A cloud platform-based power equipment monitoring and early warning system and method

Abstract

This invention discloses a cloud-based power equipment monitoring and early warning system and method, relating to the field of power equipment condition monitoring technology. The system includes the following steps: constructing and dynamically updating a power equipment fault knowledge graph; deploying monitoring units at the edge to collect multi-dimensional operational data of the target power equipment in real time; calculating the real-time entropy change rate of a high-dimensional feature vector at the edge based on the multi-dimensional operational data; comparing the real-time entropy change rate with a health benchmark entropy change rate model learned from the knowledge graph and historical data; when a continuous, minor abnormal deviation trend in the real-time entropy change rate is detected, the edge generates and uploads a lightweight early warning trigger signal to the cloud platform; after receiving the early warning trigger signal, the cloud platform generates a comprehensive diagnostic report. This solves the problems of delayed early warning and insensitivity to early, weak fault symptoms in existing power equipment monitoring and early warning technologies.

Description

Technical Field

[0001] This invention relates to the field of power equipment condition monitoring technology, specifically to a power equipment monitoring and early warning system and method based on a cloud platform. Background Technology

[0002] The safe and stable operation of the power grid is crucial to ensuring people's normal production and life. With the rapid development of smart grids, online monitoring and intelligent early warning of power equipment have become key to ensuring the safe and stable operation of the power grid.

[0003] Existing monitoring and early warning solutions mainly fall into two categories: centralized data processing solutions based on cloud platforms and distributed processing solutions based on edge computing. Cloud platform-based solutions typically transmit massive amounts of raw data to the cloud for centralized analysis and threshold judgment. While offering powerful computing capabilities, they suffer from high data transmission pressure, real-time performance is affected by network conditions, and they struggle to capture instantaneous or early, subtle anomalies. Edge computing-based solutions, while providing fast response times, are limited by the computing and storage capabilities of edge devices, typically limiting them to simple threshold alarms or feature extraction, lacking the ability to perform deep learning and correlation reasoning for complex fault modes. Summary of the Invention

[0004] This invention provides a cloud-based power equipment monitoring and early warning system and method to solve the problems of delayed early warning and insensitivity to early and subtle fault signs in existing power equipment monitoring and early warning technologies.

[0005] In a first aspect, the present invention discloses a method for monitoring and early warning of power equipment based on a cloud platform, comprising the following steps:

[0006] A knowledge graph of power equipment faults is constructed and dynamically updated. The knowledge graph takes power equipment entities as central nodes and associates their historical operating data, fault modes, environmental factors and coupling relationships between equipment as nodes and edges.

[0007] A monitoring unit is deployed at the edge to collect multi-dimensional operating data of the target power equipment in real time. The multi-dimensional operating data includes at least video data, infrared thermal imaging data, acoustic data, and partial discharge data.

[0008] Based on the multidimensional operational data, the real-time entropy change rate of the high-dimensional feature vector is calculated at the edge side. The high-dimensional feature vector is obtained by feature extraction and fusion of the multidimensional operational data.

[0009] The real-time entropy change rate is compared with the health benchmark entropy change rate model learned based on the knowledge graph and historical data. When a continuous and slight abnormal deviation trend is detected in the real-time entropy change rate, the edge side generates and uploads a lightweight early warning trigger signal containing abnormal pattern fragments and environmental context to the cloud platform.

[0010] After receiving the warning trigger signal, the cloud platform, guided by the knowledge graph, performs deep pattern matching and fault evolution path deduction in the full high-dimensional feature space to generate a comprehensive diagnostic report that includes potential fault types, evolution stages, risk probabilities, and suggested intervention measures.

[0011] The cloud platform sends the diagnostic report and optimized monitoring strategy instructions to the corresponding edge monitoring unit to update its local model and monitoring focus.

[0012] As a further improvement of this invention, a knowledge graph of power equipment faults is constructed and dynamically updated, specifically including the following steps:

[0013] From the structured and unstructured power equipment inspection corpus, equipment entities, attribute entities, and fault entities are extracted, and the relationships between entities are defined, including subordinate relationships, location coupling relationships, and fault propagation relationships.

[0014] An iterative representation enhancement graph convolutional neural network based on self-attention and cross-attention mechanisms is adopted to iteratively update and enhance the node and edge features in the knowledge graph, so that fault information can be effectively propagated and reasoned among different device nodes.

[0015] Based on the new diagnostic cases received by the cloud platform and the verification results from the edge side, the entities, relationships, and weights of the knowledge graph are continuously expanded and corrected.

[0016] As a further improvement of the present invention, the calculation of the real-time entropy change rate of the high-dimensional feature vector specifically includes the following steps:

[0017] The multidimensional operating data is preprocessed and aligned to extract time-domain, frequency-domain, and spatial-domain features, which are then fused to construct a high-dimensional feature vector characterizing the overall operating status of the equipment.

[0018] Based on the sliding time window, calculate the information entropy of the high-dimensional feature vector within the current window;

[0019] The rate of change of information entropy between continuous time windows is calculated to obtain the entropy change rate, which serves as a core indicator for measuring the orderly evolution of the system.

[0020] As a further improvement of the present invention, a lightweight early warning trigger signal is generated and uploaded at the edge, specifically as follows:

[0021] The edge monitoring unit only uploads compressed data packets that identify the anomaly start time, duration, key feature dimensions, and current environmental parameters, rather than all the original running data.

[0022] As a further improvement of the present invention, deep pattern matching and fault evolution path deduction specifically include the following steps:

[0023] The cloud platform will perform multi-dimensional similarity matching between the received abnormal pattern fragments and the historical fault case library stored in the knowledge graph;

[0024] Based on the matching results and the fault evolution logic defined in the knowledge graph, the possible development path of the anomaly is simulated, and the specific fault type and time window that it may eventually cause are predicted.

[0025] Based on the current equipment operating conditions, environmental factors, and network load, assess the risk level and probability of each predicted path.

[0026] As a further improvement of this invention, the cloud platform performs multi-dimensional similarity matching between the received abnormal pattern fragments and the historical fault case library stored in the knowledge graph, specifically including the following steps:

[0027] The current abnormal pattern fragment and historical failure cases are deconstructed into structured features, and the role set and type set of their key features are extracted respectively; the first similarity between the role sets and the second similarity between the type sets are calculated;

[0028] A deep learning time series model is used to encode temporal abnormal pattern fragments and historical failure cases to obtain their respective high-dimensional feature vectors; the cosine similarity between the high-dimensional feature vectors is calculated as the third similarity representing the deep pattern of the sequence morphology.

[0029] The current abnormal pattern fragment is taken as the subsequence to be detected and matched with the abnormal pattern profiles extracted from historical failure cases stored in the knowledge graph; the minimum distance between the subsequence to be detected and each of the abnormal pattern profiles is calculated, and the minimum distance is normalized and used as the fourth similarity to characterize the morphological similarity of the abnormal waveform.

[0030] Extract entity and relationship information from the current anomaly alert and context information; perform mixed matching between the entity and relationship information and nodes in the knowledge graph, including exact matching, fuzzy matching based on edit distance, and semantic matching based on word vectors; based on the matching results, use a graph neural network to calculate the fifth similarity between the current anomaly and historical cases in the graph structure or semantic vector space;

[0031] At least two of the first, second, third, fourth, and fifth similarities are assigned corresponding weights based on different fault types and application scenarios; the weighted similarities of each dimension are fused to obtain a final comprehensive similarity score; and the most relevant historical fault cases are recommended based on the final comprehensive similarity score according to a preset threshold.

[0032] As a further improvement of the present invention, the optimized monitoring strategy instructions include instructing the edge-side monitoring unit to focus on monitoring specific feature dimensions that are more correlated with the current abnormal pattern in the next stage, adjusting the data sampling frequency, or enabling backup sensors for cross-validation.

[0033] Secondly, this invention discloses a cloud platform-based power equipment monitoring and early warning system, the method of which is applicable to the aforementioned power equipment monitoring and early warning method, comprising:

[0034] The knowledge graph construction and maintenance module, deployed on a cloud platform, is used to construct, update, and store the power equipment fault knowledge graph.

[0035] Edge intelligent sensing modules are deployed on the side of each power equipment to perform data acquisition, feature extraction, entropy rate calculation and preliminary anomaly judgment;

[0036] The cloud-based intelligent diagnosis and simulation module, deployed on a cloud platform, is used to receive early warning signals and perform in-depth diagnosis and path simulation.

[0037] The cloud-edge collaborative optimization module is used to manage command issuance, model updates, and policy synchronization between the cloud platform and the edge.

[0038] Thirdly, the present invention discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-mentioned power equipment monitoring and early warning method.

[0039] Fourthly, the present invention discloses a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the power equipment monitoring and early warning method described above.

[0040] The present invention has the following beneficial effects:

[0041] 1. In this invention, by monitoring the small and continuous deviation of the entropy rate of change, an early warning can be issued in the early stage when the fault characteristics are not yet obvious and traditional threshold monitoring cannot be triggered, thus realizing a fundamental shift from "alarms based on obvious fault characteristics" to "prediction based on changes in the orderliness of system state".

[0042] 2. In this invention, the edge computing platform leverages its proximity to the data source to perform lightweight but crucial entropy change rate calculations and initial anomaly detection, enabling on-site data processing and reducing the burden on the cloud and network latency. The cloud, relying on its powerful computing and storage capabilities, performs knowledge graph-based deep reasoning, path simulation, and global optimization. Based on the diagnostic results, the cloud dynamically guides the edge computing platform to adjust its monitoring strategy, enabling the entire system to learn and evolve, significantly improving monitoring accuracy and resource utilization efficiency.

[0043] 3. In this invention, when an abnormal segment is reported at the edge, the cloud can perform logical deduction based on the graph, simulate the possible development path of the fault, and conduct a risk assessment by comprehensively considering factors such as operating conditions and environment. The final output is a comprehensive decision support report that includes the predicted fault type, risk level, probability, time window, and handling suggestions, which greatly improves the scientific nature and efficiency of the operation and maintenance personnel's decision-making. Attached Figure Description

[0044] Figure 1 This is a flowchart of the monitoring and early warning method of the present invention.

[0045] Figure 2 This is a flowchart of cloud-based deep pattern matching and fault evolution path deduction in this invention.

[0046] Figure 3 This is a flowchart of the entropy change rate calculation method in this invention.

[0047] Figure 4 This is the cloud-edge collaborative optimization relationship diagram in this invention. Detailed Implementation

[0048] The technical solutions of the embodiments of this specification will be explained and described below with reference to the accompanying drawings. However, the following embodiments are only preferred embodiments of this specification and not all of them. Other embodiments obtained by those skilled in the art based on the embodiments in the implementation methods without creative effort are all within the protection scope of this specification.

[0049] The terms "first," "second," "third," etc., in the description, claims, and accompanying drawings are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such processes, methods, products, or apparatus.

[0050] All data involved in this application are information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0051] Example 1

[0052] like Figure 1-4 As shown, a cloud-based power equipment monitoring and early warning method includes the following steps:

[0053] S1. Construct and dynamically update a knowledge graph of power equipment faults. This knowledge graph uses power equipment entities as central nodes, and associates their historical operating data, fault modes, environmental factors, and inter-equipment coupling relationships as nodes and edges. Specifically, this includes:

[0054] S11. Extract equipment entities, attribute entities, and fault entities from the structured and unstructured power equipment inspection corpus, and define the relationships between entities, including subordinate relationships, location coupling relationships, and fault propagation relationships.

[0055] For example, the equipment entity can be "No. 1 main transformer" or "220kV circuit breaker"; the attribute entity can be "winding hot spot temperature" or "partial discharge quantity"; and the fault entity can be "insulation aging", "winding deformation" or "bushing discharge".

[0056] For example, a dependency relationship can be represented as "#1 main transformer includes high-voltage bushings" or "cooling system belongs to #1 main transformer". Location coupling relationships are used to describe the physical connections and spatial locations between devices, such as "#1 main transformer is connected to the 220kV busbar" or "current transformer is adjacent to the circuit breaker". Fault propagation relationships are the logical chains of fault evolution, such as "insulation dampness" potentially leading to "partial discharge", and "partial discharge" potentially developing into "insulation breakdown".

[0057] S12. An iterative representation enhancement graph convolutional neural network based on self-attention and cross-attention mechanisms is adopted to iteratively update and enhance the node and edge features in the knowledge graph, so that fault information can be effectively propagated and reasoned among different device nodes.

[0058] Each entity node is assigned a feature vector. For equipment nodes, features may include their model, years of operation, number of historical failures, etc.; for fault nodes, features may include their typical symptoms, severity level, etc. In the graph, nodes interact with each other through edges. For example, when the "dissolved gas in oil" node of "#1 main transformer" detects an abnormal increase in the "hydrogen" content, this information will be transmitted to the "#1 main transformer" node through the graph neural network along the "has" attribute relationship, and also to fault nodes such as "corona discharge" and "arc discharge" along the "may cause" fault propagation relationship.

[0059] S13. Based on the new diagnostic cases received by the cloud platform and the verification results fed back from the edge side, continuously expand and correct the entities, relationships and weights of the knowledge graph.

[0060] When the cloud platform completes a deep diagnosis based on the knowledge graph, the diagnosis case will be structured and added to the knowledge graph as a new "fault case" entity and its related relationships, enriching the case library.

[0061] After receiving and executing the optimized monitoring strategy from the cloud, the edge-side monitoring unit feeds back the monitoring results and final maintenance conclusions to the cloud platform. This feedback is used to verify and correct the weights or relationship confidence of relevant fault modes in the graph. For example, if multiple feedbacks indicate a very high correlation between "vibration of a specific spectrum" and "winding looseness," the weight of this edge in the graph will be increased.

[0062] S2. Deploy a monitoring unit on the edge side to collect multi-dimensional operating data of the target power equipment in real time. The multi-dimensional operating data includes at least video data, infrared thermal imaging data, acoustic data, and partial discharge data.

[0063] Multidimensional operational data aims to construct a multimodal, holographic equipment status perception system, transcending the limitations of a single data source. The four types of data—video data, infrared thermal imaging data, acoustic wave data, and partial discharge data—form a three-dimensional, complementary monitoring network that ranges from macroscopic physical appearance, thermodynamic state, and mechanical vibration acoustic characteristics to microscopic insulation electrical characteristics, providing a rich data foundation for subsequent in-depth status analysis.

[0064] S3. Based on the multidimensional operational data, calculate the real-time entropy change rate of the high-dimensional feature vector at the edge side. The high-dimensional feature vector is obtained from the multidimensional operational data after feature extraction and fusion. Specifically, this includes:

[0065] S31. Preprocess and align the multidimensional operating data, extract time-domain, frequency-domain, and spatial-domain features, and fuse them to construct a high-dimensional feature vector representing the overall operating status of the equipment.

[0066] For example, for sensor signals, time-domain feature extraction includes: peak value, average amplitude, pulse count, rise time, fall time, and pulse repetition rate.

[0067] Frequency domain feature extraction includes performing a fast Fourier transform on the above signal to extract frequency domain features such as the dominant frequency, centroid frequency, and frequency variance.

[0068] Image feature extraction includes: for infrared thermal images, using a pre-trained convolutional neural network (CNN) model (such as ResNet) to extract its depth feature vectors, characterizing the non-uniformity of temperature distribution and the morphology of hotspot regions.

[0069] Gas feature extraction includes: calculating the ratio of each feature gas (such as C2H2 / H2) and the gas production rate from oil chromatographic data.

[0070] Feature fusion and vector construction: The 128-dimensional features extracted from different modal data (e.g., 20-dimensional UHF, 20-dimensional ultrasound, 80-dimensional infrared image, and 8-dimensional oil chromatography) are aligned according to the time window and then concatenated into a 128-dimensional high-dimensional feature vector V(t) to comprehensively characterize the operating state of the transformer at time t.

[0071] S32. Based on the sliding time window, calculate the information entropy of the high-dimensional feature vector within the current window.

[0072] The formula for calculating information entropy is: H(W) k = -Σ{s∈S}P(s)log2P(s)

[0073] Where S represents the set of all possible discretized joint states, and P(s) is the state s in the current window W. k The estimated probability of its occurrence.

[0074] W k The degree of disorder in the multidimensional operating state of power equipment within a corresponding time window was quantified. Under healthy steady state, the coupling relationship of each characteristic parameter is stable, the state is relatively concentrated, and the entropy value is low and stable. When an early fault occurs, even if an individual parameter does not exceed the standard, the disruption of its coupling relationship will lead to the divergence of the state distribution, and the entropy value will show a detectable upward trend.

[0075] S33. Calculate the rate of change of information entropy between continuous time windows to obtain the entropy change rate, which serves as the core indicator for measuring the orderly evolution of the system.

[0076] Steps S31 and S32 mention that the system extracts and fuses features from the device's multidimensional operational data (time domain, frequency domain, spatial domain) to construct a high-dimensional feature vector, and then calculates the information entropy of this vector. This is equivalent to condensing the distribution characteristics of high-dimensional, complex multi-source sensor data (such as vibration, temperature, current, etc.) into a scalar index that reflects its overall "disorder" or "orderliness". A stable and healthy device usually has a regular and stable distribution of its operating parameters, and the corresponding entropy value should also fluctuate within a stable range; however, the onset of a fault may cause abnormal disturbances in the data pattern, causing the entropy value to change. Step S33 further proposes to calculate the rate of change of information entropy between continuous time windows, i.e., the "entropy change rate". Compared with the absolute entropy value, the entropy change rate is more sensitive to small and continuous deviations in the system state. In the early stages of a fault, there may not be obvious threshold exceedances, but the inherent "orderliness" of the system has already begun to change slowly, and this change will be reflected in the trend change of the entropy value. Monitoring the rate of entropy change is like monitoring the "derivative" of a system's health status. It can detect abnormal signs earlier than traditional threshold methods, thus enabling early warnings to be "moved forward."

[0077] The monitoring system can issue an early warning through the continuous, small positive growth trend of the high-dimensional feature entropy change rate at the very early stage of power equipment failure, such as the very early stage of partial discharge and insulation degradation in transformers, when the discharge quantity has not yet reached the traditional monitoring threshold, but the discharge pattern has already begun to show statistical changes (such as slightly dispersed discharge phase distribution and pulse waveform characteristic variations). This is significantly earlier than the alarm time point of traditional methods that rely on excessive discharge quantity or single gas content, providing maintenance personnel with a valuable intervention window. At the same time, the edge side only needs to calculate and upload the entropy change rate trend and key feature dimension index, with a very small data volume, which conforms to the high-efficiency principle of cloud-edge collaboration.

[0078] S4. The real-time entropy change rate is compared with the health benchmark entropy change rate model learned based on the knowledge graph and historical data. When a continuous and slight abnormal deviation trend is detected in the real-time entropy change rate, the edge side generates and uploads a lightweight early warning trigger signal containing abnormal pattern fragments and environmental context to the cloud platform. Note that the edge-side monitoring unit only uploads a compressed data package that identifies the abnormal start time, duration, key feature dimensions, and current environmental parameters, rather than all the original running data.

[0079] For example, the process of establishing a healthy baseline entropy rate model is as follows: in the early stage of transformer commissioning or a known healthy operating cycle (such as continuous stable operation for 30 days), the entropy rate is continuously calculated, its mean and standard deviation are statistically analyzed, and its healthy baseline model is established. It usually shows a small random fluctuation with a normal distribution near the zero value.

[0080] S5. After receiving the warning trigger signal, the cloud platform, guided by the knowledge graph, performs deep pattern matching and fault evolution path deduction in the full high-dimensional feature space to generate a comprehensive diagnostic report containing potential fault types, evolution stages, risk probabilities, and suggested intervention measures.

[0081] Deep pattern matching and fault evolution path deduction specifically include the following steps:

[0082] S51. The cloud platform performs multi-dimensional similarity matching between the received abnormal pattern fragments and the historical fault case database stored in the knowledge graph. Specifically, this includes the following steps:

[0083] The current abnormal pattern fragment and historical failure cases are deconstructed into structured features, and the role set and type set of their key features are extracted respectively; the first similarity between the role sets and the second similarity between the type sets are calculated;

[0084] A deep learning time series model is used to encode temporal abnormal pattern fragments and historical failure cases to obtain their respective high-dimensional feature vectors; the cosine similarity between the high-dimensional feature vectors is calculated as the third similarity representing the deep pattern of the sequence morphology.

[0085] The current abnormal pattern fragment is taken as the subsequence to be detected and matched with the abnormal pattern profiles extracted from historical failure cases stored in the knowledge graph; the minimum distance between the subsequence to be detected and each of the abnormal pattern profiles is calculated, and the minimum distance is normalized and used as the fourth similarity to characterize the morphological similarity of the abnormal waveform.

[0086] Extract entity and relationship information from the current anomaly alert and context information; perform mixed matching between the entity and relationship information and nodes in the knowledge graph, including exact matching, fuzzy matching based on edit distance, and semantic matching based on word vectors; based on the matching results, use a graph neural network to calculate the fifth similarity between the current anomaly and historical cases in the graph structure or semantic vector space;

[0087] At least two of the first, second, third, fourth, and fifth similarities are assigned corresponding weights based on different fault types and application scenarios; the weighted similarities of each dimension are fused to obtain a final comprehensive similarity score; and the most relevant historical fault cases are recommended based on the final comprehensive similarity score according to a preset threshold.

[0088] S52. Based on the matching results and the fault evolution logic defined in the knowledge graph, simulate the possible development path of the anomaly and predict the specific fault type and time window that it may eventually cause.

[0089] "Fault Evolution Logic" is a composite knowledge system that integrates causal network topology, probabilistic conditions, and multi-factor dynamic models. It includes inference algorithms based on graph reasoning and probabilistic graphical models, inference algorithms based on dynamic simulation and modeling, inference algorithms based on search and path planning, and inference algorithms based on data-driven and pattern matching. Its inference relies on the technical means of integrating multiple algorithms such as graph reasoning, dynamic simulation, heuristic search, and data-driven pattern matching, and is organically integrated within the framework of cloud knowledge graphs. Ultimately, it achieves multi-path, probabilistic, and temporal prediction of the future development of fault initiation.

[0090] For example, based on a confirmed fault hypothesis (e.g., "initial stage of inter-turn short circuit"), the system calls the physical evolution model of the fault and historical case statistics encapsulated in the knowledge graph. The model will take into account parameters such as the current characteristic state (concentration and ratio of each gas), equipment load rate, and ambient temperature, and simulate the possible evolution path of fault characteristics (e.g., gas generation rate, partial discharge amount) at different time scales in the future (e.g., 24 hours, 72 hours, one week).

[0091] S53. Based on the current equipment operating conditions, environmental factors, and network load, assess the risk level and probability of each predicted path.

[0092] The simulation generates multiple possible evolutionary branches, each corresponding to a risk probability. This probability is calculated by weighting various factors, including: the significance of the current abnormal characteristics, the frequency of occurrence of this path in similar historical cases, and the accelerating effect of current operating conditions (such as whether it is during the estimated peak load period) on fault development. For example, the system might output: "Path A (slow development): 30% probability of developing into audible discharge within 72 hours; Path B (accelerated deterioration): 5% probability of triggering heavy gas protection due to a surge in short-circuit current within 24 hours." Based on the simulation results, the system will clearly determine which stage of the fault lifecycle the equipment is currently in. For example, it might determine it as "Inter-turn short circuit—early stage of insulation degradation, no stable discharge channel has yet formed."

[0093] S6. The cloud platform sends the diagnostic report and optimized monitoring strategy instructions to the corresponding edge monitoring unit to update its local model and monitoring focus.

[0094] The optimized monitoring strategy instructions include instructing the edge-side monitoring unit to focus on monitoring specific feature dimensions that are more correlated with the current abnormal pattern in the next stage, adjusting the data sampling frequency, or enabling backup sensors for cross-validation.

[0095] Example 2

[0096] This invention discloses a cloud platform-based power equipment monitoring and early warning system. The method is applicable to the aforementioned power equipment monitoring and early warning methods, and includes:

[0097] The knowledge graph construction and maintenance module, deployed on a cloud platform, is used to construct, update, and store the power equipment fault knowledge graph.

[0098] Edge intelligent sensing modules are deployed on the side of each power equipment to perform data acquisition, feature extraction, entropy rate calculation and preliminary anomaly judgment;

[0099] The cloud-based intelligent diagnosis and simulation module, deployed on a cloud platform, is used to receive early warning signals and perform in-depth diagnosis and path simulation.

[0100] The cloud-edge collaborative optimization module is used to manage command issuance, model updates, and policy synchronization between the cloud platform and the edge.

[0101] Taking a 110kV oil-immersed transformer as an example, the edge intelligent sensing module continuously monitors its oil chromatogram (content of gases such as H2, C2H2, and CH4), core grounding current, top oil temperature, winding hot spot temperature, vibration spectrum, and noise, among other multi-dimensional data. Under normal circumstances, although these parameters fluctuate slightly, their combined entropy change rate remains within a stable range. One day, the edge intelligent sensing module calculated and discovered that the entropy change rate began to show a small but continuous positive shift (entropy increase trend). After confirmation for three consecutive cycles (30 seconds), the edge intelligent sensing module uploaded an early warning signal.

[0102] After receiving the data, the cloud-based intelligent diagnosis and inference module matches it against the knowledge graph. The graph shows that similar weak entropy increase patterns are associated with early "multi-point grounding of the iron core" faults in 60% of historical cases and "internal localized overheating" in 30%. The graph rules further indicate that "multi-point grounding of the iron core" faults are usually accompanied by "intermittent spikes in the iron core grounding current" and "a slow increase in vibration energy in a specific frequency band (such as 100Hz)". The cloud-based intelligent diagnosis and inference module immediately instructs the transformer's edge intelligent sensing module to double the sampling and analysis frequency of the iron core grounding current and vibration data within the next 5 minutes, and to focus on extracting vibration energy in the 100Hz frequency band.

[0103] After the edge intelligent sensing module executed the command, it indeed captured several microsecond-level spikes in the core grounding current and a slow rise in 100Hz vibration energy within the next two minutes. This new evidence further strengthened the cloud's judgment of the "multi-point grounding of the core" fault path, reducing the probability of other paths. The cloud intelligent diagnosis and inference module finally generated a report: Warning device: XX substation #1 main transformer; Highly suspected fault: Early core multi-point grounding (probability 75%); Risk level: Medium; Predicted development time window: May develop into a monitorable steady-state current anomaly within the next 7-15 days; Recommended measures: Strengthen oil chromatography tracking, arrange a power outage within one week to check the core grounding lead. Based on this report, the maintenance personnel arranged preventive maintenance in advance, avoiding the escalation of the fault.

[0104] The embodiments described above are merely preferred embodiments of this specification and are not intended to limit the scope of this specification. Any modifications and improvements made by those skilled in the art to the technical solutions of this specification without departing from the spirit of this specification should fall within the protection scope defined by the claims of this specification.

Claims

1. A method for monitoring and early warning of power equipment based on a cloud platform, characterized in that, Includes the following steps: A knowledge graph of power equipment faults is constructed and dynamically updated. The knowledge graph takes power equipment entities as central nodes and associates their historical operating data, fault modes, environmental factors and coupling relationships between equipment as nodes and edges. A monitoring unit is deployed at the edge to collect multi-dimensional operating data of the target power equipment in real time. The multi-dimensional operating data includes at least video data, infrared thermal imaging data, acoustic data, and partial discharge data. Based on the multidimensional operational data, the real-time entropy change rate of the high-dimensional feature vector is calculated at the edge side. The high-dimensional feature vector is obtained by feature extraction and fusion of the multidimensional operational data. The real-time entropy change rate is compared with the health benchmark entropy change rate model learned based on the knowledge graph and historical data. When a continuous and slight abnormal deviation trend is detected in the real-time entropy change rate, the edge side generates and uploads a lightweight early warning trigger signal containing abnormal pattern fragments and environmental context to the cloud platform. After receiving the warning trigger signal, the cloud platform, guided by the knowledge graph, performs deep pattern matching and fault evolution path deduction in the full high-dimensional feature space to generate a comprehensive diagnostic report that includes potential fault types, evolution stages, risk probabilities, and suggested intervention measures. The cloud platform will send the diagnostic report and optimized monitoring strategy instructions to the corresponding edge monitoring unit to update its local model and monitoring focus; The process of constructing and dynamically updating a knowledge graph of power equipment faults includes the following steps: From the structured and unstructured power equipment inspection corpus, equipment entities, attribute entities, and fault entities are extracted, and the relationships between entities are defined, including subordinate relationships, location coupling relationships, and fault propagation relationships. An iterative representation enhancement graph convolutional neural network based on self-attention and cross-attention mechanisms is adopted to iteratively update and enhance the node and edge features in the knowledge graph, so that fault information can be effectively propagated and reasoned among different device nodes. Based on the new diagnostic cases received by the cloud platform and the verification results fed back from the edge side, the entities, relationships, and weights of the knowledge graph are continuously expanded and corrected. Calculating the real-time entropy change rate of a high-dimensional feature vector involves the following steps: The multidimensional operating data is preprocessed and aligned to extract time-domain, frequency-domain, and spatial-domain features, which are then fused to construct a high-dimensional feature vector characterizing the overall operating status of the equipment. Based on the sliding time window, calculate the information entropy of the high-dimensional feature vector within the current window; The rate of change of information entropy between consecutive time windows is calculated to obtain the entropy change rate, which serves as a core indicator for measuring the orderly evolution of the system. The optimized monitoring strategy instructions include instructing the edge-side monitoring unit to focus on monitoring specific feature dimensions that are more correlated with the current abnormal pattern in the next stage, adjusting the data sampling frequency, or enabling backup sensors for cross-validation.

2. The power equipment monitoring and early warning method based on a cloud platform according to claim 1, characterized in that, The edge side generates and uploads a lightweight alert trigger signal, specifically as follows: The edge monitoring unit only uploads compressed data packets that identify the anomaly start time, duration, key feature dimensions, and current environmental parameters, rather than all the original running data.

3. The power equipment monitoring and early warning method based on a cloud platform according to claim 1, characterized in that, Deep pattern matching and fault evolution path deduction specifically include the following steps: The cloud platform will perform multi-dimensional similarity matching between the received abnormal pattern fragments and the historical fault case library stored in the knowledge graph; Based on the matching results and the fault evolution logic defined in the knowledge graph, the possible development path of the anomaly is simulated, and the specific fault type and time window that it may eventually cause are predicted. Based on the current equipment operating conditions, environmental factors, and network load, assess the risk level and probability of each predicted path.

4. The power equipment monitoring and early warning method based on a cloud platform according to claim 3, characterized in that, The cloud platform performs multi-dimensional similarity matching between the received abnormal pattern fragments and the historical fault case database stored in the knowledge graph. This includes the following steps: The current abnormal pattern fragment and historical failure cases are deconstructed into structured features, and the role set and type set of their key features are extracted respectively; the first similarity between the role sets and the second similarity between the type sets are calculated; A deep learning time series model is used to encode temporal abnormal pattern fragments and historical failure cases to obtain their respective high-dimensional feature vectors; the cosine similarity between the high-dimensional feature vectors is calculated as the third similarity representing the deep pattern of the sequence morphology. The current abnormal pattern fragment is taken as the subsequence to be detected and matched with the abnormal pattern profiles extracted from historical failure cases stored in the knowledge graph; the minimum distance between the subsequence to be detected and each of the abnormal pattern profiles is calculated, and the minimum distance is normalized and used as the fourth similarity to characterize the morphological similarity of the abnormal waveform. Extract entity and relationship information from the current anomaly alert and context information; perform mixed matching between the entity and relationship information and nodes in the knowledge graph, including exact matching, fuzzy matching based on edit distance, and semantic matching based on word vectors; based on the matching results, use a graph neural network to calculate the fifth similarity between the current anomaly and historical cases in the graph structure or semantic vector space; At least two of the first, second, third, fourth, and fifth similarities are assigned corresponding weights based on different fault types and application scenarios; the weighted similarities of each dimension are fused to obtain a final comprehensive similarity score; and the most relevant historical fault cases are recommended based on the final comprehensive similarity score according to a preset threshold.

5. A cloud-based power equipment monitoring and early warning system, wherein the monitoring and early warning system is used to implement the power equipment monitoring and early warning method according to any one of claims 1-4, characterized in that, include: The knowledge graph construction and maintenance module, deployed on a cloud platform, is used to construct, update, and store the power equipment fault knowledge graph. Edge intelligent sensing modules are deployed on the side of each power equipment to perform data acquisition, feature extraction, entropy rate calculation and preliminary anomaly judgment; The cloud-based intelligent diagnosis and simulation module, deployed on a cloud platform, is used to receive early warning signals and perform in-depth diagnosis and path simulation. The cloud-edge collaborative optimization module is used to manage command issuance, model updates, and policy synchronization between the cloud platform and the edge.

6. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the power equipment monitoring and early warning method as described in any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the power equipment monitoring and early warning method as described in any one of claims 1 to 4.

Images