Switch cabinet state monitoring system based on cloud edge collaborative architecture

Abstract

The invention discloses a switch cabinet state monitoring system based on a cloud edge collaborative architecture, and belongs to the field of power systems. The system comprises a multi-source sensing unit, an edge computing unit, a field actuator, a cloud server and an early warning linkage terminal, and the edge computing unit is connected with the multi-source sensing unit, the field actuator and the cloud server; and the early warning linkage terminal is connected with the cloud server. According to the invention, multi-dimensional data fusion and early warning of the operation state of the substation switch cabinet are realized through cloud-side cooperation.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of power systems, and in particular, the present application relates to a switch cabinet state monitoring system based on a cloud-edge collaborative architecture. BACKGROUND

[0002] As the core equipment of the medium and low voltage distribution network, the operation state of the switch cabinet is directly related to the safety and stability of the power grid and the reliability of power supply. The traditional switch cabinet state monitoring mainly relies on periodic manual inspection, preventive test and single parameter online monitoring (such as temperature), which has the obvious disadvantages of long monitoring period, single information dimension, poor real-time performance, dependence on personnel experience, etc., and is difficult to meet the needs of intelligent operation and maintenance. In recent years, the rapid development of sensing technology, Internet of Things (IoT) and communication technology makes it possible to deploy various types of sensors (such as temperature and humidity sensors, partial discharge detection devices, zero sequence current transformers, mechanical property online monitoring devices, etc.) on the switch cabinet, realizing real-time sensing of multi-dimensional parameters such as equipment environment, electrical insulation, mechanical state and load current.

[0003] Although multi-sensor monitoring technology has been applied to switch cabinets, existing systems still have many key defects in data fusion, processing architecture and early warning mechanism, which seriously restrict the efficiency and accuracy of intelligent early warning: 1. Protocol heterogeneity and data island: Different manufacturers and different types of sensors usually use different communication protocols (such as Modbus, CAN, RS485, wireless LoRa, etc.) and data formats. Existing systems generally lack unified and efficient protocol adaptation and data standardization capabilities, making it difficult for different sources and different dimensions of sensing data to effectively collaborate, forming a "data island" and unable to perform deep correlation analysis and fusion modeling.

[0004] 2. Lack of edge intelligence: Existing edge data acquisition devices have single functions and mainly play the role of "data movers", only responsible for uploading raw data to the cloud. They usually do not have localized data processing (such as data cleaning, feature extraction, preliminary fusion), real-time computing (such as threshold discrimination, simple model reasoning) and strategy execution capabilities. This "edge weak" architecture leads to "data up to the cloud is dead" — that is, a large amount of raw, low-value and even redundant data flows to the cloud, increasing the transmission and storage burden, but failing to fully utilize the real-time processing advantages of the edge side.

[0005] 3. Cloud processing is disconnected from the field: Although the cloud platform has strong computing and storage capabilities, its rule setting and model updating often have a lag. The results of cloud analysis and the generated early warning strategies are difficult to deliver to the field edge device for execution in a timely and accurate manner. More importantly, there is a lack of effective edge-cloud linkage mechanism, and the cloud cannot dynamically adjust the edge computing logic or trigger local control instructions according to real-time data and events in the field.

[0006] 4. Insufficient warning accuracy and real-time performance: Due to the lack of effective multi-dimensional data fusion algorithms and edge-cloud collaborative processing capabilities, existing systems rely mainly on single-parameter fixed threshold alarms or simple cloud rule matching for early warning. This approach is difficult to capture the complex correlation characteristics of the evolution of device "sub-health" state to failure (such as combined abnormal situations of temperature and humidity changes, partial discharge, and sudden increase in zero sequence current), resulting in high false alarm rate, high risk of missed reports, and delayed warning, which cannot achieve accurate prediction and early intervention of failures.

[0007] 5. Weak closed-loop control capability: Existing early warning systems usually only have information push functions (such as SMS, platform alarms), lack automated linkage control capabilities with field execution devices (such as triggering cable cabinet heater dehumidification, starting partial discharge suppression devices, and executing relay protection locking), and are difficult to form an intelligent closed loop of "perception-analysis-warning-disposal".

[0008] Therefore, the present application proposes a switchgear state monitoring system based on a cloud-edge collaborative architecture. SUMMARY

[0009] The present application aims to overcome the shortcomings of the prior art and proposes a switchgear state monitoring system based on a cloud-edge collaborative architecture to achieve the following purposes: multi-dimensional data fusion and early warning of substation switchgear operating state.

[0010] To achieve the above purposes, the technical solution adopted by the present application is as follows: a switchgear state monitoring system based on a cloud-edge collaborative architecture, the system comprising a multi-source perception unit, an edge computing unit, a field executor, a cloud server, and an early warning linkage terminal, wherein the edge computing unit is connected to the multi-source perception unit, the field executor, and the cloud server; the early warning linkage terminal is connected to the cloud server.

[0011] Preferably, the multi-source perception unit comprises a temperature and humidity sensor, a partial discharge detector, a zero sequence current transformer, and a mechanical property detection device, all of which are connected to the edge computing unit.

[0012] Preferably, the early warning linkage terminal comprises a user mobile terminal and a SCADA system.

[0013] Preferably, the edge computing unit comprises a data acquisition and preprocessing module, a feature extraction module, an intelligent analysis module, and a local control module, wherein: The data acquisition and preprocessing module is configured to acquire multi-source data collected by the multi-source perception unit, and to send the preprocessed multi-source data to the feature extraction module; the preprocessing includes data standardization and data cleaning. The feature extraction module is configured to extract features from the preprocessed multi-source data and send the feature data to the intelligent analysis module and the cloud server. The intelligent analysis module is configured to use a machine learning-based event analysis model deployed locally to analyze the input feature data, and to send the event analysis result to the cloud server and the local control module. The local control module is configured to generate a control instruction based on the event analysis result and send the control instruction to the field actuator for execution.

[0014] Preferably, in the intelligent analysis module, the event analysis model is trained in advance using historical fault data after event calibration, wherein the event calibration comprises: When temperature rise, partial discharge frequency energy rise, and zero sequence current fluctuation exceed the preset threshold value simultaneously, it is calibrated as an insulation deterioration high-risk event. When mechanical vibration peak value exceeds the preset threshold value and the opening and closing time is not within the preset time range, it is calibrated as a mechanical fault symptom event.

[0015] Preferably, the event analysis model comprises a random forest model.

[0016] Preferably, after the cloud server receives the event analysis result, it reviews the result using a pre-stored knowledge graph. If the number of occurrences of the same event in the same type of equipment within a preset time before the current event occurrence time is greater than a preset number threshold value, the review is passed. When the review is passed, the cloud server sends an instruction to the early warning linkage terminal to perform linkage early warning.

[0017] Preferably, the cloud server is deployed with a deep learning-based switch cabinet health state judgment model, wherein the switch cabinet health state judgment model is established as follows: Pre-train a teacher model based on historical full-amount fault data. A student model is collectively trained through local historical fault data of multiple edge computing units, data between which is not shared, i.e., an initial student model is issued to each edge computing unit by a cloud server, each edge computing unit respectively trains the initial student model with local historical fault data thereof, and model parameters are uploaded to the cloud server after training is completed; the cloud server generates a final student model after aggregating model parameters uploaded by all edge computing units; The final student model is trained by using a pre-trained teacher model, and finally a switch cabinet health state judgment model is obtained.

[0018] Preferably, the cloud server further issues the switch cabinet health state judgment model parameters to each edge computing node, so as to locally deploy the switch cabinet health state judgment model at each edge computing node.

[0019] The technical effect of the present application is: The present application realizes end-edge-cloud integrated deployment, and can realize unified data management of the whole station, the whole process and all parameters.

[0020] The edge end of the present application is deployed with an event analysis model based on machine learning, which can identify risk events from multi-source data, so as to perform subsequent safety operations and early warning. At the same time, the event analysis result is also reviewed through a knowledge graph in the cloud to avoid false positives.

[0021] The cloud end of the present application is deployed with a switch cabinet health state judgment model based on deep learning, which can identify the health state of the switch cabinet in advance, so as to perform early warning. The switch cabinet health state judgment model is constructed by the methods of centralized training, federated learning and knowledge distillation, which realizes lightweight model while retaining the performance of large model, thereby improving the judgment accuracy. At the same time, based on the characteristics of lightweight, the switch cabinet health state judgment model can also be deployed to the edge end, so as to realize edge end self-checking, which greatly reduces the burden of the cloud end. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 A switch cabinet state monitoring system structure block diagram based on a cloud-edge collaborative architecture is provided for the embodiments of the present application. DETAILED DESCRIPTION

[0023] The specific embodiments of the present application are further described in detail below with the description of the embodiments with reference to the accompanying drawings, the purpose of which is to help the skilled in the art to have a more complete, accurate and in-depth understanding of the inventive concept and technical solution of the present application, and to help its implementation. It should be noted that the terms "first", "second" and the like in the present application are only for the convenience of describing the technical solutions and are used as a distinction between components, and the corresponding component configurations may be the same or different, and the present application is not limited to this. In order to make the technical solution of the present application more clear, the present application is explained and described by the following embodiments.

[0024] The present embodiment provides a switch cabinet state monitoring system based on a cloud-edge collaborative architecture, as shown in Figure 1 The system includes a multi-source sensing unit, an edge computing unit, a field actuator, a cloud server, and a warning linkage terminal. The edge computing unit is connected with the multi-source sensing unit, the field actuator, and the cloud server. The warning linkage terminal is connected with the cloud server.

[0025] Specifically, the multi-source sensing unit is used for monitoring the state of the switch cabinet and sending it to the edge computing unit, including a temperature and humidity sensor, a partial discharge detector, a zero sequence current transformer, a mechanical property detection device, etc. They support connection with the edge computing unit through multiple communication protocols such as Modbus, CAN, RS485, and wireless LoRa. The specific communication protocol selected can be flexibly selected according to the actual situation.

[0026] In the present embodiment, the edge computing unit is deployed near the switch cabinet or terminal box, and is used for data monitoring and local control, including a data acquisition and preprocessing module, a feature extraction module, an intelligent analysis module, and a local control module.

[0027] The data acquisition and preprocessing module is used for acquiring multi-source data collected by the multi-source sensing unit and sending it to the feature extraction module after preprocessing; the preprocessing includes data standardization and data cleaning. The multi-source data includes temperature, humidity, partial discharge, zero sequence current, mechanical vibration, etc. Due to the different specifications of various sensors and communication protocols, after receiving the multi-source data, the edge computing unit first needs to perform data standardization, that is, according to different communication protocols, the specific data content is parsed and converted into a unified data format, and a timestamp, device ID and data quality identifier are added, so as to facilitate subsequent data analysis. Further, the initially collected multi-source data often contains interference and noise, so data cleaning is also needed, including using a sliding window detection method to eliminate outliers (such as sensor transient interference), using a historical mean or linear interpolation method to fill in missing data, and finally obtaining multi-source data after preprocessing.

[0028] The feature extraction module is used for feature extraction of the pre-processed multi-source data and sends the feature data to the intelligent analysis module and the cloud server. The feature extraction includes calculating the sliding window mean, the rate of change and the cumulative temperature rise of the temperature data, performing wavelet packet transform on the partial discharge signal to extract frequency energy features; calculating the time domain statistics (peak value, root mean square value) and frequency domain harmonic components of the mechanical vibration signal. In specific implementation, the required feature data can be obtained from the multi-source data as needed.

[0029] The intelligent analysis module is used for event analysis of the input feature data by deploying a machine learning-based event analysis model locally. The event analysis result is sent to the cloud server and the local control module respectively. The event analysis model is trained in advance by historical fault data after event calibration to learn multi-parameter association rules. The event calibration includes: When the temperature rises, the partial discharge frequency energy rises, and the zero sequence current fluctuation exceeds the preset threshold value, it is calibrated as an insulation deterioration high-risk event; When the mechanical vibration peak value exceeds the preset threshold value and the closing and opening time does not occur within the preset time range, it is calibrated as a mechanical fault symptom event. After the training of the event analysis model is completed, the input of real-time feature data can realize the output of the event analysis result. In this embodiment, the random forest model is selected as the event analysis model, and the tree depth (such as ≤5 layers) and the number of trees (such as 10-20) are limited, so as to realize the lightweight of the model, facilitate the edge deployment, speed up the event analysis speed, and facilitate timely warning and measures.

[0030] The local control module is used for generating control instructions according to the time analysis result to execute the field actuators. For example, if it is determined as an insulation deterioration high-risk event, a “lock relay” instruction is generated and sent to the relay protection device through CAN bus or RS485; if it is determined as a mechanical fault symptom event, a “prohibit closing” instruction is generated and sent to the circuit breaker control loop through hardwiring or communication mode. The execution log is recorded and fed back to the cloud for subsequent strategy optimization.

[0031] In this embodiment, the cloud server obtains real-time feature data and event analysis results uploaded from multiple edge computing nodes, as well as power grid operation data of the SCADA system (such as load current, voltage, and switch opening and closing state), device account data of the production management system (such as device model, commissioning date, and maintenance history), macro weather data of the environmental monitoring system (such as station area environmental temperature and humidity), and historical fault case library. A time window sliding alignment algorithm is used to unify data of different frequencies and different time delays to the same time reference, ensuring that various types of data at the same time can be correctly associated, and then a knowledge graph with the switch cabinet as the core node is established, associating environmental parameters, electrical indicators, mechanical states, and other entities, and defining the causal relationship between entities.

[0032] For example, the core entities of the knowledge graph of this embodiment include: Device entity: such as "10kV switch cabinet 101", "circuit breaker", and "current transformer", with attributes including device ID, model, commissioning date, and health condition score.

[0033] State parameter entity: such as "temperature_101", "partial discharge_101", and "vibration_101", with attributes including parameter value, timestamp, data quality, and change trend.

[0034] Event entity: such as "2025-08-18T14:30:00_insulation deterioration warning_101", with attributes including event type, trigger time, confidence, and associated edge node ID.

[0035] Maintenance entity: such as "maintenance work order_2025001" and "inspection record_202408", with attributes including maintenance type, performer, result, and time.

[0036] Environmental entity: such as "station area environmental temperature" and "weather".

[0037] The relationship definitions between the core entities of the knowledge graph of this embodiment include: Belongs to: relationship type, such as "circuit breaker A" belongs to "switch cabinet 101".

[0038] Monitors: relationship type, such as "temperature sensor_01" monitors "circuit breaker A".

[0039] Has status: relationship type, such as "switch cabinet 101" has status "temperature_101".

[0040] Causes: relationship type, such as "continuous high temperature" causes "insulation aging".

[0041] Occurs simultaneously: relationship type, such as "partial discharge increase" occurs simultaneously with "zero sequence current anomaly".

[0042] Trigger: Relationship type, such as "High risk of insulation deterioration event" triggers "Generate maintenance work order".

[0043] Affected by the environment: Relationship type, such as "Cabinet humidity" is affected by "Station area environment humidity".

[0044] The final use graph database Graph is used as a storage engine, and the above entities are used as nodes and relationships as edges to build a dynamic knowledge graph and store it in the cloud server.

[0045] After the cloud server receives the event analysis result, it can be reviewed through the pre-stored knowledge graph. If it is found through the knowledge graph that the number of occurrences of the same type of equipment in a predetermined period of time before the current event occurrence time is greater than the predetermined number threshold, it is determined that the review is passed. When the review is passed, the cloud server sends an instruction to the early warning linkage terminal to perform linkage early warning, so that the safety operation is performed through the on-site executor at the edge, and the early warning linkage terminal is linked through the cloud to perform early warning, and multi-level early warning is realized through cloud-edge collaboration.

[0046] In order to further improve the intelligent level of the system and provide users with more switch cabinet state information for reference, the cloud server is deployed with a switch cabinet health state judgment model based on deep learning. The health state represents the health level, which is pre-calibrated in the training data of the model. The establishment method of the switch cabinet health state judgment model is as follows: A teacher model is pre-trained based on historical full-fault data. The teacher model can be a deep LSTM network or a Transformer model. This model aims to explore extremely complex fault patterns and serve as a benchmark for performance. A student model is trained by multiple edge computing units using local historical fault data. The student model is more lightweight than the teacher model (such as a lightweight LSTM network). The data between the multiple edge computing units is not shared. The cloud server issues an initial student model to each edge computing unit, and each edge computing unit trains the initial student model using its own local historical fault data. After training is complete, the model parameters are uploaded to the cloud server. The cloud server aggregates all the model parameters uploaded by the edge computing units using the FedAvg algorithm to generate a final student model. This approach not only protects data privacy but also fully utilizes the distributed data value of multiple edge computing units.

[0047] The pre-trained teacher model is used to train the final student model, i.e., knowledge distillation, to migrate the knowledge learned by the complex teacher model, such as the soft labels of the output layer or the feature representation of the intermediate layer, to the lightweight student model, and finally obtain the switch cabinet health state judgment model. This enables the small model to maintain high computational efficiency while achieving performance close to that of the large model.

[0048] At the same time, based on the characteristics of the switch cabinet health state judgment model, the cloud server also distributes the switch cabinet health state judgment model parameters to each edge computing node, so as to perform local deployment of the switch cabinet health state judgment model on each edge computing node, and thus the judgment of the switch cabinet health state can be realized on the edge side. In specific implementation, as the data obtained by the cloud server is continuously updated, the switch cabinet health state judgment model is periodically retrained to enable the model to quickly adapt to new failure modes or equipment aging trends without forgetting historical knowledge. The model parameters after retraining are also distributed to each edge computing unit.

[0049] Further, the cloud server can also generate different control signals to the early warning linkage terminal according to the switch cabinet health state for early warning. In this embodiment, the early warning linkage terminal includes a user mobile terminal, a SCADA system, etc. Through the user mobile terminal, early warning information can be directly sent to the user end to improve user experience. Through the SCADA system, a structured maintenance work order can be generated, including device ID, risk type, confidence, disposal suggestion, etc., and the monitoring picture is used to remind the early warning information.

[0050] The above describes the present application in conjunction with the drawings. Obviously, the specific implementation of the present application is not limited by the above method. As long as various non-essential improvements are made using the method concept and technical solution of the present application, or the above concept and technical solution of the present application is directly applied to other occasions without improvement, they are all within the protection scope of the present application.

Claims

1. A switchgear status monitoring system based on a cloud-edge collaborative architecture, characterized in that: The system includes a multi-source sensing unit, an edge computing unit, a field actuator, a cloud server, and an early warning linkage terminal. The edge computing unit is connected to the multi-source sensing unit, the field actuator, and the cloud server, respectively; the early warning linkage terminal is connected to the cloud server.

2. The switchgear status monitoring system based on a cloud-edge collaborative architecture according to claim 1, characterized in that: The multi-source sensing unit includes a temperature and humidity sensor, a partial discharge detector, a zero-sequence current transformer, and a mechanical characteristic detection device, all of which are connected to the edge computing unit.

3. The switchgear status monitoring system based on a cloud-edge collaborative architecture according to claim 1, characterized in that: The early warning linkage terminal includes a user mobile terminal and a SCADA system.

4. The switchgear status monitoring system based on a cloud-edge collaborative architecture according to claim 1, characterized in that: The edge computing unit includes a data acquisition and preprocessing module, a feature extraction module, an intelligent analysis module, and a local control module, wherein: The data acquisition and preprocessing module is used to acquire multi-source data collected by the multi-source sensing unit, and send it to the feature extraction module after preprocessing; the preprocessing includes data standardization and data cleaning. The feature extraction module is used to extract features from the preprocessed multi-source data and send the feature data to the intelligent analysis module and the cloud server. The intelligent analysis module is used to perform event analysis on the input feature data through a machine learning-based event analysis model deployed locally. The event analysis results are sent to the cloud server and the local control module, respectively. The local control module is used to generate control commands based on the time analysis results and send them to the field actuator for execution.

5. A switchgear status monitoring system based on a cloud-edge collaborative architecture as described in claim 4, characterized in that: In the intelligent analysis module, the event analysis model is pre-trained using historical fault data after event labeling, wherein the event labeling includes: When temperature rises, partial discharge frequency domain energy increases, and zero-sequence current fluctuation exceeds a preset threshold simultaneously, it is calibrated as a high-risk event of insulation degradation. When the peak value of mechanical vibration exceeds the preset threshold and the opening and closing times do not occur simultaneously within the preset time range, it is calibrated as a mechanical fault symptom event.

6. A switchgear status monitoring system based on a cloud-edge collaborative architecture as described in claim 4 or 5, characterized in that: The event analysis model includes the random forest model.

7. A switchgear status monitoring system based on a cloud-edge collaborative architecture as described in claim 5, characterized in that: After receiving the event analysis results, the cloud server verifies them using a pre-stored knowledge graph. If the knowledge graph query reveals that the number of times the same event occurred on similar devices within a preset time period prior to the current event's occurrence time is greater than a preset threshold, the verification is considered successful.

8. A switchgear status monitoring system based on a cloud-edge collaborative architecture according to claim 7, characterized in that: Upon successful verification, the cloud server sends a command to the early warning linkage terminal to issue a linked early warning.

9. A switchgear status monitoring system based on a cloud-edge collaborative architecture according to claim 1, characterized in that: The cloud server is equipped with a deep learning-based health status assessment model for switchgear. The method for establishing the health status assessment model for switchgear is as follows: A teacher model is pre-trained based on all historical fault data. A student model is trained by using local historical fault data from multiple edge computing units. The data between these edge computing units is not shared. That is, the cloud server distributes the initial student model to each edge computing unit, and each edge computing unit trains the initial student model using its own local historical fault data. After training, the model parameters are uploaded to the cloud server. The cloud server aggregates the model parameters uploaded by all edge computing units and generates the final student model. The pre-trained teacher model is used to train the final student model, resulting in a switch cabinet health status judgment model.

10. A switchgear status monitoring system based on a cloud-edge collaborative architecture according to claim 9, characterized in that: The cloud server also distributes the switch cabinet health status judgment model parameters to each edge computing node, thereby enabling the local deployment of the switch cabinet health status judgment model on each edge computing node.

Images