Substation automation equipment fault detection method, equipment, system and medium

Abstract

The invention provides a substation automation equipment fault detection method, equipment, a system and a medium. The method comprises the following steps: firstly, acquiring primary equipment data and secondary equipment data of a substation; then correlation analysis is carried out to obtain dominant correlation and implicit correlation; obtaining a first fault portrait of the primary equipment and a second fault portrait of the secondary equipment according to the primary equipment data, the dominant association and the implicit association; obtaining a third fault portrait of the primary equipment and a fourth fault portrait of the secondary equipment according to the secondary equipment data, the dominant association and the implicit association; determining a fault detection result of the primary equipment according to the first fault portrait and the third fault portrait; and determining a fault detection result of the secondary equipment according to the second fault portrait and the fourth fault portrait. According to the method, the association between the primary equipment data and the secondary equipment data is comprehensively mined, the equipment fault is accurately described from multiple perspectives, and the defect that a single model is poor in adaptability is overcome.

Description

Technical Field

[0001] This invention belongs to the field of substation technology, and in particular relates to a method, equipment, system and medium for fault detection of substation automation equipment. Background Art

[0002] With the continuous expansion of the power grid and the acceleration of its intelligent transformation, the number of substation equipment has surged and its structure has become increasingly complex, posing a significant challenge to ensuring its reliable operation. Primary equipment, as the core hardware for power transmission and distribution, includes key components such as transformers and circuit breakers; secondary equipment bears the heavy responsibility of monitoring, controlling, and protecting primary equipment. The coordinated operation of the two is crucial to the stability of the entire substation and even the power grid.

[0003] Existing technologies often analyze primary and secondary equipment data in isolation, failing to consider the specific connections and deep relationships between the data, resulting in low efficiency in detecting faults in automated equipment. Summary of the Invention

[0004] In view of this, the present invention provides a method, device, system and medium for fault detection of substation automation equipment, aiming to solve the problem of low fault detection efficiency of various equipment in substations in the prior art.

[0005] A first aspect of this invention provides a method for fault detection in substation automation equipment, characterized in that it includes:

[0006] Obtain primary and secondary equipment data from the substation;

[0007] A correlation analysis was performed on primary equipment data and secondary equipment data to obtain explicit and implicit correlations;

[0008] Based on primary equipment data, explicit and implicit relationships, the first fault profile of the primary equipment and the second fault profile of the secondary equipment are obtained.

[0009] Based on secondary equipment data, explicit and implicit correlations, the third fault profile of the primary equipment and the fourth fault profile of the secondary equipment are obtained.

[0010] Based on the first and third fault profiles, determine the fault detection results of the primary equipment; based on the second and fourth fault profiles, determine the fault detection results of the secondary equipment.

[0011] In one possible implementation, correlation analysis is performed on primary equipment data and secondary equipment data to obtain explicit and implicit correlations, including:

[0012] Based on the equipment knowledge graph, action events and alarm information in the primary equipment data, the primary equipment data and secondary equipment data are correlated to obtain explicit correlations; among them, the equipment knowledge graph is established based on the electrical connection topology diagram of the substation;

[0013] The pre-trained support vector machine model is adjusted based on the explicit associations, and the primary and secondary device data are input into the adjusted support vector machine model to obtain the implicit associations.

[0014] In one possible implementation, the pre-trained support vector machine model is adjusted based on explicit associations, including:

[0015] Based on explicit associations, determine gradient adjustment information;

[0016] Based on the gradient adjustment information, the gradient direction is updated, and the support vector machine model is iterated according to the updated gradient direction and a preset step size to update the model weights.

[0017] In one possible implementation, primary and secondary equipment data are input into an adjusted support vector machine model to obtain implicit associations, including:

[0018] Based on the action events and alarm information of primary equipment, the primary equipment data and secondary equipment data are divided into multiple sets of input data;

[0019] Multiple sets of input data are fed into the support vector machine model to obtain initial association features;

[0020] The initial association features are integrated to obtain implicit associations.

[0021] In one possible implementation, based on primary equipment data, explicit associations, and implicit associations, a first fault profile of the primary equipment and a second fault profile of the secondary equipment are obtained, including:

[0022] The primary equipment data, explicit and implicit correlations are input into the first deep learning model to obtain the first fault profile of the primary equipment and the second fault profile of the secondary equipment.

[0023] In one possible implementation, based on secondary equipment data, explicit associations, and implicit associations, a third fault profile of the primary equipment and a fourth fault profile of the secondary equipment are obtained, including:

[0024] The secondary equipment data, explicit and implicit associations are input into the second deep learning model to obtain the third fault profile of the primary equipment and the fourth fault profile of the secondary equipment.

[0025] In one possible implementation, the fault detection result of the primary equipment is determined based on the first fault profile and the third fault profile; the fault detection result of the secondary equipment is determined based on the second fault profile and the fourth fault profile, including:

[0026] The first and third fault profiles are aligned by feature, and the second and fourth fault profiles are aligned by feature.

[0027] The aligned first and third fault profiles are weighted and fused to obtain the fault detection results of the primary equipment.

[0028] The aligned second and fourth fault profiles are weighted and fused to obtain the fault detection results of the secondary equipment.

[0029] A second aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the substation automation equipment fault detection method of the first aspect above.

[0030] A third aspect of the present invention provides a substation automation equipment fault detection system, including a detection unit, a data acquisition terminal, and electronic equipment as described in the second aspect.

[0031] A fourth aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the substation automation equipment fault detection method of the first aspect above.

[0032] The substation automation equipment fault detection method, equipment, system, and medium provided in this invention first acquire primary and secondary equipment data of the substation; then, perform correlation analysis on the primary and secondary equipment data to obtain explicit and implicit correlations; based on the primary equipment data, explicit and implicit correlations, obtain a first fault profile of the primary equipment and a second fault profile of the secondary equipment; based on the secondary equipment data, explicit and implicit correlations, obtain a third fault profile of the primary equipment and a fourth fault profile of the secondary equipment; determine the fault detection result of the primary equipment based on the first and third fault profiles; and determine the fault detection result of the secondary equipment based on the second and fourth fault profiles. This invention comprehensively mines the correlation between primary and secondary equipment data, accurately depicting equipment faults from multiple perspectives, overcoming the shortcomings of poor adaptability of single models. When determining the fault detection result, corresponding profiles are specifically integrated to improve fault diagnosis efficiency and accuracy, and to better suit the complex and ever-changing actual operating conditions of substations. Attached Figure Description

[0033] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0034] Figure 1 This is a flowchart illustrating the implementation of the substation automation equipment fault detection method provided in this embodiment of the invention. DETAILED DESCRIPTION

[0035] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of the invention. However, those skilled in the art will understand that the invention can be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of the invention with unnecessary detail.

[0036] Substation automation systems utilize advanced computer technology, modern electronic technology, communication technology, and information processing technology to reorganize and optimize the functions of secondary equipment within a substation. Their main functions include data acquisition, operation monitoring and control, relay protection, local backup control and emergency control, and communication with a remote control center. Through integrated and modular design, substation automation systems achieve the monitoring, measurement, control, and coordination of substation equipment, thereby improving operational efficiency and management levels.

[0037] Common types of substation faults include station control layer faults, hardware faults, software faults, bay layer faults, secondary circuit faults, network equipment faults, transmission channel problems, and remote control system faults.

[0038] Figure 1 This is a flowchart illustrating the implementation of the substation automation equipment fault detection method provided in this embodiment of the invention. Figure 1 As shown, in some embodiments, a substation automation equipment fault detection method includes:

[0039] S110, acquire primary and secondary equipment data of the substation;

[0040] S120, Perform correlation analysis on primary equipment data and secondary equipment data to obtain explicit and implicit correlations;

[0041] S130, based on primary equipment data, explicit associations and implicit associations, obtain the first fault profile of the primary equipment and the second fault profile of the secondary equipment;

[0042] S140. Based on secondary equipment data, explicit associations, and implicit associations, the third fault profile of the primary equipment and the fourth fault profile of the secondary equipment are obtained.

[0043] S150, based on the first fault profile and the third fault profile, determine the fault detection result of the primary equipment; based on the second fault profile and the fourth fault profile, determine the fault detection result of the secondary equipment.

[0044] Primary equipment, as the core hardware for power transmission, needs to collect data such as transformer oil temperature, voltage, and current; circuit breaker opening and closing status and operating duration; and bus voltage amplitude. Sources include sensors installed throughout the equipment, such as voltage transformers and current transformers. Secondary equipment requires acquiring operation records and setting parameters from relay protection devices, monitoring data and communication status from measurement and control devices, and alarm information from automation systems. This primary data forms the foundation for subsequent in-depth analysis, providing raw material for understanding the health status of the equipment. Different types and dimensions of data collectively outline the initial contours of substation equipment operation.

[0045] After obtaining the basic data, correlation analysis connects the scattered data to uncover deeper relationships. Explicit relationships are based on the physical connections and established functional logic of equipment, and are relatively intuitive. They can be identified by checking the electrical connection topology diagram and equipment manuals. For example, the tripping of a circuit breaker on a certain line will inevitably trigger the corresponding protection device. This can be determined by matching real-time monitoring data with such immediate and explicit correspondences. Implicit relationships, on the other hand, are deeply hidden and require the use of big data analysis and machine learning algorithms to uncover. Long-term collected historical data is crucial. Using correlation analysis and clustering algorithms, non-intuitive potential connections can be found, such as long-term low-frequency vibration of primary equipment and a subtle increase in the communication error rate of secondary equipment. This allows maintenance personnel to discover fault clues hidden in daily operation and broaden their fault prediction capabilities.

[0046] With the help of deep learning, the above information is integrated to construct a fault profile. A suitable deep learning model, such as a convolutional neural network or a long short-term memory network, is selected to transform the primary equipment data and the clearly defined explicit and implicit relationships into a format that the model can process. The model is trained on massive amounts of historical data, learning the differences in data patterns under normal and fault conditions, and then outputting the first fault profile of the primary equipment, presenting possible fault points and fault severity as feature vectors. Similarly, a second fault profile is given for the secondary equipment, covering potential fault trends in protection and control devices, providing a key reference for subsequent accurate diagnosis and extracting core fault features from complex data.

[0047] Using secondary equipment data as a starting point, and combining explicit and implicit correlations, the data is then input into a deep learning model. This model's architecture and parameters can be flexibly adjusted compared to the previous model, with the aim of uncovering fault information from new perspectives. The output of the third fault profile for primary equipment may highlight previously overlooked fault characteristics such as mechanical wear and environmental impacts; the fourth fault profile for secondary equipment can also capture unique information such as software logic anomalies and potential communication link vulnerabilities. These two profiles enrich the overall depiction of equipment faults from different dimensions.

[0048] For primary equipment, the first and third fault profiles are integrated, and judgments are made based on feature importance weighting and specific rules. If the abnormal characteristics of key electrical parameters are significant in both profiles, combined with related information, it can be determined that the primary equipment has a corresponding fault type, and a conclusion can be given on whether a fault exists and its level. The same principle applies to secondary equipment. The second and fourth fault profiles are integrated, taking into account alarm records and communication status characteristics, to identify fault conditions such as protection malfunctions and decreased measurement and control accuracy. Reliable detection results are output, providing maintenance personnel with highly targeted maintenance guidelines to ensure the efficient and safe operation of the substation.

[0049] This invention acquires data from primary and secondary equipment, performs correlation analysis, and constructs fault profiles, enabling it to identify faults in station control layer, bay layer equipment, and secondary circuits to a certain extent. For example, if station control layer equipment exhibits abnormal operation or alarms, correlation analysis can be used to find related equipment data, thereby confirming the fault. Furthermore, this method can detect abnormal equipment operation and alarm information caused by software faults. By analyzing the correlation of equipment data, the impact of software faults on the equipment can be determined. For network equipment faults, transmission channel problems, and telecontrol system faults, when the fault affects the normal acquisition and transmission of data, correlation analysis can be used to find fault clues.

[0050] In some embodiments, correlation analysis is performed on primary equipment data and secondary equipment data to obtain explicit and implicit correlations, including:

[0051] Based on the equipment knowledge graph, action events and alarm information in the primary equipment data, the primary equipment data and secondary equipment data are correlated to obtain explicit correlations; among them, the equipment knowledge graph is established based on the electrical connection topology diagram of the substation;

[0052] The pre-trained support vector machine model is adjusted based on the explicit associations, and the primary and secondary device data are input into the adjusted support vector machine model to obtain the implicit associations.

[0053] The stable operation and maintenance of substations relies on a precise understanding of the relationships between equipment, and the discovery of explicit relationships is a crucial first step. The first step is to construct an equipment knowledge graph, created based on the substation's electrical connection topology. This topology details the physical connections of primary equipment such as transformers, circuit breakers, and busbars. Based on this, various types of equipment are abstracted as nodes in the knowledge graph, and the connecting lines between them are transformed into edges between nodes. Furthermore, key attributes of the equipment themselves are added, such as the rated parameters of primary equipment and the protection settings and operating logic of secondary equipment, thus forming a comprehensive knowledge network reflecting the internal connections between the equipment.

[0054] In daily operation, primary equipment generates various action events and alarm messages, which become important clues for uncovering explicit correlations. Once an action event is detected in the primary equipment data, such as the opening and closing action of a circuit breaker, the equipment knowledge graph can be used to quickly locate secondary equipment with connected and functional relationships, such as corresponding relay protection devices and measurement and control devices. By comparing the alarm messages and action records actually generated by these secondary devices within the same time window, if the two match, a protection action report consistent with the circuit breaker action is issued to the protection device, thus confirming an explicit correlation chain based on this event. Similarly, for alarm messages from primary equipment, such as an overheating alarm of a certain device, the knowledge graph is used to trace the chain reaction of related secondary equipment, determining a clear and intuitive correlation between the anomaly of the primary equipment and the response of the secondary equipment, and finally summarizing and integrating a comprehensive set of explicit correlations.

[0055] Once explicit associations are identified, they can be used to optimize the support vector machine model to explore implicit associations. First, feature engineering is performed on the model, extracting key features from the identified explicit associations and integrating them into the original data feature set. For example, a binary feature indicating the linkage between specific primary and secondary equipment can be added. Simultaneously, the weights of corresponding features are adjusted according to the importance of the explicit associations to highlight the key connections. Next, the kernel function and parameters are adapted, considering the changes in data distribution caused by explicit associations. A non-linear kernel function is switched as needed, and the kernel function parameters are fine-tuned using cross-validation on the validation set to ensure the model's adaptability to new data distributions. After these adjustments, the model is retrained using an updated dataset incorporating the features of explicit associations, allowing the model to learn the hidden patterns behind the explicit associations.

[0056] After training and adjusting the support vector machine model, the cleaned and integrated primary and secondary equipment data are input into it. The model processes the data based on its learning results and outputs prediction results. Special attention is paid to misclassified samples or anomalous samples whose predicted values ​​deviate from the normal range, as these often indicate undiscovered latent correlations. A thorough analysis of the equipment data characteristics corresponding to these anomalous samples is conducted to investigate subtle interrelationships between primary and secondary equipment features. For example, the synergistic relationship between long-unnoticed minor vibration changes in primary equipment and fluctuations in the communication status of secondary equipment is explored. Through multiple rounds of data verification and validation using expert knowledge and physical principles, the latent correlations are finally identified, providing a deeper understanding for substation equipment fault prediction and operation and maintenance management.

[0057] In addition to existing primary and secondary equipment data, more multi-dimensional data is introduced, such as equipment operating environment data (temperature, humidity, air pressure, etc.) and historical equipment failure data. This data is then integrated and analyzed with the existing equipment data to construct a more comprehensive knowledge graph, enabling more accurate identification of the root causes of complex failures. For example, some equipment may be more prone to failure in high-temperature environments; combining environmental data can more accurately determine the cause of the failure. After introducing new data, it needs to be cleaned to remove noise, duplication, and errors. Then, data of different formats and types is converted to a unified format for easier subsequent processing and storage. Specifically, entities related to substation equipment can be identified from the introduced data. For example, temperature and humidity from meteorological data can be considered environmental entities, and maintenance time and personnel from maintenance records can be considered maintenance entities. Relationships between entities are analyzed; for example, there may be an "influence" relationship between temperature and transformer operating status, and a "maintenance" relationship between maintenance personnel and equipment. Finally, the newly identified entities are aligned with entities in the existing knowledge graph to avoid duplication and conflicts. The newly identified relationships are added to the existing knowledge graph to expand the relationship network of the knowledge graph.

[0058] In some embodiments, adjusting a pre-trained support vector machine model based on explicit associations includes:

[0059] Based on explicit associations, determine gradient adjustment information;

[0060] Based on the gradient adjustment information, the gradient direction is updated, and the support vector machine model is iterated according to the updated gradient direction and a preset step size to update the model weights.

[0061] Having completed the initial training of the Support Vector Machine (SVM) model, explicit associations provide crucial clues for its optimization. First, explicit associations need to be quantified and features extracted, transforming the intuitive linkages between devices into computationally achievable feature forms. For example, if there is an explicit association of "circuit breaker A trips → relay protection device B operates," a binary feature can be created, marked as 1 in the corresponding sample and 0 otherwise. These features derived from explicit associations are then integrated into the overall data feature set to construct a new training sample set. Subsequently, the loss function is recalculated based on the new sample set. Based on the new loss function, partial derivatives of the weight vector and bias term are calculated to obtain gradient information.

[0062] After obtaining the gradient adjustment information, the gradient direction must first be updated. This is because the data distribution changes with the incorporation of explicit correlation features, and the initial gradient direction no longer precisely matches the new optimization requirements. By comparing the old and new gradient directions, comprehensively considering the strength of the influence of explicit correlation features on each dimension, and weighing the importance of different gradient components, the original gradient direction is corrected to better align with the current distribution of samples containing explicit correlations, guiding the weight adjustment towards a better solution.

[0063] After determining the updated gradient direction, iterative optimization begins based on a preset step size. This preset step size, often called the learning rate, controls the magnitude of each weight update. If the learning rate is too large, the model weights may easily cross the optimal solution during updates, oscillating within the loss function space and failing to converge. If the learning rate is too small, weight updates are extremely slow, significantly increasing training time costs. In each iteration, the model weights move closer to the optimal parameters after incorporating explicit correlations. After multiple iterations, the loss function converges to below a certain threshold, or the preset maximum number of iterations is reached. At this point, the model weights are updated and better adapted to device data containing explicit correlations, uncovering potential implicit correlations.

[0064] Based on the comparison of old and new gradient directions, it is necessary to weigh the importance of different gradient components. Different gradient components should be assigned different weights according to the strength of the influence of explicit correlation features on each dimension. Gradient components that are more significantly affected by explicit correlation features should be given greater importance, while those less affected can have their weights appropriately reduced. In this way, the original gradient direction is corrected to better fit the current distribution of samples containing explicit correlations, thereby more accurately guiding the model's weight adjustments and leading the model towards a better solution.

[0065] It should be noted that the convergence condition of the loss function in the above model is not constant, but can be changed according to the data characteristics and business needs of the substation.

[0066] Specifically, regarding data characteristics, the convergence threshold can be set based on the data's scale, complexity, and distribution variations. Larger scales, greater complexity, and more varied distributions result in slower convergence. For example, if a substation has a massive amount of data, such as real-time monitoring data from numerous devices, the training process may be time-consuming. In this case, the convergence threshold of the loss function can be appropriately relaxed, while setting a larger maximum number of iterations. When the data contains a large amount of noise, outliers, or complex correlations, the model's convergence difficulty increases. A dynamic convergence threshold can be used, gradually decreasing the threshold as the number of iterations increases. When the data distribution changes significantly, the convergence conditions should be readjusted, and the maximum number of iterations can be appropriately increased to allow the model to better adapt to the new data distribution.

[0067] Specifically, regarding business requirements, convergence can be set based on timeliness, accuracy, and importance. Higher timeliness, accuracy, and importance requirements will result in slower convergence. For example, for scenarios requiring rapid fault response, such as real-time protection of critical equipment, a smaller maximum number of iterations and a relatively lenient loss function threshold can be set. For scenarios with high accuracy requirements, such as fault diagnosis and analysis, stricter convergence conditions should be set. For critical equipment in substations, such as main transformers and busbars, faults can lead to serious consequences. Therefore, when training models using data from these devices, more stringent convergence conditions should be set to ensure accurate detection of faults in critical equipment.

[0068] In some embodiments, primary device data and secondary device data are input into an adjusted support vector machine model to obtain implicit associations, including:

[0069] Based on the action events and alarm information of primary equipment, the primary equipment data and secondary equipment data are divided into multiple sets of input data;

[0070] Multiple sets of input data are fed into the support vector machine model to obtain initial association features;

[0071] The initial association features are integrated to obtain implicit associations.

[0072] Primary equipment action events and alarm information are key entry points for understanding equipment operating status and important bases for dividing data groups. Action events include circuit breaker opening and closing, transformer on-load tap changing operations, etc., while alarm information includes abnormal indications such as equipment overload, overheating, and reduced insulation. Whenever such a critical event or alarm is detected, primary and secondary equipment data within a time window are collected to construct a set of input data.

[0073] For example, when a circuit breaker trips, the system traces back a few seconds and extends forward a certain timeframe, integrating the bus voltage and current data connected to the circuit breaker, as well as all monitoring data and action logs recorded by related relay protection devices and measurement and control devices, to form a targeted set of data. Following the same logic, based on different action events and alarm information, data is continuously extracted and organized, dividing the massive amounts of primary and secondary equipment data into multiple sets of clearly structured and highly correlated input data. Each set of data closely revolves around specific equipment operating status changes, providing accurate material for subsequent model mining of implicit correlations.

[0074] After data grouping is completed, these data groups are fed one by one into the adjusted support vector machine model. This model, having undergone optimization based on explicit associations, already has some "recognition" of common linkage relationships between devices. Now, when faced with these data groups divided by events and alarms, it can analyze the data from a more complex perspective.

[0075] When a set of data is input, the model performs a nonlinear mapping on the data based on a kernel function (such as the Gaussian radial basis function), transforming the low-dimensional equipment data into a high-dimensional feature space. This allows subtle relationships that were difficult to capture in the original space to be highlighted. Subsequently, by determining the decision boundary, the model outputs the corresponding data feature relationships, which are the initial correlation features. These features may manifest as potential coupling relationships between changes in certain primary equipment parameters and secondary equipment responses. Although they have not yet been fully integrated and analyzed, they have already preliminarily outlined the implicit correlation clues hidden in the data, laying the foundation for subsequent in-depth analysis.

[0076] The initial correlation features output from the model are numerous and fragmented, requiring further integration and refinement. First, correlation analysis is performed on these features, using statistical methods such as Pearson correlation coefficient and mutual information to measure the strength of the correlation between features and identify highly correlated feature clusters that may point to the same potential failure mode.

[0077] For example, it was found that among multiple initial correlation features, the increase in local vibration amplitude of primary equipment always accompanied the increase in bit error rate of secondary equipment communication modules, and the correlation was very high. Therefore, these related features were grouped together. Next, combining the physical principles of substation equipment and operation and maintenance experience, the integrated feature clusters were interpreted and verified. False correlations that occurred occasionally and did not conform to the equipment's operating logic were removed, and only well-considered feature combinations were retained, ultimately forming accurate and reliable implicit correlations. These implicit correlations reveal deep, non-intuitive mutual influence relationships between equipment, providing crucial evidence for fault diagnosis and operation and maintenance decisions.

[0078] In some embodiments, a first fault profile of the primary equipment and a second fault profile of the secondary equipment are obtained based on primary equipment data, explicit associations, and implicit associations, including:

[0079] The primary equipment data, explicit and implicit correlations are input into the first deep learning model to obtain the first fault profile of the primary equipment and the second fault profile of the secondary equipment.

[0080] The data from primary equipment comes from a wide range of sources and is quite complex, including real-time data on voltage, current, temperature, vibration, etc., collected by various sensors. The first step is to clean this data, removing obvious outliers, such as extreme values ​​caused by sudden sensor malfunctions, as these anomalies can interfere with model learning. For missing values, imputation methods using the mean or median can be employed based on the statistical patterns of historical equipment operation data. Next, key features are extracted, such as voltage fluctuation amplitude and current harmonic components, transforming continuous time-series data into feature vectors that characterize the equipment's operating state, preparing the data for input into the deep learning model.

[0081] While explicit and implicit correlations reveal the relationships between devices, they need to be quantified and encoded before they can be accepted by the model. Explicit correlations are mostly based on device connections and explicit action logic, and can be simply encoded into binary features, such as a circuit breaker being associated with a corresponding protection device, recorded as 1 when triggered and 0 when not triggered. Implicit correlations are more complex. Through feature engineering, the discovered potential relationships are transformed into numerical features. For example, the coordinated changes in the temperature of primary equipment and the communication quality of secondary equipment under long-term operation can be mapped into a numerical value of a specific dimension, which is then combined with other data to form a complete input sample.

[0082] Considering that substation equipment data has both time-series characteristics and complex nonlinear relationships, an architecture combining Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN) can be selected as the primary deep learning model. LSTM excels at capturing the temporal dependencies of data and handling dynamic changes in equipment operating parameters; CNN, on the other hand, uses convolutional layers to automatically extract local features and uncover hidden patterns in the data. The synergy between the two can better handle complex equipment data.

[0083] After establishing the basic architecture, key hyperparameters need to be determined, such as the number of hidden units in the LSTM layer, the size and number of convolutional kernels in the CNN, the learning rate, and the batch size. These hyperparameters affect the model's training speed and performance, and are often initially selected through small-scale experiments and cross-validation, followed by fine-tuning and optimization. For example, a small learning rate, such as 0.001, can be set initially to avoid gradient explosion in the early stages of the model, and then flexibly adjusted based on the results after several training rounds.

[0084] Preprocessed data, including primary equipment data, explicit associations, and implicit associations, is divided into training, validation, and test sets in a specific ratio, typically 70%, 15%, and 15%. The training set is used for model parameter learning, the validation set assists in adjusting hyperparameters and monitoring overfitting, and the test set evaluates the final model performance. Loss function and optimizer: For classification tasks like fault diagnosis, a suitable loss function is chosen, such as cross-entropy loss, which measures the difference between the model's predicted output and the true label. An optimizer, such as the Adam optimizer, adaptively adjusts the learning rate, accelerating model convergence. During training, the model calculates gradients based on the loss function, updates weights using the optimizer, and iterates repeatedly to gradually learn the fault patterns in the data.

[0085] After training, new primary equipment data to be detected, along with explicit and implicit correlation samples, are input into the model. Starting from the input layer, the model processes data sequentially through each hidden layer, extracting features, fusing correlation information, and finally outputting the corresponding results at the output layer. For primary equipment, the first fault profile is presented in vector form, with each dimension corresponding to the probability of different fault types, such as insulation fault risk and short-circuit fault probability. For secondary equipment, the second fault profile similarly depicts its potential fault state, such as the risk of protection device malfunction and the degree of accuracy degradation in measurement and control equipment, providing crucial guidance for subsequent maintenance personnel to accurately locate faults.

[0086] In some embodiments, based on secondary equipment data, explicit associations, and implicit associations, a third fault profile of the primary equipment and a fourth fault profile of the secondary equipment are obtained, including:

[0087] The secondary equipment data, explicit and implicit associations are input into the second deep learning model to obtain the third fault profile of the primary equipment and the fourth fault profile of the secondary equipment.

[0088] Secondary equipment data includes various information from relay protection devices, measurement and control devices, etc. The first step is data purification. Since this type of data contains many discrete alarm and action records, it's necessary to verify the accuracy of the data and remove duplicate or incorrectly entered information. For missing key parameters, they are reasonably filled in based on equipment manuals and historical normal operation data. Subsequently, data reflecting the essential characteristics of equipment operation are extracted. For example, features are extracted from the action duration and frequency of relay protection devices, as well as the communication delay and monitoring data deviation of measurement and control devices. The raw data is then transformed into feature vectors with analytical value to meet the model's input requirements.

[0089] For explicit and implicit correlations, they must be aligned with the input requirements of the second deep learning model. Explicit correlation information needs to be re-organized and re-encoded from the perspective of the secondary device. For example, the chain reaction relationship between a secondary device and a primary device should be converted into a dedicated binary or multi-level code based on the functional characteristics of the secondary device. When incorporating implicit correlation features, the characteristics of the secondary device data should be fully considered, and the feature dimensions and quantization methods should be optimized. The complex latent relationships previously discovered should be adjusted into numerical features that are more strongly correlated with the operational failure modes of the secondary device, and then integrated with other data to form a complete sample.

[0090] Targeted Model Selection: The architecture of the second deep learning model should be tailored to the characteristics of secondary equipment data. Given that secondary equipment data is event-driven and primarily based on logical judgment, a combination of Multilayer Perceptron (MLP) and Recurrent Neural Network (RNN) architecture can be considered. MLP can handle nonlinear mappings and address complex transformations between setpoints, parameters, and fault types; RNN and its variants can capture the sequential changes in equipment operating status over time, especially for time-series data such as continuous alarms, uncovering hidden fault evolution patterns.

[0091] Once the model architecture is determined, hyperparameters need to be precisely set. These include the number of neurons in each hidden layer of an MLP, the memory unit size of an RNN, and the learning rate and regularization coefficients during training. Preliminary experiments should be conducted using a small dataset to observe changes in model convergence speed and accuracy, gradually identifying suitable hyperparameter combinations. For example, initially setting the learning rate in the range of 0.0005-0.005, combined with regularization to prevent overfitting, ensures the model has both learning efficiency and generalization ability.

[0092] The well-trained second deep learning model receives new input samples, namely, integrated secondary equipment data, explicit correlations, and implicit correlations. The sample data passes through complex operations in the hidden layers sequentially from the model's input layer. Each layer extracts, fuses, and transforms features, finally outputting the corresponding results at the output layer. For primary equipment, a third fault profile is output, accurately outlining the degree of different fault risks across various dimensions, such as mechanical fault risk and electrical performance degradation level. For secondary equipment, a fourth fault profile clearly displays key fault information such as communication link fault risk and the possibility of control logic anomalies, providing maintenance personnel with intuitive and detailed fault diagnosis criteria.

[0093] In some embodiments, determining the fault detection result of the primary equipment based on the first fault profile and the third fault profile; determining the fault detection result of the secondary equipment based on the second fault profile and the fourth fault profile, including:

[0094] The first and third fault profiles are aligned by feature, and the second and fourth fault profiles are aligned by feature.

[0095] The aligned first and third fault profiles are weighted and fused to obtain the fault detection results of the primary equipment.

[0096] The aligned second and fourth fault profiles are weighted and fused to obtain the fault detection results of the secondary equipment.

[0097] Although both the first and third fault profiles focus on primary equipment, their feature dimensions, order, and meanings often differ due to variations in the deep learning model architecture, data processing methods, and input data emphasis. First, a detailed analysis of the features in both profiles is necessary to establish feature mapping relationships. For example, if the second dimension in the first fault profile represents "abnormal winding temperature trend," while the corresponding dimension in the third fault profile is the fourth dimension, these should be clearly labeled to ensure accurate matching of similar key features. For features present in one profile but missing in the other, if the feature has a minor impact on the overall fault assessment, it can be temporarily discarded; if it has a significant impact, it should be supplemented with zero values ​​to ensure that the number and order of feature dimensions in both profiles are completely consistent, laying a solid foundation for subsequent fusion.

[0098] Similarly, feature cleanup and standardization are also required for the second and fourth fault profiles. These two profiles cover fault features related to the protection, measurement, and control functions of secondary equipment. Due to model differences, the feature presentation formats differ. For example, the alarm feature recording method in the second profile differs from that in the fourth profile. It is necessary to unify the coding rules, remove redundant and repetitive parts, sort out the core and representative features, and then match them one by one according to the feature connotations. Missing features should be filled in or deleted as needed to ensure that the feature system of the two profiles is standardized and consistent, and can be seamlessly integrated into the subsequent fusion process.

[0099] Weight setting should comprehensively consider the past performance of both models. On one hand, evaluation methods such as cross-validation can be used to statistically analyze key indicators such as accuracy and recall of the first and third models on historical test datasets. Assuming the first model has an accuracy of 85% and the third model has 78%, the weights can be initially allocated proportionally and then normalized so that the sum of their weights is 1. Thus, the weight of the first model is approximately 0.52, and that of the third model is approximately 0.48. On the other hand, based on the experience of operations and maintenance experts, if the experts determine that the first model is more accurate in identifying electrical faults and the third model is better at capturing mechanical faults, then the weights can be flexibly fine-tuned according to the type of fault currently of interest.

[0100] Each feature is multiplied by its respective weight and then summed. All aligned features are processed in this way to form a new fused feature vector. This vector integrates fault assessment information from two models across different dimensions of the primary equipment, and is used to determine the fault detection result of the primary equipment. If a fused feature value exceeds a pre-set threshold, a corresponding fault is determined. For example, if the feature value representing an insulation fault is greater than 0.8, an insulation fault is confirmed in the primary equipment.

[0101] For the second and fourth fault profiles, the weights are determined first. In addition to considering the model's historical accuracy and recall, the fault occurrence patterns and maintenance priorities of the secondary equipment itself should also be taken into account. For example, for fault judgments related to protection devices, if past data shows that the fourth model's judgments are more consistent with actual faults, the weight of the corresponding features in the fourth model should be appropriately increased. Through in-depth analysis of past fault cases and equipment operating characteristics, the weights of the features in both profiles are precisely assigned.

[0102] Similar to the fusion of primary equipment profiles, the corresponding features of the second and fourth fault profiles are multiplied by their respective weights and then summed to obtain a fused feature vector. Then, based on the common fault threshold system for secondary equipment, the values ​​of each dimension of the fused feature vector are examined. Once the values ​​of dimensions related to communication faults exceed the critical value, combined with auxiliary information such as alarm records, it can be determined that a communication fault has occurred in the secondary equipment, thus obtaining an accurate secondary equipment fault detection result.

[0103] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0104] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for fault detection in substation automation equipment, characterized in that, The method includes: Obtain primary and secondary equipment data from the substation; A correlation analysis is performed on the primary equipment data and the secondary equipment data to obtain explicit and implicit correlations; Based on the primary equipment data, explicit associations, and implicit associations, a first fault profile of the primary equipment and a second fault profile of the secondary equipment are obtained. Based on the secondary equipment data, explicit associations, and implicit associations, a third fault profile of the primary equipment and a fourth fault profile of the secondary equipment are obtained. Based on the first fault profile and the third fault profile, the fault detection result of the primary equipment is determined; based on the second fault profile and the fourth fault profile, the fault detection result of the secondary equipment is determined.

2. The substation automation equipment fault detection method according to claim 1, characterized in that, A correlation analysis is performed on the primary equipment data and the secondary equipment data to obtain explicit and implicit correlations, including: The primary equipment data and the secondary equipment data are correlated based on the equipment knowledge graph, action events and alarm information in the primary equipment data to obtain the explicit correlation; wherein, the equipment knowledge graph is established based on the electrical connection topology of the substation; The pre-trained support vector machine model is adjusted based on the explicit association, and the primary device data and the secondary device data are input into the adjusted support vector machine model to obtain the implicit association.

3. The substation automation equipment fault detection method according to claim 2, characterized in that, Adjusting the pre-trained support vector machine model based on the explicit associations includes: Based on the explicit association, gradient adjustment information is determined; Based on the gradient adjustment information, the gradient direction is updated, and the support vector machine model is iterated according to the updated gradient direction and a preset step size to update the model weights.

4. The substation automation equipment fault detection method according to claim 2, characterized in that, The primary equipment data and the secondary equipment data are input into the adjusted support vector machine model to obtain the implicit association, including: Based on the action events and alarm information of the primary equipment, the primary equipment data and the secondary equipment data are divided into multiple sets of input data; The multiple sets of input data are input into the support vector machine model to obtain initial association features; The initial association features are integrated to obtain the implicit association.

5. The substation automation equipment fault detection method according to claim 1, characterized in that, Based on the primary equipment data, explicit associations, and implicit associations, a first fault profile of the primary equipment and a second fault profile of the secondary equipment are obtained, including: The primary equipment data, explicit associations, and implicit associations are input into the first deep learning model to obtain the first fault profile of the primary equipment and the second fault profile of the secondary equipment.

6. The substation automation equipment fault detection method according to claim 1, characterized in that, Based on the secondary equipment data, explicit and implicit associations, a third fault profile of the primary equipment and a fourth fault profile of the secondary equipment are obtained, including: The secondary equipment data, explicit associations, and implicit associations are input into the second deep learning model to obtain the third fault profile of the primary equipment and the fourth fault profile of the secondary equipment.

7. The substation automation equipment fault detection method according to claim 4, characterized in that, Based on the first fault profile and the third fault profile, determine the fault detection result of the primary equipment; Based on the second fault profile and the fourth fault profile, the fault detection results of the secondary equipment are determined, including: The first fault profile and the third fault profile are feature aligned, and the second fault profile and the fourth fault profile are feature aligned. The aligned first fault profile and the third fault profile are weighted and fused to obtain the fault detection result of the primary equipment. The aligned second fault profile and the fourth fault profile are weighted and fused to obtain the fault detection result of the secondary equipment.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the substation automation equipment fault detection method as described in any one of claims 1 to 7.

9. A fault detection system for substation automation equipment, characterized in that, It includes a detection unit, a data acquisition terminal, and the electronic device as described in claim 8 above.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the substation automation equipment fault detection method as described in any one of claims 1 to 7.

Images