Equipment edge intelligent early warning method and system based on multi-dimensional data association

Abstract

The invention provides an equipment edge intelligent early warning method and system based on multi-dimensional data association, and the method comprises the steps: obtaining a current actual health index parameter of a health index of target equipment, and obtaining a current working condition index parameter of at least one working condition index associated with the health index, the health index is used for evaluating the health condition of the target equipment, and the working condition index is an index capable of influencing the parameter size of the health index; inputting each working condition index parameter into a pre-trained health index parameter prediction model, and outputting a reference health index parameter by the health index parameter prediction model; determining a target health evaluation parameter of the target device based on a difference value between the actual health index parameter and the reference health index parameter; and if the target health evaluation parameter is greater than a preset threshold, generating first abnormal early warning information, and sending the first abnormal early warning information to operation and maintenance personnel. According to the technical scheme, the accuracy of performing abnormal early warning on the transformer substation can be improved.

Description

Technical Field

[0001] The present application relates to the field of device edge intelligent early warning technology, and specifically, to a device edge intelligent early warning method and system based on multi-dimensional data association. Background Art

[0002] Substations are critical hubs in the power system. The safe operation of their equipment (such as transformers and switchgear) relies on online monitoring systems, with temperature being a key monitoring indicator. Traditional temperature-based early warning systems rely on static thresholds (e.g., setting a fixed temperature threshold of 80°C and issuing an alert if it exceeds this threshold). While simple, this early warning approach has significant drawbacks. First, it fails to dynamically reflect the impact of actual equipment operating conditions (such as load and ambient temperature) on actual equipment temperature. Consequently, normal temperature rises under high loads frequently trigger alarms, increasing the O&M burden. Second, it is insensitive to early, slow-moving faults (such as decreased cooling system efficiency). Consequently, these faults are missed because the temperature does not reach the threshold, delaying troubleshooting. These issues make static threshold methods difficult to meet the requirements of lean O&M and preventive maintenance in modern substations. Summary of the Invention

[0003] The embodiments of the present application provide a device edge intelligent early warning method and system based on multi-dimensional data association. The technical solution provided by the present application can overcome the inherent limitations of the traditional static threshold early warning method, and can more accurately adapt to the complex and changeable operating conditions of substation equipment, thereby significantly improving the accuracy and predictability of abnormal early warning for substation equipment.

[0004] Other features and advantages of the present application will become apparent from the following detailed description, or may be learned in part by practice of the present application.

[0005] According to a first aspect of an embodiment of the present application, a device edge intelligent early warning method based on multidimensional data association is provided, which is applied to a substation in which a target device is installed. The method includes: obtaining the current actual health indicator parameters of the health indicators of the target device, and obtaining the current operating condition indicator parameters of at least one operating condition indicator associated with the health indicator, where the health indicator is used to evaluate the health status of the target device, and the operating condition indicator is an indicator that can affect the parameter size of the health indicator; inputting each operating condition indicator parameter into a pre-trained health indicator parameter prediction model, so that the health indicator parameter prediction model outputs a reference health indicator parameter; determining the target health evaluation parameter of the target device based on the difference between the actual health indicator parameter and the reference health indicator parameter; if the target health evaluation parameter is greater than a preset threshold, generating a first abnormal warning message, and sending the first abnormal warning message to the operation and maintenance personnel.

[0006] In some embodiments of the present application, based on the aforementioned solution, auxiliary equipment is further installed in the substation, and the auxiliary equipment is equipment that can affect the health status of the target equipment.

[0007] In some embodiments of the present application, based on the aforementioned scheme, the target health evaluation parameters of the target device are determined based on the difference between the actual health indicator parameters and the reference health indicator parameters, including: taking the difference between the actual health indicator parameters and the reference health indicator parameters as the initial health evaluation parameters; obtaining each first operation instruction received by the target device and / or auxiliary device within a first preset time period, and the first preset time period ends at the current moment; determining whether each first operation instruction contains a second operation instruction, and the second operation instruction is an operation instruction that can increase the health indicator parameters of the target device at the current moment; if so, adjusting the initial health evaluation parameters to obtain the target health evaluation parameters, and the target health evaluation parameters are less than the initial health evaluation parameters.

[0008] In some embodiments of the present application, based on the above-mentioned scheme, a first abnormal warning information is generated, including: obtaining a first energy consumption parameter of the auxiliary equipment within a second preset time period; if the first energy consumption parameter is less than the first energy consumption threshold, the first abnormal cause, the actual health index parameter, each operating condition index parameter, the reference health index parameter, the abnormal type corresponding to the health index, and the target health evaluation parameter are determined as the first abnormal warning information, and the first abnormal cause includes that the auxiliary equipment is not turned on or there is a fault in the control loop of the auxiliary equipment; if the first energy consumption parameter is greater than the second energy consumption threshold, the second abnormal cause, the actual health index parameter, each operating condition index parameter, the reference health index parameter, the abnormal type corresponding to the health index, and the target health evaluation parameter are determined as the first abnormal warning information, and the second abnormal cause includes that the working efficiency of the auxiliary equipment is reduced or there is an abnormality in the target equipment, and the second energy consumption threshold is greater than the first energy consumption threshold.

[0009] In some embodiments of the present application, based on the aforementioned scheme, the method also includes: obtaining a third energy consumption parameter and a target benefit parameter of the auxiliary equipment within a third preset time period, the target benefit parameter being used to characterize the degree to which the auxiliary equipment suppresses the deterioration of the health indicator parameters of the target equipment; if the ratio of the third energy consumption parameter to the target benefit parameter is greater than a preset energy efficiency ratio threshold, a second abnormal warning information is sent to the operation and maintenance personnel, and the second abnormal warning information includes that the auxiliary equipment has a decline in working efficiency.

[0010] In some embodiments of the present application, based on the aforementioned scheme, the method also includes: obtaining a first change acceleration of the energy efficiency ratio of the auxiliary equipment at the current moment, where the energy efficiency ratio is the ratio of the energy consumption parameter to the benefit parameter of the auxiliary equipment; determining whether the first change acceleration meets a preset condition; if so, generating a third abnormal warning information, and sending the third abnormal warning information to the operation and maintenance personnel, the third abnormal warning information including that the health status of the auxiliary equipment is at risk of deterioration.

[0011] In some embodiments of the present application, based on the aforementioned scheme, the first change acceleration of the energy efficiency ratio of the auxiliary equipment at the current moment is obtained, including: obtaining the target energy efficiency ratio of the auxiliary equipment at each moment within a fourth preset time period, the fourth preset time period ending at the current moment; based on each target energy efficiency ratio, fitting the correspondence between the energy efficiency ratio and time into a quadratic polynomial function through the least squares method, and taking twice the value of the quadratic term coefficient of the quadratic polynomial function as the first change acceleration of the energy efficiency ratio of the auxiliary equipment at the current moment.

[0012] In some embodiments of the present application, based on the aforementioned scheme, determining whether the first change acceleration meets the preset conditions includes: if the first change acceleration is greater than the preset acceleration threshold, determining that the first change acceleration meets the preset conditions; if the first change acceleration is in the preset acceleration range, and the energy efficiency ratio of the auxiliary equipment is in the preset acceleration range at each moment within the fifth preset time period, determining that the first change acceleration meets the preset conditions, the maximum endpoint value of the preset acceleration range is the preset acceleration threshold, and the fifth preset time period ends at the current moment.

[0013] In some embodiments of the present application, based on the aforementioned solution, generating the third abnormal warning information includes: calculating the predicted failure time of the auxiliary device by the following formula, and determining the predicted failure time as the third abnormal warning information;

[0014] in, Indicates the energy efficiency ratio failure threshold; Indicates the energy efficiency ratio of the auxiliary equipment at the current moment; Indicates the rate of change of the energy efficiency ratio of the auxiliary equipment at the current moment; represents the first change acceleration; t represents the predicted failure time.

[0015] In some embodiments of the present application, based on the aforementioned scheme, the auxiliary device includes multiple sub-devices, and generates a third abnormal warning information, including: obtaining the third change acceleration of the energy efficiency ratio of each sub-device at the current moment; taking the sub-device corresponding to the maximum value of each third change acceleration as the target sub-device; and determining that there is a fault in the target sub-device as the third abnormal warning information.

[0016] According to the second aspect of the embodiment of the present application, a device edge intelligent early warning system based on multi-dimensional data association is provided. The system is applied to a substation, and a target device is installed in the substation. The system includes an edge computing device deployed in the substation, and the edge computing device includes: an acquisition unit, which is used to obtain the current actual health indicator parameters of the health indicator of the target device, and obtain the current operating condition indicator parameters of at least one operating condition indicator associated with the health indicator. The health indicator is used to evaluate the health status of the target device, and the operating condition indicator is an indicator that can affect the parameter size of the health indicator; an output unit, which is used to input each operating condition indicator parameter into a pre-trained health indicator parameter prediction model, so that the health indicator parameter prediction model outputs a reference health indicator parameter; a determination unit, which is used to determine the target health evaluation parameter of the target device based on the difference between the actual health indicator parameter and the reference health indicator parameter; an early warning unit, which is used to generate a first abnormal early warning information if the target health evaluation parameter is greater than a preset threshold, and send the first abnormal early warning information to the operation and maintenance personnel.

[0017] The technical solution of the present application includes the following technical means for abnormal warning of target equipment in a substation: first, obtaining the current actual health indicator parameters of the health indicators of the target equipment, and obtaining the current operating condition indicator parameters of at least one operating condition indicator associated with the health indicator. The health indicator is used to evaluate the health status of the target equipment, and the operating condition indicator is an indicator that can affect the parameter size of the health indicator; secondly, inputting each operating condition indicator parameter into a pre-trained health indicator parameter prediction model, so that the health indicator parameter prediction model outputs a reference health indicator parameter; thirdly, determining the target health evaluation parameter of the target equipment based on the difference between the actual health indicator parameter and the reference health indicator parameter; finally, if the target health evaluation parameter is greater than a preset threshold, generating a first abnormal warning message, and sending the first abnormal warning message to the operation and maintenance personnel.

[0018] In this application, the expected normal value of the target device health indicator (i.e., the reference health indicator parameter) is generated by a pre-trained health indicator parameter prediction model based on the collected current operating condition indicator parameters (such as load current value, ambient temperature value, etc.), rather than a static threshold. This enables the abnormal warning threshold to be adaptively adjusted as the operating condition indicator parameters associated with the target device change, thereby bringing at least two beneficial effects: First, when the target equipment operates normally under harsh operating conditions (such as high load current or harsh ambient temperature), its actual health indicator parameters (such as the top oil temperature) may rise to a higher level, but the reference health indicator parameters calculated by the health indicator parameter prediction model will also increase accordingly, resulting in a very small deviation between the two that will not exceed the preset threshold, thereby effectively avoiding false alarms caused by fluctuations in normal operating conditions, reducing the burden on operation and maintenance personnel, and preventing the credibility of the abnormal warning system from declining.

[0019] Secondly, when the target device operates under relaxed operating conditions (such as low load current or a suitable ambient temperature), if its actual health indicator parameter experiences an abnormally slow climb due to an early potential failure of its auxiliary equipment (such as a decrease in cooler cooling efficiency), even if the actual health indicator parameter does not exceed the traditional static threshold, the reference health indicator parameter calculated by the health indicator parameter prediction model remains at a low level. This small actual health indicator parameter can lead to a significant deviation, causing it to quickly exceed the preset threshold and generate a first abnormal warning message. This enables the present method to capture early fault signs that traditional methods cannot identify, achieving a transition from post-warning to pre-warning, and saving valuable time for fault resolution. In summary, by constructing a health indicator parameter prediction model based on "actual health indicator parameter - dynamic reference health indicator parameter", the present invention fundamentally solves the problem of static threshold "inaccuracy" under complex substation operating conditions, achieving a dual improvement in the accuracy and timeliness of abnormal warnings.

[0020] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The accompanying drawings are incorporated into and constitute a part of the specification, illustrating embodiments consistent with the present application and, together with the specification, explaining the principles of the present application. Obviously, the drawings described below are only some embodiments of the present application, and those skilled in the art can derive other drawings based on these drawings without inventive effort. In the drawings: Figure 1 A schematic diagram of a process for a device edge intelligent early warning method based on multi-dimensional data association according to an embodiment of the present application is shown; Figure 2 A detailed flow chart of generating third abnormal warning information according to an embodiment of the present application is shown; Figure 3 A block diagram of a device edge intelligent early warning system based on multi-dimensional data association according to an embodiment of the present application is shown. DETAILED DESCRIPTION

[0022] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided so that this application will be thorough and complete and will fully convey the concepts of the example embodiments to those skilled in the art.

[0023] In addition, described feature, structure or characteristic can be combined in one or more embodiments in any suitable manner.In the following description, many specific details are provided so as to provide a full understanding of the embodiments of the present application. However, it will be appreciated by those skilled in the art that the technical scheme of the present application can be put into practice without one or more of the specific details, or other methods, components, devices, steps etc. can be adopted. In other cases, known methods, devices, implementations or operations are not shown or described in detail to avoid blurring the various aspects of the application.

[0024] The flowcharts shown in the accompanying drawings are for illustrative purposes only and do not necessarily include all contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps may be decomposed, while others may be combined or partially combined. Therefore, the actual execution order may vary depending on the actual situation.

[0025] It should be noted that the term "plurality" used in this document refers to two or more. "And / or" describes a relationship between associated objects, indicating that three possible relationships exist. For example, "A and / or B" can represent: A alone, A and B together, or B alone. The character " / " generally indicates an "or" relationship between the associated objects.

[0026] It should be noted that the terms "first," "second," and the like in the specification and claims of this application and the accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or precedence. It should be understood that the terms used in this manner are interchangeable where appropriate, such that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described.

[0027] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0028] The following will describe some embodiments of the present application in detail with reference to the accompanying drawings. In the absence of conflict, the following embodiments and features therein may be combined with each other.

[0029] In some embodiments, the technical solution of the present application is applied to a substation and executed on an edge computing device deployed in the substation.

[0030] It should be noted that a plurality of devices are installed in the substation, including a target device. The target device may be any device in the substation that needs to be monitored for abnormalities, such as a transformer, a switch cabinet, and the like.

[0031] Preferably, the target device is a transformer.

[0032] In some embodiments, auxiliary equipment is also installed in the substation, and the auxiliary equipment is equipment that can affect the health of the target equipment, such as coolers, cooling fans, variable frequency fans, submersible pumps, etc.

[0033] It should be noted that the following embodiments of the present application are all described by taking the target device as a transformer and the auxiliary device as a cooler as an example.

[0034] See also Figure 1 , shows a flow chart of a device edge intelligent early warning method based on multi-dimensional data association according to an embodiment of the present application, specifically including the following steps 110 to 140: Step 110, obtain the current actual health indicator parameters of the health indicator of the target device, and obtain the current operating condition indicator parameters of at least one operating condition indicator associated with the health indicator. The health indicator is used to evaluate the health status of the target device, and the operating condition indicator is an indicator that can affect the parameter size of the health indicator.

[0035] In this embodiment, health indicators include, but are not limited to, the top layer oil temperature, winding oil temperature, vibration signal, partial discharge, and the like of the target device. This application does not limit this in detail. It is understood that the actual health indicator parameter is the parameter value of the health indicator at the current moment. For example, if the health indicator is the top layer oil temperature, the actual health indicator parameter is the top layer oil temperature value at the current moment.

[0036] Preferably, the health indicator is the transformer's top oil temperature or winding oil temperature. The top oil temperature is a key macroscopic indicator of the transformer's overall thermal state, while the winding oil temperature more directly reflects the health of the winding itself, a core heat-generating component, and is more sensitive to abnormalities.

[0037] It should be noted that the following embodiments of this application all use the health indicator as the top oil temperature and the actual health indicator parameter as the actual top oil temperature value for illustration.

[0038] In this embodiment, operating condition indicators include, but are not limited to, ambient temperature, ambient humidity, target device load current, active power, reactive power, voltage, power factor, target device operational status, target device shutdown status, and the start / stop status of the target device's auxiliary devices. For example, if the operating condition indicator is ambient temperature, then the operating condition indicator parameter is the current ambient temperature value. It is understood that these operating condition indicators may affect the parameters of the health indicator; for example, higher ambient temperature values ​​indicate higher transformer top oil temperature values.

[0039] Preferably, the at least one operating condition indicator includes load current, active power, and ambient temperature. These operating condition indicators are selected because they are the most direct and core external and internal factors affecting the heat generation and heat dissipation balance of the transformer. Load current and active power jointly determine the heat generation of the transformer's internal windings and core, while ambient temperature directly affects the efficiency of heat dissipation to the outside world through the radiator.

[0040] Continue to see Figure 1 In step 120 , each operating condition indicator parameter is input into a pre-trained health indicator parameter prediction model, so that the health indicator parameter prediction model outputs a reference health indicator parameter.

[0041] It's understandable that this health indicator parameter prediction model isn't a simple function. Instead, it's trained using machine learning and other methods based on a large amount of historical health operating data. It internalizes the complex, nonlinear mathematical relationships between health indicator parameters and various operating condition parameters under normal conditions. Therefore, the model calculates the current operating condition parameter input to derive a dynamic reference health indicator parameter (i.e., the expected normal value of the health indicator at the current time) that closely matches the current operating conditions. This expected normal value isn't a fixed number; rather, it provides a theoretical health indicator parameter reference benchmark, derived by the health indicator parameter model based on the current target device operating scenario.

[0042] In this embodiment, the health indicator parameter prediction model can be a multiple linear regression model, a gradient boosted tree (GBT) model, a support vector machine (SVM) model, or, for models with strong time series characteristics, preferably a gradient boosted tree (GBT) model, or a recurrent neural network (RNN) or its variants such as a long short-term memory network (LSTM). The gradient boosted tree algorithm excels at processing tabular data and capturing high-order nonlinear relationships between variables, while the recurrent neural network is particularly suitable for processing data with time series characteristics and can capture the dynamic inertia of the transformer's thermal state over time.

[0043] The following uses the Gradient Boosting Tree (GBT) model as an example to explain the process of constructing a health indicator parameter prediction model. For example, the target device is a transformer, the health indicator is the top oil temperature, and the operating condition indicators include load current, active power, and ambient temperature. Specifically, the construction process includes steps 10 to 30: Step 10: Collection and processing of training data It is understandable that high-quality training data is the basis for building a high-precision prediction model. The training data collection and processing process of the present invention mainly includes the following steps: Step 11: Data source and multi-dimensional data collection Training data is derived from various automation and monitoring systems deployed within the substation, primarily including the substation supervisory control and data acquisition (SCADA / PMS), online equipment monitoring devices, and environmental monitoring units. The substation supervisory control system captures transformer electrical parameters such as load current, active power, reactive power, and voltage. The online equipment monitoring device captures transformer status parameters such as top oil temperature and winding oil temperature. These parameters serve as the model's prediction targets (i.e., health indicators). The environmental monitoring unit captures substation environmental parameters such as ambient temperature and humidity.

[0044] It should be noted that the data collection time span should be long enough to cover a variety of typical operating conditions and seasonal variations. Typically, historical operating data for the target equipment or similar equipment over the past 1 to 3 years is selected. The collection frequency is usually minute-by-minute (e.g., one data point every 5 or 15 minutes).

[0045] Step 12: Screening and cleaning of health status data Since historical operating data inevitably contains data from anomalies, failures, or maintenance periods, strict screening must be performed to ensure that the dataset used for training can truly reflect the operating rules of the equipment in a healthy state.

[0046] First, remove data from known abnormal periods: By combining the substation's operation and maintenance logs, maintenance records, and historical alarm information, all data from time periods where failures, maintenance, or continuous alarms have occurred can be removed from the training dataset.

[0047] Secondly, eliminate extreme operating condition data: eliminate extreme data points that are not within the scope of normal operation and are caused by special operations (such as short-circuit tests, large equipment switching, etc.) to prevent the model from learning parameter relationships under abnormal operating conditions.

[0048] Finally, preliminary data cleaning is performed: the filtered data is processed, including missing value handling and outlier / noise handling. Missing value handling involves filling in a small number of randomly missing data points using linear interpolation or the mean of adjacent points. Outlier / noise handling involves using statistical methods (such as the 3-sigma principle or the Z-score) to identify and remove isolated data points that significantly deviate from the normal fluctuation range. Filtering methods such as moving averages can also be used to smooth the data and reduce sensor noise interference.

[0049] It can be understood that after the above steps, a high-purity multi-dimensional historical data set that can represent the healthy operating status of the transformer is finally obtained.

[0050] Step 13: Data preprocessing and time alignment It is understandable that for time alignment, since the data collection frequencies of different systems may vary, it is necessary to resample the parameters of all dimensions (load current, ambient temperature, top oil temperature, etc.) to a unified time base (for example, a unified 15-minute interval) to ensure that all parameters in each data point correspond to the status at the same time.

[0051] After time alignment, feature normalization can also be performed. To eliminate the impact of different physical dimensions on model training, all numerical operating condition indicator parameters need to be normalized (for example, using Min-Max Scaling to scale them to the [0, 1] range). This helps accelerate model convergence and improve prediction accuracy.

[0052] The above is an introduction to the process of collecting and processing training data. After collecting and processing the training data, you need to continue to perform the following steps 20: Step 20: Prediction model training process After obtaining a high-quality training data set, model training can begin. The present invention preferably uses a gradient boosted tree (GBT) model or a long short-term memory network (LSTM) model. The following uses the GBT model as an example to illustrate the training process: Step 21: Feature Engineering and Dataset Partitioning Determine features and labels: Use the processed operating condition indicator parameters (such as load current, active power, and ambient temperature) as the model's input features. Use the corresponding health indicator parameters (such as top oil temperature) as the model's predicted labels.

[0053] Divide the dataset: Divide the prepared dataset into three parts in chronological order or randomly: Training set (about 70%): used to train the model and learn the intrinsic relationship between features and labels.

[0054] Validation set (about 15%): used to adjust the model's hyperparameters during training and monitor whether the model is overfitting.

[0055] Test set (approximately 15%): Used to evaluate the final generalization ability and prediction accuracy of the model after training. This data does not participate in any training process.

[0056] Step 22: Model training and hyperparameter tuning Model training: The training set is fed into the GBT algorithm. The model iteratively constructs a series of decision trees, with each new tree aiming to correct the residuals (prediction errors) of the previous trees. The training goal is to minimize a loss function (such as mean squared error (MSE)). In other words, the overall difference between the model's predicted top oil temperature and the actual top oil temperature in the training set is minimized.

[0057] Hyperparameter Tuning: GBT models contain multiple hyperparameters that affect performance (such as the number of trees, learning rate, and maximum tree depth). Using methods such as grid search or random search, perform combination experiments within the preset hyperparameter range, using validation set performance (such as MSE on the validation set) as the evaluation criterion to find the optimal hyperparameter combination.

[0058] Step 23: Model evaluation and finalization Final evaluation: Use the optimal hyperparameters found on the validation set to retrain the model on the full training set to obtain the final model. Then, input the unseen test set into the final model for prediction.

[0059] Performance metrics: The performance of the model is evaluated by calculating the difference between the predicted value and the true value on the test set. Key evaluation metrics include: Mean absolute error (MAE): The average absolute value of the prediction error, in °C, which directly reflects the average prediction accuracy.

[0060] Root mean square error (RMSE): gives higher weight to larger errors and reflects the stability of the prediction.

[0061] Coefficient of determination ( ): The range is between 0 and 1. The closer it is to 1, the stronger the model's ability to explain data changes.

[0062] Model confirmation: Only when the performance indicators of the final model on the test set meet the preset engineering requirements (for example, MAE is less than 1℃, If the training probability is greater than 0.95), the model is considered pre-trained and can be solidified and deployed to edge computing devices.

[0063] Step 24: Model deployment The health indicator parameter prediction model trained and verified through the above process is deployed to the substation's edge computing device. During actual operation, this model receives real-time operating condition indicator parameters and outputs accurate reference health indicator parameters, providing the foundation for subsequent intelligent early warning logic.

[0064] In some embodiments, a gradient boosting tree (GBT) model for constructing a health indicator parameter prediction model mainly includes the following structures: an input layer, a feature engineering module, a GBT model core, and an output layer.

[0065] Input layer: Function: Receive multi-dimensional raw data collected by the data acquisition module and time-aligned.

[0066] Parameter example: Historical values ​​of main monitoring parameters, actual value of top oil temperature at the previous moment

[0067] Reason for setting: Transformer oil temperature has huge thermal inertia, and the current temperature is highly correlated with the temperature at the previous moment. Introducing historical values ​​can help the model capture this temporal dynamic characteristic.

[0068] Core operating parameters: load current , active power , ambient temperature

[0069] Auxiliary working parameters: reactive power , power factor , cooler operating status (For example, 0 means stop, 1 means run) Data source: Raw, multi-dimensional real-time data flows into the next module.

[0070] Feature Engineering Module: Function: Preprocess and deeply process the raw input data to extract features that can more effectively reflect physical phenomena, thereby enhancing the learning ability of the model.

[0071] Processing steps and parameter examples: A. Data cleaning: Interpolate missing values ​​(such as using the mean before and after values ​​or empirical values ​​under specific working conditions).

[0072] B. Feature Derivation: Load change rate: , Setting reason: The drastic change of load can reveal the rapid change trend of temperature better than the absolute value of load.

[0073] Comprehensive product indicators: , setting reason: The heat generated by the transformer mainly comes from the copper loss of the winding (proportional to the square of the current) and the iron loss of the core (related to voltage / power). This composite feature directly simulates the total heat generation; and It is an empirical coefficient determined according to the transformer design parameters.

[0074] Temperature difference between inside and outside: , Setting reason: This temperature difference is the main driving force of the heat dissipation process and directly affects the heat dissipation efficiency.

[0075] C. Category feature coding: For "Cooler operating status" Non-numeric parameters such as are one-hot encoded to convert them into binary vectors that the model can process.

[0076] Data flow: processed and derived feature vectors: is transported to the model core.

[0077] GBT model core: Function: Receive feature vector , calculate through the integration of multiple decision trees and output the prediction results.

[0078] Data flow: predicted top oil temperature expected normal value .

[0079] Output layer: Function: Output the final expected normal value of the top oil temperature .

[0080] Data flow: Provided to the deviation analysis module for comparison with the actual collected value Make a comparison.

[0081] It should be noted that the GBT model is an ensemble learning model, and its performance is highly dependent on the settings of a series of hyperparameters. In the application scenario of this application, the typical settings of these parameters and the reasons are as follows: n_estimators (number of decision trees): Typical values: 100-500.

[0082] Reason for setting: The number of decision trees determines the complexity and fitting ability of the model. Too few may lead to underfitting and fail to fully learn the complex relationships in the data. Too many increase computational overhead and have a slight risk of overfitting. Cross-validation is used to select a number of trees that performs best on the validation set, typically in the hundreds.

[0083] learning_rate: Typical values: 0.01 -0.1.

[0084] Reason for setting this: The learner is used to reduce the contribution of each tree to prevent the model from overfitting the training data. A smaller learning rate requires more trees (n_estimators) to achieve the same fit, but generally results in a model with stronger generalization capabilities. This is crucial for noisy industrial data scenarios, as it effectively prevents the model from being biased by transient noise.

[0085] max_depth (maximum depth of decision tree): Typical value: 3-8.

[0086] Reason for setting this: To control the complexity of a single decision tree. Shallower trees (e.g., depth 3-5) learn simpler and more robust feature combinations, effectively preventing overfitting. In this scenario, while the interactions between operating parameters are complex, they generally don't require an overfitted tree to capture them. Setting a shallower depth helps improve the model's generalization performance.

[0087] subsample (subsampling ratio): Typical value: 0.7-1.0.

[0088] Reason for setting this parameter: This is a key parameter in stochastic gradient boosting. Each time a new tree is constructed, only a portion (e.g., 80%) of the training samples are randomly sampled for training. This "sampling with replacement" approach introduces randomness into the model, making it less susceptible to the influence of individual outliers, thereby significantly improving the model's robustness and generalization capabilities.

[0089] It should be noted that the GBT model is constructed incrementally using an additive model and a forward step-by-step algorithm to express the complex, nonlinear relationship between the health indicator parameter (oil temperature T) and the operating condition indicator parameter (eigenvector X). Its core mathematical concept can be expressed as follows: Initialize the model: The model starts with a simple prediction, usually all training samples measured Average value ; represents the average value of all training sample measurements.

[0090] Iteratively build decision trees : Calculate the residual: For the mth step, first calculate the current model The predicted value and the true value T The error between , that is, the residual: (For each training sample i) This residual is mathematically a direct loss function (such as mean square error )’s negative gradient; F represents the current model’s Predicted value.

[0091] Train a new tree: Next, train a new decision tree , but the training goal of this tree is not to directly predict the oil temperature T, but to fit the residual of the previous round This means that each new tree is learning how to correct the accumulated prediction errors of all the previous trees.

[0092] Update the model: add the newly trained tree At a certain learning rate Add to the overall model:

[0093] How to express nonlinear relationships: The expressive power of nonlinear relationships primarily comes from the decision tree itself, which serves as the base learner. Each decision tree divides the feature space into multiple rectangular regions through a series of "if-then" rules and assigns a predicted value to each region. For example, a tree might learn the following rules: Prediction divides the feature space into multiple rectangular regions and assigns a prediction value to each region. For example, a tree may learn the following rules: IF(load current>400A AND ambient temperature>30℃) THEN the predicted residual is +5℃ IF(Load Current < 200A AND Cooler Status = Running) THEN Prediction Residual is -2°C When hundreds or even thousands of these decision trees are combined through an additive model, they collectively divide the feature space into extremely fine and complex regions. The final prediction value is the weighted sum of the predictions from all the trees. This structure can approximate arbitrarily complex nonlinear functions with extremely high accuracy, accurately capturing the highly coupled, nonlinear dynamic relationships between transformer oil temperature and multiple operating parameters such as load, environment, and cooling status.

[0094] The following is a detailed description of the algorithm implementation process of the health indicator parameter prediction model: 1. Gradient Boosting Tree (GBT) Algorithm Core principle of the algorithm: The gradient boosting tree algorithm constructs an ensemble model of multiple weak decision tree learners. Each new tree specifically fits the residuals of all previous trees, thereby gradually approaching the expected value of the true health indicator parameter. The overall function form of the prediction model is expressed as follows:

[0095] Detailed description of each component : The final health indicator parameter prediction value (such as the expected top oil temperature value) represents the prediction result output after model calculation, and the value range is usually [20℃, 120℃].

[0096] : The initial prediction value is generally set to the mean of the health indicator parameters in the training set to provide an initial prediction benchmark for the model. The value range is mostly [40℃, 80℃].

[0097] M: The total number of decision trees. It is flexibly set based on the model's requirements for complexity and accuracy. The common value range is [50, 500].

[0098] : The weight coefficient of the mth tree, used to adjust the influence of a single tree in the overall prediction, determined by optimization methods such as line search, and the value range is generally [0.01, 1.0].

[0099] : The output of the mth decision tree, reflecting the contribution of a single tree to the prediction result, and the value range is usually [-20℃, +20℃].

[0100] : The input operating condition indicator parameter vector covers key data reflecting the equipment operating status, such as load current, active power, and ambient temperature.

[0101] The construction process of a single decision tree: The objective function of the mth tree

[0102] Description of each component : The objective function of the mth tree, which is used to measure the degree of fit between the current tree prediction result and the residual.

[0103] N: The total number of training samples, usually ranging from [1000, 10000], represents the size of the samples involved in the training.

[0104] : The true health indicator parameter value of the i-th sample is the reference benchmark for model prediction.

[0105] : The predicted value of the first m-1 trees for the i-th sample, reflecting the cumulative prediction effect of the previous trees.

[0106] : The prediction adjustment of the mth tree for the i-th sample, It can be understood as the weight of the tree. is the predicted output of the tree.

[0107] Node splitting criteria: The optimal splitting condition for each node is determined by minimizing the mean square error:

[0108] in: : The mean of the left and right child nodes. The value is determined by the specific data distribution and is used to measure the concentration of data within the child node after the split.

[0109] The maximum depth of each tree: the value range is [3,8], which controls the complexity of a single tree and affects the model fitting and generalization capabilities.

[0110] Minimum number of samples for leaf nodes: The value range is [10,100]. This limits the minimum number of samples a leaf node can contain to avoid overfitting and ensure the statistical significance of the node.

[0111] After this processing, the meaning of formulas and parameters is clearer, making it easier to understand and reference.

[0112] 2. Long Short-Term Memory Network (LSTM) Algorithm Mathematical expression of network structure:

[0113] Meaning and description of each parameter (Forget Gate): decides what information to discard from the cell state, is the sigmoid activation function, with an output range of [0,1], is the weight matrix, is the bias vector, It is the concatenation of the hidden state and the current input.

[0114] (Input gate): determines what new information is added to the cell state. The processing logic is the same as the forget gate. and Control updates.

[0115] (Candidate cell state): generated by the tanh activation function, used to generate candidate cell state update values, affected by and modulation.

[0116] (Cell state): Combine the results of the forget gate and the input gate to update the cell state, Represents element-wise multiplication.

[0117] (Output gate): determines what information is output from the cell state, by and Control, activated by sigmoid.

[0118] (Hidden state): Calculated based on the output gate and cell state, used to convey important information at the current moment.

[0119] (Final output): Through linear transformation and bias , mapping the hidden state to the output result, is the output layer weight matrix, is the output layer bias vector.

[0120] Supplementary explanation of parameter meaning and value range : sigmoid activation function, which compresses the input to the [0,1] interval and is used to control the switch state of the door.

[0121] : Weight matrix, the dimension is usually [number of hidden units × (number of hidden units + input dimension)], which is used to linearly transform the concatenation vector of input and hidden state.

[0122] : Bias vector, the initialization value needs to be set to 0, auxiliary linear transformation, affecting the output of the activation function.

[0123] Number of hidden units: This refers to the number of units used to store states within the LSTM. The value range is [32, 256]. The greater the number, the stronger the model's expressiveness, but it also increases the computational cost.

[0124] Input dimension: equal to the number of working condition indicator parameters, usually [3,10], determined by the number of features of the input data.

[0125] Time step: The size of the historical data window, with a value range of [12, 144], corresponding to 1-12 hours of historical data, is used to determine the length of the time series input into the model each time.

[0126] Loss function during training:

[0127] Explanation of the meaning of each part Loss: The overall loss value, which combines the prediction bias and regularization term, is used to measure the model prediction effect and parameter complexity.

[0128] : Mean square error (MSE) term, 1 / N is the average error of N samples, is the true value of the i-th sample, It is the model prediction value, which measures the deviation between the prediction result and the true value.

[0129] : L2 regularization term, is the regularization coefficient, It is the sum of the L2 norm of the weight parameter W in the model, which is used to avoid overfitting of the model.

[0130] N: The total number of training samples, representing the sample size involved in loss calculation.

[0131] : L2 regularization coefficient, the value range is usually [0.0001, 0.01], which is used to balance the influence of mean square error and regularization term.

[0132] 3. Multi-dimensional feature coupling relationship modeling Nonlinear coupling function: Taking into account the physical mechanism of transformer thermal characteristics, the coupling relationship between operating condition index parameters can be expressed as:

[0133] Explanation of the meaning of each part: : Expected health index parameter (top oil temperature here), that is, the expected value of the transformer top oil temperature calculated by the model.

[0134] : Basic mapping function, used to characterize load (Load), active power , ambient temperature The direct impact of major operating condition index parameters such as oil temperature on the top layer oil temperature reflects the basic thermal characteristic relationship.

[0135] : Interaction function, which describes the coupling effect between different operating condition index parameters (such as load and ambient temperature, active power and ambient temperature), and reflects the influence of the interaction between parameters on oil temperature.

[0136] : Dynamic term function, taking into account the rate of change of ambient temperature , load change rate The dynamic influence of the changing rate of parameters on the oil temperature reflects the dynamic thermal response characteristics of the system.

[0137] Feature normalization processing:

[0138] Explanation of the meaning of each part : The normalized eigenvalues ​​map the original features to the interval [0,1] to facilitate unified model processing.

[0139] : Original characteristic values ​​(such as load current, active power, ambient temperature, etc.).

[0140] : The minimum value of this feature in the training data, used to determine the lower limit of the normalization benchmark.

[0141] : The maximum value of this feature in the training data, used to determine the normalized benchmark upper limit.

[0142] Typical parameter normalization range description Load current: The normalized range is [0, 1], corresponding to the original range [0A, rated current]. The actual load current value is converted according to the above formula to eliminate the dimension effect and make the model easier to learn.

[0143] Active power: The normalized range is [0, 1], corresponding to the original range [0MW, rated power]. The active power value is mapped to a uniform range through the same normalization calculation.

[0144] Ambient temperature: The normalized range is [0, 1], corresponding to the original range [-20°C, 50°C]. This allows temperature features of different magnitudes to be used in model training in conjunction with other features.

[0145] The above algorithm implementation fully considers the physical characteristics of substation equipment operation and actual engineering constraints, and accurately describes the complex nonlinear coupling relationship between health indicator parameters and operating condition indicator parameters through mathematical modeling.

[0146] To validate the health indicator parameter prediction model proposed in this application, we conducted an 18-month field trial at a 220kV substation equipped with three main transformers, each with a capacity of 180MVA, and a complete online monitoring system and cooling device.

[0147] The experimental dataset contains 72,000 valid data samples, of which 70% (50,400 samples) are used as a training set, 15% (10,800 samples) as a validation set, and 15% (10,800 samples) as a test set. Each sample contains the actual measured value of the top oil temperature and corresponding operating parameters such as load current, active power, and ambient temperature.

[0148] Model comparison experiment We compared the health indicator parameter prediction model proposed in this application with three traditional methods: Comparison Method 1: The static threshold method uses a traditional approach with a fixed alarm threshold of 80°C. During the 18-month testing period, this method generated 1,247 alarms, of which 856 were false alarms, a false alarm rate of 68.6%. Furthermore, this method missed 23 true early failure signs, a false alarm rate of 15.3%.

[0149] Comparison method 2: Simple linear regression model. This model uses multiple linear regression to establish the relationship between oil temperature and operating parameters. This model achieves a mean absolute error (MAE) of 3.2°C and a root mean square error (RMSE) of 4.8°C. In the anomaly detection task, it achieves an accuracy of 73.4% and a recall of 68.9%.

[0150] Comparison method three: Traditional neural network model, using a three-layer feedforward neural network. This model achieved a MAE of 2.1°C, an RMSE of 3.3°C, an anomaly detection accuracy of 79.2%, and a recall of 75.6%.

[0151] This application proposes an improved gradient boosting tree model, which significantly outperforms traditional methods in terms of prediction accuracy. Specific improvements include the introduction of time window feature extraction, multi-scale feature fusion, and an adaptive weight adjustment mechanism.

[0152] Quantitative performance indicators: Through rigorous statistical analysis, this application method has achieved significant improvements in several key indicators: In terms of prediction accuracy, the improved gradient boosting tree model reduced the MAE to 1.3°C, a 38.1% improvement over traditional neural networks; the RMSE was reduced to 2.1°C, a 36.4% improvement. The correlation coefficient between the predicted and actual values ​​reached 0.94, significantly higher than the 0.76 of linear regression and the 0.87 of traditional neural networks.

[0153] Anomaly detection performance: In the fault warning task, the accuracy of this application method reached 92.8%, an increase of 13.6 percentage points compared with traditional neural networks; the recall rate reached 89.4%, an increase of 13.8 percentage points; and the F1 score reached 91.1%, which has a significant advantage over traditional methods.

[0154] False Alarm Rate Control: Most importantly, this method reduces the false alarm rate to 8.2%, a significant improvement over the 68.6% rate achieved by the static threshold method. This means that operations personnel only need to deal with an average of 3-4 false alarms per month, rather than the previous 40-50.

[0155] Practical application effect verification: We also conducted long-term practical application effect tracking: Early fault detection: During an 18-month monitoring period, the proposed method successfully detected eight cooling system efficiency degradation events, with an average lead time of 72 hours. Six of these events were promptly addressed through preventive maintenance, avoiding unplanned equipment downtime.

[0156] Verification of working condition adaptability: Through the analysis of prediction errors under different load conditions, it was found that the traditional method had a significant increase in error under high load conditions, and the standard deviation of the prediction error reached 5.2°C. However, the method of this application maintained stable prediction accuracy under various working conditions, and the standard deviation of the error was only 1.8°C.

[0157] Economic benefit analysis: Through cost-benefit analysis, the economic value brought by this application method is significant: the cost of false alarm processing is reduced by about 400,000 yuan / year, and the power outage losses avoided by early fault detection are about 1.5 million yuan / year, and the overall return on investment reaches 3.8:1.

[0158] These experimental data fully demonstrate the significant advantages of the health indicator parameter prediction model of this application in terms of accuracy, reliability and practicality, and provide strong technical support for the intelligent early warning of substation equipment.

[0159] The specific comparative data is as follows: Health indicator parameter prediction model experimental comparison data table: Table 1: Comparative analysis of prediction accuracy of each method

[0160] Table 2: Comparative analysis of anomaly detection performance

[0161] Table 3: Prediction error analysis under different working conditions

[0162] Table 4: Comparative analysis of fault warning timeliness

[0163] Table 5: Statistics of actual operation results in 18 months

[0164] Table 6: Applicability verification of different transformer capacity levels

[0165] Note: pp stands for percentage points; ↑ indicates performance improvement, ↓ indicates problem resolution; error data is formatted as "mean ± standard deviation"; all data are based on actual operating data from 18 months of continuous monitoring.

[0166] In summary, by employing the aforementioned health indicator parameter prediction model, this application is able to establish a mathematical representation of the thermal state that closely approximates the transformer's actual physical processes. Its technical benefit lies in the fact that, because the health indicator parameter prediction model inputs the most critical operating condition indicator parameters and utilizes an algorithm that can deeply reveal their inherent coupling relationships, the accuracy of the calculated reference health indicator parameters (such as the expected top oil temperature) can be significantly improved. A more accurate reference health indicator parameter forms the basis for subsequent deviation analysis, enabling the entire early warning method to more sensitively and reliably identify subtle anomalies that deviate from the normal operating trajectory, fundamentally improving the accuracy and reliability of early warning decisions.

[0167] Continue to see Figure 1 , step 130, determining a target health evaluation parameter of the target device based on the difference between the actual health indicator parameter and the reference health indicator parameter.

[0168] The specific implementation methods of this embodiment include at least the following two: In a first implementation of step 130 , the difference between the actual health indicator parameter and the reference health indicator parameter is used as the target health evaluation parameter of the target device.

[0169] The second implementation of step 130 is performed according to steps 131 to 134 as follows: Step 131 : The difference between the actual health index parameter and the reference health index parameter is used as the initial health evaluation parameter.

[0170] Step 132: Acquire each first operation instruction received by the target device and / or the auxiliary device within a first preset time period, where the first preset time period ends at the current time.

[0171] Step 133 : Determine whether each first operation instruction includes a second operation instruction, where the second operation instruction is an operation instruction that can increase the health indicator parameter of the target device at the current moment.

[0172] Step 134: If yes, adjust the initial health evaluation parameter to obtain the target health evaluation parameter, and the target health evaluation parameter is smaller than the initial health evaluation parameter.

[0173] In this embodiment, the first preset time period may end at the current time.

[0174] In this embodiment, the first operation instruction may be a switching operation, a closing operation, an opening operation, a tap adjustment, starting forced oil circulation air cooling, stopping a cooler, starting a cooler, and the like.

[0175] It is understood that among the first operation instructions, some may cause the health indicator parameters of the target device to increase, some may cause the health indicator parameters of the target device to decrease, and some may not cause the health indicator parameters of the target device to change. Furthermore, the first operation instructions are associated with information such as the effective time and the corresponding executing device. Therefore, by combining the instruction information of the first operation instructions, it is possible to determine from each first operation instruction whether there is an operation instruction that can increase the health indicator parameters of the target device at the current moment.

[0176] For example, if a switching operation instruction is detected in the first operation instruction, the operation instruction may cause the load current of the transformer to increase from 0% to 70% in a short period of time. Such a drastic change in load current will inevitably cause the top oil temperature to fluctuate rapidly and significantly, which is a normal operation. In order to avoid misjudging this normal transient process as an abnormality, the switching operation is used as the second operation instruction, and the initial health evaluation parameters are adjusted. For example, the initial health evaluation parameters can be multiplied by the suppression factor. , obtaining the target health assessment parameters. This approach can “accommodate” expected and drastic parameter fluctuations caused by normal operations, effectively preventing false alarms.

[0177] In some embodiments of the present application, if the first operation instruction is obtained to include a second operation instruction, the size of the preset threshold for determining the target health evaluation parameter mentioned later may also be adjusted, such as increasing the preset threshold.

[0178] In this embodiment, dramatic but predictable transient fluctuations in parameters caused by normal operations are actively identified and accommodated. This enables the edge computing device to effectively distinguish between real target equipment anomalies and parameter disturbances during normal operation, thereby greatly reducing the false alarm rate. Under traditional early warning logic, the sudden increase in load current and the rapid change in top oil temperature caused by the switching operation can easily be misjudged as a fault. This application intelligently adjusts its own sensitivity by "understanding" the current operation. This not only reduces the burden on operation and maintenance personnel who are disturbed by invalid warning information, but also enhances their trust in the entire abnormal warning system, ensuring that operation and maintenance resources can be focused on dealing with real equipment hazards.

[0179] Continue to see Figure 1 , step 140, if the target health evaluation parameter is greater than the preset threshold, a first abnormal warning information is generated and sent to the operation and maintenance personnel.

[0180] In this embodiment, there are at least two implementation methods for generating the first abnormality warning information: In a first implementation of step 140 , the actual health index parameter, each operating condition index parameter, the reference health index parameter, the abnormality type corresponding to the health index, and the target health evaluation parameter are determined as the first abnormality warning information.

[0181] Among them, the abnormal type corresponding to the health indicator is a qualitative description of the warning event, such as abnormally high top oil temperature.

[0182] Among them, the first abnormal warning information includes actual health indicator parameters, which can enable operation and maintenance personnel to understand the current absolute status of the target equipment, including reference health indicator evaluation parameters, which are the key basis for helping operation and maintenance personnel to judge abnormalities, including working condition indicator parameters corresponding to the target health evaluation parameters, such as the current value of the load current at that time, the temperature value of the ambient temperature, etc. Traditional warnings only tell "what happened", while the first abnormal warning information provided by this application also explains "under what working conditions it occurred" and "how much it deviated from expectations". The technical effect is that the enriched first abnormal warning information provides operation and maintenance personnel with a complete event snapshot and rich context, enabling them to quickly conduct preliminary research and judgment on the urgency and possible causes of the warning.

[0183] For example, a small deviation under high load current may not be as urgent as a large deviation under low load current. This information-rich first-error warning greatly enhances the transparency and interpretability of warnings, freeing operations and maintenance personnel from tedious data queries and providing them with direct, diagnostic-based decision-making, significantly shortening fault response and troubleshooting time.

[0184] In a second implementation of step 140 , first abnormality warning information is generated according to the following steps 141 to 143 .

[0185] Step 141: Acquire a first energy consumption parameter of the auxiliary device within a second preset time period.

[0186] In step 142, if the first energy consumption parameter is less than the first energy consumption threshold, the first abnormal cause, the actual health index parameter, the various operating condition index parameters, the abnormal type corresponding to the health index, and the target health evaluation parameter are determined as the first abnormal warning information. The first abnormal cause includes that the auxiliary equipment is not turned on or there is a fault in the control loop of the auxiliary equipment.

[0187] Step 143: If the first energy consumption parameter is greater than the second energy consumption threshold, the second abnormal cause, the actual health index parameter, each operating condition index parameter, the abnormal type corresponding to the health index, and the target health evaluation parameter are determined as the first abnormal warning information. The second abnormal cause includes a decrease in the working efficiency of the auxiliary equipment or an abnormality in the target equipment, and the second energy consumption threshold is greater than the first energy consumption threshold.

[0188] In this embodiment, the second preset time period may end at the current time.

[0189] In this embodiment, the first energy consumption parameter may be the electric energy consumed by the auxiliary device within the second preset time period.

[0190] In this embodiment, the first energy consumption threshold may be a rated operating value of the auxiliary device in a normal operating state.

[0191] For example, if analysis reveals that the actual top oil temperature is greater than a preset threshold, and the cooler's first energy consumption data is lower than the first energy consumption threshold, for example, the first energy consumption parameter is close to 0, this constitutes a strong chain of evidence indicating that the abnormally high actual top oil temperature is due to the cooler, which is supposed to provide cooling, not functioning as expected. Therefore, the first abnormality warning message includes the first abnormality cause (the cooler is not started or there is a control circuit fault), prompting maintenance personnel to check the cooler's power supply, control relay, and related wiring.

[0192] For example, if the analysis finds that the actual top oil temperature is greater than the preset threshold, and the first energy consumption parameter of the cooler is greater than the second energy consumption threshold, for example, it has reached or is close to its rated value when operating at full load. This also constitutes an important chain of evidence, indicating that the cooler is indeed working "hard", but the actual top oil temperature is still out of control, which reflects two problems: either there is a problem with the "working efficiency" of the cooler itself, or the transformer itself generates "abnormal heat" far exceeding the normal level. At this time, the enriched warning information generated by the edge computing device will also be deepened, and the second abnormal cause will be included in the first abnormal warning information, that is, the cooler's working efficiency has decreased or there is an abnormality in the transformer, so as to remind the operation and maintenance personnel to check whether the cooler's radiator is dirty and blocked, the status of the fan blades, and pay attention to whether there are signs of internal faults in the transformer.

[0193] In summary, in this embodiment, by introducing the key "circumstantial" data of auxiliary equipment energy consumption, this application can effectively distinguish whether the fault originates from the target device itself or its auxiliary equipment. This cross-validation logic greatly improves the decision-making value of the first abnormality warning information, guiding operations and maintenance personnel directly to the root cause of the problem, thereby significantly shortening the time for on-site troubleshooting and location.

[0194] The following is an overall illustration of the device edge intelligent warning method based on multi-dimensional data association provided by this application with examples 1 and 2.

[0195] The following examples 1 and 2 are described with the preset threshold being 5°C.

[0196] Example 1, assume that a static threshold of 80°C is set in the traditional operation and maintenance procedures. During a peak period in summer, due to a sharp increase in load current, the actual top oil temperature of the transformer rises to 81°C. This exceeds the traditional static threshold and triggers a traditional static warning. However, the technical solution of the present application, its health indicator parameter prediction model calculates a reference top oil temperature of 80.5°C based on the current operating condition indicator parameters of high load current value and high ambient temperature value. At this time, the target health evaluation parameter is only 0.5°C, which is much smaller than the preset threshold (5°C). Therefore, it is determined that this actual top oil temperature value is normal and the temperature rise is caused by changes in operating conditions, and no warning is generated. This effectively filters out false warnings caused by fluctuations in normal operating conditions, greatly reducing the burden on operation and maintenance personnel, allowing them to focus on real abnormal events.

[0197] Example 2: Suppose that at night, the transformer is operating at a low load current (e.g., 30%) and the ambient temperature is also low (e.g., 15°C). At this time, the top oil temperature slowly rises and reaches 65°C. This actual top oil temperature is far below the static threshold of 80°C and would not attract any attention in a traditional monitoring system. However, the health indicator parameter prediction model of this application, based on the operating condition indicator parameters of the current low load current and low ambient temperature, calculates that the reference top oil temperature should only be around 52°C. At this point, the target health assessment parameter is as high as 13°C, significantly exceeding the preset threshold (5°C). Therefore, this is considered an abnormal temperature rise, triggering an alert. This abnormal temperature rise, which deviates from the normal correlation, may indicate a cooler blockage, fan failure, or a potential, developing internal fault in the transformer. This buys valuable time for operation and maintenance personnel to investigate and resolve the issue, achieving a qualitative leap from "post-event warning" to "pre-event warning."

[0198] In some embodiments of the present application, in order to extend diagnostic capabilities from immediate faults to early identification of slow performance degradation, and to capture the key turning point from gradual to sudden fault changes, thereby achieving a higher level of predictive maintenance, a series of more refined and forward-looking abnormality warning methods may be provided. Specifically, the following steps 210 to 220 may be performed: Step 210 : Acquire a third energy consumption parameter and a target benefit parameter of the auxiliary device within a third preset time period. The target benefit parameter is used to represent the degree to which the auxiliary device suppresses deterioration of the health indicator parameter of the target device.

[0199] Step 220: If the ratio of the third energy consumption parameter to the target benefit parameter is greater than the preset energy efficiency ratio threshold, a second abnormal warning message is sent to the operation and maintenance personnel. The second abnormal warning message includes that the auxiliary equipment has a working efficiency reduction.

[0200] In this embodiment, the third preset time period may be a complete working cycle of the auxiliary device, such as a time period from start to stop.

[0201] The benefit parameter (E), energy consumption parameter (C), and energy efficiency ratio (ECR) mentioned in this embodiment are defined and explained below: Benefit parameter (E): It is defined as the actual suppression effect of the auxiliary equipment on the top oil temperature of the transformer within a certain period of time (for example, within the third preset period in this embodiment). It can be calculated by comparing the difference between the actual temperature rise rate of the top oil temperature when the cooler is running and the temperature rise rate predicted by the model when the cooler is not running. Alternatively, it can be defined as the degree of oil temperature drop per unit time (such as per hour). In this case, the unit of the benefit parameter is " ”.

[0202] Energy consumption parameter (C): defined as the total electrical energy consumed by the auxiliary equipment within the same period of time corresponding to the benefit parameter (for example, within the third preset period in this embodiment), which can be obtained by integrating its operating power over time, and the unit is "kWh".

[0203] Energy Efficiency Ratio (ECR): The energy efficiency ratio is obtained by calculating the ratio of the energy consumption parameter to the benefit parameter (for example, the ratio between the third energy consumption parameter and the target benefit parameter in this embodiment). If the unit of the benefit parameter is expressed as " The physical meaning of energy efficiency ratio is "the amount of electricity required to reduce the transformer temperature by 1°C". It should be noted that the ECR value of a healthy cooler should remain relatively stable under similar operating conditions.

[0204] In some implementations, the edge computing device can establish a multidimensional lookup table or regression model describing the baseline energy efficiency ratio under different operating conditions (e.g., ambient temperature, load current) during the initial operation or post-maintenance health of the auxiliary equipment. This table can then be used to calculate a preset energy efficiency ratio threshold, such as a preset multiple of the baseline energy efficiency ratio (which can be greater than 1, such as 1.5) as the preset energy efficiency ratio threshold.

[0205] During routine operation, the edge computing device calculates the ratio of the third energy consumption parameter to the target efficiency parameter during the third preset time period and compares it with a preset multiple (i.e., a preset energy efficiency ratio threshold) of the baseline energy efficiency ratio, retrieved from a multidimensional lookup table or regression model based on the current operating condition indicator parameters. If the energy efficiency ratio exceeds the preset threshold, even if all other parameters, such as the actual top oil temperature of the transformer, are within normal ranges, the edge computing device triggers a specific "efficiency degradation" warning (i.e., a second abnormality warning message). This second abnormality warning message notifies maintenance personnel of the "efficiency degradation" and assists them in inferring the root cause of the fault and identifying key maintenance priorities. For example, if the transformer's cooler efficiency is suspected to have decreased, the next maintenance outage can focus on checking for dust accumulation and blockage on the radiator fins and dirt or deformation on the fan blades.

[0206] The following is a detailed description of the implementation method for determining the benchmark energy efficiency ratio based on the regression model and the multidimensional lookup table: In this application, to obtain a baseline energy efficiency ratio (ECR) that dynamically matches the current operating conditions, the edge computing device may employ at least one of the following specific implementation methods, which is based on offline learning and modeling of a large amount of historical data accumulated during the long-term healthy operation of the equipment. This historical data includes at least auxiliary equipment energy consumption related to ECR calculation, main equipment efficiency parameters (such as temperature drop rate), and various operating condition indicators that affect the ECR (such as load current, active power, ambient temperature, etc.).

[0207] The implementation based on the multidimensional lookup table is as follows: This method is an intuitive and efficient way to discretize the continuous operating space and obtain the benchmark energy efficiency ratio through table lookup and interpolation. The construction and use process can specifically include the following steps: Step 1: Discretization of the working space First, select the key operating conditions that most significantly impact energy efficiency, such as load current and ambient temperature. Then, divide these continuous operating conditions into intervals (or "bins"). For example, the load current can be divided into intervals of 10% (0-10%, 10%-20%, ...), and the ambient temperature can be divided into intervals of 5°C (0-5°C, 5-10°C, ...). By performing similar divisions on all selected key operating conditions, a multi-dimensional, discrete operating condition grid is constructed.

[0208] Step 2: Fill in the baseline energy efficiency ratio. Traverse the historical healthy operation database and assign each historical data record to the corresponding cell of the above-mentioned multidimensional operating condition grid according to its operating condition indicator parameter value. Subsequently, perform statistical calculations on all historical energy efficiency ratio data points collected in each cell, such as calculating their arithmetic mean or median, and use the statistical result as the baseline energy efficiency ratio under this specific operating condition combination and fill it into the cell. After this step, a multidimensional lookup table storing the baseline energy efficiency ratios under different operating condition combinations is constructed.

[0209] Step three, online query and interpolation. When the equipment is actually running, the edge computing device obtains the current real-time operating condition index parameters. If the operating condition parameter combination happens to accurately match a cell in the lookup table, the benchmark energy efficiency ratio value of the cell is directly read. More generally, when the real-time operating condition parameters fall between two or more cells, in order to ensure the smoothness of the output benchmark value and avoid sudden changes in the benchmark value due to data points crossing the division boundary, the system will use multi-dimensional linear interpolation (for example, for two-dimensional working conditions, bilinear interpolation is used; for three-dimensional working conditions, trilinear interpolation is used) method to calculate a smooth transition benchmark energy efficiency ratio based on the "distance" between the real-time working condition point and the nearest grid points.

[0210] The implementation based on the regression model is as follows: Compared to multi-dimensional lookup tables, regression models can directly establish a continuous functional relationship between operating condition index parameters and benchmark energy efficiency ratios without discretizing the operating conditions. This has advantages in processing high-dimensional operating conditions and saving storage resources. Specific regression model algorithms include but are not limited to the following: Multivariate Polynomial Regression: This is a relatively simple but effective nonlinear regression method. It transforms the original working condition index parameters (for example, load current I and ambient temperature) into ) as independent variables, by introducing their quadratic terms, cubic terms and even cross terms (such as , , , etc.) to construct a polynomial equation that fits the baseline energy efficiency ratio (ECR). This model has the advantages of simplicity and interpretability, making it suitable for scenarios where the nonlinear relationship between operating conditions and ECR is not overly complex.

[0211] Gradient Boosting Decision Tree: This is a representative class of ensemble learning algorithms, exemplified by the widely used XGBoost and LightGBM models. This type of model is a preferred option due to its efficient processing of tabular data, the absence of data normalization requirements, and its ability to automatically learn and capture complex, high-order interactions between features (operating condition indicators). During training, the model iteratively constructs a series of weak decision tree learners, with each new tree aiming to correct the residuals of all previous trees. Ultimately, the predictions from all trees are summed to produce the final output. This model possesses powerful nonlinear fitting capabilities, enabling it to accurately characterize the complex curve of the baseline energy efficiency ratio as it varies with multi-dimensional operating conditions.

[0212] Neural networks, particularly multilayer perceptrons (MPPs), are a general-purpose function approximator and are well-suited for this task. By constructing a neural network structure consisting of an input layer (with nodes corresponding to various operating indicators), one or more hidden layers (using nonlinear activation functions such as ReLU), and an output layer (outputting a single baseline energy efficiency ratio value), and training it with a backpropagation algorithm on a large amount of historical health data, it is possible to learn the profound and complex nonlinear mapping relationship between operating indicators and baseline energy efficiency ratios.

[0213] During actual operation, the edge computing device only needs to input the real-time collected operating condition indicator parameter vector into the trained regression model. The model can then instantly calculate and output a continuous and accurate benchmark energy efficiency ratio value that is highly matched with the current operating conditions.

[0214] In summary, the above technical solutions are innovative in the following aspects: A novel problem definition and solution: Existing technologies assess equipment efficiency, either using fixed thresholds unrelated to operating conditions or relying solely on simple qualitative judgments. This application, for the first time, clearly and clearly articulates the profound insight that the healthy operating efficiency of equipment (reflected in the Energy Efficiency Ratio (ECR)) itself changes dynamically with operating conditions. Based on this insight, it proposes a novel solution: establishing a "healthy baseline" that dynamically reflects current operating conditions in real time is essential to accurately isolate anomalies caused by actual equipment performance degradation. This is a truly innovative technical concept.

[0215] Creative combination of technical means and application scenarios: The creativity of this application lies in realizing the above-mentioned "dynamic benchmark" concept by introducing specific and feasible technical means such as multi-dimensional lookup tables or advanced regression models. Although these models themselves are known, using them to construct a dynamic benchmark model of "energy efficiency ratio-multi-dimensional working conditions" for early performance degradation diagnosis of key equipment such as power transformers is a specific application that has never been disclosed or revealed in the prior art. This combination is not a simple technical grafting, but a non-obvious and original solution to a specific technical problem (how to quantify energy efficiency ratio fluctuations under normal working conditions).

[0216] An unexpected technical effect was achieved: by establishing and applying the dynamic benchmark model, the technical solution of this application achieved significant technical progress that could not be achieved by existing technologies. It can: By comparing the small but persistent deviations between the "actual energy efficiency ratio" and the "dynamic benchmark energy efficiency ratio" at an early stage, even before the absolute values ​​of equipment operating parameters reach the alarm limit, the system can accurately identify slow and gradual efficiency degradation caused by factors such as dirty radiators and dust accumulation on fan blades. This is something that traditional static threshold methods cannot achieve.

[0217] Significantly reduces false alarm rates. When load, environment, and other operating conditions change dramatically, the baseline energy efficiency ratio dynamically adjusts accordingly. The system will not mistake normal efficiency fluctuations for faults, significantly improving the signal-to-noise ratio and reliability of early warnings.

[0218] In some embodiments of the present application, in order to further achieve the transition from diagnosing slow performance degradation to predicting sudden failures, the present invention proposes a more in-depth method that can capture the key turning points of fault evolution based on the above analysis based on energy efficiency ratio (ECR). Figure 2 The method shown executes: See also Figure 2 , shows a detailed flow diagram of generating third abnormal warning information according to an embodiment of the present application, specifically including the following steps 310 to 330: Step 310 : Obtain a first change acceleration of the energy efficiency ratio of the auxiliary device at the current moment. The energy efficiency ratio is a ratio of an energy consumption parameter to a benefit parameter of the auxiliary device.

[0219] In this embodiment, specific implementation methods include at least the following two: In a first implementation of step 310, the energy efficiency ratio of the auxiliary equipment at each moment is constructed into an energy efficiency ratio time series, and the time series is subjected to a second-order numerical derivation, thereby obtaining the first change acceleration of the energy efficiency ratio at the current moment. For example, when calculating the energy efficiency ratio of the auxiliary equipment at each moment, a fixed moment is used as the time starting point, and the ratio of the energy consumption parameter to the benefit parameter of the auxiliary equipment in the time period from the fixed moment to the calculation moment is determined, thereby obtaining the energy efficiency ratio of the auxiliary equipment at each moment. It can be understood that in the present application, the energy efficiency ratio of the auxiliary equipment at each moment can be calculated in real time, and then based on the calculated continuous energy efficiency ratio, the first change acceleration of the energy efficiency ratio at the current moment can be obtained through a second-order numerical derivation.

[0220] Specifically, we can first calculate the first-order derivative (rate of change) of the energy efficiency ratio: , the rate of change reflects the speed at which the working efficiency of the auxiliary equipment deteriorates; then calculate its second-order derivative (change acceleration) , reflecting the changing trend of the deterioration rate of auxiliary equipment working efficiency.

[0221] The second implementation of step 310 can be performed according to the following steps 311 to 312: Step 311 : obtaining the target energy efficiency ratio of the auxiliary device at each moment in a fourth preset time period, where the fourth preset time period ends at the current moment.

[0222] Step 312: Based on each target energy efficiency ratio, the corresponding relationship between the energy efficiency ratio and time is fitted into a quadratic polynomial function by the least squares method, and twice the value of the quadratic term coefficient of the quadratic polynomial function is used as the first change acceleration of the energy efficiency ratio of the auxiliary equipment at the current moment.

[0223] In this embodiment, to overcome the severe interference of raw data noise on the second-order derivative calculation, a preferred method for calculating the first change acceleration is proposed, namely, using quadratic polynomial fitting. Specifically, in the energy efficiency ratio time series, the energy efficiency ratio at each moment within the fourth preset time period is determined as the target energy efficiency ratio (for example, including the past 24 target energy efficiency ratio data points). The correspondence between these target energy efficiency ratio data points and time is fitted to a quadratic polynomial function using the least squares method, as shown in Formula 1 below: Formula 1 Where t represents time. After fitting, the second-order derivative of the quadratic polynomial function is always Therefore, we can directly extract the quadratic coefficients of the fitting polynomial and double its value This approach avoids direct differential operations that are sensitive to noise, significantly improving the core algorithm's anti-interference ability and the reliability of the results in real industrial environments.

[0224] Continue to see Figure 2 , step 320, determining whether the first change acceleration meets a preset condition.

[0225] In this embodiment, the specific implementation method can be performed as follows: If the first change acceleration is greater than the preset acceleration threshold, it is determined that the first change acceleration meets the preset condition.

[0226] If the first change acceleration is within the preset acceleration range, and the energy efficiency ratio of the auxiliary equipment is within the preset acceleration range at each moment in the fifth preset time period at the second change acceleration, it is determined that the first change acceleration meets the preset conditions, the maximum endpoint value of the preset acceleration range is the preset acceleration threshold, and the fifth preset time period ends at the current moment.

[0227] In this embodiment, to improve the signal-to-noise ratio (SNR) of the third abnormality warning information and avoid nuisance warnings for temporary, recoverable fluctuations, the SNR acceleration thresholds are set in a hierarchical manner. Specifically, a lower "observation threshold" and a higher "action threshold" are set. The observation threshold is the lowest endpoint of the preset acceleration range, and the action threshold is the highest endpoint of the preset acceleration range (i.e., the preset acceleration threshold).

[0228] In this application, when it is determined that the acceleration of an auxiliary device exceeds the observation threshold but is below the action threshold, the edge computing device may mark the auxiliary device as "needing attention" in the background, but will not immediately generate an abnormality warning message. Only when the acceleration remains in the "observation" interval for longer than a preset duration (i.e., the duration of the fifth preset period, such as three consecutive hours), or when the first acceleration momentarily exceeds the "action threshold", will the edge computing device officially trigger the transmission of the high-priority third abnormality warning message.

[0229] It is understandable that this hierarchical warning mechanism effectively filters out system glitches, so that the third abnormal warning information finally issued points to a more certain deterioration trend, thereby improving the overall reliability of the warning.

[0230] In some embodiments of the present application, in order to balance system resource consumption and fault diagnosis precision, the present application also proposes an adaptive adjustment mechanism for data acquisition frequency, which dynamically adjusts the acceleration of the energy efficiency ratio of the acquisition auxiliary equipment.

[0231] During normal operation, when the calculated first acceleration change falls below a preset trigger threshold (e.g., the aforementioned "observation threshold"), the edge computing device will subsequently collect and calculate acceleration change data at a regular first frequency (e.g., every 5 minutes) to conserve computing and storage resources. When the first acceleration change exceeds the action threshold, the edge computing device will automatically switch to a high-frequency "investigation mode," increasing the subsequent collection frequency to a higher second frequency (e.g., every 10 seconds) to collect and calculate acceleration change data.

[0232] This adaptive mechanism enables edge computing devices to maintain low-power operation when the device status is stable. When signs of abnormal evolution are detected, it can automatically and temporarily increase the data "resolution" to capture high-definition dynamic data at the critical stage of fault evolution, providing valuable data support for subsequent refined diagnosis and root cause analysis.

[0233] Continue to see Figure 2 , step 330, if satisfied, then generate a third abnormal warning information, and send the third abnormal warning information to the operation and maintenance personnel, the third abnormal warning information including the health status of the auxiliary equipment is at risk of deterioration.

[0234] In some implementations, specific implementations of generating the third abnormal warning information include at least the following two: In a first implementation of step 330, the predicted failure time of the auxiliary device is calculated using the following formula, and the predicted failure time is determined as the third abnormality warning information: Formula 2 in, Indicates the energy efficiency ratio failure threshold; Indicates the energy efficiency ratio of the auxiliary equipment at the current moment; Indicates the rate of change of the energy efficiency ratio of the auxiliary equipment at the current moment; represents the first change acceleration; t represents the predicted failure time.

[0235] It can be understood that the rate of change of the energy efficiency ratio at the current moment can be obtained by taking the first-order derivative of the energy efficiency ratio time series. In addition, the predicted failure time can be calculated by reversely solving the above formula 2.

[0236] In this embodiment, by pushing the predicted failure time of the auxiliary equipment to the operation and maintenance personnel in the third abnormal warning information, a quantitative basis can be provided for the maintenance plan of the operation and maintenance personnel, thereby improving the intelligence of the device edge intelligent warning method based on multi-dimensional data association of this application.

[0237] In some embodiments of the present application, in order to improve the accuracy of the predicted failure time calculated by the failure time prediction model (i.e., the above formula 2), the failure time prediction model can also be calibrated online. It is understood that any prediction model may deviate from the actual state of the auxiliary equipment over time. The online calibration mechanism ensures the long-term accuracy and reliability of the predicted failure time. The online calibration steps include: The first step is triggering and applying active disturbances: Under pre-set safety conditions, such as when operating indicators are relatively stable and key parameters like the transformer's top oil temperature are within a sufficient safety margin, the edge computing device can trigger an online self-calibration process periodically (e.g., every week) or when the confidence score of the failure time prediction model's prediction results falls below a certain threshold. Once triggered, the edge computing device sends a small, short-term, and absolutely safe control disturbance command to a sub-device in the auxiliary system that has precise adjustment capabilities (such as a cooling fan controlled by a frequency converter). For example, "Increase the cooling fan's operating frequency by 2% from the current level and return it to the original frequency after 60 seconds."

[0238] The second step involves response capture and model verification: During the entire process of applying the disturbance and subsequent recovery, the edge computing device synchronously and accurately collects energy consumption parameters of auxiliary equipment (as input disturbances) and response changes of transformers, such as top oil temperature (as output efficiency parameters), in high-frequency mode (i.e., the aforementioned second frequency). Based on this high-frequency data, the edge computing device can calculate the precise instantaneous ECR response curve for this disturbance event. Furthermore, before applying the disturbance, the edge computing device uses the pre-calibrated failure time prediction model to perform a "virtual simulation" to predict the ECR response curve that the disturbance should theoretically cause.

[0239] The third step is self-calibration and confidence update of the failure time prediction model: the actual ECR response curve is compared with the theoretical ECR response curve predicted by the failure time prediction model before calibration.

[0240] In the fourth step, if the two are highly consistent, it proves that the current failure time prediction model accurately reflects the latest status of the auxiliary equipment. The edge computing device will increase the confidence level of the prediction result of the current failure time prediction and record a successful verification.

[0241] If there's a significant deviation between the two, it indicates that key parameters of the failure time prediction model (for example, coefficients reflecting the current aging or wear condition of the auxiliary equipment) have drifted and are no longer accurate. At this point, the edge computing device uses the precise "input-output" data pairs obtained from this active detection and, through a built-in adaptive algorithm (for example, recursive least squares (RLS)), performs online fine-tuning and correction of the relevant parameters in the failure time prediction model to realign them with the latest actual state of the auxiliary equipment. Simultaneously, the edge computing device records a "model self-calibration" event and may temporarily reduce the confidence level in the failure time prediction, which will be restored after the results of several subsequent calibrations stabilize.

[0242] Through this active detection and self-calibration method, this application constructs a closed-loop, dynamic "physical examination" and "correction" mechanism for the advanced function of predicting failure time, ensuring that the prediction results will not deviate from the actual situation of dynamic changes in auxiliary equipment due to the static nature of the failure time prediction model, thereby greatly improving the long-term reliability and authority of the entire intelligent early warning method.

[0243] In a second implementation of step 330, the auxiliary device includes multiple sub-devices, and the third abnormality warning information can be generated by executing the following steps 331 to 333: Step 331: Obtain the third change acceleration of the energy efficiency ratio of each sub-device at the current moment.

[0244] Step 332: The sub-device corresponding to the maximum value among the third acceleration changes is used as the target sub-device.

[0245] Step 333: determining that the target sub-device has a fault as third abnormality warning information.

[0246] In this embodiment, the auxiliary equipment includes multiple sub-equipment connected in parallel (for example, a large transformer has two or more independent coolers).

[0247] It should be noted that, in this embodiment, the calculation method of the third change acceleration can refer to the calculation method of the first change acceleration mentioned above, that is, the third change acceleration can be obtained by referring to the first implementation method of the above step 310 or the second implementation method of step 310, which is not elaborated in this application.

[0248] In this embodiment, the edge computing device calculates the total first acceleration change for the entire auxiliary device while also independently calculating the third acceleration change for each of its sub-devices. By comparing the third acceleration changes of each sub-device, differential diagnosis can be performed, highly likely attributing the root cause of the fault to the sub-device with the highest third acceleration change. This differential diagnosis capability significantly refines fault location, greatly facilitating rapid and accurate repairs.

[0249] In some embodiments, if it is monitored that the third change acceleration of each sub-device shows a synchronous and similar growth trend, it can be inferred that the fault is a problem with the auxiliary equipment, such as cooling medium (transformer oil) contamination, overall control logic abnormality, etc., and this inference conclusion can also be determined in the third abnormal warning information.

[0250] In summary, when the calculated first change acceleration meets the preset conditions, the edge computing device determines that the auxiliary equipment status is at the critical point of transitioning from "chronic decline" to "acute deterioration", indicating that the health status of the auxiliary equipment is at risk of deterioration, and immediately triggers the sending of the third abnormal warning information with the highest priority, so that the operation and maintenance personnel can know that the efficiency deterioration of the transformer cooler is accelerating, there is a risk of sudden failure, and emergency treatment should be taken. In this application, the introduction of the high-order dynamic feature of change acceleration enables the present invention to accurately capture the "turning point" of the health status of the auxiliary equipment, that is, the critical moment from linear and slow performance degradation to exponential and rapid deterioration. Compared with waiting for the energy efficiency ratio itself or its rate of change to reach the warning line value, this early warning strategy can gain more valuable time for emergency treatment and achieve the ultimate foresight of sudden failures.

[0251] To verify the effectiveness and superiority of the proposed multidimensional data-correlated intelligent edge warning method for devices, the applicant conducted experimental data collection and comparative analysis on an oil-immersed power transformer (hereinafter referred to as the target transformer). The experiment compared the warning results of the proposed method (hereinafter referred to as the inventive method) with those of a traditional static threshold warning method (hereinafter referred to as the comparison technique), which uses a fixed top oil temperature threshold of 80°C.

[0252] In this method, the target device is the target transformer, the health indicator is the top oil temperature, and operating condition indicators include load current, active power, and ambient temperature. The health indicator parameter prediction model uses a pre-trained gradient boosting tree model. The preset threshold for the target health assessment parameter is set at 5°C.

[0253] Scenario 1: False alarm suppression verification under high load conditions Experimental background: Simulating the peak electricity consumption period in summer, the target transformer operates normally under high load and high ambient temperature.

[0254] Data recording and analysis: Data from 14:00 to 14:30 on a certain day were selected for analysis, and the records are shown in Table 7 below.

[0255] Table 7: Comparative analysis of high load condition data

[0256] Conclusion Analysis: Table 7 shows that between 2:10 PM and 2:30 PM, the actual top oil temperature of the target transformer exceeded the static threshold of 80°C due to the combined effects of high load and high ambient temperature. Consequently, the comparison technique frequently triggered false alarms, increasing the subsequent identification burden for personnel.

[0257] This discovery method, using a parameter prediction model, dynamically calculated a similarly high reference top oil temperature (79.9°C, 80.6°C, and 80.2°C, respectively) based on the current operating parameters (high load current and high ambient temperature). Based on this, the calculated target health assessment parameter (deviation) consistently remained at a low level of 0.4°C-0.6°C, well below the preset threshold of 5°C. Therefore, this discovery method correctly determined that this was a normal operating temperature rise and did not generate any warning information.

[0258] Experiments have shown that this discovered method can effectively distinguish normal operating fluctuations from real anomalies, significantly reducing the false alarm rate and improving the credibility and practicality of the early warning system.

[0259] Scenario 2: Early warning verification of fault acceptance Experimental background: Simulating low-load conditions at night, a cooling fan of the target transformer experiences a speed drop due to a bearing problem, leading to a slow decrease in cooling efficiency and an early fault scenario.

[0260] Data recording and analysis: Data from 23:00 to 23:30 on another day were selected for analysis, and the records are shown in Table 8 below.

[0261] Table 8: Comparative analysis of early acceptance failure tracking

[0262] Table 8 shows that in this scenario, although the actual top oil temperature of the target transformer exhibited an abnormally slow upward trend due to an incipient fault, its absolute value (60.1°C-64.0°C) was far below the static threshold of 80°C. Therefore, the comparison technique was completely unable to detect this potential fault, posing a serious risk of underreporting and potentially leading to further failure.

[0263] However, our discovery method, based on the current low-load, low-ambient-temperature operating conditions, calculated a reference top-oil temperature of approximately 51°C. At this point, the deviation between the actual top-oil temperature and the reference top-oil temperature (the target health assessment parameter) was as high as 8.6°C to 13.1°C, significantly exceeding the preset threshold of 5°C. Therefore, our discovery method successfully generated and issued the first abnormality warning at the early stages of the fault.

[0264] Experiments have shown that this discovery method is extremely sensitive to early-stage gradual faults hidden within the normal operating range, and can achieve the transformation from post-warning to pre-warning, thus winning valuable time for preventive maintenance.

[0265] Scenario 3: Diagnostic verification of reduced efficiency of auxiliary equipment Experimental background: To verify the diagnostic capability of the present invention for the decline in the efficiency of auxiliary equipment (cooler), the energy consumption and benefits of the cooler during its operation cycle were tracked over a long period of time, and the energy efficiency ratio (ECR) was calculated in units of (including the energy consumed to reduce the oil temperature), set the preset energy efficiency ratio threshold to 0.6 .

[0266] Data recording and analysis: Record data for three consecutive months, as shown in Table 9 below:

[0267] As shown in Table 9, although the target transformer's top oil temperature remained within normal limits in the third month, its cooler's energy efficiency ratio (ECR) had deteriorated to 0.68 kWh / °C due to severe dust accumulation on the heat sink, exceeding the preset ECR threshold of 0.6 kWh / °C. At this point, the method of the present invention proactively sends a second abnormality warning message, alerting maintenance personnel to a decline in auxiliary equipment efficiency. Traditional methods, however, focus solely on oil temperature parameters and are completely unaware of these performance issues.

[0268] Experiments have shown that by introducing energy efficiency ratio analysis, the present invention can discover potential problems in equipment performance from the new dimension of "basic energy consumption to maintain operating status", thereby achieving a higher level of predictive maintenance.

[0269] In the technical solutions provided in some embodiments of the present application, the technical means adopted for abnormal warning of target equipment in the substation include, first, obtaining the current actual health indicator parameters of the health indicators of the target equipment, and obtaining the current operating condition indicator parameters of at least one operating condition indicator associated with the health indicator, the health indicator is used to evaluate the health status of the target equipment, and the operating condition indicator is an indicator that can affect the parameter size of the health indicator; secondly, inputting each operating condition indicator parameter into a pre-trained health indicator parameter prediction model, so that the health indicator parameter prediction model outputs a reference health indicator parameter; thirdly, based on the difference between the actual health indicator parameter and the reference health indicator parameter, determining the target health evaluation parameter of the target equipment; finally, if the target health evaluation parameter is greater than the preset threshold, generating a first abnormal warning information, and sending the first abnormal warning information to the operation and maintenance personnel.

[0270] In this application, the expected normal value of the target device health indicator (i.e., the reference health indicator parameter) is generated by a pre-trained health indicator parameter prediction model based on the collected current operating condition indicator parameters (such as load current value, ambient temperature value, etc.), rather than a static threshold. This enables the abnormal warning threshold to be adaptively adjusted as the operating condition indicator parameters associated with the target device change, thereby bringing at least two beneficial effects: First, when the target equipment operates normally under harsh operating conditions (such as high load current or harsh ambient temperature), its actual health indicator parameters (such as the top oil temperature) may rise to a higher level, but the reference health indicator parameters calculated by the health indicator parameter prediction model will also increase accordingly, resulting in a very small deviation between the two that will not exceed the preset threshold, thereby effectively avoiding false alarms caused by fluctuations in normal operating conditions, reducing the burden on operation and maintenance personnel, and preventing the credibility of the abnormal warning system from declining.

[0271] Secondly, when the target device operates under relaxed operating conditions (such as low load current or a suitable ambient temperature), if its actual health indicator parameter experiences an abnormally slow climb due to an early potential failure of its auxiliary equipment (such as a decrease in cooler cooling efficiency), even if the actual health indicator parameter does not exceed the traditional static threshold, the reference health indicator parameter calculated by the health indicator parameter prediction model remains at a low level. This small actual health indicator parameter can lead to a significant deviation, causing it to quickly exceed the preset threshold and generate a first abnormal warning message. This enables the present method to capture early fault signs that traditional methods cannot identify, achieving a transition from post-warning to pre-warning, and saving valuable time for fault resolution. In summary, by constructing a health indicator parameter prediction model based on "actual health indicator parameter - dynamic reference health indicator parameter", the present invention fundamentally solves the problem of static threshold "inaccuracy" under complex substation operating conditions, achieving a dual improvement in the accuracy and timeliness of abnormal warnings.

[0272] Based on the same inventive concept, embodiments of the present invention provide a device edge intelligent warning system based on multidimensional data association, which can be used to implement the device edge intelligent warning method based on multidimensional data association in the above-mentioned embodiments of this application. For details not disclosed in the embodiments of this application, please refer to the embodiments of the device edge intelligent warning method based on multidimensional data association in the above-mentioned embodiments of this application.

[0273] See also Figure 3 , shows a block diagram of a device edge intelligent early warning system based on multidimensional data association according to an embodiment of the present application.

[0274] like Figure 3 As shown, according to an embodiment of the present application, a device edge intelligent early warning system 400 based on multi-dimensional data association is applied to a substation, and a target device is installed in the substation. The system includes an edge computing device deployed in the substation, and the edge computing device includes: an acquisition unit 401, an output unit 402, a determination unit 403, and an early warning unit 404.

[0275] Among them, the acquisition unit 401 is used to obtain the current actual health indicator parameters of the health indicators of the target device, and to obtain the current operating condition indicator parameters of at least one operating condition indicator associated with the health indicator. The health indicator is used to evaluate the health status of the target device, and the operating condition indicator is an indicator that can affect the parameter size of the health indicator; the output unit 402 is used to input each operating condition indicator parameter into a pre-trained health indicator parameter prediction model, so that the health indicator parameter prediction model outputs a reference health indicator parameter; the determination unit 403 is used to determine the target health evaluation parameter of the target device based on the difference between the actual health indicator parameter and the reference health indicator parameter; the early warning unit 404 is used to generate a first abnormal early warning information if the target health evaluation parameter is greater than a preset threshold, and send the first abnormal early warning information to the operation and maintenance personnel.

[0276] In some embodiments of the present application, based on the aforementioned solution, auxiliary equipment is further installed in the substation, and the auxiliary equipment is equipment that can affect the health status of the target equipment.

[0277] In some embodiments of the present application, based on the aforementioned scheme, the determination unit 403 is also used to: use the difference between the actual health indicator parameter and the reference health indicator parameter as the initial health evaluation parameter; obtain each first operation instruction received by the target device and / or auxiliary device within a first preset time period, and the first preset time period ends at the current moment; determine whether each first operation instruction contains a second operation instruction, and the second operation instruction is an operation instruction that can increase the health indicator parameter of the target device at the current moment; if so, adjust the initial health evaluation parameter to obtain the target health evaluation parameter, and the target health evaluation parameter is less than the initial health evaluation parameter.

[0278] In some embodiments of the present application, based on the above-mentioned scheme, the early warning unit 404 is also used to: obtain a first energy consumption parameter of the auxiliary equipment within a second preset time period; if the first energy consumption parameter is less than the first energy consumption threshold, the first abnormal cause, the actual health index parameter, each operating condition index parameter, the abnormal type corresponding to the health index, and the target health evaluation parameter are determined as the first abnormal warning information, and the first abnormal cause includes that the auxiliary equipment is not turned on or there is a fault in the control loop of the auxiliary equipment; if the first energy consumption parameter is greater than the second energy consumption threshold, the second abnormal cause, the actual health index parameter, each operating condition index parameter, the abnormal type corresponding to the health index, and the target health evaluation parameter are determined as the first abnormal warning information, and the second abnormal cause includes that the working efficiency of the auxiliary equipment is reduced or there is an abnormality in the target equipment, and the second energy consumption threshold is greater than the first energy consumption threshold.

[0279] In some embodiments of the present application, based on the aforementioned scheme, the early warning unit 404 is also used to: obtain a third energy consumption parameter and a target benefit parameter of the auxiliary equipment within a third preset time period, the target benefit parameter being used to characterize the degree to which the auxiliary equipment suppresses the deterioration of the health indicator parameters of the target equipment; if the ratio of the third energy consumption parameter to the target benefit parameter is greater than a preset energy efficiency ratio threshold, a second abnormal early warning information is sent to the operation and maintenance personnel, and the second abnormal early warning information includes that the auxiliary equipment has a decline in work efficiency.

[0280] In some embodiments of the present application, based on the aforementioned scheme, the early warning unit 404 is also used to: obtain the first change acceleration of the energy efficiency ratio of the auxiliary equipment at the current moment, where the energy efficiency ratio is the ratio of the energy consumption parameter of the auxiliary equipment to the benefit parameter; determine whether the first change acceleration meets the preset conditions; if so, generate a third abnormal early warning information, and send the third abnormal early warning information to the operation and maintenance personnel, where the third abnormal early warning information includes that the health status of the auxiliary equipment is at risk of deterioration.

[0281] In some embodiments of the present application, based on the aforementioned scheme, the early warning unit 404 is also used to: obtain the target energy efficiency ratio of the auxiliary device at each moment within a fourth preset time period, and the fourth preset time period ends at the current moment; based on each target energy efficiency ratio, the correspondence between the energy efficiency ratio and time is fitted into a quadratic polynomial function through the least squares method, and twice the value of the quadratic term coefficient of the quadratic polynomial function is used as the first change acceleration of the energy efficiency ratio of the auxiliary device at the current moment.

[0282] In some embodiments of the present application, based on the aforementioned scheme, the early warning unit 404 is further used to: if the first change acceleration is greater than the preset acceleration threshold, determine that the first change acceleration meets the preset condition; if the first change acceleration is in the preset acceleration range, and the energy efficiency ratio of the auxiliary equipment is in the preset acceleration range at each moment within the fifth preset time period, determine that the first change acceleration meets the preset condition, the maximum endpoint value of the preset acceleration range is the preset acceleration threshold, and the fifth preset time period ends at the current moment.

[0283] In some embodiments of the present application, based on the aforementioned solution, the warning unit 404 is further configured to calculate the predicted failure time of the auxiliary device using the following formula, and determine the predicted failure time as the third abnormal warning information:

[0284] in, Indicates the energy efficiency ratio failure threshold; Indicates the energy efficiency ratio of the auxiliary equipment at the current moment; Indicates the rate of change of the energy efficiency ratio of the auxiliary equipment at the current moment; represents the first change acceleration; t represents the predicted failure time.

[0285] In some embodiments of the present application, based on the aforementioned scheme, the early warning unit 404 is also used to: obtain the second change acceleration of the energy efficiency ratio of each sub-device at the current moment; take the sub-device corresponding to the maximum value of each second change acceleration as the target sub-device; and determine the existence of a fault in the target sub-device as the third abnormal early warning information.

[0286] The foregoing is merely an embodiment of the present application and is not intended to limit the present application. Various modifications and variations are possible for those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present application shall be included within the scope of the claims of the present application.

Claims

1. A device edge intelligent early warning method based on multi-dimensional data association, characterized in that: The method is applied to a substation in which a target device is installed, and includes: Obtaining a current actual health indicator parameter of a health indicator of the target device, and obtaining a current operating condition indicator parameter of at least one operating condition indicator associated with the health indicator, wherein the health indicator is used to evaluate the health status of the target device, and the operating condition indicator is an indicator that can affect the parameter size of the health indicator; Inputting each of the operating condition indicator parameters into a pre-trained health indicator parameter prediction model, so that the health indicator parameter prediction model outputs a reference health indicator parameter; determining a target health evaluation parameter of the target device based on a difference between the actual health indicator parameter and the reference health indicator parameter; If the target health evaluation parameter is greater than a preset threshold, first abnormal warning information is generated and sent to operation and maintenance personnel.

2. The method according to claim 1, characterized in that Auxiliary equipment is also installed in the substation, and the auxiliary equipment is equipment that can affect the health status of the target equipment.

3. The method according to claim 2, characterized in that The determining, based on the difference between the actual health indicator parameter and the reference health indicator parameter, a target health evaluation parameter of the target device includes: taking the difference between the actual health index parameter and the reference health index parameter as an initial health assessment parameter; Acquire each first operation instruction received by the target device and / or the auxiliary device within a first preset time period, where the first preset time period ends at the current time; determining whether each of the first operation instructions includes a second operation instruction, where the second operation instruction is an operation instruction capable of increasing a health indicator parameter of the target device at a current moment; If so, the initial health evaluation parameter is adjusted to obtain a target health evaluation parameter, and the target health evaluation parameter is smaller than the initial health evaluation parameter.

4. The method according to claim 2, characterized in that The generating of the first abnormal warning information includes: Acquiring a first energy consumption parameter of the auxiliary device within a second preset time period; If the first energy consumption parameter is less than a first energy consumption threshold, the first abnormality cause, the actual health index parameter, each of the operating condition index parameters, the reference health index parameter, the abnormality type corresponding to the health index, and the target health evaluation parameter are determined as the first abnormality warning information, where the first abnormality cause includes that the auxiliary device is not turned on or there is a fault in the control circuit of the auxiliary device; If the first energy consumption parameter is greater than the second energy consumption threshold, the second abnormal cause, the actual health index parameter, each of the operating condition index parameters, the reference health index parameter, the abnormality type corresponding to the health index, and the target health evaluation parameter are determined as the first abnormal warning information. The second abnormal cause includes a decrease in the working efficiency of the auxiliary equipment or an abnormality in the target equipment, and the second energy consumption threshold is greater than the first energy consumption threshold.

5. The method according to claim 2, characterized in that The method further comprises: Acquiring a third energy consumption parameter and a target benefit parameter of the auxiliary device within a third preset time period, wherein the target benefit parameter is used to represent the degree to which the auxiliary device suppresses deterioration of a health indicator parameter of the target device; If the ratio of the third energy consumption parameter to the target benefit parameter is greater than a preset energy efficiency ratio threshold, a second abnormal warning message is sent to the operation and maintenance personnel, where the second abnormal warning message includes that the auxiliary equipment has a work efficiency attenuation.

6. The method according to claim 2, characterized in that The method further comprises: Obtaining a first change acceleration of an energy efficiency ratio of the auxiliary device at a current moment, where the energy efficiency ratio is a ratio of an energy consumption parameter to a benefit parameter of the auxiliary device; determining whether the first change acceleration satisfies a preset condition; If the conditions are met, a third abnormal warning message is generated and sent to the operation and maintenance personnel. The third abnormal warning message includes that the health status of the auxiliary equipment is at risk of deterioration.

7. The method according to claim 6, characterized in that The obtaining of a first change acceleration of the energy efficiency ratio of the auxiliary device at a current moment includes: Obtaining a target energy efficiency ratio of the auxiliary device at each moment in a fourth preset time period, where the fourth preset time period ends at the current moment; Based on each of the target energy efficiency ratios, the corresponding relationship between the energy efficiency ratio and time is fitted into a quadratic polynomial function by the least squares method, and twice the value of the quadratic term coefficient of the quadratic polynomial function is used as the first change acceleration of the energy efficiency ratio of the auxiliary equipment at the current moment.

8. The method according to claim 6, characterized in that The determining whether the first change acceleration satisfies a preset condition includes: If the first change acceleration is greater than a preset acceleration threshold, determining that the first change acceleration meets a preset condition; If the first change acceleration is within a preset acceleration range, and the energy efficiency ratio of the auxiliary equipment is within the preset acceleration range at each moment within a fifth preset time period at a second change acceleration, it is determined that the first change acceleration meets the preset condition, the maximum endpoint value of the preset acceleration range is the preset acceleration threshold, and the fifth preset time period ends at the current moment.

9. The method according to claim 6, characterized in that Generating the third abnormal warning information includes: Calculate the predicted failure time of the auxiliary equipment by the following formula, and determine the predicted failure time as the third abnormality warning information; in, Indicates the energy efficiency ratio failure threshold; Indicates the energy efficiency ratio of the auxiliary equipment at the current moment; Indicates the rate of change of the energy efficiency ratio of the auxiliary equipment at the current moment; represents the first change acceleration; t represents the predicted failure time.

10. The method according to claim 6, characterized in that The auxiliary device includes a plurality of sub-devices, and the generating of the third abnormal warning information includes: Obtaining a third change acceleration of the energy efficiency ratio of each of the sub-devices at the current moment; The sub-device corresponding to the maximum value of each third change acceleration is used as the target sub-device; The existence of a fault in the target sub-device is determined as the third abnormality warning information.

11. A device edge intelligent early warning system based on multi-dimensional data association, characterized in that: The system is applied to a substation, in which a target device is installed. The system includes an edge computing device deployed in the substation, and the edge computing device includes: an acquisition unit, configured to acquire a current actual health indicator parameter of a health indicator of the target device, and acquire a current operating condition indicator parameter of at least one operating condition indicator associated with the health indicator, wherein the health indicator is used to evaluate the health status of the target device, and the operating condition indicator is an indicator that can affect the parameter size of the health indicator; an output unit, configured to input each of the operating condition indicator parameters into a pre-trained health indicator parameter prediction model, so that the health indicator parameter prediction model outputs a reference health indicator parameter; a determining unit, configured to determine a target health evaluation parameter of the target device based on a difference between the actual health indicator parameter and the reference health indicator parameter; The early warning unit is configured to generate a first abnormal early warning message if the target health evaluation parameter is greater than a preset threshold value, and send the first abnormal early warning message to the operation and maintenance personnel.

Images