Power equipment detection method and device based on big data analysis
By building a dynamic metadata database and integrating device knowledge graphs with deep learning, the data compatibility problem of heterogeneous equipment in power equipment detection is solved, the feature distribution alignment and privacy protection of new and old equipment are achieved, and the detection accuracy and dynamic optimization capability of the model are improved.
Patent Information
- Application Number
- CN202510810536.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-17
- Publication Date
- 2025-09-19
AI Technical Summary
In power equipment testing, the differences in data characteristics caused by mixed scenarios of heterogeneous equipment affect the accuracy and reliability of the analysis model. Existing technologies are difficult to effectively solve the problems of data compatibility and comparability.
Build a dynamic metadata database, unify the data format through protocol conversion and edge gateway, use interpolation to align time series to generate proxy variables, integrate device knowledge graph and deep learning, use transfer learning and federated learning to achieve feature distribution alignment and privacy protection of new and old devices, conduct hierarchical evaluation based on device health index, dynamically adjust thresholds based on service life, and integrate online learning to update model parameters in real time.
It improves the data utilization rate of power equipment detection and the accuracy of anomaly detection, dynamically adapts to the aging characteristics of equipment, and realizes the advance warning of faults and continuous optimization of models.
Smart Images

Figure CN120670782A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of data processing, and in particular relates to a method and device for detecting electric power equipment based on big data analysis. Background Art
[0002] During the analysis of big data from power equipment inspections, the presence of heterogeneous equipment (e.g., the coexistence of old and new transformers, or equipment from different manufacturers or models operating simultaneously) can lead to significant differences in the characteristics of the collected data. This heterogeneity (including inconsistencies in technical parameters, operating age, and monitoring standards) can lead to data compatibility and comparability issues, which in turn affects the accuracy of the analysis model, ultimately raising questions about the reliability of the analysis results and necessitating improvement. Summary of the Invention
[0003] Based on this, it is necessary to provide a power equipment detection method and device based on big data analysis to address the above problems.
[0004] The embodiment of the present invention is implemented as follows: a power equipment detection method based on big data analysis includes the following steps:
[0005] Build a dynamic metadata database to record device models, sensor configurations, and physical parameters. Unify data formats with edge gateways through protocol conversion, align time series using interpolation, and generate proxy variables for missing sensors (e.g., estimating discharge intensity from oil chromatography), addressing protocol and sampling frequency differences.
[0006] The device knowledge graph (such as insulation levels and manufacturer specifications) is integrated with deep learning (1D-CNN + Transformer) to extract common features across devices. Transfer learning is used to pre-train a universal model, and domain adaptation (MMD loss) and federated learning are used to align the feature distribution of new and old devices and protect privacy.
[0007] Based on a hierarchical assessment of equipment health index (HI), thresholds are dynamically adjusted based on service life, and relative anomalies are identified through group Z-score ranking (a statistical method). Fault propagation maps are linked to historical case libraries to locate fault causes and recommend maintenance strategies (e.g., prioritizing repairs for equipment at the bottom of the HI ranking).
[0008] Integrate online learning to update model parameters in real time and use operation and maintenance feedback data to optimize false positives / missing negatives; aggregate cross-device knowledge through federated learning and dynamically expand knowledge graph rules (such as adding new fault modes to the database).
[0009] In one embodiment, the present invention provides a method for detecting power equipment based on big data analysis. The method includes building a dynamic metadata database to record equipment models, sensor configurations, and physical parameters; unifying data formats through protocol conversion and edge gateways, aligning time series using interpolation, generating proxy variables for missing sensors, and resolving protocol and sampling frequency differences. The method specifically includes:
[0010] Build a dynamic metadata database to collect structured data, including equipment model, rated parameters, sensor type (such as oil temperature, vibration), and historical maintenance records, and associate unstructured data, including manufacturer design documents (PDF / drawings). Connect different device interfaces through the OPC UA protocol, install edge intelligent gateways on legacy equipment, and convert traditional protocols (such as Modbus) to MQTT / HTTP protocols (enabling unified access to multiple source protocols and resolving heterogeneous communication layer issues).
[0011] For the high-frequency data of new equipment (10,000 points per second) and the low-frequency data of old equipment (1 point per minute, new and old equipment are determined based on the factory date and service life), sliding window cubic spline interpolation is used to generate equally spaced time series; for missing sensor data (such as the lack of partial discharge detection in old equipment), proxy variables are generated based on physical relationships (for example, a regression model is established by combining dissolved gas in oil with winding temperature to infer the virtual discharge intensity index to ensure the integrity of the multidimensional data matrix).
[0012] In one embodiment, the present invention provides a power equipment detection method based on big data analysis. The method integrates equipment knowledge graphs (such as insulation levels and manufacturer specifications) with deep learning (1D-CNN + Transformer) to extract common features across devices. The method also uses transfer learning to pre-train a universal model, and achieves feature distribution alignment and privacy protection for new and old devices through domain adaptation (MMD loss) and federated learning. The steps specifically include:
[0013] Build a knowledge graph based on equipment manuals and industry standards (such as IEEE C57.91) to define entity relationships (such as "Model A transformer - maximum load rate - 130%"). Embed the knowledge graph into the model training process (for example, by introducing the insulation material temperature resistance threshold as a feature boundary constraint in the loss function) to prevent the model from outputting predictions that violate physical laws.
[0014] Design a physical-data dual-channel hybrid network: The physical channel extracts time-frequency domain features (such as the Mel spectrum of the vibration signal and the oil temperature change rate) and overlays design parameters (such as radiator efficiency) from the knowledge graph to generate composite features. The data channel uses a 1D-CNN (a feature extractor) to capture local fluctuation patterns and a Transformer encoder (a feature extractor) to focus on key frequency bands (such as 2-40kHz for partial discharge) through a self-attention mechanism, outputting device-independent deep features.
[0015] Pre-train the Contrastive Predictive Coding time series model on the entire data set of new devices to generate a universal feature representation. For old devices, use MMD (maximum mean difference) loss to align the distribution of new and old features, and introduce a Bayesian neural network layer to quantify prediction uncertainty. Through federated learning (federated learning uses a horizontally / vertically segmented distributed training architecture to enable participants, such as devices from different manufacturers, to collaboratively train models without sharing original data), aggregate the model gradients of each node and transmit them encrypted to the cloud for fusion (enabling cross-vendor knowledge sharing without leaking original data).
[0016] In one embodiment, the present invention provides a power equipment detection method based on big data analysis. The method includes the following steps: hierarchical assessment based on equipment health index, dynamic adjustment of thresholds based on service life, identification of relative anomalies through group Z-score ranking, associating fault propagation maps with historical case libraries, locating fault causes, and recommending maintenance strategies.
[0017] Hierarchical health index (HI) modeling is performed: the device-level health index is constructed by weighted aggregation of feature deviations (e.g., temperature variance × 0.3 + vibration energy × 0.5), with weights dynamically adjusted by random forest feature importance. The group-level health index is calculated by calculating the Z-score of similar devices to identify the bottom 5% of relatively degraded devices in the health index ranking. For devices with a service life of more than 10 years, the Weibull failure distribution (a probabilistic model) is superimposed to predict the health index threshold attenuation curve.
[0018] Dynamically adjust alarm thresholds based on equipment age and environmental load (using a 3σ statistical threshold for new equipment and exponential smoothing to predict threshold drift for older equipment). Construct a fault propagation map (e.g., "cooling failure → oil temperature rise → insulation aging"), combine it with a causal discovery algorithm (PC algorithm) to locate the root cause of the fault, and recommend a response strategy based on historical cases (e.g., "prioritize replacement of bushings on equipment with the lowest HI level").
[0019] It integrates real-time equipment data, health index trends, and similar comparative analysis to automatically generate diagnostic reports, annotate abnormal probability heat maps, and remaining life prediction curves (based on LSTM degradation models). It also highlights the locations of high-risk equipment through GIS maps and outputs them to the cloud.
[0020] In one embodiment, the present invention provides a power equipment detection method based on big data analysis. The integrated online learning method updates model parameters in real time and uses operation and maintenance feedback data to optimize false positives and false negatives. The method aggregates cross-device knowledge through federated learning and dynamically expands the knowledge graph rule step, specifically including:
[0021] Deploy a lightweight TinyLSTM model on the edge to detect device data in real time, receive feedback from the cloud (such as manually confirmed false positives / missed negatives), and update samples (prioritizing high-value samples) using a weighted loss function using importance sampling.
[0022] After each edge node trains the model locally, it uploads encrypted gradients (the parameter update direction calculated during each model training process, indicating how to adjust model parameters to reduce prediction error) to the cloud for aggregation. Differential privacy technology is used to add Gaussian noise (σ = 0.1) to prevent reverse derivation of the original data. (Shared knowledge is integrated into characteristic patterns, such as "overheating fault → high-frequency vibration + sudden increase in CO2," improving the ability to identify similar faults across devices.)
[0023] Automatically parse the manual of the newly connected equipment, extract parameter constraints (such as "Model B maximum load rate 140%") and convert them into graph entity relationships; convert the potential patterns discovered by federated learning (such as "Insulation aging accelerates when ambient humidity > 80%) into interpretable rules. After manual review, the interpretable rules are added to the database.
[0024] In one embodiment, the present invention provides an electric power equipment detection device based on big data analysis, comprising:
[0025] The heterogeneous data integration module is used to build a dynamic metadata database to record device models, sensor configurations, and physical parameters. It unifies data formats with edge gateways through protocol conversion, aligns time series using interpolation, and generates proxy variables for missing sensors (such as estimating discharge intensity from oil chromatograms), addressing protocol and sampling frequency differences.
[0026] A model training module is used to integrate device knowledge graphs (such as insulation levels and manufacturer specifications) with deep learning (1D-CNN + Transformer) to extract common features across devices. Transfer learning is used to pre-train a universal model, and domain adaptation (MMD loss) and federated learning are used to align the feature distribution of new and old devices and protect privacy.
[0027] The abnormality diagnosis decision module is used to evaluate equipment health index (HI) hierarchically, dynamically adjust thresholds based on service life, and identify relative anomalies through group Z-score ranking (a statistical method). It also links fault propagation maps with historical case libraries to locate fault causes and recommend maintenance strategies (such as prioritizing repair of equipment with the lowest HI score).
[0028] The diagnostic logic optimization module is used to integrate online learning to update model parameters in real time and optimize false positives / missing negatives using operation and maintenance feedback data; it aggregates cross-device knowledge through federated learning and dynamically expands knowledge graph rules (such as adding new fault modes to the database).
[0029] In one embodiment, the present invention provides a power equipment detection device based on big data analysis, wherein the heterogeneous data integration module includes:
[0030] The database establishment unit is used to build a dynamic metadata database, collect structured data, including equipment model, rated parameters, sensor type (such as oil temperature, vibration), and historical maintenance records, and associate unstructured data, including manufacturer design documents (PDF / drawings). It connects to different device interfaces through the OPC UA protocol, and installs edge intelligent gateways on legacy equipment to convert traditional protocols (such as Modbus) to MQTT / HTTP protocols (enabling unified access to multiple source protocols and resolving heterogeneous communication layer issues).
[0031] The timing alignment and missing compensation unit is used to generate equally spaced time series using sliding window cubic spline interpolation for the high-frequency data of new equipment (10,000 points per second) and the low-frequency data of old equipment (1 point per minute, with new and old equipment determined based on their production date and years of service). For missing sensor data (such as the lack of partial discharge detection in old equipment), proxy variables are generated based on physical relationships (for example, a regression model is established between dissolved gas in oil and winding temperature to infer a virtual discharge intensity index to ensure the integrity of the multidimensional data matrix).
[0032] In one embodiment, the present invention provides a power equipment detection device based on big data analysis, wherein the model training and establishment module includes:
[0033] A knowledge graph construction and embedding unit is used to build a knowledge graph based on equipment manuals and industry standards (such as IEEE C57.91) and define entity relationships (such as "Model A transformer - maximum load rate - 130%"). The knowledge graph is embedded in the model training process (for example, introducing the insulation material temperature resistance threshold as a feature boundary constraint in the loss function) to prevent the model from outputting prediction results that violate physical laws.
[0034] A dual-channel feature extraction unit is used to design a physical-data dual-channel hybrid network: the physical channel extracts time-frequency domain features (such as the Mel spectrum of the vibration signal and the oil temperature change rate) and overlays design parameters (such as radiator efficiency) from the knowledge graph to generate composite features. The data channel uses a 1D-CNN (a feature extractor) to capture local fluctuation patterns and a Transformer encoder (a feature extractor) to focus on key frequency bands (such as 2-40kHz for partial discharge) through a self-attention mechanism, outputting device-independent deep features.
[0035] The cross-device processing unit is used to pre-train the Contrastive Predictive Coding time series model on the full data of the new device to generate a universal feature representation; the MMD (maximum mean difference) loss is used to align the new and old feature distributions of the old device, and a Bayesian neural network layer is introduced to quantify the prediction uncertainty; through federated learning (federated learning uses a distributed training architecture with horizontal / vertical segmentation to enable various participants, such as devices from different manufacturers, to collaboratively train models without sharing the original data), the model gradients of each node are aggregated and encrypted and transmitted to the cloud for fusion (to achieve cross-vendor knowledge sharing without leaking the original data).
[0036] In one embodiment, the present invention provides a power equipment detection device based on big data analysis, wherein the abnormality diagnosis decision module includes:
[0037] The health index modeling unit is used to perform hierarchical health index (HI) modeling. The device-level health index is constructed by weighted aggregation of feature deviations (e.g., temperature variance × 0.3 + vibration energy × 0.5), with weights dynamically adjusted by random forest feature importance. The group-level health index is calculated by calculating the Z-score of similar devices to identify the bottom 5% of relatively degraded devices in the health index ranking. For devices in service for more than 10 years, the Weibull failure distribution (a probabilistic model) is superimposed to predict the health index threshold attenuation curve.
[0038] The dynamic threshold and fault source location unit dynamically adjusts alarm thresholds based on equipment age and environmental load (new equipment uses a 3σ statistical threshold, while older equipment uses exponential smoothing to predict threshold drift). It also constructs a fault propagation map (e.g., "cooling failure → oil temperature rise → insulation aging"), uses a causal discovery algorithm (PC algorithm) to locate the root cause of the fault, and recommends a response strategy based on historical cases (e.g., "prioritize replacement of bushings on equipment with the lowest HI level").
[0039] The diagnostic report generation unit is used to integrate real-time equipment data, health index trends, and similar comparative analysis to automatically generate diagnostic reports, annotate abnormal probability heat maps, and remaining life prediction curves (based on the LSTM degradation model). It also highlights the locations of high-risk equipment through GIS maps and outputs them to the cloud.
[0040] In one embodiment, the present invention provides a power equipment detection device based on big data analysis, wherein the diagnostic logic optimization module includes:
[0041] The sample update unit is used to deploy a lightweight TinyLSTM model on the edge to detect device data in real time. It receives feedback from the cloud (such as manually confirmed false positives / missed negatives) and updates samples using a weighted loss function based on importance sampling (prioritizing high-value samples).
[0042] The knowledge aggregation unit is used to train the model locally on each edge node and then upload the encrypted gradient (the parameter update direction calculated during each model training process, indicating how to adjust model parameters to reduce prediction error) to the cloud for aggregation. Differential privacy technology is used to add Gaussian noise (σ = 0.1) to prevent reverse derivation of the original data. (Shared knowledge is integrated into characteristic patterns, such as "overheating fault → high-frequency vibration + sudden increase in CO2", to improve the ability to identify similar faults across devices.)
[0043] The knowledge graph update unit is used to automatically parse the instructions of newly connected equipment, extract parameter constraints (such as "maximum load rate of model B is 140%") and convert them into graph entity relationships; convert potential patterns discovered by federated learning (such as "insulation aging is accelerated when ambient humidity is >80%) into interpretable rules. After manual review, the interpretable rules are added to the database.
[0044] Compared with the existing technology, the beneficial effects of the present invention are: the present invention is aimed at the mixed scenario of heterogeneous power equipment, with deep compatibility of multi-source data. Through dynamic metadata database, proxy variable generation and sliding interpolation alignment, it solves the protocol heterogeneity and sampling frequency differences and improves data utilization; physical-data dual-drive modeling, integrating knowledge graph constraints and dual-channel deep learning (1D-CNN+Transformer), realizes cross-device feature distribution alignment under the transfer learning framework, and ensures accurate anomaly detection of new and old equipment; dynamic closed-loop optimization, based on health index (HI) hierarchical evaluation and federated learning encrypted gradient aggregation, realizes dynamic drift of thresholds and incremental update of models, and early fault warning. BRIEF DESCRIPTION OF THE DRAWINGS
[0045] Figure 1 A flowchart of a method for detecting power equipment based on big data analysis is provided in an embodiment of the present invention.
[0046] Figure 2 A schematic diagram of a flow chart for resolving protocol and sampling frequency differences provided by an embodiment of the present invention.
[0047] Figure 3 A schematic diagram of a process for aligning the feature distribution of new and old devices provided in an embodiment of the present invention.
[0048] Figure 4 A schematic diagram of the process of abnormality diagnosis and recommended maintenance strategy provided by an embodiment of the present invention.
[0049] Figure 5 A schematic diagram of a process for continuous optimization of diagnostic logic provided by an embodiment of the present invention.
[0050] Figure 6 A schematic diagram of an electric power equipment detection device based on big data analysis provided by an embodiment of the present invention.
[0051] Figure 7 A schematic diagram of a heterogeneous data integration module provided in an embodiment of the present invention.
[0052] Figure 8 A schematic diagram of a model training and establishment module provided in an embodiment of the present invention.
[0053] Figure 9 A schematic diagram of an abnormality diagnosis decision module provided in an embodiment of the present invention.
[0054] Figure 10 A schematic diagram of a diagnostic logic optimization module provided in an embodiment of the present invention. DETAILED DESCRIPTION
[0055] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present 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 only used to explain the present invention and are not intended to limit the present invention.
[0056] In one embodiment, Figure 1 As shown, a power equipment detection method based on big data analysis includes the following steps:
[0057] Step S1: Build a dynamic metadata database to record device models, sensor configurations, and physical parameters. Unify data formats with the edge gateway through protocol conversion, align time series using interpolation, and generate proxy variables for missing sensors (e.g., estimating discharge intensity from oil chromatograms) to address protocol and sampling frequency differences.
[0058] In step S2, the device knowledge graph (such as insulation level and manufacturer specifications) is integrated with deep learning (1D-CNN + Transformer) to extract common features across devices. Transfer learning is used to pre-train a universal model, and domain adaptation (MMD loss) and federated learning are used to achieve feature distribution alignment and privacy protection between new and old devices.
[0059] Step S3: Based on a hierarchical assessment of the equipment health index (HI), the threshold is dynamically adjusted based on the service life, and relative anomalies are identified through group Z-score ranking (a statistical method). The fault propagation map is linked to the historical case library to locate the cause of the fault and recommend a maintenance strategy (such as prioritizing the repair of the equipment with the lowest HI).
[0060] Step S4 integrates online learning to update model parameters in real time, and uses operation and maintenance feedback data to optimize false positives / missing negatives; aggregates cross-device knowledge through federated learning, and dynamically expands knowledge graph rules (such as adding new fault modes to the database).
[0061] Step S1 (dynamic metadata database construction) aims to address the heterogeneity of power equipment data. Due to significant differences in equipment models, sensor protocols, and sampling frequencies, it is necessary to unify the data interface through protocol conversion (such as Modbus to MQTT), and fill in missing data through interpolation and proxy variable generation (such as oil chromatogram to estimate discharge intensity) to ensure the spatiotemporal alignment and integrity of multi-source data, providing standardized input for subsequent analysis. Step S2 (knowledge graph + deep learning fusion) enhances the interpretability of the model by introducing equipment knowledge graphs (such as insulation level constraints), while combining 1D-CNN and Transformer to extract common features across devices, and using transfer learning and federated learning to achieve feature distribution alignment and privacy protection for new and old equipment, taking into account the advantages of physical laws and data-driven. Step S3 (layered health assessment) uses dynamic thresholds (such as Z-score ranking and Weibull failure distribution) to distinguish differences in equipment aging, and locates the root cause in combination with fault propagation maps to avoid misjudgment of old equipment by a single threshold, thereby improving alarm accuracy. Step S4 (online learning and knowledge expansion) ensures that the system adapts to new equipment and new failure modes through federated learning and feedback optimization model iteration, forming a closed-loop optimization.
[0062] In one embodiment, Figure 2 As shown, a power equipment detection method based on big data analysis, said step S1, builds a dynamic metadata database to record equipment model, sensor configuration and physical parameters; unifies data format through protocol conversion and edge gateway, uses interpolation to align time series, generates proxy variables for missing sensors, and solves the protocol and sampling frequency differences. The steps specifically include:
[0063] Step S11: Build a dynamic metadata database to collect structured data, including equipment model, rated parameters, sensor type (such as oil temperature, vibration), and historical maintenance records, and associate unstructured data, including manufacturer design documents (PDF / drawings). Connect different device interfaces via the OPC UA protocol, install edge intelligent gateways on legacy equipment, and convert traditional protocols (such as Modbus) to MQTT / HTTP protocols (to achieve unified access to multiple source protocols and address heterogeneous communication layers).
[0064] Step S12: For the high-frequency data of new equipment (10,000 points per second) and the low-frequency data of old equipment (1 point per minute, new and old equipment are determined based on the factory date and service life), sliding window cubic spline interpolation is used to generate an equally spaced time series; for missing sensor data (such as no partial discharge detection in old equipment), proxy variables are generated based on physical relationships (for example, a regression model is established by combining dissolved gas in oil with winding temperature to infer a virtual discharge intensity index to ensure the integrity of the multidimensional data matrix).
[0065] Step S11 focuses on multi-source data integration and protocol unification. The association of structured data (such as device models) and unstructured data (such as design drawings) provides complete context, while the OPC UA protocol and edge gateway (such as Modbus to MQTT) solve the heterogeneous communication layer issues of legacy devices and achieve unified access for all devices. Step S12 addresses the sampling rate differences between new and old devices by aligning time series through sliding window cubic spline interpolation and generating proxy variables based on physical relationships (such as oil gas and temperature regression models) to ensure the integrity of the data matrix for low-frequency / missing sensor devices and avoid model deviations due to missing data.
[0066] In one embodiment, Figure 3 As shown, a power equipment detection method based on big data analysis is described. In step S2, the device knowledge graph (such as insulation level and manufacturer specifications) is integrated with deep learning (1D-CNN+Transformer) to extract common features across devices. The general model is pre-trained using transfer learning, and the feature distribution alignment and privacy protection of new and old devices are achieved through domain adaptation (MMD loss) and federated learning. Specifically, the following steps are included:
[0067] Step S21: Build a knowledge graph based on equipment manuals and industry standards (e.g., IEEE C57.91) and define entity relationships (e.g., "Model A Transformer - Maximum Load Rate - 130%"). This knowledge graph is embedded in the model training process (e.g., by introducing the insulation material temperature resistance threshold as a feature boundary constraint in the loss function) to prevent the model from outputting predictions that violate physical laws.
[0068] Step S22: Design a physical-data dual-channel hybrid network: The physical channel extracts time-frequency domain features (vibration signal Mel spectrum, oil temperature change rate) and superimposes design parameters (such as radiator efficiency) in the knowledge graph to generate composite features; the data channel uses 1D-CNN (a feature extractor) to capture local fluctuation patterns and uses a Transformer encoder (a feature extractor) to focus on key frequency bands (such as 2-40kHz for partial discharge) through a self-attention mechanism, outputting device-independent deep features.
[0069] In step S23, a Contrastive Predictive Coding time series model is pre-trained on the full data of the new device to generate a universal feature representation. The MMD (maximum mean difference) loss is used to align the new and old feature distributions of the old device, and a Bayesian neural network layer is introduced to quantify the prediction uncertainty. Through federated learning (federated learning uses a distributed training architecture with horizontal / vertical segmentation to enable all participants, such as devices from different manufacturers, to collaboratively train models without sharing original data), the model gradients of each node are aggregated and encrypted and transmitted to the cloud for fusion (to achieve cross-manufacturer knowledge sharing without leaking original data).
[0070] Step S21 constrains the model output through knowledge graphs (such as IEEE standards) to prevent violations of physical laws (such as super-insulation temperature resistance thresholds). Step S22 designs a dual-channel network: the physical channel integrates time-frequency features and design parameters (such as heat dissipation efficiency), and the data channel extracts deep features through 1D-CNN+Transformer, taking into account both domain knowledge and data patterns. Step S23 uses contrastive predictive coding (CPC) to pre-train a general model, aligns the feature distributions of new and old devices through MMD loss, and introduces federated learning to aggregate encrypted gradients to achieve cross-vendor knowledge sharing without leaking original data, thereby improving model generalization.
[0071] In one embodiment, Figure 4 As shown, a power equipment detection method based on big data analysis, said step S3, based on the hierarchical assessment of equipment health index, dynamically adjusts the threshold in combination with the service life, identifies relative anomalies through group Z-score ranking; associates the fault propagation map with the historical case library, locates the cause of the fault and recommends a maintenance strategy, specifically including:
[0072] Step S31: Hierarchical health index (HI) modeling is performed: the device-level health index is constructed by weighted aggregation of feature deviations (e.g., temperature variance × 0.3 + vibration energy × 0.5), with weights dynamically adjusted by random forest feature importance. The group-level health index is constructed by calculating the Z-score of similar devices to identify the bottom 5% of relatively degraded devices in the health index ranking. For devices with a service life of more than 10 years, the Weibull failure distribution (a probabilistic model) is superimposed to predict the health index threshold attenuation curve.
[0073] Step S32: Dynamically adjust the alarm threshold based on the equipment's service life and environmental load (a 3σ statistical threshold is used for new equipment, and exponential smoothing is introduced to predict threshold drift for older equipment). Construct a fault propagation map (e.g., "cooling failure → oil temperature rise → insulation aging"), combine it with a causal discovery algorithm (PC algorithm) to locate the root cause of the fault, and recommend a treatment strategy based on historical cases (e.g., "prioritize replacement of bushings on equipment with the lowest HI level").
[0074] Step S33 integrates real-time equipment data, health index trends, and similar comparative analysis to automatically generate a diagnostic report, annotate the abnormal probability heat map, and the remaining life prediction curve (based on the LSTM degradation model), and highlight the location of high-risk equipment through the GIS map and output it to the cloud.
[0075] Step S31: Hierarchical health index (device-level weight aggregation + group Z-score ranking) balances individual differences with group comparisons. Overlaying the Weibull distribution dynamically adjusts thresholds to adapt to device aging characteristics. Step S32: Using fault propagation maps (e.g., "cooling failure → oil temperature rise") and a causal discovery algorithm (PC algorithm) to locate the root cause of the fault, combined with historical case recommendation strategies, improves diagnostic operability. Step S33: LSTM lifespan prediction and GIS visualization are integrated to generate multidimensional reports (e.g., heat maps, remaining life curves), supporting efficient operation and maintenance decision-making.
[0076] In one embodiment, Figure 5 As shown, a power equipment detection method based on big data analysis, said step S4, integrating online learning to update model parameters in real time, using operation and maintenance feedback data to optimize false positives / missing negatives; aggregating cross-device knowledge through federated learning, and dynamically expanding the knowledge graph rule step, specifically including:
[0077] Step S41: Deploy a lightweight TinyLSTM model on the edge to detect device data in real time, receive feedback information (such as manually confirmed false positives / missed negatives) through the cloud, and update samples (prioritize high-value samples) through the importance sampling weighted loss function.
[0078] Step S42: After each edge node has trained the model locally, it uploads encrypted gradients (the parameter update direction calculated during each model training process, indicating how to adjust model parameters to reduce prediction error) to the cloud for aggregation. Differential privacy technology is used to add Gaussian noise (σ = 0.1) to prevent reverse derivation of the original data (integrating shared knowledge into characteristic patterns, such as "overheating fault → high-frequency vibration + CO2 surge", to improve the ability to identify similar faults across devices).
[0079] Step S43 automatically parses the manual of the newly connected device, extracts parameter constraints (such as "maximum load rate of model B is 140%) and converts them into graph entity relationships; converts the potential patterns discovered by federated learning (such as "insulation aging is accelerated when ambient humidity is >80%) into interpretable rules, and after manual review, adds the interpretable rules to the database.
[0080] Step S41 deploys a lightweight edge model (TinyLSTM) for real-time detection and optimizes sample updates through importance sampling, prioritizing high-value feedback data. Step S42 uses federated learning to encrypt gradients and add differential privacy noise (σ = 0.1) to prevent data back-propagation and achieve secure cross-device knowledge aggregation. Step S43 automatically parses new device document extraction rules and converts patterns discovered by federated learning (such as accelerated aging when humidity > 80%) into interpretable rules. After review, the knowledge graph is updated to form a dynamic knowledge expansion closed loop.
[0081] In one embodiment, Figure 6 As shown, a power equipment detection device based on big data analysis includes:
[0082] Heterogeneous data integration module 1 is used to build a dynamic metadata database to record device models, sensor configurations, and physical parameters. It unifies data formats with edge gateways through protocol conversion, aligns time series using interpolation, and generates proxy variables for missing sensors (e.g., estimating discharge intensity from oil chromatograms), addressing protocol and sampling frequency differences.
[0083] Model training and establishment module 2 is used to integrate the device knowledge graph (such as insulation level and manufacturer specifications) with deep learning (1D-CNN + Transformer) to extract common features across devices. It uses transfer learning to pre-train a universal model and achieves feature distribution alignment and privacy protection for new and old devices through domain adaptation (MMD loss) and federated learning.
[0084] Abnormal diagnosis decision module 3 is used to evaluate equipment health index (HI) hierarchically, dynamically adjust thresholds based on service life, and identify relative abnormalities through group Z-score ranking (a statistical method). It also links fault propagation maps with historical case libraries to locate fault causes and recommend maintenance strategies (such as prioritizing repair of equipment with the lowest HI score).
[0085] Diagnostic logic optimization module 4 is used to integrate online learning to update model parameters in real time and use operation and maintenance feedback data to optimize false positives / missing negatives; it aggregates cross-device knowledge through federated learning and dynamically expands knowledge graph rules (such as adding new fault modes to the database).
[0086] In the model training and establishment module 2, a meta-learning framework (such as MAML, a learning framework) can also be introduced in transfer learning to quickly adapt to extreme environments (such as high-altitude substations) through a small number of samples, thereby reducing the cold start cost of new scenarios.
[0087] In one embodiment, Figure 7 As shown, a power equipment detection device based on big data analysis, the heterogeneous data integration module 1 includes:
[0088] The database establishment unit 11 is used to build a dynamic metadata database, collect structured data, including equipment model, rated parameters, sensor type (such as oil temperature, vibration), and historical maintenance records, and associate unstructured data, including manufacturer design documents (PDF / drawings). It connects different device interfaces via the OPC UA protocol, installs edge intelligent gateways on legacy equipment, and converts traditional protocols (such as Modbus) to MQTT / HTTP protocols (to achieve unified access to multiple source protocols and resolve heterogeneous communication layer issues).
[0089] The timing alignment and missing compensation unit 12 is used to generate equally spaced time series using sliding window cubic spline interpolation for the high-frequency data of new equipment (10,000 points per second) and the low-frequency data of old equipment (1 point per minute, with new and old equipment determined based on the factory date and service life). For missing sensor data (such as the lack of partial discharge detection in old equipment), proxy variables are generated based on physical relationships (for example, a regression model is established between dissolved gas in oil and winding temperature to infer a virtual discharge intensity index to ensure the integrity of the multidimensional data matrix).
[0090] In the timing alignment and missing compensation unit 12, the interpolation method can be further dynamically selected according to the device type (such as using Kalman filtering to complete the low-frequency data of old devices) to reduce the cumulative error of the sliding window cubic spline interpolation in high-noise scenarios.
[0091] In one embodiment, Figure 8 As shown, a power equipment detection device based on big data analysis, the model training and establishment module 2 includes:
[0092] A knowledge graph construction and embedding unit 21 is used to construct a knowledge graph based on equipment manuals and industry standards (such as IEEE C57.91), define entity relationships (such as "Model A transformer - maximum load rate - 130%"); embed the knowledge graph into the model training process (for example, introducing the insulation material temperature resistance threshold as a feature boundary constraint in the loss function) to prevent the model from outputting prediction results that violate physical laws;
[0093] The dual-channel feature extraction unit 22 is used to design a physical-data dual-channel hybrid network: the physical channel extracts time-frequency domain features (vibration signal Mel spectrum, oil temperature change rate), superimposes design parameters in the knowledge graph (such as radiator efficiency) to generate composite features; the data channel uses 1D-CNN (a feature extractor) to capture local fluctuation patterns, and uses the Transformer encoder (a feature extractor) to focus on key frequency bands (such as 2-40kHz for partial discharge) through a self-attention mechanism, outputting device-independent deep features;
[0094] The cross-device processing unit 23 is used to pre-train the Contrastive Predictive Coding time series model on the full data of the new device to generate a universal feature representation; the MMD (maximum mean difference) loss is used to align the new and old feature distributions of the old device, and a Bayesian neural network layer is introduced to quantify the prediction uncertainty; through federated learning (federated learning uses a distributed training architecture with horizontal / vertical segmentation to enable all participants, such as devices from different manufacturers, to collaboratively train models without sharing original data), the model gradients of each node are aggregated and encrypted and transmitted to the cloud for fusion (to achieve cross-manufacturer knowledge sharing without leaking original data).
[0095] It is also possible to introduce multimodal alignment mechanisms (such as cross-modal association between text instructions and sensor data) in federated learning to improve the credibility of cross-vendor knowledge fusion.
[0096] In one embodiment, Figure 9 As shown, a power equipment detection device based on big data analysis, the abnormality diagnosis decision module 3 includes:
[0097] Health index modeling unit 31 is used to perform hierarchical health index (HI) modeling: the device-level health index is constructed by weighted aggregation of feature deviations (e.g., temperature variance × 0.3 + vibration energy × 0.5), with weights dynamically adjusted by random forest feature importance; the group-level health index is calculated by calculating the Z-score of similar devices to identify the bottom 5% of relatively degraded devices in the health index ranking; for devices with a service life of more than 10 years, the Weibull failure distribution (a probabilistic model) is superimposed to predict the health index threshold attenuation curve;
[0098] The dynamic threshold and fault source determination unit 32 dynamically adjusts alarm thresholds based on equipment age and environmental load (new equipment uses a 3σ statistical threshold, while older equipment uses exponential smoothing to predict threshold drift). It constructs a fault propagation map (e.g., "cooling failure → oil temperature rise → insulation aging"), uses a causal discovery algorithm (PC algorithm) to locate the root cause of the fault, and recommends a treatment strategy based on historical cases (e.g., "prioritize replacement of bushings on equipment with the lowest HI level").
[0099] The diagnostic report generation unit 33 is used to integrate real-time equipment data, health index trends, and similar comparative analysis to automatically generate a diagnostic report, annotate the abnormal probability heat map, and the remaining life prediction curve (based on the LSTM degradation model), and highlight the location of high-risk equipment through the GIS map and output it to the cloud.
[0100] In the diagnostic report generation unit 33, the operation and maintenance cost and the failure risk can also be combined to design a reward function to train an intelligent agent (such as DQN, a reinforcement learning method) to generate the optimal maintenance sequence, replacing the static priority rule (such as HI last digit first).
[0101] In one embodiment, Figure 10 As shown, a power equipment detection device based on big data analysis, the diagnostic logic optimization module 4 includes:
[0102] The sample update unit 41 is used to deploy a lightweight TinyLSTM model on the edge to detect device data in real time, receive feedback information from the cloud (such as manually confirmed false positives / missed negatives), and update samples (prioritizing high-value samples) through an importance sampling weighted loss function.
[0103] The knowledge aggregation unit 42 is used to encrypt and upload the gradient (the parameter update direction calculated during the model training process, indicating how to adjust the model parameters to reduce prediction error) to the cloud for aggregation after training the model locally on each edge node. Gaussian noise (σ = 0.1) is added using differential privacy technology to prevent reverse derivation of the original data. (Shared knowledge is integrated into characteristic patterns, such as "overheating fault → high-frequency vibration + sudden increase in CO2", to improve the ability to identify similar faults across devices.)
[0104] The knowledge graph update unit 43 is used to automatically parse the instructions of the newly connected equipment, extract parameter constraints (such as "the maximum load rate of model B is 140%) and convert them into graph entity relationships; convert the potential patterns discovered by federated learning (such as "insulation aging is accelerated when the ambient humidity is >80%) into interpretable rules, and add the interpretable rules to the database after manual review.
[0105] In the sample updating unit 41, the feedback false positives / missed negatives are recorded and the authenticity of the data is verified to prevent malicious tampering and enhance the data reliability of the model optimization.
[0106] It should be understood that, although the various steps in the flow chart of each embodiment of the present invention are shown in sequence according to the indication of the arrows, these steps are not necessarily performed in sequence according to the order indicated by the arrows. Unless otherwise specified herein, the execution of these steps is not strictly limited in order, and these steps can be performed in other orders. Moreover, at least a portion of the steps in each embodiment may include a plurality of sub-steps or a plurality of stages, and these sub-steps or stages are not necessarily performed at the same time, but can be performed at different times, and the execution order of these sub-steps or stages is not necessarily performed in sequence, but can be performed in turn or alternately with at least a portion of other steps or sub-steps or stages of other steps.
[0107] The technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above-mentioned embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0108] The above-described embodiments merely illustrate several implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.
[0109] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.
[0110] In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.
Claims
1. A power equipment detection method based on big data analysis, characterized in that: The power equipment detection method based on big data analysis includes the following steps: Build a dynamic metadata database to record device models, sensor configurations, and physical parameters. Unify data formats with edge gateways through protocol conversion, align time series using interpolation, generate proxy variables for missing sensors, and address protocol and sampling frequency differences. Integrate device knowledge graphs with deep learning to extract common features across devices. Use transfer learning to pre-train a universal model, and achieve feature distribution alignment and privacy protection for new and old devices through domain adaptation and federated learning. Based on a hierarchical assessment of equipment health index, thresholds are dynamically adjusted based on service life, and relative anomalies are identified through group Z-score ranking. Fault propagation maps are linked to historical case libraries to locate fault causes and recommend maintenance strategies. Integrate online learning to update model parameters in real time, and use operation and maintenance feedback data to optimize false positives / missing negatives; aggregate cross-device knowledge through federated learning and dynamically expand knowledge graph rules.
2. The power equipment detection method based on big data analysis according to claim 1, characterized in that: The steps of constructing a dynamic metadata database to record device models, sensor configurations, and physical parameters; unifying data formats with edge gateways through protocol conversion, aligning time series using interpolation, generating proxy variables for missing sensors, and resolving protocol and sampling frequency differences specifically include: Build a dynamic metadata database to collect structured data, including device models, rated parameters, sensor types, and historical maintenance records, and associate unstructured data, including manufacturer design documents. Connect different device interfaces through the OPC UA protocol, add edge intelligent gateways to legacy devices, and convert legacy protocols to MQTT / HTTP. For the high-frequency data of new equipment and the low-frequency data of old equipment, sliding window cubic spline interpolation is used to generate equally spaced time series; for missing sensor data, proxy variables are generated based on physical relationships.
3. The power equipment detection method based on big data analysis according to claim 1, characterized in that: The device knowledge graph and deep learning are integrated to extract common features across devices; Using transfer learning to pre-train a general model, domain adaptation and federated learning are used to achieve feature distribution alignment and privacy protection for new and old devices. Specifically, the following steps are involved: Build a knowledge graph based on equipment manuals and industry standards to define entity relationships; embed the knowledge graph into the model training process to prevent the model from outputting prediction results that violate physical laws; Design a physical-data dual-channel hybrid network: The physical channel extracts time-frequency domain features and overlays design parameters from the knowledge graph to generate composite features. The data channel uses a 1D-CNN to capture local fluctuation patterns and a Transformer encoder to focus on key frequency bands through a self-attention mechanism, outputting device-independent deep features. Pre-train the Contrastive Predictive Coding time series model on the entire data of new devices to generate a universal feature representation. For old devices, use MMD loss to align the distribution of new and old features, and introduce a Bayesian neural network layer to quantify prediction uncertainty. Aggregate the model gradients of each node through federated learning and transmit them encrypted to the cloud for fusion.
4. The power equipment detection method based on big data analysis according to claim 1, characterized in that: The equipment health index is based on a hierarchical assessment, and the threshold is dynamically adjusted in combination with the service life, and relative anomalies are identified through group Z-score ranking; The steps of associating the fault propagation map with the historical case library, locating the fault cause and recommending a maintenance strategy include: Hierarchical health index modeling is performed: the device-level health index is constructed by weighted aggregation of feature deviations, with weights dynamically adjusted by random forest feature importance. The group-level health index is calculated by calculating the Z-score of similar devices to identify the bottom 5% of relatively degraded devices in the health index ranking. For devices with more than 10 years of service, the Weibull failure distribution is superimposed to predict the health index threshold attenuation curve. Dynamically adjust alarm thresholds based on equipment service life and environmental load; construct fault propagation maps, use causal discovery algorithms to locate fault root causes, and recommend treatment strategies based on historical cases; Integrate real-time equipment data, health index trends, and similar comparative analysis to automatically generate diagnostic reports, annotate abnormal probability heat maps and remaining life prediction curves, and highlight the locations of high-risk equipment through GIS maps and output them to the cloud.
5. The power equipment detection method based on big data analysis according to any one of claims 1 to 4, characterized in that: The integrated online learning updates model parameters in real time and uses operation and maintenance feedback data to optimize false positives and false negatives. The steps of aggregating cross-device knowledge through federated learning and dynamically expanding the knowledge graph rules specifically include: Deploy a lightweight TinyLSTM model on the edge to detect device data in real time, receive feedback from the cloud, and update samples using a weighted loss function using importance sampling. After training the model locally on each edge node, the gradient is encrypted and uploaded to the cloud for aggregation. Differential privacy technology is used to add Gaussian noise to prevent reverse derivation of the original data. Automatically parse the manuals of newly connected devices, extract parameter constraints and convert them into graph entity relationships; convert the potential patterns discovered by federated learning into interpretable rules, and add the interpretable rules to the database after manual review.
6. A power equipment detection device based on big data analysis, characterized in that: include: Heterogeneous data integration module, used to build a dynamic metadata database to record device models, sensor configurations, and physical parameters; Unify data formats through protocol conversion and edge gateways, align time series using interpolation, generate proxy variables for missing sensors, and resolve protocol and sampling frequency differences; A model training module is used to integrate device knowledge graphs with deep learning to extract common features across devices. It uses transfer learning to pre-train a universal model and achieves feature distribution alignment and privacy protection for new and old devices through domain adaptation and federated learning. The abnormality diagnosis decision module is used to evaluate equipment health indexes in a hierarchical manner, dynamically adjust thresholds based on service life, and identify relative abnormalities through group Z-score ranking. Associate fault propagation maps with historical case libraries to locate fault causes and recommend maintenance strategies; The diagnostic logic optimization module is used to integrate online learning to update model parameters in real time and use operation and maintenance feedback data to optimize false positives and false negatives; it aggregates cross-device knowledge through federated learning and dynamically expands knowledge graph rules.
7. The power equipment detection device based on big data analysis according to claim 6, characterized in that: The heterogeneous data integration module includes: The database establishment unit is used to build a dynamic metadata database, collect structured data, including device model, rated parameters, sensor type, and historical maintenance records, and associate unstructured data, including manufacturer design documents. It connects to different device interfaces through the OPC UA protocol, and installs edge intelligent gateways on legacy devices to convert traditional protocols to MQTT / HTTP. The timing alignment and missing compensation unit is used to generate equally spaced time series using sliding window cubic spline interpolation for high-frequency data from new devices and low-frequency data from old devices; for missing sensor data, proxy variables are generated based on physical relationships.
8. The power equipment detection device based on big data analysis according to claim 6, characterized in that: The model training and building modules include: The knowledge graph construction and embedding unit is used to build a knowledge graph based on equipment manuals and industry standards and define entity relationships. The knowledge graph is embedded in the model training process to prevent the model from outputting prediction results that violate physical laws. A dual-channel feature extraction unit is used to design a physical-data dual-channel hybrid network: the physical channel extracts time-frequency domain features and overlays design parameters from the knowledge graph to generate composite features; the data channel uses a 1D-CNN to capture local fluctuation patterns and a Transformer encoder to focus on key frequency bands through a self-attention mechanism, outputting device-independent deep features. The cross-device processing unit is used to pre-train the Contrastive Predictive Coding time series model on the full data of new devices to generate a universal feature representation; MMD loss is used to align the distribution of new and old features on old devices, and a Bayesian neural network layer is introduced to quantify prediction uncertainty; the gradients of the model of each node are aggregated through federated learning and encrypted and transmitted to the cloud for fusion.
9. The power equipment detection device based on big data analysis according to claim 6, characterized in that: The abnormal diagnosis and decision-making module includes: The health index modeling unit is used to perform hierarchical health index modeling: the device-level health index is constructed by weighted aggregation of feature deviations, with weights dynamically adjusted by random forest feature importance. The group-level health index is calculated by calculating the Z-score of similar devices to identify the bottom 5% of relatively degraded devices in the health index ranking. For devices with a service life of more than 10 years, the Weibull failure distribution is superimposed to predict the health index threshold attenuation curve. Dynamic threshold and fault source determination unit dynamically adjusts alarm thresholds based on equipment service life and environmental load; constructs fault propagation maps, uses causal discovery algorithms to locate the root cause of faults, and recommends handling strategies based on historical cases; The diagnostic report generation unit is used to integrate real-time equipment data, health index trends, and similar comparative analysis to automatically generate diagnostic reports, annotate abnormal probability heat maps and remaining life prediction curves, and highlight the locations of high-risk equipment through GIS maps and output them to the cloud.
10. The power equipment detection device based on big data analysis according to any one of claims 6 to 9, characterized in that: The diagnostic logic optimization module includes: The sample update unit is used to deploy a lightweight TinyLSTM model on the edge to detect device data in real time, receive feedback from the cloud, and update samples using a weighted loss function using importance sampling. The knowledge aggregation unit is used to encrypt and upload the gradients to the cloud for aggregation after training the model locally on each edge node. It uses differential privacy technology to add Gaussian noise to prevent reverse deduction of the original data. The knowledge graph update unit is used to automatically parse the instructions of newly connected devices, extract parameter constraints and convert them into graph entity relationships; convert the potential patterns discovered by federated learning into interpretable rules, and add the interpretable rules to the database after manual review.
Citation Information
Cited By
Electric meter metering error correction method and system based on time sequence anomaly detection and electric meter
CN120951224A
Smart power grid cloud edge collaborative heterogeneous data security access method
CN121151097A
Needle report data analysis system based on artificial intelligence algorithm
CN121455735A
A Needle Report Data Analysis System Based on Artificial Intelligence Algorithms
CN121455735B
