the following units: The data segmentation unit is used to segment the initial device's multi-dimensional data into a time series according to a time window, and add position information to the time series using sinusoidal position coding to obtain the target time series. The weight calculation unit is used to calculate the query, key, and value matrix of the target time series, and calculate the weights by scaling dot product attention to obtain the time series features; The feature splicing unit is used to splice the time-series features and the device collaborative features, and output the failure probability and the estimated remaining life of the device at the current moment through a fully connected network.

7. The power equipment fault prediction system based on big data analysis as described in claim 1, characterized in that, The fault classification and early warning module includes the following units: The relation mapping unit is used to acquire environmental parameters through sensors, establish a mapping relationship model between the equipment health index and environmental parameters using random forest, and dynamically adjust the fault discrimination threshold according to changes in environmental parameters. The fault judgment unit is used to judge the health index of the equipment based on the adjusted fault judgment threshold, generate the fault prediction result of the equipment, and design a graded early warning strategy according to the fault risk level of the equipment.

8. A method for implementing the power equipment fault prediction system based on big data analysis as described in claim 1, characterized in that, The method includes the following steps: The device multi-dimensional data of the power equipment is obtained. The device multi-dimensional data is then subjected to noise reduction, missing value filling, normalization and time-frequency domain transformation in sequence to obtain the initial device multi-dimensional data. A dynamic topology network of power equipment is constructed, including inter-equipment dependencies and fault propagation paths. A graph neural network (GNN) is used to embed the dynamic topology network and extract the equipment collaboration features from the initial multi-dimensional equipment data. A Transformer encoder is used to process time-series data, extract long-term dependency features, and combine the device collaborative features to construct a multi-task learning framework to predict the probability of device failure and remaining service life, thereby obtaining the device health index. The system acquires environmental parameters, dynamically adjusts the fault discrimination threshold based on the equipment health index and environmental parameters, generates prediction results, integrates the prediction results with the SCADA system, and triggers tiered early warnings.

9. A method for implementing the power equipment fault prediction system based on big data analysis as described in claim 1, characterized in that, The method includes the following steps: Based on power grid design drawings and SCADA system data, static physical connections between devices are constructed, and the direct dependence strength between devices is quantified through topology analysis. The Apriori association rule mining algorithm is used to extract fault propagation patterns from fault logs and generate propagation paths; combined with the current device status and physical topology, potential fault propagation paths are predicted using a Bayesian network. The sliding window analysis method is used to detect the mutation characteristics of abnormal events and mark them as high-priority propagation nodes.

10. A method for implementing the power equipment fault prediction system based on big data analysis as described in claim 1, characterized in that, The method includes the following steps: Based on the GAT graph attention network, an attention mechanism is introduced to dynamically adjust the weights of dependencies between different devices; The initial device's multi-dimensional data is used as node feature vectors, and neighbor node information is iteratively aggregated through a GNN layer to calculate the node embedding vector. The node embedding vector is concatenated with the device's own features to obtain the device collaborative features.

Images