An edge-computing-based power equipment monitoring and early warning method

Abstract

This invention provides a power equipment monitoring and early warning method based on edge computing. The method involves collecting multi-source status data of the target power equipment from edge nodes at the equipment side or station end, forming a time-stamped monitoring data sequence; clustering the data within a baseline monitoring period to establish a baseline clustering state for the equipment nodes; matching newly collected data with the baseline clustering state, determining the trigger time of an abnormal event when the data exceeds the corresponding data distribution range, and extracting continuous monitoring data before and after that time to form an abnormal event window; and generating normal upload data or abnormal upload data based on whether an abnormal event occurs within the current monitoring period, and performing equipment status early warning analysis accordingly. This invention retains the data reduction effect of edge clustering while reducing the weakening of short-term anomalies during data compression, and provides data basis for early warning judgment of equipment nodes and related components.

Description

Technical Field

[0001] This invention relates to the field of power equipment monitoring and early warning technology, and in particular to a power equipment monitoring and early warning method based on edge computing. Background Technology

[0002] With the continuous advancement of digitalization in power systems, power equipment such as transformers, switchgear, and transmission lines are typically equipped with monitoring units for temperature, vibration, partial discharge, and current to continuously acquire equipment operation data. In actual operation, a substation often contains multiple monitoring nodes, and the sampling frequencies of various sensors differ. If all the collected raw data were uploaded to the cloud, the communication links and servers would bear a significant data processing burden. Based on this situation, existing solutions are beginning to offload some computational tasks to the equipment side or substation, where edge nodes first perform data clustering, feature extraction, or status determination before uploading the processed results.

[0003] For example, Chinese invention patent application CN114298175A discloses a power equipment status monitoring and fault early warning method based on edge computing. It clusters equipment operation data within a certain period of time at the edge and uses node identifiers and cluster centers for subsequent early warning, so as to reduce the amount of data involved in transmission and calculation.

[0004] The equipment status trend perception method disclosed in Chinese invention patent application CN118152970A extracts equipment operating characteristics at the edge and forms local status assessment results, which are then uploaded to the cloud for further analysis. This method effectively reduces data transmission burden and demonstrates that edge computing has become a common processing method in online monitoring of power equipment.

[0005] In summary, cluster centers or comprehensive status indicators are, after all, generalizations of data over a period of time. For changes such as gradually increasing temperature and continuously increasing vibration, they can usually reflect the corresponding trends. However, partial discharge pulses, instantaneous vibration impacts, or short-term current surges have short durations and account for a small proportion of a large amount of normal data. After clustering or feature compression, the abnormal features may be weakened. Existing processing results usually cannot retain the continuous change process of each monitoring quantity before and after the occurrence of an anomaly, which will affect the identification of early fault symptoms.

[0006] To address this, a power equipment monitoring and early warning method based on edge computing is proposed. Summary of the Invention

[0007] In view of this, the present invention provides a power equipment monitoring and early warning method based on edge computing to solve or alleviate the technical problems existing in the prior art, and at least provides a beneficial option.

[0008] The technical solution of this invention is implemented as follows: A power equipment monitoring and early warning method based on edge computing, comprising the following steps: Step 1: The edge node collects multi-source status data of the target power equipment, and organizes the multi-source status data according to the equipment node, monitoring quantity category and sampling time to obtain a monitoring data sequence with time stamp; Step 2: Select data corresponding to the baseline monitoring period from the monitoring data sequence and perform clustering to obtain the baseline cluster of the corresponding device node. Each baseline cluster is represented by the cluster center and the data distribution range, and together they constitute the baseline clustering state of the device node. First, the data within the baseline monitoring period are normalized according to the monitoring quantity category. Then, the data of different monitoring quantity categories are clustered separately. The distance distribution between each monitoring data and the cluster center within the same baseline cluster is used to determine the data distribution range of that baseline cluster.

[0009] Step 3: After the newly collected monitoring data reaches the edge node, it is matched with the baseline clustering state. First, the baseline cluster closest to the monitoring data is determined. When the monitoring data exceeds the data distribution range of the baseline cluster, the corresponding sampling time is determined as the abnormal event trigger time. The trigger time for an abnormal event is determined as follows: calculate the deviation distance between the newly collected monitoring data and the cluster center of the nearest benchmark cluster, and calculate the change of the monitoring data relative to the monitoring data at the previous sampling time; when the monitoring data is outside the corresponding data distribution range, and the change corresponding to the consecutive sampling times is either positive or negative, the trigger time for the abnormal event is determined.

[0010] Step 4: Using the trigger time of the abnormal event as the boundary, retrieve the continuous monitoring data before and after that time from the monitoring data sequence cached by the edge node, and arrange them according to the sampling time to form abnormal event window data; Step 5: If no abnormal event trigger occurs during the current monitoring period, cluster the data within that monitoring period to form normal upload data; if an abnormal event trigger occurs during the current monitoring period, retain the clustering results for that monitoring period and associate them with the abnormal event window data to form abnormal upload data. Step 6: Perform device status early warning analysis based on the normal or abnormal uploaded data, and output the early warning results for the corresponding device nodes; By combining the monitoring quantity category that triggers the abnormal event and the correspondence between the power equipment component and the monitoring quantity category, the power equipment component associated with the abnormal event is determined; the equipment node identifier and the power equipment component identifier are recorded in the early warning result.

[0011] More preferably, a correspondence between power equipment components and monitoring quantity categories is established in advance; when retrieving continuous monitoring data in step four, in addition to retaining the monitoring quantity data that triggered the abnormal event, data of other monitoring quantities corresponding to the same power equipment component within the same time period are also retrieved.

[0012] More preferably, the abnormal event window data includes a pre-triggered data segment, a triggered data segment, and a state change data segment; the pre-triggered data segment is located before the abnormal event triggering time, the triggered data segment contains the abnormal event triggering time, and the state change data segment is located after the abnormal event triggering time, and the three are arranged continuously according to the sampling time.

[0013] More preferably, the abnormal uploaded data records at least the device node identifier, the time when the abnormal event was triggered, and the monitoring quantity category that triggered the abnormal event; the deviation data of the current monitoring data relative to the most recent benchmark cluster, the clustering results of the current monitoring period, and the abnormal event window data are uploaded in association according to the same abnormal event identifier.

[0014] More preferably, the data uploaded under normal circumstances is marked with a device node identifier to indicate the data source, and the cluster centers formed during the current monitoring period are retained; the data quantity and data distribution range corresponding to each cluster center are uploaded together with the cluster center.

[0015] More preferably, it also includes an update judgment process for the early warning model: comparing the distribution difference between the current clustering state and the baseline clustering state for each monitoring period; if the distribution difference persists in the continuous monitoring period, the corresponding device node is recorded as having experienced data distribution drift, and the early warning model update is initiated.

[0016] More preferably, when data distribution drift occurs only in one device node, the edge-side local model used by that device node is updated; when the cluster centers of the corresponding monitoring quantities in multiple device nodes of the same device type all shift in the same direction of numerical change relative to their respective baseline cluster states, the model update information of each device node is sent to the cloud, and the cloud updates the global model accordingly.

[0017] The embodiments of the present invention have the following advantages due to the adoption of the above technical solutions: First, this invention, after clustering monitoring data at edge nodes, determines the trigger time of anomalies based on the deviation of newly collected data from the baseline clustering state, and extracts continuous monitoring data before and after that time. This preserves the data reduction effect of clustering while preventing short-term anomalies from being weakened by a large amount of normal data.

[0018] Second, this invention generates normal upload data and abnormal upload data based on whether abnormal events occur within the monitoring period. Under normal conditions, only clustering results are uploaded; under abnormal conditions, abnormal event window data is also uploaded, reducing normal communication volume while preserving relatively complete data for anomaly analysis.

[0019] Third, by establishing a correspondence between power equipment components and monitoring quantity categories, this invention can synchronously retrieve relevant monitoring quantity data when an anomaly occurs, and identify the equipment nodes and components associated with the anomaly, which facilitates subsequent fault location and maintenance. Attached Figure Description

[0020] Figure 1 This is a flowchart of the steps of the method of the present invention. Detailed Implementation

[0021] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0022] like Figure 1 As shown, this embodiment of the invention provides a power equipment monitoring and early warning method based on edge computing, mainly used for transformers, switchgear, transmission line joints, cable terminals, and other power equipment that requires continuous monitoring of their operating status. In practical applications, edge nodes can be configured near the equipment, or existing edge gateways at the station can be used to handle data access, caching, and front-end analysis tasks. The edge nodes are connected to temperature, vibration, partial discharge, current, and voltage sensors. The specific monitoring parameters connected can be determined according to the type of equipment being monitored, and it is not required that all equipment use the exact same sensor combination.

[0023] In this embodiment, after receiving the data uploaded by each sensor, the edge node first aggregates it according to the device node, and then organizes it into a monitoring data sequence by combining the monitoring quantity category and sampling time. The device node can be an independent device or a region within the device that needs to be monitored separately. Taking a transformer as an example, the transformer body, bushings, and on-load tap changers can be regarded as different device nodes. For switchgear, it can be divided according to the busbar connection part, circuit breaker mechanism, and cable joint. Each monitoring data has a corresponding time stamp, so that when an anomaly occurs later, the edge node can accurately retrieve data forward and backward, rather than just storing a single isolated anomaly value.

[0024] After the system is put into operation, a segment of monitoring data in normal operating condition is selected as the baseline data. The data is then normalized according to categories such as temperature, vibration, current, and partial discharge amplitude to avoid the difference in the dimensions of different monitoring quantities directly affecting the clustering results. After normalization, the data of each category are clustered to obtain one or more baseline clusters.

[0025] This embodiment uses the K-means clustering algorithm to process the normalized monitoring data. The initial cluster centers are determined using the K-means++ method to reduce the impact of random selection of initial cluster centers on the clustering results. The number of clusters can be preset according to the number of equipment operating normally, or it can be determined in combination with the silhouette coefficient. After clustering is completed, the cluster centers of each benchmark cluster are recorded, and the distance distribution from the monitoring data within the cluster to the corresponding cluster center is statistically analyzed to determine the data distribution range of each benchmark cluster. Multiple benchmark clusters together constitute the benchmark clustering state of the corresponding equipment node.

[0026] After the baseline cluster state is established, the edge nodes continuously receive new monitoring data. Whenever a new set of data is obtained, the data is first matched with the baseline clusters of the corresponding device nodes to find the nearest baseline cluster. Then, it is determined whether the data falls within the data distribution range of the cluster. If it is still within the distribution range, it is processed as normal data. If it has exceeded the distribution range, it indicates that the current data has deviated from the existing normal state.

[0027] In this embodiment, the change of the current monitoring data relative to the data at the previous sampling time is calculated, and the subsequent sampling results are continuously observed. When the change corresponding to the consecutive sampling times is positive or negative, it indicates that the monitoring quantity maintains the same direction of change over a continuous period of time. At this time, if the monitoring data is located outside the data distribution range of the nearest benchmark cluster, the corresponding sampling time is determined as the trigger time of the abnormal event. For sudden noise that occurs only once and then returns to the original distribution range, an abnormal event may not be formed immediately.

[0028] After an abnormal event is triggered, the edge node does not only upload a single data point at the trigger time, but uses that time as a time base to retrieve continuous data before and after from the locally cached monitoring data sequence. The data before the trigger time constitutes the pre-trigger data segment, which is used to reflect the state changes of the device before the abnormality occurs. Data containing the trigger moment constitutes the trigger data segment; data collected and retained after the trigger moment is used as the state change data segment.

[0029] The above three parts are arranged consecutively according to the sampling time to form the abnormal event window data.

[0030] With this processing method, even if a partial discharge pulse, short-term vibration impact, or current surge has a short duration, it will not be weakened by normal data due to its subsequent participation in large-scale clustering. The server or early warning model receives an abnormal data window with a preceding and following process, thus showing how the anomaly occurred, whether it continued to develop, and whether it recovered, rather than only seeing a compressed single clustering result.

[0031] In one implementation, the edge node also pre-stores the correspondence between power equipment components and monitoring quantity categories. Taking a switchgear cable joint as an example, the monitoring quantities associated with the component may include joint temperature, partial discharge signal and corresponding loop current. When the temperature data triggers an abnormal event, the edge node, in addition to extracting the temperature data, also simultaneously retrieves the partial discharge data and loop current data within the same time period, placing the monitoring quantities related to the same component in the same abnormal event window, providing a basis for subsequent judgment of the source of the abnormality.

[0032] In this embodiment, when no abnormal event trigger occurs during the current monitoring period, the edge node performs routine clustering on the data for that period, uploading only the device node identifier, cluster center, and the data quantity and distribution range corresponding to the cluster center. During normal operation, it is not necessary to repeatedly transmit all the original data to the server, thereby controlling the communication volume.

[0033] If an abnormal event has already occurred within the current monitoring period, the original clustering results for that period will be retained. At the same time, the abnormal event window data, the abnormal event trigger time, the monitoring quantity category that triggered the abnormality, and its deviation from the most recent benchmark cluster will be associated to form abnormal upload data.

[0034] The above content can be bound using the same abnormal event identifier to avoid mismatch after the server receives multiple types of monitoring data; the amount of data uploaded in an abnormal state will increase accordingly, but the increase is concentrated in a limited time period before and after the abnormality occurs, and will not be transformed into a long-term continuous upload of all original data.

[0035] After receiving normal uploaded data, the server or the early warning analysis unit located at the station uses it to judge the overall operating trend of the equipment. After receiving abnormal uploaded data, it combines the clustering results and the abnormal event window for analysis. Based on the monitoring quantity category that triggered the abnormality and the pre-established correspondence between equipment components, it can further determine the equipment components that may be associated with the abnormality. The output results record at least the equipment node identifier and the equipment component identifier. It can also add the abnormal time, associated monitoring quantity and processing level according to the actual configuration of the early warning system, so as to facilitate maintenance personnel to quickly locate the problem.

[0036] During long-term operation of equipment, the original baseline clustering state may also change. For example, equipment aging, load structure adjustments, or seasonal environmental changes can all cause the overall distribution of monitoring data to gradually shift. In this case, the current clustering state can be compared with the baseline clustering state for each monitoring period. If this distribution difference only occurs within a short period, it is still treated as an abnormal event. If it persists for multiple consecutive monitoring periods, the corresponding equipment node is recorded as having experienced data distribution drift, and the early warning model is updated to make a judgment.

[0037] When data distribution drift occurs only in a single device node, the local edge model corresponding to that device node is updated first, without needing to adjust other devices simultaneously. If, among multiple device nodes of the same type, the cluster centers for the same monitored quantity all shift in the same numerical direction relative to their respective baseline states, it indicates that the change may not be an accidental fluctuation of a single device. In this case, the model update information for each device node is sent to the cloud, which then updates the global model based on the changes across multiple nodes. After the model update, new operational data can be used for verification. If the verification results do not meet the early warning requirements, the model before the update is retained to avoid directly learning sensor drift or short-term anomalies as a new normal state.

[0038] This embodiment retains the role of edge clustering in reducing the amount of data uploaded, but data that deviates from the baseline clustering state is no longer directly incorporated into the normal clustering results. Instead, continuous data before and after the anomaly is extracted separately. During normal operation, the clustering results are still uploaded in a simplified manner. When a short-term anomaly occurs, the corresponding change process is retained. The two data processing methods are switched according to the actual operating status of the device.

[0039] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in the present invention, and these should all be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for monitoring and early warning of power equipment based on edge computing, characterized in that, Includes the following steps: Step 1: The edge node collects multi-source status data of the target power equipment, and organizes the multi-source status data according to the equipment node, monitoring quantity category and sampling time to obtain a monitoring data sequence with time stamp; Step 2: Select data corresponding to the baseline monitoring period from the monitoring data sequence and perform clustering to obtain the baseline cluster of the corresponding device node. Each baseline cluster is represented by the cluster center and the data distribution range, and together they constitute the baseline clustering state of the device node. Step 3: After the newly collected monitoring data reaches the edge node, it is matched with the baseline clustering state to first determine the baseline cluster closest to the monitoring data. When the monitoring data exceeds the data distribution range of the benchmark cluster, the corresponding sampling time is determined as the abnormal event trigger time; Step 4: Using the trigger time of the abnormal event as the boundary, retrieve the continuous monitoring data before and after that time from the monitoring data sequence cached by the edge node, and arrange them according to the sampling time to form abnormal event window data; Step 5: If no abnormal event trigger occurs during the current monitoring period, cluster the data within that monitoring period to form routine upload data; If an abnormal event is triggered during the current monitoring period, the clustering results for that monitoring period are retained and associated with the abnormal event window data to form abnormal uploaded data. Step 6: Perform device status early warning analysis based on the normal or abnormal uploaded data, and output the early warning results for the corresponding device nodes.

2. The power equipment monitoring and early warning method based on edge computing according to claim 1, characterized in that, In step two, the data within the baseline monitoring period are first normalized according to the monitoring quantity category, and then the data of different monitoring quantity categories are clustered separately; the distance distribution between each monitoring data and the cluster center within the same baseline cluster is used to determine the data distribution range of the baseline cluster.

3. The power equipment monitoring and early warning method based on edge computing according to claim 1, characterized in that, In step three, the trigger time of the abnormal event is determined as follows: calculate the deviation distance between the newly collected monitoring data and the cluster center of the nearest benchmark cluster, and calculate the change of the monitoring data relative to the monitoring data at the previous sampling time; when the monitoring data is outside the corresponding data distribution range, and the change corresponding to the consecutive sampling times is either positive or negative, the trigger time of the abnormal event is determined.

4. The power equipment monitoring and early warning method based on edge computing according to claim 1, characterized in that, Establish a correspondence between power equipment components and monitoring quantity categories in advance; when retrieving continuous monitoring data in step four, in addition to retaining the monitoring quantity data that triggered the abnormal event, retrieve the data of other monitoring quantities corresponding to the same power equipment component within the same time period.

5. The power equipment monitoring and early warning method based on edge computing according to claim 1, characterized in that, The abnormal event window data includes a pre-trigger data segment, a trigger data segment, and a state change data segment; the pre-trigger data segment is located before the abnormal event trigger time, the trigger data segment contains the abnormal event trigger time, and the state change data segment is located after the abnormal event trigger time, and the three are arranged continuously according to the sampling time.

6. The power equipment monitoring and early warning method based on edge computing according to claim 1, characterized in that, The abnormal uploaded data shall at least record the device node identifier, the time when the abnormal event was triggered, and the monitoring quantity category that triggered the abnormal event; the deviation data of the current monitoring data from the most recent benchmark cluster, the clustering results of the current monitoring period, and the abnormal event window data shall be uploaded in association according to the same abnormal event identifier.

7. The power equipment monitoring and early warning method based on edge computing according to claim 1, characterized in that, The data uploaded under normal circumstances is marked with the device node identifier to indicate the data source, and the cluster centers formed during the current monitoring period are retained; the data quantity and data distribution range corresponding to each cluster center are uploaded together with the cluster center.

8. The power equipment monitoring and early warning method based on edge computing according to claim 1, characterized in that, It also includes the process of updating the early warning model: comparing the distribution difference between the current clustering state and the baseline clustering state for each monitoring period; if the distribution difference persists in the continuous monitoring period, the corresponding device node is recorded as having experienced data distribution drift, and the early warning model is updated.

9. A power equipment monitoring and early warning method based on edge computing according to claim 8, characterized in that, When data distribution drift occurs only in one device node, the edge-side local model used by that device node is updated; when the cluster centers of the corresponding monitoring quantities in multiple device nodes of the same device type all shift in the same direction of change relative to their respective baseline cluster states, the model update information of each device node is sent to the cloud, and the cloud updates the global model accordingly.

10. A power equipment monitoring and early warning method based on edge computing according to claim 4, characterized in that, In step six, the power equipment component associated with the abnormal event is determined by combining the monitoring quantity category that triggers the abnormal event and the correspondence between the power equipment component and the monitoring quantity category; the warning result records the equipment node identifier and the power equipment component identifier.

Images