Power equipment intelligent inspection abnormity identification method based on multi-dimensional data fusion

Abstract

The invention provides a power equipment intelligent inspection abnormity identification method based on multi-dimensional data fusion, and the method comprises the steps: obtaining historical operation data and real-time monitoring data, fusing the historical operation data and the real-time monitoring data through a time sequence alignment method, generating a unified data set containing voltage, current, temperature and load information, and obtaining multi-dimensional equipment operation features; aiming at the outlier feature vector, adopting an adaptive neural network to adjust a weight coefficient of the health state evaluation model, generating an optimized weight parameter, and determining an optimized model structure; and performing health state evaluation on the unified data set through the optimized model structure, generating a health index scoring result of the equipment in combination with load over-limit period counting and a temperature abnormal fluctuation range, and determining an evaluation basis for guiding local maintenance of the miniature equipment.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of information technology, and in particular to a multi-dimensional data fusion-based power equipment intelligent inspection abnormality identification method. BACKGROUND

[0002] Power equipment health state evaluation is a key field to ensure the safe and stable operation of the power grid, and its importance lies in accurately predicting equipment failure, reducing power outage risk, and optimizing asset management decisions. In the power system, the equipment is subjected to complex working conditions for a long time, and the health state evaluation directly affects the maintenance strategy and economic benefits. However, the existing methods have obvious limitations in processing multi-dimensional data fusion, mainly manifested in the insufficient integration of historical operation data and real-time monitoring data, and the inability to accurately capture the dynamic changes of the equipment after overload or short-circuit impact. This leads to the evaluation model being unable to adapt to the nonlinear changes of the equipment health state, thereby affecting the accuracy of preventive maintenance. Currently, the core challenge of power equipment health state evaluation lies in the effective fusion of time dimension data. After overload or short-circuit impact, the health index often shows accelerated decay, and the aging rate of key components also increases nonlinearly. This phenomenon makes it difficult for traditional static models to accurately depict the dynamic evolution of equipment state. The deeper problem is that the equipment after impact may exhibit abnormal out-of-group characteristics in the similar equipment group, which further exacerbates the complexity of model weight coefficient adjustment. Since the weight coefficient fails to dynamically adapt to the nonlinear changes of the equipment state, the evaluation results often deviate from the actual situation, leading to misjudgment of maintenance decisions. These challenges are closely related in logic: the accelerated decay of the health index directly drives the nonlinear change of the aging rate, and the abnormality of the aging rate amplifies the influence of the out-of-group characteristics in the equipment group, making the weight adjustment a bottleneck for model accuracy. Therefore, how to capture the accelerated decay of the health index, the nonlinear increase of the aging rate of key components, and the abnormal out-of-group characteristics after overload or short-circuit impact of the equipment on the basis of integrating historical operation data and real-time monitoring data has become a key problem in power equipment health state evaluation and preventive maintenance decision optimization. SUMMARY

[0003] The present application provides a multi-dimensional data fusion-based power equipment intelligent inspection abnormality identification method, mainly including:

[0004] Obtain historical operation data and real-time monitoring data, fuse them through a time series alignment method, generate a unified data set containing voltage, current, temperature and load information, and obtain multi-dimensional equipment operation characteristics;

[0005] Principal component analysis is used to reduce the dimensionality of the unified dataset, extract key feature vectors reflecting the health status of the equipment, analyze the cumulative effect of overload impact and the frequency statistics of current surge, and calculate the abnormal stress on the equipment by combining the duration of voltage instability. If the cumulative effect of overload impact exceeds the preset threshold range, it is determined that the equipment is subjected to overload or short circuit impact, and impact event markers are generated to obtain the time series of impact impact.

[0006] By combining the time series of impact effects with the range of abnormal temperature fluctuations caused by poor internal heat dissipation and the count of overload cycles caused by peak business periods, a dynamic trajectory of equipment health index is constructed. A long short-term memory network is used to perform time series modeling of the dynamic trajectory, and an equipment health status assessment model is established to obtain the short-term decay prediction curve of the health index.

[0007] Based on the short-term decay prediction curve of the health index, the fatigue damage index of the component is analyzed and compared with the historical decay rate. The particle filter algorithm is used to estimate the aging rate distribution of the key components. If the dynamic distribution of the aging rate deviates from the preset statistical range of the equipment group, it is determined that there are abnormal outlier characteristics, and an outlier feature vector is generated.

[0008] For outlier feature vectors, an adaptive neural network is used to adjust the weight coefficients of the health status assessment model, generate optimized weight parameters, and determine the optimized model structure.

[0009] The health status of the unified dataset is assessed by optimizing the model structure. The health index score of the equipment is generated by combining the overload cycle count and the abnormal temperature fluctuation range, and the assessment basis for guiding the local maintenance of micro equipment is determined.

[0010] Based on the health index score, the duration of voltage instability and the frequency of current mutations are analyzed. The conditional random field algorithm is used to predict the maintenance demand time window of micro devices under specific operating scenarios, generate maintenance decision sequences for locally overloaded components, and determine the preventive maintenance execution plan.

[0011] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0012] The application discloses a kind of based on multi-dimension data fusion power equipment intelligent inspection abnormality identification method.The method is by fusing historical operation data and real-time monitoring data, constructs the unified data set including voltage, current, temperature and load information, extracts the key features reflecting equipment health status.Combining overload impact cumulative effect, current mutation frequency and other factors, the abnormal stress received by equipment is analyzed, and the dynamic change trajectory of health index is constructed.Long short-term memory network is used to model the trajectory in time sequence, and the short-term decay trend of health index is predicted.The aging rate of key components is estimated by particle filtering algorithm, and abnormal outlier characteristics are identified.Adaptive neural network is used to optimize health status evaluation model, and equipment health index score is generated.Finally, based on conditional random field algorithm, maintenance requirement time window is predicted, and targeted preventive maintenance plan is made.The application realizes the accurate evaluation and timely maintenance of micro device health status, effectively prolongs the service life of equipment and improves operation reliability. BRIEF DESCRIPTION OF DRAWINGS

[0013] Figure 1 A flow chart of the application of a kind of based on multi-dimension data fusion power equipment intelligent inspection abnormality identification method.

[0014] Figure 2 A schematic diagram of the application of a kind of based on multi-dimension data fusion power equipment intelligent inspection abnormality identification method. DETAILED DESCRIPTION

[0015] For further understanding of the content of the application, the application is described in detail in conjunction with the drawings and examples.The application is further described in detail in conjunction with the drawings and examples.It can be understood that the specific embodiments described herein are only used to explain the related application, and not limited to the application.In addition, it should be noted that, for the convenience of description, only the parts related to the application are shown in the drawings.

[0016] As Figure 1 -2, the application of a kind of based on multi-dimension data fusion power equipment intelligent inspection abnormality identification method specifically can include:

[0017] S101, historical operation data and real-time monitoring data are acquired, and the two are fused by time series alignment method, to generate the unified data set including voltage, current, temperature and load information, to obtain multi-dimension equipment operation characteristics.

[0018] The device operation record is extracted from the historical database, including voltage, current, temperature value sequence, the current sampled voltage value, current value, temperature value and load rate data are read from the real-time monitoring port, the two types of data obtained from the historical database and the monitoring port are sorted according to the time stamp, the difference between the historical data timestamp and the real-time data timestamp is calculated, the time offset and the sampling frequency ratio are determined, and the time sequence alignment parameters are obtained. According to the time offset in the time sequence alignment parameter, the historical data timestamp is adjusted, the historical data is resampled by using the linear interpolation method, so that the sampling rate of the historical data is consistent with the real-time data, for the missing data points, the missing data points are filled by weighted average of the adjacent values, and the time continuous complete data sequence is obtained. The voltage value, current value and temperature value in the time continuous complete data sequence are divided by the maximum value of each parameter respectively to carry out normalization processing, the maximum value and the minimum value of each parameter under different load rates are counted as the distribution range, the data set is divided according to the load rate interval, and a multi-dimensional feature matrix containing time stamp, normalized voltage, current, temperature and load rate is constructed, and the fused multi-dimensional device operation feature is obtained.

[0019] Specifically, when extracting the device operation record from the historical database, the system reads the time series data of voltage, current, temperature and other parameters stored in the database. These historical data are usually recorded at fixed sampling intervals, such as once every 5 minutes. While the real-time monitoring port continuously receives real-time data stream from the sensor, the sampling frequency may be higher, such as updating once every 30 seconds. Because the sampling frequencies of the two data sources are different, direct merging will cause misalignment on the time axis, so time sequence alignment processing is needed.

[0020] Specifically, the calculation process of the time sequence alignment parameters includes two key elements. First, the determination of the time offset, by comparing the difference between the last timestamp of the historical data and the first timestamp of the real-time data, the system can accurately locate the relative position of the two data sets on the time axis. Second, the calculation of the sampling frequency ratio, by counting the number of sampling points of the historical data and the real-time data in unit time, the frequency ratio relationship of the two is obtained. These two parameters together constitute the basis for subsequent data fusion.

[0021] In one possible implementation, the application of linear interpolation method can effectively solve the problem of inconsistent sampling rate. When the sampling interval of historical data is 5 minutes and the real-time data is 30 seconds, the system will insert new value points between adjacent historical data points. The interpolation process is based on the linear assumption that the parameters change linearly in a short time. For the processing of missing data, the weighted average method considers the influence of time distance, the data points closer to the missing point have larger weight, so as to ensure the continuity and reasonableness of the data.

[0022] It should be noted that the normalization plays a key role in multi-dimensional data fusion. The voltage value may fluctuate around 380V, the current value may vary between dozens to hundreds of amperes, and the temperature value is usually in the range of tens of degrees Celsius. If not normalized, the parameters with larger values will dominate in subsequent analysis, masking the variation characteristics of other parameters. By dividing by the respective maximum value, all parameters are mapped into the unified interval of 0 to 1, ensuring the equality of each dimension feature.

[0023] Exemplarily, when constructing the multi-dimensional feature matrix, the system divides the data into different intervals according to the load rate, such as the low load interval 0-30%, the medium load interval 30-70%, and the high load interval 70-100%. This division helps to identify the running characteristics of the device under different working conditions. Each data point contains a timestamp, normalized voltage, current, temperature value, and corresponding load rate information, forming a five-dimensional feature vector. Multiple feature vectors are arranged in chronological order to ultimately form a complete multi-dimensional feature matrix, providing a comprehensive data basis for subsequent device state analysis and fault prediction.

[0024] S102, using principal component analysis method to reduce dimension of the unified data set, extracting key feature vectors reflecting the health status of the device, analyzing the cumulative effect of overload impact and the statistical frequency of current sudden change, combining the duration of voltage instability to calculate the abnormal stress received by the device, if the cumulative effect of overload impact exceeds the preset threshold range, it is judged that the device is subjected to overload or short circuit impact, generating impact event markers, and obtaining the time sequence of impact influence.

[0025] A covariance matrix is constructed for the unified data set, the eigenvalues and corresponding eigenvectors of the matrix are calculated, sorted in descending order of eigenvalues, and the first several eigenvectors with cumulative variance contribution rate reaching the preset proportion are selected as principal components. Through matrix multiplication operation of the original data and the principal components, the key feature vectors after dimension reduction are obtained. According to the key feature vectors after dimension reduction, the current component sequence is extracted, the number of times the current value exceeds the current rated value within a unit time window is counted, and the product of the amplitude and duration of each time exceeding the rated value is accumulated as the cumulative effect value of overload impact. At the same time, the frequency of current change rate exceeding the preset mutation threshold between adjacent sampling points is detected to obtain the statistical results of current sudden change frequency and the corresponding mutation time records. For each time point in the mutation time records, the voltage component data before and after the time point is extracted from the key feature vectors, the amplitude and duration of the voltage deviation from its rated value are calculated, and the abnormal stress value is obtained by multiplying the product of the cumulative effect value of overload impact and the current sudden change frequency by the duration of voltage instability. If the abnormal stress value exceeds the preset threshold range, it is judged that the device is subjected to overload or short circuit impact in that period, and impact event markers containing impact type, starting time and ending time are generated to form the time sequence of impact influence.

[0026] Specifically, the application of principal component analysis in equipment health monitoring is based on the core concept of data dimensionality reduction. The unified dataset contains information in multiple dimensions such as voltage, current, temperature, and load rate. There is often a correlation between these parameters. The construction of the covariance matrix is to quantify this correlation. Each element in the matrix reflects the degree of covariance between two different parameters. By calculating the eigenvalues and eigenvectors of the covariance matrix, the system can identify the main direction of data change.

[0027] Specifically, the size of the eigenvalue represents the degree of dispersion of the data in the corresponding eigenvector direction. A larger eigenvalue means that this direction contains more information. The cumulative variance contribution rate is calculated by arranging the eigenvalues in descending order, then accumulating and dividing by the sum of all eigenvalues. When the cumulative contribution rate reaches 85% or 90%, it means that the selected principal components have included most of the information of the original data. This dimensionality reduction not only reduces the computational load of data processing, but more importantly, extracts the key features that reflect the health status of the equipment.

[0028] In one possible implementation, the calculation of the cumulative effect of overload impact needs to consider both the strength and duration of the impact. When the current value exceeds 120% of the rated value, the system starts timing and records the overload amplitude. If the current reaches 150% of the rated value and lasts for 5 seconds, its cumulative effect is much greater than that of 130% for 2 seconds. This cumulative effect is quantified by the product of amplitude and time, reflecting the actual stress level of the equipment. The statistics of current mutation frequency focus on transient changes, and a current change rate exceeding 10% of the rated value per second between adjacent sampling points is considered as a mutation event.

[0029] It should be noted that the calculation of abnormal stress value considers the interaction of multiple factors. The cumulative effect of overload impact reflects the persistent damage, the current mutation frequency reflects the impact damage, and the duration of voltage instability reflects the impact of power supply quality on the equipment. The product of these three factors constitutes the comprehensive abnormal stress index. When the transformer experiences multiple current mutations in a short period of time, accompanied by voltage fluctuations, even if a single impact does not exceed the protection threshold, the cumulative effect may cause irreversible damage to the insulation material.

[0030] Exemplarily, the generation process of the impact event label contains two key links of type determination and time positioning. The system distinguishes the two impact types of overload and short circuit by analyzing the rising rate of abnormal stress value. Overload impact usually shows progressive rise, while short circuit impact presents steep jump characteristics. The construction of time series not only records the time of impact occurrence, but also contains the device state change trajectory before and after the impact, providing complete data support for subsequent fault diagnosis and life evaluation. Compared with the traditional single threshold judgment, this impact recognition method based on multi-dimensional feature fusion can more accurately capture abnormal events in device operation, and improve the accuracy and reliability of power equipment health management.

[0031] S103, through the time series of impact influence, combined with the temperature abnormal fluctuation range caused by poor internal heat dissipation of the device and the load overrun period count caused by business peak period, the dynamic change trajectory of the device health index is constructed, the long short-term memory network is used for time series modeling of the dynamic change trajectory, the device health state evaluation model is established, and the short-term decay prediction curve of the health index is obtained.

[0032] According to the time series of impact influence, the temperature data of each impact event occurrence period is extracted, the amplitude value of temperature deviation from the preset normal working temperature range is calculated, the time length required for temperature to fall back to stable value from peak value is identified as the heat dissipation time constant, the number of times that temperature fluctuation exceeds the preset threshold in unit time is counted, and the temperature abnormal fluctuation range parameter is determined in combination with the heat dissipation time constant. For the period corresponding to the temperature abnormal fluctuation range parameter, the time period exceeding the preset rated load value is identified from the load data, the continuous overrun time and interval time are counted, the periodic count of load overrun is calculated according to day and week dimensions respectively, the weight coefficient is assigned according to the impact event label, the temperature abnormal fluctuation range parameter is multiplied by the decay factor, the load overrun periodic count is multiplied by the influence coefficient, and the sum of the three is normalized to construct the health index sequence, and the dynamic change trajectory of the device health index is obtained. The long short-term memory network is used for time series modeling of the dynamic change trajectory, the health index sequence is divided into training samples according to the time window, the network learns the change mode and trend characteristics of historical health index, the mapping relationship between input sequence and output prediction value is established, the health index values at future multiple time points are generated through iterative prediction, and the short-term decay prediction curve of the health index is obtained by connecting each prediction point.

[0033] Specifically, the determination of the temperature abnormal fluctuation range parameter involves the comprehensive analysis of multiple key elements. When the equipment is subjected to a shock event, its internal temperature will rise sharply, and this rise directly reflects the severity of the shock. The preset normal working temperature range is usually determined according to the design specifications of the equipment, such as the upper limit of the normal working temperature of the transformer being 85°C. When the temperature is detected to exceed this range, the system records the magnitude of the excess. The introduction of the heat dissipation time constant makes the temperature analysis more comprehensive, which reflects the actual situation of the equipment's heat dissipation capacity. The equipment with poor heat dissipation needs a longer time to fall back to the stable value from the peak value, which means that the equipment is under high temperature stress for a longer duration.

[0034] In one possible implementation, the periodic analysis of load overruns reveals the regularity characteristics of equipment operation. Daily dimension statistics can identify the daily peak power consumption period, usually during the morning 9-11 am and afternoon 2-4 pm industrial production peak period. The weekly dimension analysis can find the load difference between weekdays and weekends, and the load of weekdays is generally higher than that of weekends. The identification of this periodic regularity is of great significance for predicting the future health status of the equipment, because periodic overload can cause cumulative damage to the equipment.

[0035] Specifically, the construction process of the health index embodies the idea of multi-factor comprehensive evaluation. The shock event marker is given a higher weight coefficient, because the sudden impact on the equipment is usually instantaneous and severe. The temperature abnormal fluctuation range parameter is multiplied by a decay factor, which decreases over time, reflecting the time-dependent characteristics of temperature influence. The influence coefficient of load overrun period count takes into account the relationship between load level and equipment bearing capacity. The weighted sum of the three factors is not simply added, but is normalized to make the final health index fall within the standard interval of 0 to 100, which is convenient for intuitive understanding and comparison.

[0036] It should be noted that the application of long short-term memory network in time series modeling is based on its unique memory mechanism. Compared with traditional time series analysis methods, this network can automatically learn and remember long-term dependencies. The health index sequence is divided into fixed-length time windows, each window containing consecutive historical health index values. The network analyzes these historical patterns to identify the trend characteristics of the health index decline, such as linear decay, exponential decay, or stepwise decline.

[0037] Exemplarily, the generation of the short-term decay prediction curve adopts an iterative prediction manner. The network first predicts the health index at the next time point, and then adds this predicted value to the input sequence for predicting the value at the time point after that. This rolling prediction manner can capture the dynamic change trend of the health index. The prediction curve usually presents a smooth downward trend, but there will be an accelerated downward feature during the expected high load period. This prediction result provides a quantitative basis for equipment maintenance decision, so that the maintenance personnel can take preventive measures before the device health index falls to a dangerous level, thereby effectively prolonging the service life of the equipment and reducing the risk of failure.

[0038] S104, according to the short-term decay prediction curve of the health index, the comparison between the component fatigue damage index and the historical decay rate is analyzed, the particle filtering algorithm is used to estimate the aging rate distribution of the key component, and if the dynamic distribution of the aging rate deviates from the preset statistical range of the device group, it is judged that there is an abnormal out-of-group characteristic, and an out-of-group feature vector is generated.

[0039] According to the short-term decay prediction curve of the health index, the difference between the health indexes at adjacent time points is calculated as the instantaneous decay rate, the average value of the decay rate in the prediction period is calculated, the average value of the historical decay rate of the same type of equipment at the same running time is extracted from the historical database, and the ratio of the current average decay rate to the historical average decay rate is calculated. Get the decay rate ratio and the instantaneous decay rate sequence. The particle filtering algorithm is used to estimate the state of the instantaneous decay rate sequence, the particle population representing the possible values of the aging rate is initialized, the decay rate ratio is used as the observation value to update the weight of each particle, and the high weight particles are preserved by resampling and iteratively updated. The mean and variance of the particle population are calculated to obtain the distribution parameters of the aging rate of the key component. According to the distribution parameters of the aging rate, the statistical mean and standard deviation of the aging rate of the same type of equipment group are read from the database, the standard deviation multiple of the current equipment aging rate mean deviating from the group mean is calculated, and if the deviation multiple exceeds the preset threshold, it is judged that there is an abnormal out-of-group characteristic. The aging rate mean, the aging rate variance, the deviation multiple and the decay rate ratio are combined to generate an out-of-group feature vector.

[0040] Specifically, the calculation of the instantaneous decay rate is based on the change characteristics of the health index on the time axis. The short-term decay prediction curve of the health index presents a continuous downward trend, and the difference between adjacent time points reflects the degradation speed of the device health state. When the health index decreases from 85 at a certain time to 83 at the next time, the instantaneous decay rate is 2. This point-by-point calculation method can capture the subtle fluctuations of the health index change and provide accurate basic data for subsequent statistical analysis.

[0041] In one possible implementation, the extraction of historical degradation rate involves data mining of a large number of similar devices. The database stores the health index change records of devices with different service life, and by screening the historical data close to the current device service life, the average degradation rate of the life segment is calculated. For example, a transformer that has been running for 5 years has an average historical degradation rate of 1.5 health index points per month. The current device's degradation rate is compared with this historical benchmark. If the current degradation rate is 2.5 per month, the degradation rate ratio is 1.67, indicating that the current device is aging significantly faster than the average level of similar devices.

[0042] Specifically, the particle filter algorithm has unique advantages in estimating the aging rate. The algorithm initializes a large number of particles, each representing a possible aging rate value. These particles are uniformly distributed in the range of 0.5 to 3.0, covering various possibilities from slow aging to rapid aging. The observation value, i.e., the degradation rate ratio, is used to adjust the weight of each particle. The closer the ratio is to the aging rate represented by a particle, the greater the weight of the particle. Through multiple iterations and resampling, particles with smaller weights are eliminated, and the remaining particles gradually concentrate around the true aging rate.

[0043] It should be noted that the calculation process of the aging rate distribution parameters directly affects the accuracy of anomaly detection. After sufficient iteration of the particle population, its distribution shows a clear concentration trend. The mean reflects the most likely aging rate of the device, while the variance represents the uncertainty of this estimate. A smaller variance indicates that the aging rate estimate is more reliable, while a larger variance suggests that the device's operating state has greater volatility.

[0044] Illustratively, the judgment of out-of-gauge characteristics is based on statistical principles. The aging rate of the device population usually follows a normal distribution, with most devices' aging rates concentrated around the mean. When the aging rate of a device deviates from the group mean by more than 2 standard deviations, this is a statistically small probability event, suggesting that the device may have special degradation factors. The construction of the out-of-gauge feature vector integrates information from multiple dimensions: the aging rate mean directly reflects the degradation degree, the variance reflects the running stability, the deviation multiple quantifies the degree of abnormality, and the degradation rate ratio provides a time dimension comparison. The feature vector formed by these four elements provides a comprehensive quantitative basis for subsequent fault warning and maintenance decisions, enabling maintenance personnel to accurately identify abnormal devices that require special attention.

[0045] S105, for the out-of-gauge feature vector, an adaptive neural network is used to adjust the weight coefficients of the health state evaluation model, generate optimized weight parameters, and determine the optimized model structure.

[0046] According to the aging rate mean, the aging rate variance, the deviation multiple and the decay rate ratio in the outlier feature vector, the four components are input into the adaptive adjustment module of the health state evaluation model, the deviation of the current model output and the actual health index of the outlier device is calculated, and the error signal required for model adjustment is obtained. The gradient value of the connection weight of each layer is calculated by using the error signal through the back propagation algorithm, the step size parameter of gradient descent is set according to the size of the deviation multiple in the outlier feature vector, the larger the deviation multiple is, the larger the step size is, the weight matrix in the health state evaluation model is updated through iteration, and the iteration is stopped when the error signal is lower than the preset threshold value, and the optimized weight parameter adapted to the outlier feature is obtained. According to the optimized weight parameter, the network parameters of the health state evaluation model are reconfigured, the connection weight value and the bias item value between each layer are updated, the original number of layers and nodes are kept unchanged, the model structure adjusted for the outlier device characteristics is formed, and the optimized model structure is determined.

[0047] Specifically, the outlier feature vector is the core input of adaptive adjustment, and the four components each carry abnormal information of different dimensions. The aging rate mean reflects the speed of overall degradation of the device, and when this value is significantly higher than the average level of the group, it indicates that the device has an accelerated aging trend. The aging rate variance reveals the stability of the degradation process, and a high variance means that the device state fluctuates dramatically and may have intermittent failures. The deviation multiple directly quantifies the degree of abnormality, and the decay rate ratio provides a time-dimension reference. After the four-dimensional information is input into the adaptive adjustment module, the system can comprehensively evaluate the abnormal characteristics of the device.

[0048] In one possible implementation, the adaptive adjustment module of the health state evaluation model plays a key role. After receiving the outlier feature vector, the module first compares and analyzes it with the current output of the model. The model is originally trained based on the data of the normal device group, and when facing the outlier device, it often produces a large prediction deviation. This deviation is the basis for the model to adjust. The calculation process of the error signal involves the difference between the health index predicted by the model and the actual health index of the outlier device, which includes not only the numerical value but also the difference in change trend.

[0049] Specifically, the application of the back propagation algorithm in weight adjustment embodies the principle of gradient optimization. After the error signal is determined, the algorithm starts from the output layer and calculates the partial derivative of the error with respect to each connection weight, i.e., the gradient value, layer by layer. These gradient values indicate the direction and amplitude of weight adjustment. The key innovation lies in the dynamic setting of the step size parameter: the traditional method uses a fixed step size, while here it is dynamically adjusted according to the deviation multiple. When the deviation multiple is 2, the step size may be set to 0.1; when the deviation multiple reaches 3, the step size increases to 0.15. This adaptive mechanism enables the model to quickly respond to and adjust to severely abnormal devices.

[0050] It's important to note that the iterative update process of the weight matrix needs to balance convergence speed and stability. In each iteration, the new weights are equal to the old weights minus the product of the step size and the gradient. As the iterations proceed, the error signal gradually decreases. When it falls below a preset threshold, such as 0.01, the model is considered to have adequately adapted to the outlier characteristics. This process typically requires tens to hundreds of iterations, depending on the initial error magnitude and step size setting.

[0051] For example, determining the optimized model structure does not change the network topology, but rather updates the internal parameters while maintaining the architecture. The original model may contain an input layer, two hidden layers, and an output layer, with each layer connected by a weight matrix. The optimization process updates the values ​​of these weight matrices and adjusts the bias terms of each node. This parameter-level optimization allows the model to more accurately capture the special operating modes of outlier devices while retaining its ability to evaluate normal devices. Although the optimized model structure is formally the same as the original model, its internal mapping relationships have been specifically adjusted according to the outlier characteristics, achieving personalized adaptation of the model.

[0052] S106. The health status of the unified dataset is assessed through the optimized model structure. The health index score of the equipment is generated by combining the overload cycle count and the abnormal temperature fluctuation range, and the assessment basis for guiding the local maintenance of micro equipment is determined.

[0053] The unified dataset is processed using an optimized model structure. The model outputs a health status value at each time point, while simultaneously extracting the load over-limit cycle count and temperature anomaly fluctuation range values ​​for the corresponding time period. The health status value is multiplied by a base weighting coefficient, the load over-limit cycle count by a load influence coefficient, and the temperature anomaly fluctuation range by a temperature influence coefficient. These three values ​​are then summed and normalized to obtain a comprehensive health score sequence. Based on the numerical distribution in the comprehensive health score sequence, the percentage of time periods with scores below a first preset threshold is used as the degradation rate. The difference in scores between adjacent time points is calculated to identify periods of rapid deterioration. The mean of the comprehensive health score sequence, the degradation rate, and the deterioration period information are combined to generate an equipment health index score. For the mean score in the equipment health index score, if it is below the emergency maintenance threshold, it is marked as an emergency maintenance level; if it is between the emergency maintenance threshold and the planned maintenance threshold, it is marked as a planned maintenance level; and if it is above the planned maintenance threshold, it is marked as a routine inspection level. The maintenance level is combined with the corresponding recommended maintenance time interval to determine the evaluation criteria used to guide the local maintenance of micro-equipment.

[0054] Specifically, the optimized model structure plays the role of the core evaluation engine when processing the unified dataset. The unified dataset contains multi-dimensional features that have been time-aligned and dimension-reduced, which are input into the model in a standardized format. The neural network layers inside the model have been self-adaptively adjusted in advance, and the weight parameters have been optimized for the characteristics of outlier devices, so that they can more accurately capture the health state changes under abnormal patterns. The health state value, as the direct output of the model, is usually in the range of 0 to 100, reflecting the overall health level of the device at the current time.

[0055] In one possible implementation, the calculation of the comprehensive health score embodies the idea of multi-factor fusion evaluation. The basic weight coefficient is usually set to 0.6, indicating that the health state value output by the model is dominant. The load influence coefficient is determined according to the rated capacity and actual operation of the device. When the load overrun period count is high, it means that the device is frequently in an overload state, and its coefficient may reach 0.25. The temperature influence coefficient takes into account the cumulative damage of thermal stress to the service life of the device. The larger the range of abnormal temperature fluctuations, the higher the coefficient, and the typical value is 0.15. After the weighted sum of the three components, the final score is normalized to ensure that it falls within the standard interval.

[0056] Specifically, the statistics of the degradation rate provide a key basis for maintenance decisions. When 30% of the time points in the sequence of comprehensive health scores are below the first preset threshold of 60 points, the degradation rate is 0.3. This indicator directly reflects the proportion of time that the device is in a poor state. The identification of the rapid deterioration period focuses on short-term rapid degradation, such as a decrease in health score by more than 10 points within 5 consecutive sampling points. Such abnormal decline often indicates an impending failure. The score mean provides an overview of the overall health level, while the degradation rate and deterioration period information reveal the dynamic characteristics of the health state.

[0057] It should be noted that the division of maintenance levels adopts the concept of hierarchical management. The emergency maintenance threshold is usually set at 40 points. When the average of the device health index score is below this value, it indicates that the device is in a high-risk state and needs to be immediately arranged for maintenance. The planned maintenance threshold is set at 60 points. Devices between 40 and 60 points can still operate, but need to be maintained preventively during the planned downtime. Devices above 60 points are in a relatively healthy state and only need regular inspections.

[0058] Exemplarily, the determination of the maintenance suggestion time interval is closely related to the maintenance level. The device of the emergency maintenance level is suggested to arrange a maintenance window within 24 hours to prevent greater loss caused by sudden failure. The device of the planned maintenance level can be maintained in the future 7 to 14 days to take advantage of production gaps or low-load periods. The device of the routine inspection level is executed according to the standard monthly or quarterly inspection plan. Such differentiated maintenance strategies realize the optimal allocation of resources, which not only ensures that high-risk devices are timely processed, but also avoids excessive maintenance of healthy devices. The formation of the evaluation basis not only contains the quantitative health index and maintenance level, but also provides specific time suggestions, which provides clear guidance for on-site maintenance personnel to develop work plans.

[0059] S107, according to the health index score result, analyze the voltage instability duration and current mutation frequency statistics, adopt the conditional random field algorithm to predict the maintenance demand time window of the micro device in the specific running scene, generate the maintenance decision sequence for the local overload component, and determine the preventive maintenance execution plan.

[0060] According to the health index score result, extract the time period data with a score lower than a preset maintenance threshold, obtain the voltage instability duration and current mutation frequency statistics from the monitoring records of the corresponding period, identify the load level, environmental temperature and running time parameters of the running scene, arrange the health score, voltage instability duration, current mutation frequency, load level, environmental temperature and running time in time sequence, and construct an observation sequence. The conditional random field algorithm is used to model the observation sequence, the difference between the health scores of adjacent time points is taken as the state transition feature input, and the voltage instability duration, current mutation frequency and load level at the current time point are taken as the observation feature input. The feature weight parameters are learned by maximum likelihood estimation, and the probability value of the device from the current state to the fault state is calculated. When the probability value exceeds a preset threshold, the starting point of the maintenance demand time window is recorded, and the maintenance demand time window is obtained. For the maintenance demand time window, analyze the components corresponding to the monitoring points with the highest current mutation frequency in the window, sort the components according to the number of current mutations from high to low, generate a component maintenance priority sequence, combine the start and end time of the maintenance demand time window, the component priority sequence and the standard maintenance process corresponding to the component type, arrange the maintenance decision sequence according to time and priority, and determine the preventive maintenance execution plan.

[0061] Specifically, the construction of the observation sequence embodies the time series integration of multi-dimensional information. The periods with health index scores below the maintenance threshold in the result often correspond to abnormal operating states of the equipment, and the monitoring data of these periods contains key information of fault evolution. The duration of voltage instability reflects the continuous impact of power supply quality on the equipment. When the voltage deviates from the rated value beyond the allowed range and lasts for more than 10 minutes, it will cause cumulative damage to the insulation materials of the equipment. The frequency statistics of current mutation capture the occurrence law of instantaneous impact. Frequent current mutations mean that the equipment is subjected to repeated electromagnetic stress. The introduction of the operating scene parameter makes the analysis more comprehensive. High load level combined with high environmental temperature will accelerate the aging process of the equipment.

[0062] In one possible implementation, the conditional random field algorithm has unique advantages in time series prediction. Unlike the simple Markov model, it can consider both the observation features and the state transition relationship at multiple times. The state transition feature is obtained by calculating the difference between the health scores at adjacent times. When the health score decreases from 85 to 80, the difference is -5, indicating that the equipment state is deteriorating. The observation features include the actual measurement values at the current time, such as the duration of voltage instability reaching 15 minutes, the frequency of current mutation being 8 times per hour, and the load level being 90% of the rated capacity. These features are mapped to numerical values through a feature function and participate in probability calculation.

[0063] Specifically, maximum likelihood estimation plays a key role in parameter learning. The algorithm learns feature weights from known state transition cases in historical data, so that the model prediction results are most consistent with actual observations. Feature weights reflect the influence of each factor on state transition. The weight of current mutation frequency is usually high because it is directly related to the electrical stress of the equipment. When the model calculates that the probability of the equipment transferring to a fault state within the next 24 hours reaches 0.7, exceeding the preset threshold of 0.6, the system determines that maintenance needs to be arranged within this time window.

[0064] It should be noted that the determination of component maintenance priority is based on the fault risk assessment principle. The monitoring point with the highest current mutation frequency usually corresponds to the component that bears the most stress, such as the circuit breaker contact, the transformer winding connection point, and other key positions. By counting the number of mutations of each monitoring point within the time window, the damage degree of each component can be quantitatively evaluated. A monitoring point records 50 current mutations, while other monitoring points have an average of only 10, indicating that the component corresponding to this point needs urgent maintenance.

[0065] Exemplarily, the generation of the maintenance decision sequence comprehensively considers time constraints and resource optimization. The maintenance demand time window determines the urgency boundary of the maintenance, such as the maintenance must be completed within the next 48 hours. The component priority sequence guides the maintenance sequence, and the components with the highest risk are processed first. The standard maintenance process is predefined according to the component type, and the circuit breaker maintenance includes contact cleaning, spring mechanism inspection and other operations, which is expected to take 2 hours; the transformer local maintenance involves insulation testing, fastener inspection and other operations, which takes 3 hours. The maintenance decision sequence integrates these information to form an executable plan: the main contact of the circuit breaker is maintained in the first period, and the high-voltage side connection of the transformer is inspected in the second period. This fine maintenance arrangement minimizes the equipment downtime while ensuring that critical components are timely processed.

[0066] The above description is merely exemplary of one or more embodiments of the present description and is not intended to limit one or more embodiments of the present description. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of one or more embodiments of the present description should be included in the scope of protection of one or more embodiments of the present description.

Claims

1. A method for identifying power equipment inspection anomalies based on multi-dimensional data fusion, characterized in that, The method comprises: Obtaining historical operation data and real-time monitoring data, aligning and fusing through time series to generate a unified data set, obtaining multi-dimensional equipment operation characteristics; performing dimension reduction processing on the unified data set, extracting a key feature vector, analyzing the cumulative effect of overload impact, current mutation frequency and voltage instability duration, generating an impact event label, and forming a time series of impact influence; constructing a dynamic change trajectory of the equipment health index by combining the time series of impact influence, the temperature abnormal fluctuation range and the load overrun period count, establishing an equipment health state evaluation model, and generating a short-term decay prediction curve of the health index; according to the short-term decay prediction curve, analyzing the component fatigue damage index and the historical decay rate comparison, estimating the aging rate distribution of the key components, and generating an outlier feature vector; generating an optimized health state evaluation model structure by adaptively adjusting the weight coefficient of the health state evaluation model; performing health state evaluation on the unified data set through the optimized health state evaluation model structure, combining the load overrun period count and the temperature abnormal fluctuation range, and generating an equipment health index score result; according to the health index score result, the voltage instability duration and the current mutation frequency, predicting the maintenance requirement time window of the micro device in a specific operation scene, and generating a maintenance decision sequence. 2.The method according to claim 1, characterized in that, The method comprises: Extracting operation records containing voltage, current and temperature from a historical database, reading current voltage, current, temperature and load rate data from a real-time monitoring port, sorting according to time stamp, calculating the time stamp difference value of historical data and real-time data, determining the time offset and sampling frequency ratio, and generating time series alignment parameters; adjusting the historical data timestamp according to the time series alignment parameters, resampling the historical data by linear interpolation to make the sampling rate consistent with the real-time data, filling the missing data points with weighted average of the previous and next values, and generating a complete time-continuous data sequence; normalizing the voltage, current and temperature of the complete data sequence, calculating the distribution range of each parameter under different load rates, dividing according to load rate intervals, and constructing multi-dimensional equipment operation characteristics. 3.The method according to claim 1, characterized in that, The method comprises: Constructing a covariance matrix for the unified data set, calculating eigenvalues and eigenvectors, sorting by eigenvalue size, selecting eigenvectors with a cumulative variance contribution rate reaching a preset proportion as principal components, and generating a key feature vector through matrix multiplication. 4.The method of claim 1, wherein, The method comprises: According to the time sequence of the impact effect, temperature data of the impact event period is extracted, the amplitude of temperature deviation from the normal range and the length of time for returning to a stable value are calculated, the number of times that the temperature fluctuation exceeds a threshold value is counted, and a temperature abnormal fluctuation range parameter is generated; from the load data, an over-limit period is identified, the over-limit length and interval length are counted, the load over-limit period count is calculated in daily and weekly dimensions, and the dynamic change trajectory of the device health index is generated in combination with the temperature abnormal fluctuation range parameter and the impact event label.

5. The method according to claim 1, wherein, According to the short-term attenuation prediction curve, the component fatigue damage index and the historical attenuation rate are compared and analyzed, the key component aging rate distribution is estimated, and an outlier feature vector is generated, including: According to the short-term attenuation prediction curve, the health index difference value at adjacent time points is calculated as the instantaneous attenuation rate, the average value of the attenuation rate in the prediction period is counted, the historical attenuation rate average value of the same type of device is extracted from the historical database, and the ratio of the instantaneous attenuation rate to the historical average value is calculated; the state of the instantaneous attenuation rate sequence is estimated, the particle population representing the aging rate is initialized, the particle weight is updated, the average value and variance are calculated, and the aging rate distribution parameter is generated; according to the aging rate distribution parameter, the statistical average value and standard deviation of the device group aging rate are read, the deviation multiple is calculated, and the aging rate average value, variance and deviation multiple are combined to generate an outlier feature vector. 6.The method of claim 1, wherein, The weight coefficient of the health state evaluation model is adjusted adaptively, and an optimized health state evaluation model structure is generated, including: According to the outlier feature vector, the deviation between the health state evaluation model output and the actual health index is calculated, and an error signal is generated; the weight gradient value is calculated according to the error signal, the gradient step size is set, the weight matrix is updated, and the optimized weight parameter is generated; the connection weight and bias term of the model are updated according to the optimized weight parameter, and the number of layers and nodes is kept unchanged. 7.The method of claim 1, wherein, The health state of the unified data set is evaluated through the optimized health state evaluation model structure, and a device health index score result is generated, including: The unified data set is processed through the optimized health state evaluation model structure, and the health state value is output, the load over-limit period count and the temperature abnormal fluctuation range are combined, and a comprehensive health score sequence is calculated; the proportion of the period in the comprehensive health score sequence that is below the threshold value and the score difference value are counted, the score average value and the degradation information are combined, and a health index score result is generated. 8.The method according to claim 1, characterized in that, According to the health index score result, the voltage instability duration and the current mutation frequency, the maintenance requirement time window of the micro device in a specific operating scenario is predicted, and a maintenance decision sequence is generated, including: According to the health index score result, the period data below the maintenance threshold value is extracted, the voltage instability duration and the current mutation frequency are obtained, the load level and the environment parameters are combined, and an observation sequence is constructed; the observation sequence is modeled, the state transition probability is calculated, and the time point that exceeds the threshold value is recorded as the starting point of the maintenance requirement time window; according to the maintenance requirement time window, the number of component current mutations is sorted, and a component maintenance priority sequence and a maintenance decision sequence are generated.

Images