Real-time monitoring and fault prediction method and system for intelligent medium-voltage switch cabinet

Abstract

The invention belongs to the technical field of power system monitoring and maintenance, and relates to an intelligent medium-voltage switch cabinet real-time monitoring and fault prediction method and system. The data acquisition and processing module is used for acquiring real-time data of the interior of the switch cabinet and the surrounding environment of the switch cabinet and performing data preprocessing; the anomaly detection and fault prediction module is used for screening out abnormal data based on the preprocessed data, calculating a reconstruction error value and judging whether the reconstruction error value exceeds a normal range or not; the alarm module is used for giving an alarm by adopting a grading alarm mechanism when abnormal data or an error value is detected to exceed a normal range; by collecting various operation data such as temperature, voltage and current of equipment in real time and combining with an intelligent algorithm, anomaly detection and fault prediction are realized, and the system can provide multi-parameter monitoring, intelligent diagnosis and fault early warning functions for the switch cabinet in a high-voltage electric power environment, so that the safety, stability and maintenance efficiency of electric power equipment are effectively improved.

Description

Technical Field

[0001] The invention belongs to the technical field of power system monitoring and maintenance, and relates to an intelligent medium-voltage switch cabinet real-time monitoring and fault prediction method and system. Background Art

[0002] In modern power systems, medium voltage switchgear is a vital component for the control and distribution of power. These switchgears typically handle voltages up to 35KV, and their design must take into account operational safety, maintenance convenience, and long-term reliability.

[0003] However, the traditional medium-voltage switchgear is relatively simple in design and is usually equipped with only basic electrical monitoring devices, such as single-color status indicators and simple alarm devices, which cannot provide detailed feedback on the multi-dimensional status of the equipment. This simple monitoring method is difficult to support complex operating environment requirements, especially in high-load, rapidly changing power grid environments, where the monitoring capabilities of traditional switchgear are difficult to meet. In addition, traditional medium-voltage switchgear also has obvious deficiencies in fault handling. Most devices rely on the "post-alarm" mode, that is, the alarm will only be issued after the fault occurs and reaches a certain severity, and usually only limited fault information can be provided. This method makes it impossible to perform preventive maintenance on the equipment before the fault occurs, making maintenance operations passive. In addition, due to the lack of effective fault prediction methods, operators often find it difficult to intervene in time at the early abnormal stage of the equipment, increasing the possibility of the expansion of the fault impact. This lagging fault handling mode not only increases the risk of power grid operation, but also significantly reduces the service life and reliability of the equipment. In addition, current maintenance and fault diagnosis work usually relies on the experience of operators, especially the judgment and maintenance decision of the internal operating status of the switchgear. Due to the lack of intelligent diagnosis and automatic monitoring methods, operators need to regularly check the equipment and manually collect data to determine the equipment status. This method is not only time-consuming and labor-intensive, but also has high errors and subjectivity. Summary of the invention

[0004] The purpose of the present invention is to overcome the above-mentioned shortcomings of the prior art and to provide an intelligent medium voltage switch cabinet real-time monitoring and fault prediction method and system.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] An intelligent medium voltage switchgear real-time monitoring and fault prediction method comprises the following steps:

[0007] Step 1: Obtain real-time data inside the switch cabinet and the environment around the switch cabinet and perform data preprocessing;

[0008] Step 2: Based on the preprocessed data, filter out abnormal data and calculate the reconstruction error value to determine whether the reconstruction error value exceeds the normal range; wherein: the abnormal data is screened by using a multi-scale GEN network and a multi-layer embedded high-dimensional data modeling method;

[0009] Step 3: When abnormal data or error values ​​are detected that exceed the normal range, a graded alarm mechanism is used to issue an alarm;

[0010] Step 4: The alarm information generated by the alarm system is synchronously sent to the remote management system and saved in the log.

[0011] Furthermore, the real-time data in step one is obtained through a temperature sensor, a humidity sensor, a voltage sensor, and a current sensor; the temperature sensor is used to collect temperature data of the core heating components of the switch cabinet, the current sensor and the voltage sensor are installed at the input and output ports of the switch cabinet, and the humidity sensor is installed inside the switch cabinet.

[0012] Furthermore, the sampling frequency of the temperature sensor, the voltage sensor, and the current sensor is once per second, and the sampling frequency of the humidity sensor is once per minute.

[0013] Furthermore, the high-dimensional data modeling method of the multi-scale GEN network and multi-layer embedding mainly includes graph structure construction, multi-scale graph embedding, multi-scale feature fusion, anomaly detection and scoring mechanism, model training and optimization.

[0014] Furthermore, the graph structure construction includes the weight w for the edge ij The calculation is as follows:

[0015]

[0016] x i is the feature vector of node i; x j is the feature vector of node j; σ is the bandwidth parameter of the kernel function;

[0017] Furthermore, the multi-scale feature fusion is a node representation obtained by multi-layer embedding is integrated into the final node embedding H i , specifically:

[0018]

[0019] Among them: concat is a concatenation operation; is the node obtained by multi-layer embedding; L is the number of layers of the graph embedding network.

[0020] Furthermore, the anomaly detection and scoring mechanism is based on the node embedding H iThe density of each node is calculated, and then the anomaly score is defined. The node is judged as abnormal according to the preset threshold τ, specifically:

[0021] The density of each node is calculated as:

[0022]

[0023] Where: Ν(i) is the neighborhood set of node i; w ij is the similarity weight between nodes i and j; j is a node;

[0024] The calculation formula for the anomaly score is:

[0025]

[0026] Where: ρi is the density of each node;

[0027] According to the preset threshold τ, the calculation formula for abnormal judgment of node i is:

[0028]

[0029] Furthermore, the reconstruction error value is achieved by constructing an autoencoder model, and the reconstruction error of the autoencoder is defined as:

[0030]

[0031] Where: x is the data point; is the square difference.

[0032] Furthermore, the triggering condition of the hierarchical alarm mechanism is based on the output of the anomaly detection and fault prediction module. If the value of is lower than the first alarm threshold, the alarm system triggers a yellow alarm; if the anomaly score Score or the reconstruction error is higher than the second alarm threshold, the alarm system triggers a red alarm.

[0033] The present invention also provides an intelligent medium voltage switch cabinet real-time monitoring and fault prediction system, comprising:

[0034] Data acquisition and processing module: used to collect real-time data inside the switch cabinet and the environment around the switch cabinet and perform data preprocessing;

[0035] Anomaly detection and fault prediction module: Based on the preprocessed data, filter out abnormal data and calculate the reconstruction error value to determine whether the reconstruction error value exceeds the normal range;

[0036] Alarm module: used to use a graded alarm mechanism to issue an alarm when abnormal data or error values ​​beyond the normal range are detected;

[0037] Record saving module: used to synchronously send the alarm information generated by the alarm system to the remote management system and save it to the log.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] 1. This invention applies the Internet of Things, artificial intelligence and automatic control technologies to traditional medium-voltage switchgear, aiming to improve the equipment's operation monitoring and fault prediction capabilities. By real-time collection of equipment's temperature, voltage, current and other operating data, combined with intelligent algorithms to achieve abnormality detection and fault prediction, the system can provide switchgear with multi-parameter monitoring, intelligent diagnosis and fault warning functions in a high-voltage power environment, thereby effectively improving the safety, stability and maintenance efficiency of power equipment, without the need for manual detection, saving time and effort, and with high accuracy.

[0040] 2. The present invention has a dual protection mechanism of anomaly detection method of multi-scale GEN network and autoencoder, which ensures that the system can actively carry out prediction and early warning work in both abnormal and abnormal situations, thereby improving the intelligent operation and maintenance level of power equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The accompanying drawings are incorporated in and constitute a part of this specification and, together with the description, serve to explain the principles of the present invention.

[0042] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0043] Figure 1 It is a schematic diagram of the process steps of the present invention;

[0044] Figure 2 It is a schematic diagram of the self-encoding fault detection model of the present invention. DETAILED DESCRIPTION

[0045] Here, exemplary embodiments will be described in detail, and the embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Instead, they are only examples consistent with some aspects of the present invention described in detail in the appended claims.

[0046] Example

[0047] like Figures 1-2 As shown, a method for real-time monitoring and fault prediction of an intelligent medium-voltage switch cabinet includes the following steps:

[0048] Step 1: Obtain real-time data inside the switch cabinet and the environment around the switch cabinet and perform data preprocessing; (This step is completed through the data acquisition and processing module)

[0049] Step 2: Based on the preprocessed data, filter out abnormal data and calculate the reconstruction error value to determine whether the reconstruction error value exceeds the normal range; wherein: the abnormal data is filtered out using a multi-scale GEN network and a multi-layer embedded high-dimensional data modeling method; (filtering out abnormal data is achieved through a real-time anomaly detection module, and calculating the reconstruction error value is achieved through a fault prediction module)

[0050] Step 3: When abnormal data is detected or the error value exceeds the normal range, a graded alarm mechanism is used to alarm; (This step is implemented through the alarm module)

[0051] Step 4: Save record module: The alarm information generated by the alarm system is synchronously sent to the remote management system and saved to the log. (This step is implemented by saving the record module)

[0052] In this embodiment: the overall process of the system is as follows:

[0053] Data acquisition: The sensor collects the real-time data of the equipment and transmits it to the anomaly detection and fault prediction module after preprocessing.

[0054] Real-time anomaly detection: Multi-scale GEN network and multi-layer embedded high-dimensional data modeling method perform real-time detection of data and quickly identify data anomalies.

[0055] Fault prediction: The autoencoder calculates the reconstruction error of real-time data and triggers an early warning when the error exceeds the normal range.

[0056] Alarm and display: When an abnormality is detected or a failure trend is predicted, the system automatically triggers an alarm on the display screen and alarm system.

[0057] Remote monitoring and log recording: Alarm information is synchronously sent to the remote management terminal and saved in the log system, and operators can perform remote management and tracking.

[0058] The intelligent 35KV medium-voltage switchgear system of the present invention realizes all-round intelligent monitoring, abnormality detection and fault warning of equipment, effectively improving the operation safety, management efficiency and response speed of the equipment.

[0059] Anomaly detection and fault prediction are two key links in intelligent power equipment monitoring. Although their core tasks are different, they are closely linked and work together in practical applications. Anomaly detection aims to analyze equipment operation data in real time and identify abnormal patterns that may deviate from the normal state; fault prediction evaluates the possibility of anomalies developing into specific faults and the time point when they occur. However, this collaborative relationship is not limited to the input results of anomaly detection. The fault prediction model can also capture potential trend changes and make forward-looking fault inferences through continuous operation data analysis in the absence of obvious anomalies.

[0060] In the system, the anomaly detection method based on the multi-scale GEN network first monitors the multi-dimensional data in real time. If an abnormal point is detected, the abnormal characteristics and related data will be input into the autoencoder model for fault prediction. The autoencoder uses its latent space modeling ability to evaluate whether the abnormality may lead to a specific fault, or predict the possible time and type of fault. If no abnormality is detected, the fault prediction model will still receive real-time operation data as input, and through the learning and analysis of long-term trends and historical patterns, it will discover fault trends that may not yet show obvious abnormalities but have potential risks. This dual protection mechanism ensures that the system can actively carry out prediction and early warning work in both the absence and presence of abnormalities, thereby improving the intelligent operation and maintenance level of power equipment.

[0061] Furthermore, the data acquisition and processing module includes a temperature sensor, a humidity sensor, a voltage sensor, and a current sensor; the temperature sensor is used to collect temperature data of the core heating components of the switch cabinet, the current sensor and voltage sensor are installed at the input and output line ports of the switch cabinet, and the humidity sensor is installed inside the switch cabinet.

[0062] In this embodiment: the data acquisition module is responsible for collecting real-time data inside the switch cabinet and the surrounding environment, and is the basis for system operation.

[0063] (1) Sensor selection and layout: Temperature sensors, current sensors, voltage sensors and humidity sensors are arranged inside and outside the equipment.

[0064] Temperature sensors should be installed near the core heat-generating components of the switch cabinet (such as switches and busbars) to monitor temperature changes; current sensors and voltage sensors should be installed at the input and output ports to ensure accurate measurement of current and voltage values.

[0065] Humidity sensors should be installed in the air circulation area of ​​the switch cabinet to monitor the humidity changes in the internal environment in real time and avoid insulation problems caused by excessive humidity.

[0066] (2) Data collection frequency setting:

[0067] The sampling frequency of temperature, current and voltage is set to once per second to ensure real-time monitoring of high-frequency data, especially when there are large load changes or ambient temperature fluctuations.

[0068] The humidity sampling frequency is set to once per minute because it changes slowly and does not require high-frequency updates.

[0069] (3) Data preprocessing process:

[0070] De-noising: De-noising is performed on the collected raw data. The temperature and current data are smoothed by the moving average method to filter out short-term fluctuations and reduce noise interference.

[0071] Outlier processing: Use median filtering to process obvious outliers. For example, if the current or voltage data point deviates from the previous data point by more than a set threshold (such as 3 times the standard deviation), it is determined to be an outlier and corrected.

[0072] Data format conversion: Data is uniformly formatted as floating point type to ensure data transmission accuracy, especially between sensors and data processing modules.

[0073] (4) Data transmission:

[0074] Use high-speed communication protocols (such as Modbus, CAN bus) to transmit processed data to ensure anti-interference during transmission. Data is transmitted to the real-time anomaly detection module and fault prediction module via wired or wireless networks.

[0075] Furthermore, the high-dimensional data modeling method of the multi-scale GEN network and multi-layer embedding mainly includes graph structure construction, multi-scale graph embedding, multi-scale feature fusion, anomaly detection and scoring mechanism, model training and optimization.

[0076] In this embodiment:

[0077] (1) High-dimensional data modeling based on multi-scale graph embedding network GEN

[0078] The present invention proposes a high-dimensional data modeling method based on a multi-scale graph embedding network (GEN) to cope with the multi-dimensional complexity and nonlinear relationship problems of medium-voltage switchgear operation data. By constructing a graph structure, the multi-dimensional features of the equipment operation data are mapped to nodes, the similarity between nodes is defined as the weight of the edge, and the multi-scale graph embedding network is used to extract local and global correlation information between data. The multi-scale design captures the characteristics of the node neighborhood at different levels through a hierarchical aggregation mechanism, which not only retains local detail information, but also enhances the perception of global structural relationships. Compared with traditional linear or single-scale feature-based modeling methods, the multi-scale graph embedding network can dynamically learn the potential structure of high-dimensional data, comprehensively improve the ability to characterize complex operation modes, and provide a more accurate and efficient basic support for anomaly detection.

[0079] (2) Distribution drift correction

[0080] In order to solve the problem of dynamic distribution changes of multidimensional data in time or space, this method introduces a distribution drift correction mechanism, evaluates the difference between the source data distribution and the target data distribution through the maximum mean difference (MMD), and uses adversarial learning to dynamically adjust the distribution to achieve correction. Unlike traditional static anomaly detection methods, this module can capture changes in data distribution in real time, ensure the adaptability of the model in different time periods and environments, and thus improve the robustness of anomaly detection. Through this mechanism, the interference of data distribution changes on anomaly detection results can be effectively reduced, which is suitable for complex multidimensional scenarios.

[0081] (3) Neighborhood Aggregation Mechanism

[0082] When there are nonlinear relationships between data, traditional methods are difficult to fully extract features. This method uses the neighborhood aggregation characteristics of graph embedding networks to construct a similarity-based graph structure and dynamically extract the relationship information between nodes and neighborhoods. By converting high-dimensional data into a graph structure, the neighborhood aggregation mechanism can capture the local characteristics and nonlinear associations of the data. Compared with the traditional method of directly analyzing a single data point, neighborhood aggregation can more comprehensively express the internal relationship of the data, thereby improving the accuracy of detection.

[0083] (4) Multi-layer embedded structure

[0084] In order to improve the feature extraction capability of the model, this method designs a multi-layer embedding structure to extract local and global features layer by layer and construct a multi-scale expression. Multi-layer embedding can effectively capture the deep feature relationship of the data and ensure the detection model's ability to recognize complex patterns. Unlike shallow models, multi-layer embedding not only improves the sensitivity to local features, but also enhances the expression of global features, providing stronger support for high-dimensional data anomaly detection.

[0085] (5) Real-time optimization

[0086] Low-latency anomaly detection is crucial in large-scale real-time data processing scenarios. This method achieves efficient streaming processing by designing a distributed processing framework, optimizing the data processing process and computing efficiency. The optimization framework can reduce data processing latency while maintaining detection accuracy, ensuring that the system outputs detection results in real time under dynamic environments. The introduction of real-time optimization makes this method not only suitable for offline data analysis, but also able to meet the needs of real-time monitoring and rapid response.

[0087] The graph structure construction includes the weight w for the edge ij The calculation is as follows: Equipment operation data X = {x 1 ,x 2 ,...,x n} samples, each with d-dimensional features. We transform the data into a graph G = (V, E), where:

[0088] V={v 1 ,v,...,v n} is a node set, each node corresponds to a sample;

[0089] E is an edge set, which is used to define the similarity relationship between samples.

[0090] The weight of the edge w ij Calculated using the Gaussian kernel function:

[0091]

[0092] x i is the feature vector of node i; x j is the feature vector of node j; σ is the bandwidth parameter of the kernel function;

[0093] Each element of the weight matrix W is:

[0094]

[0095] The multi-scale graph embedding integrates the information of the neighboring nodes into the target node through the neighborhood aggregation mechanism (in order to fully extract the potential features of multi-dimensional data, a multi-scale graph embedding strategy is adopted. The graph embedding network integrates the information of the neighboring nodes into the target node through the neighborhood aggregation mechanism). The node embedding of the lth layer is expressed as:

[0096]

[0097] in: is the feature embedding of node i in the (l)th layer; Ν(i) is the neighborhood set of node i; w ijis the similarity weight between nodes i and j; c i is the normalization factor of node i, usually defined as c_i=\sum_{j\in\mathcal{N}(i)}w_{ij}; W (l) is the learnable weight matrix of the (l)th layer; σ is the nonlinear activation function.

[0098] The multi-scale design extracts features of different scales through multiple embedding layers: the embeddings in the lower layers capture local neighborhood information; the embeddings in the higher layers extract global structure information.

[0099] Furthermore, the multi-scale feature fusion is a node representation obtained by multi-layer embedding is integrated into the final node embedding H i (The node representation obtained by multi-layer embedding is integrated into the final node embedding H i ), specifically:

[0100]

[0101] Among them: concat is a concatenation operation; is the node obtained by multi-layer embedding; L is the number of layers of the graph embedding network. Multi-scale feature fusion can simultaneously express the local characteristics and global patterns of nodes, significantly improving the ability to model complex nonlinear relationships.

[0102] Furthermore, the anomaly detection and scoring mechanism is based on the node embedding H i The density of each node is calculated, and then the anomaly score is defined. The node is judged as abnormal according to the preset threshold τ, specifically:

[0103] According to the node embedding H i Calculate the density of each node. The density represents its "normality" in the neighborhood. The calculation formula for the density of each node is:

[0104]

[0105] Where: Ν(i) is the neighborhood set of node i; w ij is the similarity weight between node i and node j;

[0106] Nodes with lower density are considered more likely to be abnormal. The calculation formula for abnormality score is:

[0107]

[0108] Where: i is the density of each node.

[0109] According to the preset threshold τ, the calculation formula for abnormal judgment of node i is:

[0110]

[0111] Model training and optimization: To optimize the graph embedding network parameters W^{(l)}, supervised learning loss based on node classification is used:

[0112]

[0113] Where: y ik is the true label of node i belonging to category k; p ik is the probability predicted by the model.

[0114] Technical advantages:

[0115] Multi-scale expression capability: Through a hierarchical embedding mechanism, local and global features are captured simultaneously.

[0116] Dynamic adaptability: Based on graph structure modeling, the neighborhood range and embedding depth can be automatically adjusted.

[0117] Accurate detection: The anomaly scoring mechanism combined with graph embedding features improves the accuracy and robustness of detection.

[0118] Real-time support: The optimized computing process makes the system suitable for real-time processing scenarios of large-scale data.

[0119] Experimental results and evaluation:

[0120] Experimental setup

[0121] (1) Experimental Dataset:

[0122] 1. Synthetic dataset: simulates multidimensional time series data with known outliers, containing 10,000 samples, each with 20 features.

[0123] 2. Actual data set: medium-voltage switchgear operation data, including 15 dimensional features such as temperature, voltage, current, etc., totaling 50,000 data, with known abnormal events (such as overload, short circuit, abnormal temperature rise, etc.) marked.

[0124] (2) Comparison method:

[0125] Isolation Forest(IF): A random tree based anomaly detection method.

[0126] One-Class SVM(OCSVM): A one-class classifier based on support vector machine.

[0127] AutoEncoder(AE): An unsupervised anomaly detection model based on deep learning.

[0128] Proposed Method(GEN): This paper proposes a method based on multi-scale graph embedding network.

[0129] (3) Evaluation indicators:

[0130] Accuracy: The ratio of detected anomalies to true anomalies.

[0131] Recall: The ratio of detected anomalies to actual anomalies.

[0132] Precision: The ratio of true outliers among the detected outliers.

[0133] F1-Score: The harmonic mean of recall and precision.

[0134] Experimental results:

[0135]

[0136] Results Analysis

[0137] (1) Accuracy

[0138] The accuracy of this method on both datasets is significantly better than the comparison method. On the synthetic dataset, the accuracy reaches 96.4%, which is 5.6% higher than AutoEncoder. On the actual dataset, the accuracy reaches 95.7%, which is 8.2% higher than Isolation Forest.

[0139] (2) Robustness

[0140] In terms of recall and precision indicators, the GEN method performs well, with recall rates reaching 94.7% (synthetic dataset) and 93.3% (real dataset), indicating that this method can effectively detect most anomalies while avoiding too many false positives.

[0141] (3) Capability of capturing complex anomalies

[0142] Since the multi-scale embedding network can capture the nonlinear relationship of high-dimensional data, this method performs well in detecting complex abnormal patterns (such as high-dimensional interactive features). This feature is particularly effective on real data sets.

[0143] (4) Real-time

[0144] The inference time of the GEN method is only 65% ​​of that of AutoEncoder, indicating that while improving detection performance, it optimizes computational efficiency and can support real-time anomaly detection of large-scale data.

[0145] advantage:

[0146] (1) This method is based on a multi-scale graph embedding network and has strong modeling capabilities for high-dimensional complex data, resulting in more accurate and comprehensive detection results.

[0147] (2) The distribution drift correction mechanism improves the adaptability of the method and can cope with dynamic changes in data distribution.

[0148] (3) The designed neighborhood aggregation and multi-layer embedding structure enhance the model's ability to extract local and global features.

[0149] Practical significance:

[0150] In the actual application scenarios of medium-voltage switchgear, this method can detect potential anomalies in real time and significantly improve the equipment's operation monitoring and fault prediction capabilities.

[0151] It provides an efficient and reliable solution for the intelligent maintenance of power equipment.

[0152] Evaluation of the actual effect of the algorithm:

[0153] The multi-dimensional data real-time anomaly detection method based on multi-scale graph embedding network (GEN) proposed in the present invention has demonstrated excellent performance in actual tests. In multiple experimental data sets, especially in medium-voltage switchgear operation data, this method is significantly better than the comparison method in terms of accuracy, recall and precision. Compared with traditional anomaly detection algorithms (such as Isolation Forest and AutoEncoder), the GEN method can more accurately capture the complex nonlinear correlation features in high-dimensional data and effectively identify potential abnormal patterns. This shows that GEN has obvious advantages in solving anomaly detection tasks in multi-dimensional complex scenarios.

[0154] This method achieves comprehensive extraction of local and global features through a multi-scale embedding mechanism. Compared with traditional shallow models, its ability to characterize complex abnormal patterns has been greatly improved. In the detection task of actual medium-voltage switchgear operation data, the GEN method has a recall rate of 93.3% and a precision rate of 94.1%, which significantly improves the accuracy of fault prediction. In addition, the introduction of the distribution drift correction module enables the model to dynamically adapt to changes in data distribution under different times and environments, further improving the detection robustness and ensuring the wide applicability of the method.

[0155] In terms of real-time performance, the GEN method has been optimized to have high computational efficiency. Experimental results show that its reasoning speed is significantly better than traditional deep learning methods such as AutoEncoder, which is only 65% ​​of its computational time, and can meet the high requirements for real-time anomaly detection in practical applications. This high efficiency is not only suitable for offline data analysis, but also supports real-time monitoring scenarios in power systems such as medium-voltage switchgear, providing a guarantee for rapid response.

[0156] In general, this method not only solves the shortcomings of traditional methods in high-dimensional data anomaly detection through innovative multi-scale graph embedding design and efficient real-time processing framework, but also shows high practical value in actual power equipment monitoring. Its advantage is that it can accurately identify anomalies and output results in real time, providing strong technical support for intelligent power systems. In the future, the complexity of the model can be further optimized to enhance its processing capabilities for larger data sets, providing possibilities for applications in more complex scenarios.

[0157] This technology can effectively solve the complexity problems in multi-dimensional data anomaly detection and provide efficient and reliable technical support for the operation and maintenance of medium-voltage switchgear.

[0158] Furthermore, the reconstruction error value is achieved by constructing an autoencoder model, and the reconstruction error of the autoencoder is defined as:

[0159]

[0160] Where: x is the data point; is the square difference.

[0161] In this embodiment: the reconstruction error value is implemented based on the fault prediction module.

[0162] The present invention realizes intelligent prediction through an autoencoder model. The autoencoder efficiently compresses and reconstructs the normal operation data of the equipment through the process of "encoding-decoding", and establishes a behavioral model of the normal state of the equipment. In actual operation, when the sensor data is input into the autoencoder model for reconstruction, if a reconstruction error higher than the normal range occurs, it indicates that the operating state of the equipment may be offset, and there is a potential risk of failure. The system will trigger an early warning at this time to notify the operator to inspect or maintain the equipment. This predictive early warning mechanism can significantly reduce unplanned downtime caused by sudden failures, not only ensuring the normal operation of the equipment, but also gaining time for operators to perform maintenance in advance, extending the service life of the equipment, and improving the overall reliability of the power system.

[0163] (1) Autoencoder model construction:

[0164] Model structure design: The autoencoder consists of an encoder and a decoder. The encoder compresses the input data to a low-dimensional hidden layer representation, and the decoder reconstructs the hidden layer data back to the original format. The number of layers and neurons of the encoder and decoder are set according to the complexity of the data in order to accurately reconstruct the normal data of the device.

[0165] Model training: Use historical data of the equipment under normal operating conditions as the training set, optimize the model parameters, and enable the autoencoder to minimize the error between the input data and the reconstructed data, thereby achieving effective modeling of normal state data.

[0166] Error threshold setting: Perform statistical analysis on the reconstruction error of the normal state and set an abnormal error threshold (such as 3 times the standard deviation) as the basis for judging whether the device state deviates from the normal value.

[0167] (2) Real-time fault prediction:

[0168] Data input and reconstruction error calculation: Real-time data is input into the autoencoder model, and the model calculates the reconstruction error value. If the error is within the normal range, it is judged as a normal state; if the error exceeds the threshold, it is judged as abnormal.

[0169] Early warning triggering and abnormality classification: Based on the error exceeding the threshold, the fault prediction module classifies the abnormality as minor or major. Minor abnormalities are only prompted on the display, while major abnormalities trigger a full alarm.

[0170] Update the model regularly: The autoencoder model is retrained at a specific period (such as every month) to adapt to changes in data distribution. This update mechanism enables the model to maintain high accuracy in predicting device status over the long term.

[0171] Specifically, the present invention realizes fault prediction of switch cabinet through autoencoder algorithm. The autoencoder model uses data reconstruction error to detect abnormal trends of equipment status, identifies possible faults in advance, and is suitable for predicting potential equipment faults.

[0172] The autoencoder consists of an encoder and a decoder. The encoder maps the input data x to a low-dimensional space and generates a hidden representation h(x):

[0173] h(x)=f(W e x+b e ) (10)

[0174] Where W e and b e are the weights and biases of the encoder, and f is the activation function. The decoder then reconstructs the hidden representation into the original data

[0175]

[0176] Where W d and b d are the weights and biases of the decoder, and g is the activation function.

[0177] The reconstruction error of the autoencoder is defined as:

[0178]

[0179] in is the data point x and the reconstruction value The square error between . Minimize the reconstruction error during training To ensure that the autoencoder can accurately reconstruct the data in normal state.

[0180] In practical applications, if the reconstruction error of real-time data is If it exceeds the normal value (such as more than 3 times the standard deviation), the system considers the equipment to be in an abnormal state and triggers a fault prediction warning.

[0181] The autoencoder algorithm achieves accurate modeling of the device state by compressing and reconstructing data, and its reconstruction error can directly reflect the normality of the device state. The autoencoder is particularly suitable for identifying changes in nonlinear patterns, and its early prediction effect on faults is significant. By analyzing the reconstruction error of real-time data, the autoencoder can quickly identify when the data deviates from the normal state and effectively reduce the false alarm rate before a fault occurs.

[0182] Fault prediction based on the autoencoder model provides an efficient preventive maintenance tool for equipment. The autoencoder model can identify small abnormal trends before the signs of fault appear, providing operators with sufficient maintenance time. Compared with the traditional "after-the-fact alarm" mode, the fault prediction mechanism of the present invention reduces the risk of downtime caused by sudden faults, improves the continuity and safety of equipment operation, and realizes efficient management and operation of the power system.

[0183] Furthermore, the triggering condition of the hierarchical alarm mechanism is based on the output of the anomaly detection and fault prediction module. If the value of is lower than the first alarm threshold, the alarm system triggers a yellow alarm; if the anomaly score Score or the reconstruction error is higher than the second alarm threshold, the alarm system triggers a red alarm.

[0184] In this embodiment:

[0185] (1) Status display:

[0186] Real-time data display: A display screen is equipped on the front panel of the switch cabinet to display the sensor data (such as temperature, current, voltage, humidity) and system status information in real time. The screen can display a variety of status colors, such as green for normal operation, yellow for minor abnormalities, and red for serious abnormalities.

[0187] Touch operation: The display screen is equipped with a touch function, allowing operators to view equipment historical data, alarm records or make system settings by clicking.

[0188] Data playback and log query: Operators can query historical data on the screen to analyze abnormal trends of equipment, facilitate predictive maintenance and correct settings.

[0189] (2) Multi-level alarm system:

[0190] Alarm triggering: When the anomaly detection module or fault prediction module identifies an equipment anomaly, the system triggers a multi-level alarm based on the anomaly level. Minor anomalies are indicated by a yellow warning light and screen, while serious anomalies trigger a red warning light and sound alarm.

[0191] Sound alarm: When a serious abnormality occurs, the system will sound an alarm prompt, and the sound level can be adjusted according to the ambient noise to ensure that the operator can hear it clearly in multiple environments.

[0192] Visual warning: The alarm light uses different colors and flashing frequencies to indicate the abnormality level. The yellow light flashes slowly when there is a minor abnormality, and the red light flashes quickly when there is a serious abnormality, which is convenient for operators to make quick judgments.

[0193] (3) Remote alarm notification:

[0194] Remote communication settings: The system supports sending alarm information to the manager's mobile terminal (mobile phone or tablet) via the Internet to achieve remote monitoring.

[0195] Information push method: Remote alarm can be notified by email, SMS or App. For serious faults, the system will automatically send a detailed alarm report to the responsible personnel, including information such as device name, fault type and time.

[0196] Alarm log storage: All alarm information is automatically stored in the log system for subsequent tracking and analysis.

[0197] The remote alarm and graded alarm mechanism of the present invention ensures that the equipment responds quickly to abnormal situations of different severity through a multi-level alarm system and remote notification. The mechanism combines visual, sound and remote push functions to help operators take timely action.

[0198] The triggering conditions of the hierarchical alarm are based on the outputs of the anomaly detection module and the fault prediction module, and are defined as follows:

[0199] (1) If the abnormal score or reconstruction error If the value is lower than the alarm threshold 1 (slight abnormality), the system triggers a yellow alarm;

[0200] (2) If the abnormal score or reconstruction error Above threshold 2 (serious abnormality), the system triggers a red alarm.

[0201] The alarm classification formula is as follows:

[0202]

[0203] where s yellow and L yellow is the slight abnormality threshold, s red and L red is the severe abnormality threshold.

[0204] In addition, the system sends alarm information to mobile terminals or remote monitoring systems through remote push notifications. The push information includes the type of abnormality, severity and timestamp, which is convenient for remote operators to view in real time and respond quickly.

[0205] The hierarchical alarm mechanism designed in the present invention can provide multi-level alarm prompts according to the severity of the abnormality, and ensure the operator's rapid response by combining multiple means such as sound, vision, and remote push. The system sets different threshold levels corresponding to minor and serious abnormalities, thereby achieving hierarchical early warning. At the same time, the remote alarm function can ensure that the operator can obtain alarm information in a timely manner whether on-site or remote, further enhancing the system's responsiveness.

[0206] Remote alarm and hierarchical alarm mechanism provide a convenient solution for real-time monitoring and remote management of power systems. This mechanism not only ensures that on-site operators can respond quickly, but also notifies remote managers in emergency situations, improving the flexibility of power equipment monitoring. The hierarchical alarm mechanism effectively reduces the frequency of false alarms, avoids excessive intervention due to minor anomalies, and improves operation and maintenance efficiency. This multi-level alarm design improves the system's fault response speed and management efficiency, providing operators and managers with a full range of monitoring experience.

[0207] The present invention also provides an intelligent medium voltage switch cabinet real-time monitoring and fault prediction system, comprising:

[0208] Data acquisition and processing module: used to collect real-time data inside the switch cabinet and the environment around the switch cabinet and perform data preprocessing;

[0209] Anomaly detection and fault prediction module: Based on the preprocessed data, filter out abnormal data and calculate the reconstruction error value to determine whether the reconstruction error value exceeds the normal range;

[0210] Alarm module: used to use a graded alarm mechanism to issue an alarm when abnormal data or error values ​​beyond the normal range are detected;

[0211] Record saving module: used to synchronously send the alarm information generated by the alarm system to the remote management system and save it to the log.

[0212] In summary, the intelligent 35KV medium-voltage switchgear monitoring and fault prediction system of the present invention provides an advanced solution for the equipment management of the power system through four key innovations: multi-dimensional real-time anomaly detection, progressive fault detection, intelligent prediction and hierarchical alarm. First, multi-dimensional real-time anomaly detection is achieved through a multi-scale GEN network and a multi-layer embedded high-dimensional data modeling method, which accurately analyzes multi-dimensional data such as temperature, current, and voltage, effectively identifies abnormal conditions of the equipment under any operating state, and significantly improves the monitoring accuracy of the equipment, so that the equipment can be warned before an obvious fault occurs, ensuring the continuous operation and reliability of the power system.

[0213] Secondly, the fault prediction model based on the autoencoder can predict the potential failure of the equipment in advance through intelligent data reconstruction error analysis, especially in nonlinear mode. The introduction of this innovation has transformed equipment management from the traditional "after-the-fact maintenance" mode to "predictive maintenance", effectively reducing the downtime caused by sudden failures, providing operators with sufficient response time, and reducing the economic losses caused by equipment failures. Through the application of intelligent prediction models, the efficiency of equipment management has been significantly improved, providing strong support for the safe operation of the power system, and is suitable for various power equipment scenarios that require efficient and stable operation.

[0214] Finally, the remote alarm and graded alarm mechanism of this system provides multi-level protection for the convenience and response speed of equipment management. The system performs graded alarms according to the severity of the abnormality, and ensures the transmission of alarm information through sound, vision and remote push, so that operators can obtain the equipment status in time whether on-site or remote, which is convenient for rapid response and adjustment. In addition, the graded alarm mechanism avoids frequent alarms due to minor abnormalities, improves the effectiveness of alarms and operation and maintenance efficiency. This design not only meets the safety requirements of on-site operations, but also meets the convenience of remote monitoring, and provides a comprehensive solution for the efficient and safe management of modern power systems. This system can be widely used in power systems, industrial equipment monitoring, and other equipment fields that require intelligent management, and has extremely high application value.

[0215] The above description is only a specific embodiment of the present invention, so that those skilled in the art can understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention.

[0216] It should be understood that the present invention is not limited to what has been described above and that various modifications and changes may be made without departing from its scope. The scope of the present invention is limited only by the appended claims.

Claims

1. An intelligent medium voltage switchgear real-time monitoring and fault prediction method, characterized in that: The following steps are involved: Step 1: Obtain real-time data inside the switch cabinet and the environment around the switch cabinet and perform data preprocessing; Step 2: Based on the preprocessed data, filter out abnormal data and calculate the reconstruction error value to determine whether the reconstruction error value exceeds the normal range; wherein: the abnormal data is screened by using a multi-scale GEN network and a multi-layer embedded high-dimensional data modeling method; Step 3: When abnormal data or error values ​​are detected that exceed the normal range, a graded alarm mechanism is used to issue an alarm; Step 4: The alarm information generated by the alarm system is synchronously sent to the remote management system and saved in the log.

2. The method for real-time monitoring and fault prediction of an intelligent medium-voltage switch cabinet according to claim 1 is characterized in that: The real-time data in step 1 is obtained through a temperature sensor, a humidity sensor, a voltage sensor, and a current sensor; The temperature sensor is used to collect temperature data of the core heating components of the switch cabinet, the current sensor and the voltage sensor are installed at the inlet and outlet ports of the switch cabinet, and the humidity sensor is installed in the switch cabinet.

3. The method for real-time monitoring and fault prediction of an intelligent medium-voltage switch cabinet according to claim 2 is characterized in that: The sampling frequency of the temperature sensor, voltage sensor and current sensor is once per second, and the sampling frequency of the humidity sensor is once per minute.

4. The method for real-time monitoring and fault prediction of an intelligent medium-voltage switch cabinet according to claim 1 is characterized in that: The multi-scale GEN network and multi-layer embedded high-dimensional data modeling method includes graph structure construction, multi-scale graph embedding, multi-scale feature fusion, anomaly detection and scoring mechanism, model training and optimization.

5. The method for real-time monitoring and fault prediction of an intelligent medium-voltage switch cabinet according to claim 4 is characterized in that: The graph structure construction includes the weight w for the edge ij The calculation is as follows: x i is the feature vector of node i; x j is the feature vector of node j; σ is the bandwidth parameter of the kernel function.

6. The method for real-time monitoring and fault prediction of an intelligent medium-voltage switch cabinet according to claim 4, characterized in that: The multi-scale feature fusion is a node representation obtained by multi-layer embedding is integrated into the final node embedding H i , specifically: Among them: concat is a concatenation operation; is the node obtained by multi-layer embedding; L is the number of layers of the graph embedding network.

7. The method for real-time monitoring and fault prediction of an intelligent medium voltage switch cabinet according to claim 6, characterized in that: The anomaly detection and scoring mechanism is based on the node embedding H i The density of each node is calculated, and then the anomaly score is defined. The node is judged as abnormal according to the preset threshold τ, specifically: The density of each node is calculated as: Where: Ν(i) is the neighborhood set of node i; w ij is the similarity weight between nodes i and j; j is a node; The calculation formula for the anomaly score is: Where: i is the density of each node; The calculation formula for abnormal judgment of node i is: Among them: τ is the preset threshold, Score i is the abnormality score value.

8. The method for real-time monitoring and fault prediction of intelligent medium voltage switch cabinet according to claim 1, characterized in that: The reconstruction error value is achieved by constructing an autoencoder model, and the reconstruction error of the autoencoder is defined as: Where: x is the data point; is the square difference.

9. The method for real-time monitoring and fault prediction of intelligent medium-voltage switchgear according to claim 1, characterized in that: The triggering condition of the hierarchical alarm mechanism is based on the output of the anomaly detection and fault prediction module. If the value of is lower than the first alarm threshold, the alarm system triggers a yellow alarm; if the value of the anomaly score Score or the reconstruction error is higher than the second alarm threshold, the alarm system triggers a red alarm.

10. An intelligent medium voltage switchgear real-time monitoring and fault prediction system, characterized in that: include: Data acquisition and processing module: used to collect real-time data inside the switch cabinet and the environment around the switch cabinet and perform data preprocessing; Anomaly detection and fault prediction module: Based on the preprocessed data, filter out abnormal data and calculate the reconstruction error value to determine whether the reconstruction error value exceeds the normal range; Alarm module: used to use a graded alarm mechanism to issue an alarm when abnormal data or error values ​​beyond the normal range are detected; Record saving module: used to synchronously send the alarm information generated by the alarm system to the remote management system and save it to the log.

Images