A high-voltage circuit breaker fault diagnosis model optimization method
By constructing a multi-deep neural network model, integrating key signal features of high-voltage circuit breakers, and utilizing a distributed data storage architecture, the problems of insufficient fault identification capability and low data utilization rate of online monitoring models for high-voltage circuit breakers are solved, achieving efficient fault diagnosis and model updates.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CRSC (CHANGSHA) RAILWAY TRAFFIC CONTROL TECH CO LTD
- Filing Date
- 2022-09-27
- Publication Date
- 2026-05-26
AI Technical Summary
Existing online monitoring models for high-voltage circuit breakers rely on single monitoring signal characteristics, have limited fault identification capabilities, and cannot fully utilize actual online monitoring data, resulting in poor performance of the models in practical applications.
A multi-depth neural network model is constructed, which integrates the current signals of the opening and closing coils and the linear displacement signals of the moving iron core. The model is deployed on a server cluster using Spark technology, and incremental extraction and updates are performed through the Hadoop distributed data storage architecture. The model is continuously iterated and optimized based on real-time running data.
This improved the robustness and accuracy of the fault diagnosis model, made reasonable use of computing resources, alleviated data storage pressure, and enabled continuous optimization and updating of the model.
Smart Images
Figure CN115456106B_ABST
Abstract
Description
Technical Field
[0001] This disclosure belongs to the interdisciplinary field of power equipment fault diagnosis, artificial intelligence algorithms, and Internet technology, and specifically relates to an optimization method for a fault diagnosis model of a high-voltage circuit breaker. Background Technology
[0002] High-voltage circuit breakers, as electrical devices in power systems that carry and interrupt normal operating currents and cut off fault currents, play a crucial role in ensuring the stable operation of the power grid. Most circuit breaker faults, such as failure to open or close, and accidental opening or closing, are caused by malfunctions in the operating mechanism. Therefore, research on circuit breaker operating mechanism faults is beneficial for promoting improvements in circuit breaker design structure, achieving intelligent online monitoring of circuit breakers, and developing predictive maintenance strategies.
[0003] Currently, thanks to the development of smart hardware and improved computing power, most online monitoring fault diagnosis models for high-voltage circuit breakers adopt big data-based artificial intelligence models. These models can, to some extent, simulate human expert experience and reasoning processes, effectively identify and establish the mapping relationship between signal characteristics and faults, and realize the automation and intelligence of online monitoring systems. However, most of these artificial intelligence models use single monitoring signal characteristics as input, resulting in limited fault identification capabilities. Furthermore, the accumulation of online sensing data for high-voltage circuit breakers in the power industry is severely insufficient and processing is difficult. Academia and industry mostly build their models based on offline simulation experimental data. However, due to economic and safety considerations, the databases formed by offline fault simulation experiments are relatively simple, making it difficult to identify complex faults in actual circuit breaker operation, resulting in low confidence levels. On the other hand, these artificial intelligence models are usually deployed at the edge and are not dynamically updated, failing to fully utilize the value of actual online monitoring data.
[0004] The aforementioned problems result in poor practical performance of artificial intelligence models in real-world applications, still relying heavily on the experience and judgment of technical personnel. This invention proposes a hot-start incremental neural network-based online monitoring algorithm for high-voltage circuit breakers, effectively addressing these shortcomings. This invention trains a multi-deep neural network using experimental, simulated, and generated circuit breaker opening and closing coil current signals and moving iron core linear displacement signals as the training set, effectively achieving a hot start for the incremental AI model. Furthermore, Spark technology is used to deploy the base classifier of this multi-deep neural network on a cluster of multiple servers, improving computational efficiency. After actual deployment, the Hadoop distributed data storage architecture is used to incrementally extract operational data from all circuit breakers within a certain period as the training set, updating the multi-deep neural network fault diagnosis model, improving the utilization rate of actual online monitoring data, and continuously enhancing the model's practical application performance. Summary of the Invention
[0005] To address the shortcomings of existing technologies, the purpose of this disclosure is to provide a method for optimizing a fault diagnosis model for high-voltage circuit breakers. This method can continuously iterate and optimize the model using real-time operating data of the circuit breaker, thereby improving fault diagnosis efficiency.
[0006] To achieve the above objectives, this disclosure provides the following technical solutions:
[0007] A method for online monitoring of high-voltage circuit breakers includes the following steps:
[0008] S100: Building a data warehouse;
[0009] S200: Use the constructed data warehouse as the training set to train the high-voltage circuit breaker fault diagnosis model to obtain a trained high-voltage circuit breaker fault diagnosis model.
[0010] S300: Continuously collects real-time operating data of the high-voltage circuit breaker under test and expands it to the data warehouse. Based on the real-time operating data, it continuously trains the trained high-voltage circuit breaker fault diagnosis model online to iteratively optimize the model.
[0011] Preferably, in step S200, training the fault diagnosis model for the high-voltage circuit breaker includes the following steps:
[0012] S201: Divide the data warehouse into a training set, a validation set, and a test set;
[0013] S202: Based on the stochastic gradient descent algorithm, the training set is randomly divided into several mini-batch sample sets, which are then input into the model for feedforward training. The training error is calculated based on the cross-entropy loss function.
[0014] S203: Calculate the partial derivative of the cross-entropy loss function with respect to the network connection weights based on the training error;
[0015] S204: Update the parameters of the network connection weights based on the partial derivatives of the network connection weights. After the update is completed, the model training is finished.
[0016] S205: Validate the trained model using the validation set. If the model's prediction accuracy on the validation set reaches 90%, the validation is successful, and the trained model is obtained; otherwise, the model's hyperparameters are adjusted and the model is retrained.
[0017] S206: Test the trained model using the test set. If the model's prediction accuracy on the test set reaches 90%, the test passes; otherwise, adjust the model's hyperparameters and retrain.
[0018] Preferably, in step S100, the data warehouse includes offline test data and simulation data of high-voltage circuit breakers, as well as generated signal data obtained based on the offline test data and simulation data.
[0019] Preferably, in step S202, the cross-entropy loss function is expressed as follows:
[0020]
[0021] Where x represents the sample feature, y represents the sample true label, w represents the weight, N represents the number of samples, M represents the number of classes, and d ic Let p represent the sign function, where i represents a sample, c represents a class, and the value is 1 if the true class of i is c, and 0 otherwise. ic This represents the predicted probability that sample i belongs to category c.
[0022] Preferably, in step S203, the partial derivative of the cross-entropy loss function with respect to the network connection weights is calculated using the following formula:
[0023]
[0024] Where, x i y represents the feature vector of the sample. i This represents the true label of the sample. It is expressed as the partial derivative of the cross-entropy loss function, where N represents the number of samples.
[0025] Preferably, in step S204, the network connection weights are updated using the following formula:
[0026]
[0027] Where ε represents the learning rate.
[0028] This disclosure also provides a high-voltage circuit breaker fault diagnosis model optimization device, including:
[0029] Build modules are used to build data warehouses;
[0030] The training module is used to train the high-voltage circuit breaker fault diagnosis model using the constructed data warehouse as the training set, so as to obtain the trained high-voltage circuit breaker fault diagnosis model.
[0031] The data acquisition module is used to continuously collect real-time operating data of the high-voltage circuit breaker under test;
[0032] The iterative optimization module is used to continuously train the trained high-voltage circuit breaker fault diagnosis model online based on real-time running data, so as to iteratively optimize the model.
[0033] This disclosure also provides a high-voltage circuit breaker fault diagnosis model optimization device, including:
[0034] Build modules are used to build data warehouses;
[0035] The training module is used to train the high-voltage circuit breaker fault diagnosis model using the constructed data warehouse as the training set, so as to obtain the trained high-voltage circuit breaker fault diagnosis model.
[0036] The data acquisition module is used to continuously collect real-time operating data of the high-voltage circuit breaker under test;
[0037] The iterative optimization module is used to train the pre-trained high-voltage circuit breaker fault diagnosis model online based on real-time operating data, so as to iteratively optimize the model.
[0038] Compared with existing technologies, the beneficial effects of this disclosure are as follows: This invention integrates the opening and closing coil current signals and the linear displacement signals of the moving iron core of a high-voltage circuit breaker as typical state-sensing information of the circuit breaker, and constructs a data warehouse based on experimental data, simulation data, and generated data. The multi-deep neural network model trained based on this data warehouse is robust and can be deployed as a good hot-start model on a regional server cluster. The base classifiers of the multi-deep neural network model are deployed on different node servers, and after deployment, online monitoring data of all circuit breakers in the region within a certain period of time are integrated for training and updating in a distributed incremental extraction manner. This invention not only improves the utilization rate of actual online monitoring data and enables the fault diagnosis model to be optimized and updated over time, but also makes reasonable use of computing resources and alleviates data storage pressure. Attached Figure Description
[0039] Figure 1 This is a flowchart of a high-voltage circuit breaker fault diagnosis model optimization method provided in one embodiment of this disclosure;
[0040] Figure 2 This is a schematic diagram of offline test data for a circuit breaker provided in another embodiment of this disclosure;
[0041] Figure 3 This is a schematic diagram of circuit breaker simulation data provided in another embodiment of this disclosure;
[0042] Figure 4 This is a schematic diagram illustrating the principle of an adversarial generative network provided in another embodiment of this disclosure;
[0043] Figure 5 This is a schematic diagram of the structure of a multi-deep neural network provided in another embodiment of this disclosure;
[0044] Figure 6This is a schematic diagram illustrating the model training results obtained using a data warehouse and simply using experimental data, provided in another embodiment of this disclosure;
[0045] Figure 7 This is a schematic diagram of the model test error curve provided in another embodiment of this disclosure;
[0046] Figure 8 This is a schematic diagram of a neural network. Detailed Implementation
[0047] The following will refer to the appendix. Figures 1 to 8 Specific embodiments of this disclosure are described in detail. While specific embodiments of this disclosure are shown in the accompanying drawings, it should be understood that this disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of this disclosure to those skilled in the art.
[0048] It should be noted that certain terms are used in the specification and claims to refer to specific components. Those skilled in the art will understand that different terms may be used to refer to the same component. This specification and claims do not distinguish components based on differences in terminology, but rather on differences in function. The terms "comprising" or "including" used throughout the specification and claims are open-ended and should be interpreted as "comprising but not limited to." The following descriptions are preferred embodiments for carrying out this disclosure; however, these descriptions are for the purpose of understanding the general principles of the specification and are not intended to limit the scope of this disclosure. The scope of protection of this disclosure is determined by the appended claims.
[0049] To facilitate understanding of the embodiments of this disclosure, further explanations and descriptions will be provided below with reference to the accompanying drawings and specific embodiments. The accompanying drawings do not constitute a limitation on the embodiments of this disclosure.
[0050] In one embodiment, such as Figure 1 As shown, this disclosure proposes an optimization method for a fault diagnosis model of a high-voltage circuit breaker, including the following steps:
[0051] S100: Building a data warehouse;
[0052] In this step, the data warehouse includes offline test data and simulation data of high-voltage circuit breakers, as well as generated signal data obtained based on the offline test data and simulation data, such as... Figure 2As shown, the offline test data is obtained from simulation tests of various common electromechanical faults of circuit breakers conducted on the circuit breaker itself. Specifically, based on a statistically analyzed list of common circuit breaker fault types, a circuit breaker fault simulation platform is built in the laboratory to simulate faults of different degrees and types, obtaining test waveforms of the circuit breaker's opening and closing coil current signals and the linear displacement travel signal of the moving iron core. The simulated electromechanical fault types include: poor contact in the coil power supply circuit, abnormal driving voltage of the operating mechanism, abnormal air gap in the moving iron core of the operating mechanism, core movement jamming, weak opening spring, and sticking of the closing contactor.
[0053] like Figure 3 As shown, the simulation data is based on the physical process of high-voltage circuit breaker opening. A multi-domain physical simulation model of the permanent magnet operating mechanism was built using the Simulink function of MATLAB software. This simulation model consists of a parameter input subsystem, a control loop subsystem, a measurement subsystem, a coil circuit electrical subsystem, and a core movement mechanical subsystem connected via ports. Subsequently, this simulation model was imported into the pre-developed "High-Voltage Circuit Breaker Fault Simulation Platform" software for batch simulation of combined faults in the circuit breaker operating mechanism. This fault simulation software employs parallel technology, enabling it to obtain a large amount of simulation data under combined circuit breaker fault conditions in a short time.
[0054] Based on existing offline experimental and simulation data, a generative adversarial network (GAN) is trained to generate realistic signal data with random noise. GANs are a type of generative model, such as... Figure 4As shown, it consists of a generator and a discriminator. The generator is responsible for capturing the latent distribution of real data samples and generating new data samples. The discriminator is a binary classifier that distinguishes whether the input is real data or a generated sample. The training objective of GAN is to make the generator learn the real sample distribution as closely as possible, i.e., minimize the generator's loss; on the other hand, it is necessary to make the discriminator unable to effectively distinguish whether the input data comes from real data or the generator, i.e., maximize the discriminator's loss. Once the GAN achieves its training objective, it can obtain simulated fault data by inputting random vectors and fault labels into the generator. GAN can generate a large number of datasets, which can be combined with offline experimental data and simulation data to train high-voltage circuit breaker fault diagnosis models. This is beneficial to enhancing the model's generalization ability because the existing artificial intelligence model training process has a large difference in the mapping relationship between data features and faults collected in offline experiments and in actual industrial scenarios, resulting in poor performance of models trained only on offline experimental data in practical applications. Simulation models are computer simulations of physical models, and their signal outputs differ from reality. GANs, on the other hand, generate feature data with the same distribution as the original dataset, incorporating random errors. Both types of errors are within a reasonable range and have certain physical and mathematical meanings. Training neural networks with features containing these errors can improve overfitting, enhance noise resistance, and thus strengthen generalization ability. When the input features of a model differ from the feature distribution of the training set, models with weak generalization ability will experience a significant drop in recognition accuracy, while models with strong generalization ability can maintain high recognition accuracy.
[0055] S200: Use the constructed data warehouse as the training set to train the high-voltage circuit breaker fault diagnosis model to obtain a trained high-voltage circuit breaker fault diagnosis model.
[0056] In this step, the high-voltage circuit breaker fault diagnosis model employs a multi-depth neural network, such as... Figure 5 As shown, a multi-deep neural network includes multiple base classifiers with the same structure, each of which includes an input layer, a hidden layer, and an output layer.
[0057] S300: Continuously collects real-time operating data of the high-voltage circuit breaker under test and expands it to the data warehouse. Based on the real-time operating data, it continuously trains the trained high-voltage circuit breaker fault diagnosis model online to iteratively optimize the model.
[0058] In this step, after obtaining the trained high-voltage circuit breaker fault diagnosis model, the base classifier of the trained model is deployed on the node servers of the regional server cluster using the Spark computing architecture. Finally, in practical applications, the recognition results of the base classifier are weighted and averaged using the Reduce module of the Spark architecture to obtain the final fault diagnosis result.
[0059] When deploying the model, the high-voltage circuit breaker fault diagnosis model is deployed on a regional server cluster. The pre-training database is distributed across various nodes of the storage server cluster based on a Hadoop distributed data storage rack. Whenever a circuit breaker in the region trips, the circuit breaker's controller transmits real-time online monitoring data to the model server cluster and outputs the fault diagnosis results. The tripping data is stored on different storage server nodes depending on the circuit breaker's location. Periodically, the model server cluster extracts newly added or modified data from the database tables since the last extraction from the data storage server cluster. This data is used to train and update the neural network model, thereby achieving iterative optimization of the model.
[0060] In summary, by continuously training the model online, the utilization rate of real-time operating data of circuit breakers is improved, enabling the fault diagnosis model to be continuously optimized and updated over time. Moreover, since online training uses an incremental approach to extract training data from real-time operating data, this method avoids using all sample data in the database for each training session, thus improving training speed. Furthermore, when data storage space is limited, data that has already participated in incremental extraction can be deleted, thereby making reasonable use of computing resources and alleviating the pressure on data storage.
[0061] In another embodiment, step S200, training the high-voltage circuit breaker fault diagnosis model includes the following steps:
[0062] S201: Divide the data warehouse into a training set, a validation set, and a test set;
[0063] S202: Based on the stochastic gradient descent algorithm, the training set is randomly divided into several mini-batch sample sets, which are then input into the model for feedforward training. The training error is calculated based on the cross-entropy loss function.
[0064] In this step, the cross-entropy loss function is expressed as follows:
[0065]
[0066] Where x represents the sample feature, y represents the sample true label, w represents the weight, N represents the number of samples, M represents the number of classes, and d ic Let p represent the sign function, where i represents a sample, c represents a class, and the value is 1 if the true class of i is c, and 0 otherwise. ic This represents the predicted probability that sample i belongs to category c.
[0067] S203: Calculate the partial derivative of the cross-entropy loss function with respect to the network connection weights based on the training error;
[0068] In this step, the partial derivative of the cross-entropy loss function with respect to the network connection weights is calculated using the following formula:
[0069]
[0070] Where, x i y represents the feature vector of the sample. i This represents the true label of the sample. It is expressed as the partial derivative of the cross-entropy loss function, where N represents the number of samples.
[0071] S204: Update the parameters of the network connection weights based on the partial derivatives of the network connection weights. After the update is completed, the model training is finished.
[0072] In this step, the network connection weights are updated using the following formula:
[0073]
[0074] Where ε represents the learning rate.
[0075] In this step, the weights of network connectors are explained as follows: Figure 8 As shown, the network connections between the layers of the neural network actually correspond to the output of the j-th neuron in the i-th layer. Input to the m-th neuron in the (i+1)-th layer Calculation path: Weight It can also be called network connection weight.
[0076] S205: Validate the trained model using a validation set. If the model's prediction accuracy on the validation set reaches 90%, the validation is successful, and the trained model is obtained. Otherwise, adjust the model's hyperparameters, such as the number of base classifiers, the number of neural network layers, the number of hidden layer neurons, etc.
[0077] S206: Test the trained model using a test set. If the model's prediction accuracy on the test set reaches 90%, the test is passed; otherwise, adjust the model's hyperparameters, such as the number of base classifiers, the number of neural network layers, the number of hidden layer neurons, etc.
[0078] The technical effects obtained by the above method will now be described in conjunction with specific embodiments.
[0079] Taking the circuit breaker closing situation as an example, the feature vector of the dataset consists of eight time-domain features of the circuit breaker operating coil current signal and six time-domain features of the angular displacement signal. There are seven categories of labels, including the normal operation state of the circuit breaker, two single electrical fault states and four single mechanical fault states. This dataset is obtained entirely from circuit breaker fault simulation experiments in the laboratory. In addition, sample data with the same distribution is generated using a single Gaussian distribution model based on the training data. Consider the following two cases: (1) using the experimental data as the training set to train the model; (2) using the training set composed of the experimental data and the generated data to train the model. The models all adopt the same multi-deep neural network structure. The results of 50 training and testing are as follows: Figure 6 As shown, by Figure 6 It is evident that models trained using training sets containing generated data exhibit higher accuracy and greater stability.
[0080] Furthermore, Figure 7 This is a graph showing the test error of the iteratively optimized model. Figure 7 It can be seen that the test error decreases as the number of iterations increases, indicating that by continuously collecting real-time operating data of the circuit breaker and iteratively optimizing the model, the model can become more and more accurate in diagnosing circuit breaker faults.
[0081] In another embodiment, this disclosure also provides a high-voltage circuit breaker fault diagnosis model optimization device, comprising:
[0082] Build modules are used to build data warehouses;
[0083] The training module is used to train the high-voltage circuit breaker fault diagnosis model using the constructed data warehouse as the training set, so as to obtain the trained high-voltage circuit breaker fault diagnosis model.
[0084] The data acquisition module is used to continuously collect real-time operating data of the high-voltage circuit breaker under test;
[0085] The iterative optimization module is used to continuously train the trained high-voltage circuit breaker fault diagnosis model online based on real-time running data, so as to iteratively optimize the model.
[0086] The foregoing general description of the invention and its specific embodiments should not be construed as a limitation on the technical solution of the invention. Those skilled in the art, based on the disclosure of this application, can add, reduce, or combine the disclosed technical features in the foregoing general description and / or specific embodiments (including examples) without departing from the constituent elements of the invention, to form other technical solutions within the scope of protection of this application.
Claims
1. A method for optimizing a fault diagnosis model for high-voltage circuit breakers, comprising the following steps: S100: Building a data warehouse; S200: Use the constructed data warehouse as the training set to train the high-voltage circuit breaker fault diagnosis model to obtain a trained high-voltage circuit breaker fault diagnosis model. S300: Continuously collects real-time operating data of the high-voltage circuit breaker under test and expands it to the data warehouse. Based on the real-time operating data, it continuously trains the trained high-voltage circuit breaker fault diagnosis model online to iteratively optimize the model. In step S100, the data warehouse includes offline test data and simulation data of high-voltage circuit breakers, as well as generated signal data obtained based on the offline test data and simulation data; The method integrates the opening and closing coil current signals and the linear travel displacement signals of the moving iron core of the high-voltage circuit breaker as typical state-sensing information of the circuit breaker, and constructs a data warehouse based on experimental data, simulation data, and generated data. The multi-deep neural network model trained on this data warehouse is robust and can be deployed as a good hot-start model on the regional server cluster. The base classifiers of the multi-deep neural network model are deployed on different node servers. After deployment, the online monitoring data of all circuit breakers in the region within a certain period of time are integrated for training and updating in a distributed incremental extraction manner. in, In step S200, training the fault diagnosis model for the high-voltage circuit breaker includes the following steps: S201: Divide the data warehouse into a training set, a validation set, and a test set; S202: Based on the stochastic gradient descent algorithm, the training set is randomly divided into several mini-batch sample sets, which are then input into the model for feedforward training. The training error is calculated based on the cross-entropy loss function. S203: Calculate the partial derivative of the cross-entropy loss function with respect to the network connection weights based on the training error; S204: Update the parameters of the network connection weights based on the partial derivatives of the network connection weights. After the update is completed, the model training is finished. S205: Validate the trained model using the validation set. If the model's prediction accuracy on the validation set reaches 90%, the validation is successful, and the trained model is obtained; otherwise, the model's hyperparameters are adjusted and the model is retrained. S206: Test the trained model using the test set. If the model's prediction accuracy on the test set reaches 90%, the test passes; otherwise, adjust the model's hyperparameters and retrain. In step S202, the cross-entropy loss function is expressed as follows: ， Where x represents the sample feature, y represents the sample true label, w represents the weight, N represents the number of samples, and M represents the number of classes. Represents a symbolic function. Indicates a sample, Indicates the category, and if The true category is If the result is positive, then take 1; otherwise, take 0. Indicates sample Category The predicted probability; In step S203, the partial derivative of the cross-entropy loss function with respect to the network connection weights is calculated using the following formula: ， Where, x i y represents the feature vector of the sample. i This represents the true label of the sample. It is expressed as the partial derivative of the cross-entropy loss function, where N represents the number of samples; The offline test data is obtained from simulation tests of various common electromechanical faults of circuit breakers conducted on the circuit breaker itself. Specifically, based on the statistically analyzed common fault types of circuit breakers, a circuit breaker fault simulation platform is built in the laboratory to simulate faults of different degrees and types, and to obtain the test waveforms of the circuit breaker opening and closing coil current signals and the linear displacement travel signals of the moving iron core. The simulated electromechanical fault types include: poor contact in the coil power supply circuit, abnormal driving voltage of the operating mechanism, abnormal air gap of the moving iron core of the operating mechanism, iron core movement jamming, weak opening spring, and sticking of the closing contactor. The simulation data is based on the physical process of high-voltage circuit breaker opening. A multi-domain physical simulation model of the permanent magnet operating mechanism was built using the Simulink function of MATLAB software. Based on existing offline experimental and simulation data, a generative adversarial network is trained to generate realistic signal data with random noise.
2. The method according to claim 1, wherein, In step S204, the network connection weights are updated using the following formula: ， in, This represents the learning rate.
3. An apparatus for performing the high-voltage circuit breaker fault diagnosis model optimization method according to any one of claims 1 to 2, comprising: Build modules are used to build data warehouses; The training module is used to train the high-voltage circuit breaker fault diagnosis model using the constructed data warehouse as the training set, so as to obtain the trained high-voltage circuit breaker fault diagnosis model. The data acquisition module is used to continuously collect real-time operating data of the high-voltage circuit breaker under test; The iterative optimization module is used to continuously train the trained high-voltage circuit breaker fault diagnosis model online based on real-time running data, so as to iteratively optimize the model.