Intelligent diagnosis system for power equipment

Abstract

The invention provides an intelligent diagnosis system for power equipment, and the system comprises an equipment family management module which is used for dividing a plurality of pieces of power equipment into at least one equipment family based on the static attributes, dynamic working conditions and time dimension information of the power equipment; the group baseline construction module is used for constructing a group health baseline based on the historical state monitoring data of the equipment family; the collaborative deviation detection module is used for acquiring real-time state data of target member equipment in the equipment family, calculating a first deviation degree between the real-time state data and an individual historical baseline of the target member equipment, calculating a second deviation degree between the real-time state data and a group health baseline, when the first deviation degree and / or the second deviation degree meet a preset condition, a diagnosis conclusion is generated; and the intelligent early warning module is used for analyzing the diagnosis conclusion and generating early warning information. According to the method, systematic risks are eliminated in advance, the overall safety level and the risk resisting capacity of the power grid are enhanced, and the reliability of social power utilization is guaranteed.

Description

Technical Field

[0001] This invention relates to the field of power equipment diagnostic technology, and in particular to an intelligent diagnostic system for power equipment. Background Technology

[0002] With the rapid development of society and economy, the power system has become the core infrastructure supporting industrial production, people's livelihood and social operation. As key components of the power system, the operating status of power equipment such as transformers, circuit breakers and switchgear directly determines the safety, stability and reliability of power supply.

[0003] In recent years, with the development of sensor technology, the Internet of Things and big data technology, the operation and maintenance of power equipment has entered the era of condition-based maintenance. Various sensors installed on the equipment collect status data in real time, providing a data foundation for fault diagnosis. However, how to extract valuable information from the entire network, multi-dimensional and heterogeneous data, and accurately judge the health status and fault trends of the equipment has become a new challenge. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent diagnostic system for power equipment, which enhances the overall safety level and risk resistance of the power grid by eliminating systemic risks in advance, and ensures the reliability of electricity use in society.

[0005] To achieve the above objectives, the present invention provides an intelligent diagnostic system for power equipment, the system comprising: The device family management module is used to divide multiple power devices into at least one device family based on the static attributes, dynamic operating conditions and time dimension information of the power devices. The group baseline construction module is used to construct a group health baseline for a device family based on historical status monitoring data of multiple member devices in the device family. The collaborative deviation detection module is used to acquire real-time status data of target member devices in the device family, calculate the first deviation between the real-time status data and the individual historical baseline of the target member device, calculate the second deviation between the real-time status data and the group health baseline, and generate a diagnostic conclusion when the first deviation and / or the second deviation meet the preset conditions. The intelligent early warning module is used to analyze diagnostic conclusions and generate early warning information when multiple member devices in the device family deviate from their synchronization characteristics.

[0006] Furthermore, the device family management module includes: The first clustering unit is used to divide devices into an initial device family based on the static attributes of the devices using a graph clustering algorithm. The second clustering unit is used to perform secondary clustering based on the dynamic operating conditions and time dimension information of the devices in the initial device cluster, forming at least one device family.

[0007] Furthermore, the first clustering unit is specifically used to perform the following operations: S11. Map each power device as a node in the graph data model; S12. Based on the static attributes of power equipment, establish connection edges at the corresponding equipment nodes according to preset rules to construct an initial graph data model; S13. Traverse each node of the initial graph data model and calculate the modularity gain caused by moving each node to the community of an adjacent node. S14. Aggregate nodes within the same community into a new node until the modularity gain of the new node reaches a preset threshold, forming an initial device family. Each aggregation of a new node forms an initial device family.

[0008] Furthermore, the second clustering unit is specifically used to perform the following operations: S21. For each device in the initial device family, extract features from dynamic operating conditions and time dimensions to form a feature vector; S22. Based on the correlation between the dynamic operating conditions and time-dimensional features of the equipment and the health status of the equipment, assign adaptive weights to each feature vector. S23. Preset the number of target families and define the corresponding membership matrix and cluster center vector; S24. Based on adaptive weights, calculate the weighted distance from each device in the initial device family to the corresponding target family cluster center; S25. With minimizing the weighted sum of squared distances of all devices as the objective function, update the membership matrix and cluster center vector of each device to the corresponding target family until the change in the membership matrix is ​​less than a preset threshold. S26. Assign each device in the initial device family to the target family with the highest membership degree to form at least one device family.

[0009] Furthermore, establish a health baseline for the device family, specifically including: S31. Obtain historical status monitoring data and corresponding historical operating condition data of each member device in the device family under healthy operating conditions. S32. For at least one key status indicator, predict the theoretical value of the key status indicator based on historical operating condition data using a regression model, and calculate the residual between the theoretical value and the actual monitored value. S33. Fit the probability distribution of the residuals calculated for all members of the equipment family, and use the fitted probability distribution as the group health baseline of the key status indicators of the equipment family.

[0010] Furthermore, the first deviation between the real-time status data and the individual historical baseline of the target member device is calculated, specifically including: S41. Based on the historical health data of the target member equipment, establish an individual regression model between key status indicators and operating condition variables; S42. Obtain the historical residuals of the individual regression model and fit the probability distribution to obtain the individual health baseline of the key state indicators. S43. Obtain real-time status data and real-time operating condition data of key status indicators of target member equipment, calculate the predicted status value under real-time operating conditions through individual regression model, and calculate the real-time residual between real-time status data and predicted status value. S44. Based on the individual health baseline, the real-time residuals are standardized, and the absolute value of the processing result is used as the first deviation.

[0011] Furthermore, a second deviation between the real-time status data and the population health baseline is calculated, specifically including: S51. Obtain the real-time residual of at least one key status indicator of the target member device and convert it into a real-time residual vector. S52. Calculate the Mahalanobis distance between the real-time residual vector and the mean vector of the population health baseline, as the second deviation.

[0012] Furthermore, the generation of diagnostic conclusions specifically includes: If the first deviation meets the preset condition but the second deviation does not meet the preset condition, a first diagnostic conclusion indicating that the target member device has an individual-specific abnormality is generated. If the second deviation meets the preset conditions while the first deviation does not meet the preset conditions, a second diagnostic conclusion is generated indicating that the target member device and the device family are not in the same operating state. If both the first deviation and the second deviation meet the preset conditions, a third diagnostic conclusion indicating that the target member device has a serious abnormality is generated.

[0013] Further analysis of the diagnostic conclusions includes: S61. Perform structured parsing on each diagnostic conclusion to extract key information including at least device identifier, anomaly type, deviation severity, and timestamp. Based on the extracted key information, classify multiple member devices into different anomaly type groups. S62. For the same abnormal type group, determine whether the number of member devices exceeds the first threshold, determine whether the timestamps of the diagnostic conclusions of the member devices are concentrated in the second preset time window, and determine whether the average deviation severity of the member devices is lower than the third threshold. S63. If all the above conditions are met, it is determined that multiple member devices in the device family have a synchronization characteristic deviation mode.

[0014] Compared with the prior art, the beneficial effects of the present invention are: This invention provides an intelligent diagnostic system for power equipment. The equipment family management module scientifically and multidimensionally divides power equipment into families, enabling systematic management of equipment groups and laying the foundation for subsequent intelligent group analysis. The group baseline construction module utilizes historical data from multiple healthy devices within the family to construct a group health baseline, obtaining stable health standards and improving the reliability of status assessment. The collaborative deviation detection module synchronously calculates and collaboratively verifies the deviation between real-time data and both individual historical baselines and the group health baseline, improving the accuracy of fault identification and reducing false alarms and missed alarms. The intelligent early warning module analyzes the collaborative detection results and identifies group synchronous deviation characteristics, achieving early and accurate warnings of potential family-related defects, thus elevating the operation and maintenance mode from passive response to proactive defense. This invention enhances the overall safety level and risk resistance capability of the power grid by eliminating systemic risks in advance, ensuring the reliability of electricity supply to society. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, 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 embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of the structure of an intelligent diagnostic system for power equipment provided in an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the specific execution flow of the first clustering unit provided in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the specific execution flow of the second clustering unit provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the process for building a family of health baselines for devices provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the process for calculating the first deviation line provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the process for calculating the second deviation line provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of the analytical diagnostic conclusion process provided in an embodiment of the present invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0018] Reference Figure 1 This embodiment provides an intelligent diagnostic system for power equipment, the system comprising: The device family management module is used to divide multiple power devices into at least one device family based on the static attributes, dynamic operating conditions, and time-related information of the power devices.

[0019] The group baseline construction module is used to construct a group health baseline for a device family based on historical status monitoring data of multiple member devices in the device family.

[0020] The collaborative deviation detection module is used to acquire real-time status data of target member devices in the device family, calculate the first deviation between the real-time status data and the individual historical baseline of the target member device, calculate the second deviation between the real-time status data and the group health baseline, and generate a diagnostic conclusion when the first deviation and / or the second deviation meet preset conditions.

[0021] The intelligent early warning module is used to analyze diagnostic conclusions and generate early warning information when multiple member devices in the device family deviate from their synchronization characteristics.

[0022] In this embodiment, the system first classifies all power equipment in the network into different equipment families based on static, dynamic, and time-related attributes such as equipment model, specifications, operating environment, and commissioning time. This ensures that highly similar and comparable equipment are grouped together, guaranteeing the scientific validity and effectiveness of subsequent group comparative analysis and resolving the problem of difficulty in directly comparing equipment due to individual differences. The system aggregates historical health status monitoring data of member equipment within an equipment family and constructs a group health baseline representing the overall health level of the family through machine learning or statistical models. Compared to a single equipment baseline, the group baseline effectively smooths out random fluctuations and noise of individual equipment, more clearly reflecting the common patterns of health status of this type of equipment, and providing a reliable reference system for anomaly detection. For the target equipment in the family, the system acquires its status data in real time and simultaneously calculates its deviation from its own individual historical baseline to obtain a first deviation, and compares it with the group health baseline of its family to obtain a second deviation.

[0023] By judging whether the first and second deviations meet preset alarm conditions, dual verification is achieved, improving diagnostic accuracy. Simultaneously, it can keenly detect early faults where individual baselines are still within the normal range but have begun to deviate from the overall health level of the group. The system not only alarms for anomalies in individual devices but also intelligently analyzes the diagnostic conclusions of the entire device family. When multiple member devices in the family show synchronous characteristic deviations, it generates early warning information, achieving proactive early warning of potential family-wide defects. This invention elevates the operational perspective from a point-to-surface approach, identifying systemic risks caused by factors such as material defects and design problems during the incubation period, thereby shifting from reactive emergency repairs to proactive prevention and control, avoiding large-scale failures.

[0024] In a preferred embodiment, the device family management module includes: The first clustering unit is used to divide devices into an initial device family based on the static attributes of the devices using a graph clustering algorithm.

[0025] The second clustering unit is used to perform secondary clustering based on the dynamic operating conditions and time dimension information of the devices in the initial device cluster, forming at least one device family.

[0026] In this embodiment, the first clustering unit, based on the static attributes of the devices, such as device model, manufacturer, rated voltage, capacity, design specifications, and component materials, uses a graph clustering algorithm to divide all network devices into several initial device families. Among the static attributes of the devices, devices of the same model and batch exhibit high similarity in their design life, theoretical aging curves, and potential failure modes. The graph clustering algorithm, by processing the complex relationships between devices based on static attributes, quickly and accurately categorizes all network devices into different initial device families, providing a clear and stable initial framework for subsequent analysis.

[0027] The second clustering unit, within statically similar groups, further identifies homogeneous subgroups with similar aging drivers and health states based on dynamic and temporal characteristics that have the greatest impact on health status. This provides a data foundation for subsequently constructing a group health baseline. Since the aging rate and operating status of equipment are influenced by its individual operating experience—for example, a piece of equipment operating at full load year-round will have drastically different health baselines and degradation trajectories than a piece of the same model operating at light load for extended periods—mixing the two together to establish a unified group baseline would result in a blurred, inaccurate baseline, failing to sensitively detect anomalies. After the first clustering unit categorizes all network equipment into different initial equipment families, the second clustering unit uses secondary clustering to group equipment with similar origins and experiences together, forming a single equipment family. For example, a type A transformer that has been in operation for approximately 10 years and consistently operates at 70%-80% load is grouped into one family.

[0028] In a preferred embodiment, the first clustering unit is specifically used to perform the following operations: S11. Map each power device as a node in the graph data model.

[0029] S12. Based on the static attributes of power equipment, establish connection edges at the corresponding equipment nodes according to preset rules to construct an initial graph data model.

[0030] S13. Traverse each node of the initial graph data model and calculate the modularity gain caused by moving each node to the community of the adjacent node.

[0031] S14. Aggregate nodes within the same community into a new node until the modularity gain of the new node reaches a preset threshold, forming an initial device family. Each aggregation of a new node forms an initial device family.

[0032] In this embodiment, a single device may share different attributes with multiple devices simultaneously. For example, A and B may be the same model, B and C may be from the same manufacturer, and A and C may be from the same substation. This complex many-to-many relationship is difficult to represent using simple tree or list structures, while graph structures can naturally represent such complex networks. Graph clustering is an unsupervised learning method that can automatically discover community structures based on the tightness of connections between nodes. For instance, devices of different models that use the same batch of key components or come from the same supplier's secondary contract manufacturer may have potential common defect risks. Graph models can capture these hidden relationships through connecting edges. Therefore, by employing graph clustering algorithms, we can uncover the groups of devices with the tightest static attribute associations, providing a data foundation for constructing a group health baseline.

[0033] Specifically, each power device is mapped to a node in the graph. Based on static attributes, connecting edges are established between device nodes. The establishment of edges is based on rules: strong connection edges: if two devices have the same production batch number, a strong connection edge is established between them, because devices in the same batch are highly consistent in raw materials, processes, and quality inspection, and have the highest risk of family defects; medium connection edges: if two devices have the same model and manufacturer, but different batches, a medium-weight connection edge is established, because they have the same design but different manufacturing times, and still have certain common risks; weak connection edges: if two devices only have the same model but different manufacturers, a weak connection edge is established for auxiliary analysis.

[0034] Once the graph model is constructed, graph clustering algorithms are executed to automatically identify tightly connected subgraphs. For example, community detection algorithms, including the Louvain and Leiden algorithms, can identify community structures in large networks without pre-specifying the number of communities. By continuously optimizing the community affiliation of nodes, the connection density within a community is made much higher than the connection density between communities. Each node is assigned a community label, and the set of devices with the same community label forms the initial device cluster (initial device family).

[0035] The goal of the community detection algorithm is to maximize global modularity by adjusting the communities to which nodes belong. In other words, it evaluates the quality of node partitioning through a quantified objective function (modularity). Modularity is a metric used to measure the quality of community partitioning, ranging from -0.5 to 1; a higher value indicates tighter connections within a community and sparser connections between communities. First, each device node is assigned to an independent community. Then, each node in the graph is traversed. For the current node, consider moving it to the community of each of its neighbors. After each move attempt, the change in global modularity is calculated. The node is moved to the community that produces the largest positive change. If no moves produce a positive change, the node remains in its original community. This process is repeated until no move increases the modularity, at which point a local optimum is reached.

[0036] Secondly, each community formed in the first stage is aggregated into a new node to construct a new graph. The edge weights between the new nodes are the sum of the weights of all edges between nodes from two different communities; the weights of self-loop edges in the new node are the sum of the weights of all edges within the community. Using the newly constructed graph as input, the algorithm returns to the first stage and continues to move and optimize nodes on the new graph. The algorithm stops when the modularity no longer increases, or when the increase is less than a preset threshold. Finally, a community assignment list is output, where each device node is assigned a final community number. The set of device nodes with the same community number is identified as an initial device family.

[0037] In a preferred embodiment, the second clustering unit is specifically used to perform the following operations: S21. For each device in the initial device family, extract features from dynamic operating conditions and time dimensions to form a feature vector.

[0038] S22. Based on the correlation between the dynamic operating conditions and time-dimensional features of the equipment and the health status of the equipment, assign adaptive weights to each feature vector.

[0039] S23. Preset the number of target families and define the corresponding membership matrix and cluster center vector.

[0040] S24. Based on adaptive weights, calculate the weighted distance from each device in the initial device family to the corresponding target family cluster center.

[0041] S25. Using minimizing the weighted sum of squared distances of all devices as the objective function, update the membership matrix and cluster center vector of each device to the corresponding target family until the change in the membership matrix is ​​less than a preset threshold.

[0042] S26. Assign each device in the initial device family to the target family with the highest membership degree to form at least one device family.

[0043] In this embodiment, the first clustering unit completes the initial grouping based on static attributes. However, devices within the same initial device family still exhibit significant differences in their health status and degradation trajectories. For example, devices of the same model may have different load rates, start-stop frequencies, and electrothermal stresses, which could directly lead to different aging rates. Differences in commissioning time and maintenance cycles place devices at different stages of their life cycle. If devices with different operating histories and conditions are directly mixed together to construct a unified group health baseline, the group health baseline will become ambiguous and unrepresentative, thereby reducing the sensitivity of subsequent deviation detection. The second clustering unit further identifies homogeneous subgroups with similar aging driving factors and health status within statically similar groups, based on the dynamic and time-dimensional characteristics that have the greatest impact on health status, thereby constructing device families.

[0044] Specifically, for each device in the initial equipment family, the system extracts features representing dynamic operating conditions and the time dimension from historical data, forming a numerical feature vector. Dynamic operating conditions include average load rate, load fluctuation variance, and overload count; the time dimension includes years of operation and operating time since the last maintenance. Weights are calculated based on the correlation between each feature and the device's health status. Health status is typically quantified using one or more key status indicators, such as insulation aging index and gas production rate. Correlation is assessed using correlation coefficients, mutual information, or model-based feature importance analysis. Features with higher correlation to health status are assigned greater weights and play a more significant role in subsequent distance calculations.

[0045] The membership matrix and cluster center vectors are randomly initialized by setting a preset number of target families. The membership matrix represents the degree to which each device belongs to each target family. In each iteration, the weighted Euclidean distance from the device to the cluster center is calculated. Based on the weighted distance, the membership matrix is ​​updated; the closer a device is to a cluster center, the higher its degree of belonging to that family. Based on the current membership, the cluster center vector of each target family is recalculated, making it a weighted average of the feature vectors of all devices in that family. The algorithm iterates with the objective function of minimizing the sum of squared weighted distances of all devices until the change in the membership matrix is ​​less than a preset threshold, indicating that the algorithm has converged to a stable solution. For each device, the target family with the highest membership is selected as its final affiliation, thus dividing the initial device families into more homogeneous final device families.

[0046] As a preferred embodiment, establishing a group health baseline for the device family specifically includes: S31. Obtain historical status monitoring data and corresponding historical operating condition data of each member device in the device family under healthy operating conditions.

[0047] S32. For at least one key status indicator, predict the theoretical value of the key status indicator based on historical operating condition data using a regression model, and calculate the residual between the theoretical value and the actual monitored value.

[0048] S33. Fit the probability distribution of the residuals calculated for all members of the equipment family, and use the fitted probability distribution as the group health baseline of the key status indicators of the equipment family.

[0049] In this embodiment, traditional health baselines may directly perform statistical modeling on condition monitoring data, such as the hydrogen content in transformer oil, leading to equipment condition monitoring values ​​being highly dependent on its operating conditions. For example, a transformer operating under high temperature and heavy load conditions will have significantly higher winding hot spot temperatures and gas production rates than under light load conditions. If the influence of operating conditions is ignored and thresholds are set directly, either a large number of false alarms will occur under heavy load conditions, or early anomalies will not be detected under light load conditions. Therefore, by establishing a dynamic baseline model that considers the equipment's operating conditions, the normal influence of operating condition changes on condition indicators is eliminated, and the remaining residual sequence more purely represents abnormal changes caused by the equipment's own health degradation; the equipment's own health degradation includes insulation material aging and component wear.

[0050] Specifically, historical status monitoring data and corresponding historical operating condition data for each member device in the equipment family are acquired under healthy operating conditions. Healthy operating condition typically refers to a period of time without alarms or faults. Historical status monitoring data includes key status indicators such as hydrogen concentration and vibration amplitude; historical operating condition data includes characteristic variables such as load current, ambient temperature, and operating duration. For each key status indicator, a regression model is established to describe how operating conditions normally affect the corresponding status indicator under healthy conditions. For each data point under healthy conditions, the trained regression model is used to predict the theoretical value of the status indicator under the operating condition, and the residual is the difference between the actual monitored value and the theoretical predicted value.

[0051] In a healthy state, residuals are randomly distributed around zero, representing random fluctuations and measurement noise that the model cannot explain. By aggregating the residuals calculated for all members of the equipment family in a healthy state, and fitting the residual dataset using statistical methods, a probability distribution function describing the residual behavior of the equipment family in a healthy state is obtained, thus providing the group health baseline for key state indicators. Since the group health baseline is constructed based on the health data of the entire family, it represents the common health characteristics of the family. If multiple devices in the family exhibit similar types of residual anomalies, such as residuals all shifting in the positive direction, even if the absolute shift of each device is small, it strongly suggests the existence of common defects affecting the entire family, such as material problems, thereby initiating a family-wide early warning.

[0052] As a preferred embodiment, calculating a first deviation between the real-time status data and the individual historical baseline of the target member device specifically includes: S41. Based on the historical health data of the target member equipment, establish an individual regression model of key status indicators and operating condition variables.

[0053] S42. Obtain the historical residuals of the individual regression model and fit the probability distribution to obtain the individual health baseline of the key state indicators.

[0054] S43. Obtain real-time status data and real-time operating condition data of key status indicators of target member equipment, calculate the predicted status value under real-time operating conditions through individual regression model, and calculate the real-time residual between real-time status data and predicted status value.

[0055] S44. Based on the individual health baseline, the real-time residuals are standardized, and the absolute value of the processing result is used as the first deviation.

[0056] In this embodiment, simply comparing real-time status data with historical data ignores the current operating conditions of the equipment. Instead, a dynamic benchmark model considering the operating history of each target device is established. This model dynamically predicts normal values ​​under different operating conditions through regression models, and the evaluation benchmark is adaptively adjusted according to the operating conditions. The calculation of the first deviation involves constructing an individual regression model for each device and standardizing the residuals. This transforms equipment status assessment into a statistical hypothesis testing process, enhancing the ability to detect early faults while suppressing interference caused by fluctuations in operating conditions.

[0057] Specifically, historical data of the target equipment under known health conditions is used, including key status indicators and corresponding operating condition variables. Key status indicators include temperature, vibration, and gas content, while operating condition variables include load, ambient temperature, and rotational speed. An individual regression model is trained using a regression algorithm. This model describes how operating condition variables determine the normal values ​​of the status indicators under healthy conditions. Each operating condition point from the historical health data is input into the individual regression model to obtain a predicted value, and historical residuals are calculated. A probability distribution is fitted to all historical residuals to obtain a probability density function. The probability density function describes the normal fluctuation range of the residuals when the equipment is healthy, i.e., the individual health baseline. It is typically assumed that the probability density function follows a normal distribution.

[0058] The system acquires real-time status data and real-time operating condition data of the equipment. This real-time operating condition data is then input into an individual regression model to obtain a predicted status value under the current operating condition. The real-time residual between the real-time status data and the predicted status value is calculated. The real-time residual reflects the deviation of the equipment's actual status from its historical health model under the current operating condition. To eliminate the influence of dimensions and achieve a uniform scale of assessment, the real-time residual is standardized using statistics from the individual health baseline. Specifically, the absolute value of the standardized value is taken as the first deviation. The first deviation is a dimensionless value; a larger value indicates a more significant deviation between the current status and the equipment's historical health behavior.

[0059] As a preferred embodiment, calculating a second deviation between real-time status data and the population health baseline specifically includes: S51. Obtain the real-time residual of at least one key status indicator of the target member device and convert it into a real-time residual vector.

[0060] S52. Calculate the Mahalanobis distance between the real-time residual vector and the mean vector of the population health baseline, as the second deviation.

[0061] In this embodiment, the group health baseline is a multivariate probability distribution describing the joint fluctuation behavior of the residuals of healthy devices within the family. These residuals of key status indicators are often correlated; for example, when the methane residual in transformer oil increases, the ethylene residual may also increase synchronously. Traditional Euclidean distance only calculates the straight-line distance between points, completely ignoring the correlation between variables and their respective variances. The second deviation, by calculating Mahalanobis distance, measures the comprehensive deviation between multiple monitored residual values ​​and the group health baseline. This achieves a precise measurement of the multidimensional, correlated comprehensive deviation between device status and the group health baseline, elevating anomaly detection from univariate, independent threshold judgment to a multivariate statistical inference level considering joint distribution, thereby enabling the system to achieve collaborative diagnosis.

[0062] Specifically, suppose the system monitors several key status indicators, such as the residuals of gas contents like H2, CH4, and C2H4. For the target device, at a certain moment, the real-time residuals of these key status indicators are calculated and transformed into a real-time residual vector. The group health baseline is obtained by fitting the residual data of a family of healthy devices to a multivariate probability distribution, characterized by two key parameters: the mean vector and the covariance matrix. The mean vector represents the average value of each residual variable under healthy conditions, ideally close to zero. The diagonal elements of the covariance matrix are the variances of each residual variable, and the off-diagonal elements are the covariances between different residual variables, quantifying their correlation and joint fluctuations. By calculating the difference vector between the residual vector and the mean vector, multiplying the difference vector by the inverse of the covariance matrix, multiplying the result by the transpose of the difference vector, and finally taking the square root, the Mahalanobis distance, i.e., the second deviation, is obtained.

[0063] Mahalanobis distance is a relative value, measured in standard deviations. This eliminates the influence of differences in dimensions and variances between different state indicators. For example, a small change in an indicator with very low variance will contribute more to the deviation than a similar change in an indicator with very high variance. This allows for a fair and comprehensive assessment of state indicators of different natures. Familial defects often lead to specific, multi-indicator synergistic anomalies in equipment status. Mahalanobis distance is extremely sensitive to these synergistic patterns. Even if the deviation of each individual indicator is small, the specific pattern formed by their combination may indicate a significant deviation from the overall distribution of the healthy population, thus enabling early warning of potential familial risks.

[0064] As a preferred embodiment, the generation of diagnostic conclusions specifically includes: If the first deviation meets the preset conditions but the second deviation does not meet the preset conditions, a first diagnostic conclusion is generated indicating that the target member device has an individual-specific abnormality.

[0065] If the second deviation meets the preset conditions while the first deviation does not, a second diagnostic conclusion is generated indicating that the target member device and the device family are not in the same operating state.

[0066] If both the first deviation and the second deviation meet the preset conditions, a third diagnostic conclusion indicating that the target member device has a serious abnormality is generated.

[0067] In this embodiment, the first diagnostic conclusion indicates that the current state of the equipment has significantly deviated from its own health history pattern, but its state is still within the normal fluctuation range of the equipment family. That is, the anomaly may be caused by factors specific to this equipment, such as occasional component failures, unique installation problems, or local environmental stresses affecting only the equipment. The second diagnostic conclusion indicates that the current state of the equipment appears normal within its own historical range, but has significantly deviated from the collective behavior of the equipment family, which is a typical early signal of potential family defects. At the same time, it may also indicate that the individual health baseline of the equipment has drifted due to long-term slow degradation and has become inaccurate, requiring reassessment. The third diagnostic conclusion indicates that the equipment anomaly is significant both in the individual historical dimension and prominent in cross-group comparisons, that is, the equipment anomaly is a strong and mature fault signal that requires immediate intervention.

[0068] As a preferred embodiment, the analysis of diagnostic conclusions specifically includes: S61. Perform structured parsing on each diagnostic conclusion to extract key information including at least device identifier, anomaly type, deviation severity, and timestamp. Based on the extracted key information, classify multiple member devices into different anomaly type groups.

[0069] S62. For the same anomaly type group, determine whether the number of member devices exceeds the first threshold, whether the timestamps of the diagnostic conclusions of the member devices are concentrated within the second preset time window, and whether the average deviation severity of the member devices is lower than the third threshold.

[0070] S63. If all the above conditions are met, it is determined that multiple member devices in the device family have a synchronization characteristic deviation mode.

[0071] In this embodiment, the system parses each generated diagnostic conclusion, extracts key structured fields, and forms a standardized abnormal event record. The abnormal event record includes the device identifier, the abnormality type (second diagnostic conclusion), the severity of the deviation, and a timestamp. The system groups all abnormal event records by abnormality type to focus on anomalies of the same nature. For example, all events with "second diagnostic conclusion" are grouped into the same group. Within the same abnormality type group, the system executes three independent judgment conditions; only when all three are met is the final warning triggered.

[0072] Condition one involves a group-based assessment, counting whether the number of different member devices involved in the group exceeds a first threshold to ensure that the anomaly is not isolated but rather has a certain degree of prevalence within the group. The first threshold is set based on the total family size and is used to filter out occasional, unrelated individual failures; for example, 10% of the total family size, or an absolute number such as 3 units. Condition two involves a synchronicity assessment, analyzing whether the timestamps of diagnostic conclusion events within the group are concentrated within a second preset time window to determine whether these anomalies occur synchronously. The time window is usually set to a relatively short range, such as 72 hours or one week; this is used to confirm that these anomalies were detected within a similar time period, thus suggesting that they may be triggered by the same common cause, rather than occurring randomly or asynchronously, such as batch material problems simultaneously entering their expiration period.

[0073] Condition three involves potential fault assessment, specifically calculating whether the average deviation severity of all events within the group is below a third threshold to identify potential anomalies. This threshold is set to a low deviation value, for example, a second deviation between 2 and 4. The goal is to ensure that only early, latent signals that have not yet developed into serious faults are captured. If the average deviation is already high, it indicates a significant fault, falling outside the scope of early warning and triggering an emergency fault alarm. The system determines that a synchronization characteristic deviation pattern exists within the device family only if all three conditions are met. This determination triggers the system to generate an early warning message regarding a potential family-wide defect, clearly indicating the type of anomaly, the list of devices involved, the time concentration, and the average deviation degree, providing the operations and maintenance team with a clear decision-making basis.

[0074] The above description is only a preferred embodiment of the present invention and is 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 diagnostic system for power equipment, characterized in that, The system includes: The device family management module is used to divide multiple power devices into at least one device family based on the static attributes, dynamic operating conditions and time dimension information of the power devices. The group baseline construction module is used to construct a group health baseline for a device family based on historical status monitoring data of multiple member devices in the device family. The collaborative deviation detection module is used to acquire real-time status data of target member devices in the device family, calculate the first deviation between the real-time status data and the individual historical baseline of the target member device, calculate the second deviation between the real-time status data and the group health baseline, and generate a diagnostic conclusion when the first deviation and / or the second deviation meet the preset conditions. The intelligent early warning module is used to analyze diagnostic conclusions and generate early warning information when multiple member devices in the device family deviate from their synchronization characteristics.

2. The intelligent diagnostic system for power equipment according to claim 1, characterized in that, The device family management module includes: The first clustering unit is used to divide devices into an initial device family based on the static attributes of the devices using a graph clustering algorithm. The second clustering unit is used to perform secondary clustering based on the dynamic operating conditions and time dimension information of the devices in the initial device cluster, forming at least one device family.

3. The intelligent diagnostic system for power equipment according to claim 2, characterized in that, The first clustering unit is specifically used to perform the following operations: S11. Map each power device as a node in the graph data model; S12. Based on the static attributes of power equipment, establish connection edges at the corresponding equipment nodes according to preset rules to construct an initial graph data model; S13. Traverse each node of the initial graph data model and calculate the modularity gain caused by moving each node to the community of an adjacent node. S14. Aggregate nodes within the same community into a new node until the modularity gain of the new node reaches a preset threshold, forming an initial device family. Each aggregation of a new node forms an initial device family.

4. The intelligent diagnostic system for power equipment according to claim 3, characterized in that, The second clustering unit is specifically used to perform the following operations: S21. For each device in the initial device family, extract features from dynamic operating conditions and time dimensions to form a feature vector; S22. Based on the correlation between the dynamic operating conditions and time-dimensional features of the equipment and the health status of the equipment, assign adaptive weights to each feature vector. S23. Preset the number of target families and define the corresponding membership matrix and cluster center vector; S24. Based on adaptive weights, calculate the weighted distance from each device in the initial device family to the corresponding target family cluster center; S25. With minimizing the weighted sum of squared distances of all devices as the objective function, update the membership matrix and cluster center vector of each device to the corresponding target family until the change in the membership matrix is ​​less than a preset threshold. S26. Assign each device in the initial device family to the target family with the highest membership degree to form at least one device family.

5. The intelligent diagnostic system for power equipment according to claim 1, characterized in that, Establish a health baseline for the entire device family, specifically including: S31. Obtain historical status monitoring data and corresponding historical operating condition data of each member device in the device family under healthy operating conditions. S32. For at least one key status indicator, predict the theoretical value of the key status indicator based on historical operating condition data using a regression model, and calculate the residual between the theoretical value and the actual monitored value. S33. Fit the probability distribution of the residuals calculated for all members of the equipment family, and use the fitted probability distribution as the group health baseline of the key status indicators of the equipment family.

6. The intelligent diagnostic system for power equipment according to claim 1, characterized in that, Calculate the first deviation between the real-time status data and the individual historical baseline of the target member device, specifically including: S41. Based on the historical health data of the target member equipment, establish an individual regression model between key status indicators and operating condition variables; S42. Obtain the historical residuals of the individual regression model and fit the probability distribution to obtain the individual health baseline of the key state indicators. S43. Obtain real-time status data and real-time operating condition data of key status indicators of target member equipment, calculate the predicted status value under real-time operating conditions through individual regression model, and calculate the real-time residual between real-time status data and predicted status value. S44. Based on the individual health baseline, the real-time residuals are standardized, and the absolute value of the processing result is used as the first deviation.

7. The intelligent diagnostic system for power equipment according to claim 1, characterized in that, Calculating the second deviation between real-time status data and the population health baseline, specifically including: S51. Obtain the real-time residual of at least one key status indicator of the target member device and convert it into a real-time residual vector. S52. Calculate the Mahalanobis distance between the real-time residual vector and the mean vector of the population health baseline, as the second deviation.

8. The intelligent diagnostic system for power equipment according to claim 1, characterized in that, The generation of diagnostic conclusions specifically includes: If the first deviation meets the preset condition but the second deviation does not meet the preset condition, a first diagnostic conclusion indicating that the target member device has an individual-specific abnormality is generated. If the second deviation meets the preset conditions while the first deviation does not meet the preset conditions, a second diagnostic conclusion is generated indicating that the target member device and the device family are not in the same operating state. If both the first deviation and the second deviation meet the preset conditions, a third diagnostic conclusion is generated indicating that the target member device has a serious abnormality.

9. The intelligent diagnostic system for power equipment according to claim 1, characterized in that, The diagnostic conclusions are analyzed, specifically including: S61. Perform structured parsing on each diagnostic conclusion to extract key information including at least device identifier, anomaly type, deviation severity, and timestamp. Based on the extracted key information, classify multiple member devices into different anomaly type groups. S62. For the same abnormal type group, determine whether the number of member devices exceeds the first threshold, determine whether the timestamps of the diagnostic conclusions of the member devices are concentrated in the second preset time window, and determine whether the average deviation severity of the member devices is lower than the third threshold. S63. If all the above conditions are met, it is determined that multiple member devices in the device family have a synchronization characteristic deviation mode.

Images