Intelligent drop-out fuse monitoring system
The intelligent drop-out fuse monitoring system utilizes gradient boosting decision trees and gated cyclic unit network models for real-time fault diagnosis and trend warning, solving the problems of lag and high false alarm rate in traditional fuse monitoring, and achieving accurate monitoring and predictive maintenance.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-25
- Publication Date
- 2026-04-03
AI Technical Summary
Traditional drop-out fuses lack condition monitoring and information exchange capabilities, resulting in delayed operation and maintenance, high false alarm rate, lack of predictive ability, and inability to achieve accurate monitoring and trend early warning.
The system uses a data acquisition module to collect sensor data in real time, combines a gradient boosting decision tree model for fault diagnosis, and utilizes a gated cyclic unit network model for trend early warning. The results are then uploaded to the backend management system via wireless IoT.
It enables accurate fault identification and trend early warning, significantly reduces false alarm rate, improves the pertinence of operation and maintenance response, shortens fault diagnosis and power restoration time, and reduces operation and maintenance costs.
Smart Images
Figure CN121784533A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of power grid monitoring technology and relates to an intelligent drop-out fuse monitoring system. Background Technology
[0002] Drop-out fuses are widely used protection and control components in power distribution networks, and their operational reliability directly affects the safety and stability of power distribution lines. However, traditional drop-out fuses are essentially mechanical and passive protection devices, lacking condition monitoring and information exchange capabilities.
[0003] Currently, monitoring the status of fuses mainly relies on regular inspections by maintenance personnel and user complaints after a fault. This method has significant limitations, failing to provide real-time information on the fuse's operating conditions (such as three-phase imbalance or overload trends), and making it difficult to quickly locate and determine the type of fault when one occurs, resulting in prolonged power outages and low maintenance efficiency.
[0004] In existing technologies, some solutions have been proposed to add monitoring units to fuses. However, most of these solutions are single-function, typically only monitoring current or status signals and issuing alarms through simple threshold comparisons. These methods have the following inherent drawbacks:
[0005] High false alarm rate: Simple threshold judgment cannot effectively distinguish between different types of faults (such as single-phase grounding and phase-to-phase short circuit), nor can it distinguish between transient overload and continuous fault, which easily generates a large number of false alarms and interferes with operation and maintenance work.
[0006] Lack of predictive capabilities: It can only provide post-event alarms for faults that have already occurred, and cannot perform trend analysis on the health status of the line, nor can it perform predictive maintenance. For example, it cannot issue a warning before the fuse breaks due to aging from long-term overload.
[0007] Therefore, there is an urgent need in this field for an intelligent fuse monitoring system that can achieve accurate monitoring and trend early warning. Summary of the Invention
[0008] This invention proposes an intelligent drop-out fuse monitoring system, which can accurately and quickly identify faults and significantly reduce the false alarm rate.
[0009] The technical solution of this invention is implemented as follows:
[0010] A smart drop-out fuse monitoring system includes:
[0011] The data acquisition module is used to collect sensor data installed on the drop-out fuse in real time. The data includes three-phase current, voltage, load current and three-phase imbalance.
[0012] The fault diagnosis module is communicatively connected to the data acquisition module. It is used to receive the sensor data and perform real-time analysis of the data using a pre-trained gradient boosting decision tree model to identify and output the fault type.
[0013] The trend warning module is communicatively connected to the data acquisition module. It is used to receive continuous time-series data and use a pre-trained gated recurrent unit network model to perform trend prediction, so as to output latent fault risk warning information that characterizes the deterioration of line health.
[0014] The data communication module, connected to the fault diagnosis module and the trend early warning module, is used to upload fault identification results and risk warning information to the background management system via wireless Internet of Things.
[0015] Preferably, the fault types include overload, single-phase grounding fault, phase-to-phase short circuit fault, and open circuit fault.
[0016] Preferably, the establishment of the gradient boosting decision tree model includes model initialization and model training, wherein the model initialization establishes a gradient boosting framework with a classification decision tree as a weak learner as the initial model.
[0017] Preferably, the model training step includes:
[0018] Prepare a training sample set, calculate a multi-dimensional electrical feature vector for each sample including the real-time effective values of three-phase current and voltage, three-phase unbalance, total harmonic distortion rate of current, and high-frequency transient energy features of current signal extracted based on wavelet transform, and label the actual fault type.
[0019] The initialized model is iteratively trained using the sample set, and the model parameters are optimized by minimizing the multivariate logistic loss function and using gradient descent.
[0020] Training stops when the model's loss function on the validation set no longer decreases significantly or the classification accuracy meets a predetermined threshold, resulting in the final gradient boosting decision tree model.
[0021] Preferably, the establishment of the gated recurrent unit network model includes model initialization and model training, wherein model initialization includes: constructing an initial model of the recurrent neural network, the structure of which includes: an input layer, at least one GRU layer, a fully connected output layer, and using the Sigmoid activation function.
[0022] Preferably, the establishment of the gated recurrent unit network model further includes a model training step:
[0023] Prepare training samples;
[0024] The initialized model is trained using the training samples, and the model parameters are optimized by minimizing the binary cross-entropy loss function and using backpropagation and time-series backpropagation algorithms.
[0025] Training stops when the model's prediction accuracy on the validation set is stable and the binary cross-entropy loss converges, resulting in the final gated recurrent unit network model.
[0026] The beneficial effects of the present invention using the above technical solution are as follows:
[0027] This invention provides an intelligent drop-out fuse monitoring system. By introducing multi-dimensional electrical characteristics and machine learning models, the system can accurately distinguish various faults, greatly reducing false alarms and missed alarms caused by the traditional threshold comparison method, making operation and maintenance responses more targeted. This invention utilizes deep learning to model time-series data, enabling it to capture the slow deterioration trend of line and equipment status, preventing problems before they occur and improving the reliability of the distribution network. The system can monitor in real time, accurately locate faults, and report quickly, supplemented by intelligent decision support, allowing operation and maintenance personnel to quickly understand the fault situation, locate the fault point, and take the best handling measures, significantly shortening the time for fault diagnosis and power restoration, and reducing operation and maintenance costs. Attached Figure Description
[0028] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.
[0029] Figure 1 This is a flowchart illustrating the working steps of the intelligent drop-out fuse monitoring system of the present invention. Detailed Implementation
[0030] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.
[0031] The intelligent drop-out fuse monitoring system provided by this invention includes:
[0032] The data acquisition module is used to collect sensor data installed on the drop-out fuse in real time. The data includes three-phase current, voltage, load current and three-phase imbalance.
[0033] The fault diagnosis module is communicatively connected to the data acquisition module. It is used to receive the sensor data and perform real-time analysis of the data using a pre-trained gradient boosting decision tree model to identify and output the fault type, which includes overload, single-phase grounding fault, phase-to-phase short circuit fault and open circuit fault.
[0034] The trend warning module is communicatively connected to the data acquisition module. It is used to receive continuous time-series data and use a pre-trained gated recurrent unit network model to perform trend prediction, so as to output latent fault risk warning information that characterizes the deterioration of line health.
[0035] The data communication module, connected to the fault diagnosis module and the trend early warning module, is used to upload fault identification results and risk warning information to the background management system via wireless Internet of Things.
[0036] The establishment of the gradient boosting decision tree model includes model initialization and model training;
[0037] The model initialization steps are as follows:
[0038] A gradient boosting framework with a classification decision tree as a weak learner is established as the initial model.
[0039] Set the structural hyperparameters of the model, including: the number of weak learners ranging from 100 to 200, the learning rate ranging from 0.05 to 0.2, and the maximum depth of the decision tree ranging from 3 to 6;
[0040] The model training steps include:
[0041] Prepare a training sample set, calculate a multi-dimensional electrical feature vector for each sample including the real-time effective values of three-phase current and voltage, three-phase unbalance, total harmonic distortion rate of current, and high-frequency transient energy features of current signal extracted based on wavelet transform, and label the actual fault type.
[0042] The initialized model is iteratively trained using the sample set, and the model parameters are optimized by minimizing the multivariate logistic loss function and using gradient descent.
[0043] Training stops when the model's loss function on the validation set no longer decreases significantly or the classification accuracy meets a predetermined threshold, resulting in the final gradient boosting decision tree model.
[0044] The establishment of the gated recurrent unit network model includes model initialization and model training. Model initialization constructs the initial model of the recurrent neural network. The initial model structure of the recurrent neural network includes: an input layer for receiving input vectors of dimension 2; at least one GRU layer with the number of hidden units between 16 and 64; a fully connected output layer using the sigmoid activation function; and then the weight matrix and bias terms of the model are initialized.
[0045] Model training steps:
[0046] Prepare training samples, each sample being a time series sliding window with a window length L ranging from 60 to 180 time steps. Each time step within the window contains the average load current and three-phase imbalance data. Label positive sample windows with a label 1 indicating a fault will occur within a specific future time period, and label negative sample windows with a label 0 indicating no fault will occur.
[0047] The initialized model is trained using the training samples, and the model parameters are optimized by minimizing the binary cross-entropy loss function and using backpropagation and time-series backpropagation algorithms.
[0048] Training stops when the model's prediction accuracy on the validation set is stable and the binary cross-entropy loss converges, resulting in the final gated recurrent unit network model.
[0049] The gradient boosting decision tree model in the fault diagnosis module is as follows:
[0050] 1. Model Initialization
[0051] This invention employs a gradient boosting decision tree as the classifier. The initial model is an untrained GBDT framework. In a Python environment, it can be initialized using the GradientBoostingClassifier class from the scikit-learn library. Key structural parameters set during initialization include:
[0052] The initial number of weak learners (decision trees) is 100.
[0053] The learning rate is 0.1, which controls the contribution weight of each weak learner.
[0054] The maximum depth of each decision tree is 3, which controls the complexity of the model.
[0055] Use the log loss function for multivariate classification.
[0056] 2. Training Data Preparation
[0057] Data source: Collect a large amount of historical operating data of 10kV lines from distribution network automation systems or deployed monitoring terminals, covering various states such as normal, overload, single-phase grounding, phase-to-phase short circuit and open circuit.
[0058] Feature engineering (compiling input feature vectors): For each frame of raw data (e.g., one sample point per second), a multi-dimensional electrical feature vector is calculated, which serves as the input to the model. Specific features include:
[0059] I a ,I b ,I c Real-time effective values of the three-phase currents A, B, and C;
[0060] Three-phase current imbalance.
[0061] Total harmonic distortion of phase A current;
[0062] Based on the db4 wavelet basis, the A-phase current is decomposed into three layers of wavelet packets, and the sum of squares of the detail coefficients of each node in the third layer is calculated as the transient energy feature.
[0063] Data labeling: Based on fault recording information and protection action signals, domain experts label each feature vector sample with its corresponding real fault type label, such as 0 (normal), 1 (overload), etc.
[0064] Dataset partitioning: The processed sample set is randomly divided into training set, validation set and test set in a ratio of 7:2:1.
[0065] 3. Model Training
[0066] The initialized GBDT model is trained using the prepared training set. The training process involves iteratively optimizing the parameters of 100 decision trees using gradient descent. The specific steps are as follows:
[0067] Input the training set features and labels into the model.
[0068] The model begins to iterate: the first tree fits the initial residuals of the data, the second tree fits the residuals of the first tree, and so on.
[0069] In each iteration, the generation of a new tree is guided by calculating the negative gradient (i.e., the residual) of the loss function.
[0070] During training, the model's performance (such as accuracy and F1 score) can be monitored on the validation set to prevent overfitting.
[0071] Training stops when the iteration is completed (100 rounds) or the performance on the validation set no longer improves significantly, resulting in the final gradient boosting decision tree model with determined parameters.
[0072] 4. Model Application
[0073] The trained model file is deployed in the fault diagnosis module of the monitoring system. During online operation, the front-end terminal collects current and voltage data in real time and calculates the aforementioned multidimensional feature vector. This vector is then input into the trained GBDT model, which outputs the corresponding fault type classification result.
[0074] The specific network model of the gated recurrent unit in the trend early warning module is as follows:
[0075] 1. Model Initialization
[0076] This invention employs a GRU network as the time series prediction model. The initial model is constructed using the PyTorch or TensorFlow framework. Its network structure includes:
[0077] A GRU layer with an input size of 2 (corresponding to [load current, unbalance]) and a hidden layer cell count of 32.
[0078] A fully connected layer with an input dimension of 32 (GRU layer output) and an output dimension of 1.
[0079] The fully connected layer is followed by the Sigmoid activation function, which maps the output to probability values between 0 and 1.
[0080] 2. Training Data Preparation
[0081] Constructing time-series samples: Samples are constructed from historical data using a sliding window approach. The window length is set to L = 120 (i.e., 120 consecutive minutes of data), and the prediction step size is T = 30 (predicting whether a failure will occur within the next 30 minutes).
[0082] Positive sample: The data segment extracted L minutes before the fault occurred, with the label set to 1.
[0083] Negative samples: Randomly selected normal data segments of length L outside of any fault period, with the label set to 0.
[0084] Each sample is a matrix of shape (120,2), where 120 is the time step and 2 is the feature dimension (average load current and three-phase imbalance).
[0085] 3. Model Training
[0086] The initial GRU model was trained using the constructed time-series sample set.
[0087] Loss function: Binary cross-entropy loss function.
[0088] Optimizer: Adam optimizer, with an initial learning rate set to 0.001.
[0089] Training process: Samples are input into the model in batches (batch size = 32). Forward propagation is used to calculate the predicted probabilities and loss, followed by backpropagation (BPTT) to calculate the gradient. The Adam optimizer then updates the weight parameters of the GRU network and fully connected layers. Training continues for multiple epochs until the loss function converges on the validation set.
[0090] 4. Model Application
[0091] The trained GRU model is deployed in the system backend. The trend warning module continuously receives real-time data streams and maintains a current time series window of length 120. Every minute, the data in this window is fed into the GRU model for inference, outputting a fault probability. When the probability value exceeds a preset threshold (e.g., 0.7), the system immediately generates and uploads a risk warning message.
[0092] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. An intelligent drop-out fuse monitoring system, characterized in that, include: The data acquisition module is used to collect sensor data installed on the drop-out fuse in real time. The data includes three-phase current, voltage, load current and three-phase imbalance. The fault diagnosis module is communicatively connected to the data acquisition module. It is used to receive the sensor data and perform real-time analysis of the data using a pre-trained gradient boosting decision tree model to identify and output the fault type. The trend warning module is communicatively connected to the data acquisition module. It is used to receive continuous time-series data and use a pre-trained gated recurrent unit network model to perform trend prediction, so as to output latent fault risk warning information that characterizes the deterioration of line health. The data communication module, connected to the fault diagnosis module and the trend early warning module, is used to upload fault identification results and risk warning information to the background management system via wireless Internet of Things.
2. The intelligent drop-out fuse monitoring system according to claim 1, characterized in that, The fault types include overload, single-phase grounding fault, phase-to-phase short circuit fault, and open circuit fault.
3. The intelligent drop-out fuse monitoring system according to claim 1, characterized in that, The establishment of the gradient boosting decision tree model includes model initialization and model training, wherein the model initialization establishes a gradient boosting framework with a classification decision tree as a weak learner as the initial model.
4. The intelligent drop-out fuse monitoring system according to claim 3, characterized in that, The model training steps include: Prepare a training sample set, calculate a multi-dimensional electrical feature vector for each sample including the real-time effective values of three-phase current and voltage, three-phase unbalance, total harmonic distortion rate of current, and high-frequency transient energy features of current signal extracted based on wavelet transform, and label the actual fault type. The initialized model is iteratively trained using the sample set, and the model parameters are optimized by minimizing the multivariate logistic loss function and using gradient descent. Training stops when the model's loss function on the validation set no longer decreases significantly or the classification accuracy meets a predetermined threshold, resulting in the final gradient boosting decision tree model.
5. The intelligent drop-out fuse monitoring system according to claim 1, characterized in that, The establishment of the gated recurrent unit network model includes model initialization and model training. Model initialization includes: constructing an initial model of the recurrent neural network, the structure of which includes: an input layer, at least one GRU layer, a fully connected output layer, and a Sigmoid activation function.
6. The intelligent drop-out fuse monitoring system according to claim 5, characterized in that, The establishment of the gated recurrent unit network model also includes a model training step: Prepare training samples; The initialized model is trained using the training samples, and the model parameters are optimized by minimizing the binary cross-entropy loss function and using backpropagation and time-series backpropagation algorithms. Training stops when the model's prediction accuracy on the validation set is stable and the binary cross-entropy loss converges, resulting in the final gated recurrent unit network model.