Power equipment fault prediction and early warning method based on edge calculation

Abstract

The invention discloses a power equipment fault prediction and early warning method based on edge calculation, and the method comprises the following steps: S1, collecting voltage, current, temperature, frequency and other parameters, carrying out the data cleaning and normalization, and generating standardized time series data; s2, dividing the time sequence data into a plurality of subsequences through a sliding window; s3, extracting trend characteristics of each sub-sequence, and calculating a disturbance value reflecting trend change; s4, constructing a cutting model with a priority division strategy in combination with the disturbance value, and obtaining an abnormal degree score of each sub-sequence; s5, performing weighted fusion on the trend disturbance value and the abnormal score, and outputting a prediction score; s6, according to a comparison result of the prediction score and a threshold value, generating early warning information containing time, a serial number and an abnormal level; and S7, packing and packaging the early warning information on the equipment site, and sending the early warning information to a superior system or a local terminal. According to the invention, on-site rapid prediction and graded early warning of potential faults of power equipment can be realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for power systems, and in particular to a method for predicting and warning of power equipment faults based on edge computing. Background Technology

[0002] With the continuous improvement of the intelligence and digitalization of power systems, the monitoring of the operating status and fault early warning of power equipment have become important means to ensure the stability and security of power supply. Against the backdrop of the development of multi-source sensing and distributed computing technologies, edge computing-based equipment management has gradually gained attention. It possesses real-time on-site processing and rapid response capabilities, making it suitable for power application scenarios with high requirements for timeliness and stability.

[0003] In existing technologies, power equipment fault prediction and early warning methods mostly rely on centralized architectures and rule-based models. These traditional methods still have the following limitations in application: Limited response timeliness: Traditional methods generally rely on transmitting large amounts of raw data back to the cloud or central platform for processing. Limited by network bandwidth and transmission latency, they cannot meet the needs of rapid on-site early warning. Limited dimensions of data anomaly identification: Most existing methods rely on fixed thresholds or statistical models, which cannot fully characterize the trend changes and potential disturbances in equipment operating parameters, making it difficult to identify some early fault signals. Insufficient model robustness: Under complex operating conditions, anomaly detection models are easily affected by fluctuations in single anomaly indicators, lacking the ability to jointly model trend and structural anomalies, affecting prediction accuracy. Incomplete early warning information structure: Some methods output relatively simple results, only providing a judgment of whether an anomaly exists, lacking more detailed anomaly level classification and structured output mechanisms, limiting the targeting and efficiency of subsequent response strategies.

[0004] Therefore, how to provide a method for predicting and warning power equipment faults based on edge computing is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a power equipment fault prediction and early warning method based on edge computing. This invention combines trend disturbance analysis and an improved random cutting model, and describes in detail the entire process of multi-parameter time series data processing, fault trend identification, and anomaly level determination at edge nodes. It has the advantages of strong real-time performance, high identification accuracy, and high degree of structured early warning information.

[0006] A method for predicting and warning power equipment faults based on edge computing according to an embodiment of the present invention includes the following steps:

[0007] S1. Collect voltage, current, temperature and frequency parameters, perform filtering and normalization processing, and generate a continuous time series data sequence;

[0008] S2. Perform a sliding window segmentation operation on the continuous time series data sequence to form a set of observation subsequences of fixed length;

[0009] S3. Input the set of observed subsequences into the Hearst exponent anomaly detection algorithm, calculate the Hearst exponent for each observed subsequence, and output the trend perturbation value that represents the degree of trend change.

[0010] S4. Input the set of observed subsequences and the corresponding trend perturbation values ​​into the improved robust random cut forest model, adjust the cutting priority of data points according to the trend perturbation values, construct a density-weighted cutting tree, and output the isolation score of each observed subsequence.

[0011] S5. Extract the trend disturbance value and isolation score, perform normalization processing on each, perform weighted fusion according to the set weight, and output the prediction score used to characterize potential faults.

[0012] S6. Based on the comparison between the predicted score and the preset threshold, generate fault warning information including timestamp, observation subsequence number and anomaly level;

[0013] S7. In the edge computing node, the fault warning information is encapsulated and processed, and then output to the external system or local device through the communication interface.

[0014] This invention proposes a power equipment fault prediction and early warning method based on edge computing. It establishes a complete process from multi-parameter time-series data acquisition, sliding window segmentation, trend disturbance mining, improved random forest anomaly modeling, trend and structure fusion prediction, to edge early warning encapsulation and output. This effectively solves the problems of insufficient early detection of equipment trend anomalies, inaccurate structural fault identification, and limited real-time and local response capabilities in existing technologies. This method innovatively uses the Hearst exponent to capture changes in equipment operating trends and introduces a disturbance-driven mechanism to optimize the forest modeling process. It improves anomaly judgment accuracy through multi-dimensional index normalization and weighting, and finally achieves rapid output of multi-level, structured fault early warning information at edge nodes. It possesses advantages such as high accuracy, strong adaptability, and low latency, significantly improving the intelligent operation and maintenance level and equipment fault prevention capabilities of power systems.

[0015] Optionally, S1 specifically includes:

[0016] S11. Set up edge acquisition terminals at the power equipment operation site to synchronously acquire four types of physical quantities, namely three-phase voltage, three-phase current, ambient temperature and power grid frequency, through wired or wireless means. Collect continuously at a fixed sampling frequency and generate the original parameter data sequence by arranging them in ascending order of timestamp.

[0017] S12. Divide the original parameter data sequence into multi-channel sub-sequences according to parameter type, and perform median filtering on each channel to filter out abrupt changes and high-frequency discrete points in each sampling window, forming a filtered and smoothed parameter channel sequence.

[0018] S13. For each channel data in the filter smoothing parameter channel sequence, calculate the minimum and maximum values ​​of the channel within a preset observation time window, perform linear normalization according to the proportional conversion rule, compress all channel values ​​to a uniform numerical scale range, and form a normalized channel data sequence set.

[0019] S14. Using each sampling time as a reference, concatenate all the normalized channel data corresponding to the sampling time into a feature vector, arrange all feature vectors in chronological order, and construct a standardized continuous time series data sequence.

[0020] By deploying edge acquisition terminals at the power equipment site, multi-source physical quantities such as three-phase voltage, three-phase current, ambient temperature, and grid frequency are acquired in real time synchronously using wired or wireless methods. Data is continuously acquired at a fixed sampling frequency. Claim 2 specifies the data type and acquisition order in detail to ensure the comprehensiveness and temporal consistency of the original data. Combining multi-channel partitioning and median filtering effectively removes abrupt changes and high-frequency noise, improving data stability and accuracy. Based on linear normalization and feature vector concatenation, all sampled data points are uniformly mapped to a standardized scale range, eliminating the influence of dimensions between channels and ensuring the effectiveness of subsequent feature modeling and trend analysis. This implementation provides a high-quality, low-noise, and standardized time-series input data foundation for subsequent anomaly detection algorithms, significantly improving the system's sensitivity to changes in the operating status of power equipment and the reliability of fault detection.

[0021] Optionally, S2 specifically includes:

[0022] S21. Extract the parameter feature vector sequence from the standardized continuous time series data sequence in chronological order, set a fixed window length and step size parameter, and determine the starting index sequence of the sliding window;

[0023] S22. Based on the starting index of each sliding window, select a corresponding number of feature vectors to form an observation subsequence. All observation subsequences have the same length and the parameter arrangement order is consistent, forming a set of observation subsequences with temporal continuity.

[0024] S23. During each slide, advance the starting index of the previous window by one step, and repeat the feature vector combination operation until all standardized continuous time series data are traversed to form a complete set of observation subsequences, with some overlap between the subsequences in the set.

[0025] S24. Number each observation subsequence according to the sampling order and mark its starting time point in the original time series data to generate a set of observation subsequences.

[0026] By segmenting standardized continuous time-series data into a sliding window with a fixed window length and step size, the length and parameter arrangement order of each observation subsequence are strictly limited, ensuring the temporal continuity and structural consistency of the data segments. During the window sliding process, partial overlap of subsequences is achieved, and each observation subsequence is uniquely numbered and time-stamped, effectively preserving the dynamic evolution information and location attributes of the time-series data. This implementation not only improves the utilization rate of the original data but also provides a high-resolution, multi-granular input foundation for subsequent trend analysis and anomaly detection, enhancing the model's ability to capture subtle state changes and effectively supporting refined modeling of power equipment operating status and early fault identification.

[0027] Optionally, S3 specifically includes:

[0028] S31. Extract the parameter sequences from the three-phase voltage and three-phase current channels for each observation subsequence in the observation subsequence set to form a trend analysis input sequence set. Each sequence in the trend analysis input sequence set maintains its original time order.

[0029] S32. For each parameter sequence in the trend analysis input sequence set, set three sets of unequally spaced increasing scale segments, perform segmented calculations in sequence, determine the sampling length of each scale segment, and form a scale segment set {n1,n2,n3}.

[0030] S33. Within each scale segment, divide the sample into segments, calculate the range R and standard deviation S within each segment, and obtain their ratio B = R / S. Take the logarithm of the scale length and the ratio to construct a point set (log(n), log(B)). Perform least-squares linear fitting on all point sets, and define the slope as the single-channel Hurst exponent H of the parameter sequence. j ;

[0031] S34. For each observed subsequence, calculate the set of Hearst exponents {H} of all its trend analysis input sequences. j} Calculate the channel maximum value, channel average value, and mean square deviation value D. i Construct the trend disturbance value T i The calculation formula is as follows:

[0032] T i =(max(H) j )-mean(H j ))+D i ;

[0033] S35. Match the observation subsequence number i with the corresponding trend disturbance value T. i Composition of trend disturbance label pairs (i,T) i All trend disturbance identifiers constitute a set of trend disturbance values ​​in chronological order.

[0034] By extracting time-series data of three-phase voltage and three-phase current for each observation subsequence and setting three sets of unequally spaced increasing scale segments, the range and standard deviation at each scale are calculated segment by segment. Claim 4 achieves a refined modeling of the long-term dependence and trend changes of multi-channel data. Least-squares linear fitting is performed using a logarithmic point set of scale length and range ratio to obtain the Hearst exponent reflecting the trend of the sequence. Furthermore, the maximum value, average value, and mean square deviation of each channel are statistically analyzed to comprehensively construct trend perturbation values. By associating the trend perturbation values ​​with the observation subsequence numbers, a trend perturbation set arranged in time series is formed, effectively improving the sensitivity to slow-changing anomalies and trend-type faults in power equipment. This lays a solid data foundation for subsequent models to accurately capture the evolution of potential faults and enhances the system's early warning capability for trend anomalies under complex operating conditions.

[0035] Optionally, S4 specifically includes:

[0036] S41. Construct each observation subsequence into a fixed-length feature vector X containing all parameter channels. i and its corresponding trend disturbance value T i Added as a perturbation factor, generating perturbation samples (X) i ,T i This forms a set of perturbed samples;

[0037] S42. Arrange the disturbance sample set according to the trend disturbance value T. i The samples are sorted from largest to smallest. The sorted samples are used to determine the processing order when building the tree. Samples with larger perturbation values ​​are processed first and assigned to the left sub-branch first, forming a trend-driven left-side splitting path.

[0038] S43. Select samples sequentially from the sorted perturbation sample set to construct a single cutting tree. The construction process of the single cutting tree includes the following operations:

[0039] S431. Extract all samples to be split in the current tree node, set all parameter channels as candidate cutting dimensions, and calculate the perturbation-weighted variance V for each dimension. d The definition is as follows:

[0040]

[0041] Among them, X i,d Let be the value of sample i in the d-th dimension. This represents the weighted mean along the d-dimensional axis;

[0042] S432. Select the dimension with the largest perturbation weighted variance as the current cutting axis, and randomly generate a cutting point p between the maximum and minimum values ​​of the dimension. d ;

[0043] S433, All samples with cutting axis values ​​less than or equal to p d The sample value is assigned to the left child node, and the remaining samples are assigned to the right child node; when the sample value is within the tolerance range near the cut point [p d -∈,p d When the value of the trend disturbance is within the range of +∈], it is preferentially assigned to the left child node according to the magnitude of the trend disturbance value, forming a trend-dominated leftward propagation structure;

[0044] S434. Repeat the feature vector extraction, perturbation weighted variance calculation, cutting axis selection, cutting point generation, and sample partitioning operations for the left and right child nodes respectively, until the leaf node contains only one sample, thus completing the construction of a single cutting tree.

[0045] S44. Based on the same perturbation sample set, repeat the pruning tree construction process to generate k pruning trees with perturbation sorting and left propagation mechanisms, and combine them to form an improved robust random pruning forest model.

[0046] S45. In the robust random cut forest model, record the path length traversed by each observed subsequence in all cut trees, denoted as . Where j is the pruning tree number, the average path length is calculated as follows:

[0047]

[0048] S46. Calculate the isolation score S of the observed subsequence. i The definition is as follows: Where n represents the total number of samples, and log2(n) is the logarithmic baseline of the theoretical expected path length;

[0049] S47. Number each observation subsequence and associate it with the corresponding isolation score S. i The isolation degree output pairs are combined to form the isolation degree score set.

[0050] By transforming each observed subsequence into a perturbed sample containing all parameter channels and corresponding trend perturbation values, and sorting the samples according to the magnitude of the perturbation values, this invention innovatively introduces a trend-driven splitting priority in the forest construction stage. Samples with larger perturbation values ​​participate in leftward splits first, effectively focusing on potential trend anomalies. During the splitting process, perturbation-weighted variance is used as the selection criterion for splitting dimensions, making the splitting decision more sensitive to trend features. Furthermore, perturbation values ​​are prioritized for classification within the neighborhood of the splitting point, realizing a trend-driven leftward propagation mechanism. This process forms an overall forest structure after multiple rounds of tree construction. The isolation score is calculated by combining path length and theoretical expectation, comprehensively reflecting the degree of anomaly isolation of each observed subsequence in multidimensional space. The above implementation significantly enhances the model's ability to identify combined trend and structural anomalies, improves the efficiency and accuracy of capturing potential faults in power equipment, and enhances the targeting and reliability of early warning systems.

[0051] Optionally, S5 specifically includes:

[0052] S51. Extract the trend perturbation value and isolation score of each observation subsequence, arrange them in order according to the observation subsequence number, and form the trend index sequence and the structure index sequence respectively.

[0053] S52. Perform normalization processing on the trend indicator sequence and the structure indicator sequence respectively to generate normalized trend components and normalized structure components.

[0054] S53. Set the weight coefficients of the trend component and the weight coefficients of the structure component, and the sum of the two is 1. Multiply them by the normalized trend component and the normalized structure component respectively, perform weighted summation, and generate the fusion prediction score corresponding to each observation subsequence.

[0055] S54. Set a prediction score threshold for judging whether the observed subsequence is abnormal, compare each fusion prediction score with the threshold, and if the fusion prediction score is greater than or equal to the prediction score threshold, the observed subsequence is marked as abnormal; otherwise, it is marked as normal.

[0056] S55. Output the fused prediction score as a numerical result to characterize the potential fault state.

[0057] By normalizing the trend perturbation value and isolation score of each observed subsequence and setting weighting coefficients to weight and fuse the trend and structural components, this invention achieves information integration of multi-dimensional indicators and dynamic adjustment of prediction scores. By setting flexible anomaly detection thresholds, subsequences with fused scores exceeding the threshold are marked as anomalies, enabling the fault identification process to consider both trend-type and structural anomalies while possessing adaptability and interpretability. This implementation improves the sensitivity and accuracy of fault detection, effectively reduces the probability of false alarms and missed alarms, and provides a more scientific and refined quantitative basis for early warning and risk classification of power equipment.

[0058] Optionally, S6 specifically includes:

[0059] S61. Extract the fusion prediction score p for each observation subsequence. i Subsequence number i and its start timestamp t in the standardized continuous time series data sequence i This constitutes the set of fusion prediction mapping triples (i,t) i ,p i );

[0060] S62. Set three level thresholds θ1, θ2, and θ3 locally on the edge computing node, satisfying θ1 > θ2 > θ3, and apply these thresholds to each fusion prediction score p. i The abnormality level L is assigned as follows: i :

[0061]

[0062] S63, transfer the corresponding number i and timestamp t i and abnormality level L i Combined construction of structured fault early warning record triples (i,t) i ,L i );

[0063] S64. Sort all generated fault warning record triplets in ascending order of number to generate a fault warning information set containing timestamp, observation subsequence number and anomaly level.

[0064] This invention extracts and maps the fusion prediction score, number, and start timestamp of each observed subsequence, and sets multi-level judgment thresholds at the edge computing nodes to assign the prediction scores of each subsequence to different anomaly levels. This generates structured fault warning record triples, which are then systematically aggregated into a fault warning information set according to their numbers. This implementation achieves multi-level classification and localized judgment of fault risks, facilitating rapid subsequent tracing and batch management. It improves the response speed and discrimination capability of the early warning system to changes in equipment status, making power equipment anomaly warnings more detailed, practical, and efficient.

[0065] Optionally, S7 specifically includes:

[0066] S71. Configure a local exception cache structure inside the edge computing node, and record a triplet (i,t) for each fault warning. i ,L i The system receives data in real time and writes it as a key-value pair (t) with key i as the number and value as the value. i ,L i In the hash cache table structure;

[0067] S72, Read the current system local timestamp t sys The device unique identifier code e in the edge node configuration file id As global identifier metadata, it is appended to each cache record to construct a 5-tuple (i,t). i ,L i ,t sys ,e id );

[0068] S73. Set the task cycle of the timer scheduler to ΔT. In each scheduling cycle, extract all five-tuple records from the hash cache table in batches and sort them in ascending order according to the number i to generate a structured output queue of abnormal events.

[0069] S74. Encode each sorted quintuple into a data frame of a uniform format. The data frame structure includes a frame header identifier field, a timestamp field, a subsequence number field, an anomaly level field, and a node identifier field. Each field has a fixed length and is encapsulated into a binary transmission unit using byte alignment.

[0070] S75. Call the physical communication interface controller configured in the edge computing node, write the generated data frames one by one into the transmission buffer register, and transmit them frame by frame through the preset serial port or Ethernet port. The target address includes the upper-level monitoring platform and the local control module. The transmission method is unicast or multicast.

[0071] By constructing a local anomaly cache and hash table structure within the edge computing node, fault warning records are received and stored in real time. Timestamps and unique device identifiers are integrated into a five-tuple, claim 8 achieves standardized and efficient management of multi-source anomaly information. Combined with timed batch extraction, unified format data frame encoding, and byte-aligned binary encapsulation, the reliability of data transmission and system compatibility are greatly improved. The frame-by-frame output of warning data and efficient integration with the upper-level platform and control module via a physical communication interface ensures real-time push of anomaly warnings, enhances the system's network adaptability and engineering availability, and provides solid support for intelligent edge operation and maintenance and monitoring of large-scale power equipment.

[0072] The beneficial effects of this invention are:

[0073] This invention introduces the Hearst exponent anomaly detection method, which can perform trend disturbance analysis on key parameter sequences such as voltage and current, quantify the degree of trend change in equipment operation, effectively identify early slow-change anomalies, enhance the forward-looking judgment ability of potential fault symptoms, and make up for the weakness of traditional methods in identifying trend faults.

[0074] This invention designs an improved robust random cutting forest model, dynamically adjusts the data cutting order and weights by combining trend perturbation values, and constructs a cutting tree structure with a trend-driven mechanism. This not only captures the isolation characteristics of data, but also improves the model's adaptability to complex data structures and enhances the stability and accuracy of anomaly detection results.

[0075] This invention constructs a multi-factor weighted fault prediction scoring mechanism by integrating trend perturbation values ​​and isolation scores, overcoming the false alarm and false negative problems caused by judging a single abnormal indicator, and improving the generalization ability and robustness of the prediction model in different scenarios. Attached Figure Description

[0076] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0077] Figure 1 This is a flowchart of a power equipment fault prediction and early warning method based on edge computing proposed in this invention;

[0078] Figure 2 This is a schematic diagram of the calculation process of the Hearst exponent anomaly detection algorithm proposed in this invention.

[0079] Figure 3 This is a schematic diagram illustrating the construction of the improved robust random cutting forest model proposed in this invention. Detailed Implementation

[0080] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0081] refer to Figure 1-3 A method for predicting and warning power equipment faults based on edge computing includes the following steps:

[0082] S1. Collect voltage, current, temperature and frequency parameters, perform filtering and normalization processing, and generate a continuous time series data sequence;

[0083] S2. Perform a sliding window segmentation operation on the continuous time series data sequence to form a set of observation subsequences of fixed length;

[0084] S3. Input the set of observed subsequences into the Hearst exponent anomaly detection algorithm, calculate the Hearst exponent for each observed subsequence, and output the trend perturbation value that represents the degree of trend change.

[0085] S4. Input the set of observed subsequences and the corresponding trend perturbation values ​​into the improved robust random cut forest model, adjust the cutting priority of data points according to the trend perturbation values, construct a density-weighted cutting tree, and output the isolation score of each observed subsequence.

[0086] S5. Extract the trend disturbance value and isolation score, perform normalization processing on each, perform weighted fusion according to the set weight, and output the prediction score used to characterize potential faults.

[0087] S6. Based on the comparison between the predicted score and the preset threshold, generate fault warning information including timestamp, observation subsequence number and anomaly level;

[0088] S7. In the edge computing node, the fault warning information is encapsulated and processed, and then output to the external system or local device through the communication interface.

[0089] In this embodiment, S1 specifically includes:

[0090] S11. Set up edge acquisition terminals at the power equipment operation site to synchronously acquire four types of physical quantities, namely three-phase voltage, three-phase current, ambient temperature and power grid frequency, through wired or wireless means. Collect continuously at a fixed sampling frequency and generate the original parameter data sequence by arranging them in ascending order of timestamp.

[0091] S12. Divide the original parameter data sequence into multi-channel sub-sequences according to parameter type, and perform median filtering on each channel to filter out abrupt changes and high-frequency discrete points in each sampling window, forming a filtered and smoothed parameter channel sequence.

[0092] S13. For each channel data in the filter smoothing parameter channel sequence, calculate the minimum and maximum values ​​of the channel within a preset observation time window, perform linear normalization according to the proportional conversion rule, compress all channel values ​​to a uniform numerical scale range, and form a normalized channel data sequence set.

[0093] S14. Using each sampling time as a reference, concatenate all the normalized channel data corresponding to the sampling time into a feature vector, arrange all feature vectors in chronological order, and construct a standardized continuous time series data sequence.

[0094] In this embodiment, S2 specifically includes:

[0095] S21. Extract the parameter feature vector sequence from the standardized continuous time series data sequence in chronological order, set a fixed window length and step size parameter, and determine the starting index sequence of the sliding window;

[0096] S22. Based on the starting index of each sliding window, select a corresponding number of feature vectors to form an observation subsequence. All observation subsequences have the same length and the parameter arrangement order is consistent, forming a set of observation subsequences with temporal continuity.

[0097] S23. During each slide, advance the starting index of the previous window by one step, and repeat the feature vector combination operation until all standardized continuous time series data are traversed to form a complete set of observation subsequences, with some overlap between the subsequences in the set.

[0098] S24. Number each observation subsequence according to the sampling order and mark its starting time point in the original time series data to generate a set of observation subsequences.

[0099] In this embodiment, S3 specifically includes:

[0100] S31. Extract the parameter sequences from the three-phase voltage and three-phase current channels for each observation subsequence in the observation subsequence set to form a trend analysis input sequence set. Each sequence in the trend analysis input sequence set maintains its original time order.

[0101] S32. For each parameter sequence in the trend analysis input sequence set, set three sets of unequally spaced increasing scale segments, perform segmented calculations in sequence, determine the sampling length of each scale segment, and form a scale segment set {n1,n2,n3}.

[0102] S33. Within each scale segment, divide the sample into segments, calculate the range R and standard deviation S within each segment, and obtain their ratio B = R / S. Take the logarithm of the scale length and the ratio to construct a point set (log(n), log(B)). Perform least-squares linear fitting on all point sets, and define the slope as the single-channel Hurst exponent H of the parameter sequence. j ;

[0103] S34. For each observed subsequence, calculate the set of Hearst exponents {H} of all its trend analysis input sequences. j} Calculate the channel maximum value, channel average value, and mean square deviation value D. i Construct the trend disturbance value T i The calculation formula is as follows:

[0104] T i =(max(H) j)-mean(H j ))+D i ;

[0105] S35. Match the observation subsequence number i with the corresponding trend disturbance value T. i Composition of trend disturbance label pairs (i,T) i All trend disturbance identifiers constitute a set of trend disturbance values ​​in chronological order.

[0106] In this embodiment, S4 specifically includes:

[0107] S41. Construct each observation subsequence into a fixed-length feature vector X containing all parameter channels. i and its corresponding trend disturbance value T i Added as a perturbation factor, generating perturbation samples (X) i ,T i This forms a set of perturbed samples;

[0108] S42. Arrange the disturbance sample set according to the trend disturbance value T. i The samples are sorted from largest to smallest. The sorted samples are used to determine the processing order when building the tree. Samples with larger perturbation values ​​are processed first and assigned to the left sub-branch first, forming a trend-driven left-side splitting path.

[0109] S43. Select samples sequentially from the sorted perturbation sample set to construct a single cutting tree. The construction process of the single cutting tree includes the following operations:

[0110] S431. Extract all samples to be split in the current tree node, set all parameter channels as candidate cutting dimensions, and calculate the perturbation-weighted variance V for each dimension. d The definition is as follows:

[0111]

[0112] Among them, X i,d Let be the value of sample i in the d-th dimension. This represents the weighted mean along the d-dimensional axis;

[0113] S432. Select the dimension with the largest perturbation weighted variance as the current cutting axis, and randomly generate a cutting point p between the maximum and minimum values ​​of the dimension. d ;

[0114] S433, All samples with cutting axis values ​​less than or equal to p d The sample value is assigned to the left child node, and the remaining samples are assigned to the right child node; when the sample value is within the tolerance range near the cut point [p d -∈,p dWhen the value of the trend disturbance is within the range of +∈], it is preferentially assigned to the left child node according to the magnitude of the trend disturbance value, forming a trend-dominated leftward propagation structure;

[0115] S434. Repeat the feature vector extraction, perturbation weighted variance calculation, cutting axis selection, cutting point generation, and sample partitioning operations for the left and right child nodes respectively, until the leaf node contains only one sample, thus completing the construction of a single cutting tree.

[0116] S44. Based on the same perturbation sample set, repeat the pruning tree construction process to generate k pruning trees with perturbation sorting and left propagation mechanisms, and combine them to form an improved robust random pruning forest model.

[0117] S45. In the robust random cut forest model, record the path length traversed by each observed subsequence in all cut trees, denoted as . Where j is the pruning tree number, the average path length is calculated as follows:

[0118]

[0119] S46. Calculate the isolation score S of the observed subsequence. i The definition is as follows: Where n represents the total number of samples, and log2(n) is the logarithmic baseline of the theoretical expected path length;

[0120] S47. Number each observation subsequence and associate it with the corresponding isolation score S. i The isolation degree output pairs are combined to form the isolation degree score set.

[0121] In this embodiment, S5 specifically includes:

[0122] S51. Extract the trend perturbation value and isolation score of each observation subsequence, arrange them in order according to the observation subsequence number, and form the trend index sequence and the structure index sequence respectively.

[0123] S52. Perform normalization processing on the trend indicator sequence and the structure indicator sequence respectively to generate normalized trend components and normalized structure components.

[0124] S53. Set the weight coefficients of the trend component and the weight coefficients of the structure component, and the sum of the two is 1. Multiply them by the normalized trend component and the normalized structure component respectively, perform weighted summation, and generate the fusion prediction score corresponding to each observation subsequence.

[0125] S54. Set a prediction score threshold for judging whether the observed subsequence is abnormal, compare each fusion prediction score with the threshold, and if the fusion prediction score is greater than or equal to the prediction score threshold, the observed subsequence is marked as abnormal; otherwise, it is marked as normal.

[0126] S55. Output the fused prediction score as a numerical result to characterize the potential fault state.

[0127] In this embodiment, S6 specifically includes:

[0128] S61. Extract the fusion prediction score p for each observation subsequence. i Subsequence number i and its start timestamp t in the standardized continuous time series data sequence i This constitutes the set of fusion prediction mapping triples (i,t) i ,p i );

[0129] S62. Set three level thresholds θ1, θ2, and θ3 locally on the edge computing node, satisfying θ1 > θ2 > θ3, and apply these thresholds to each fusion prediction score p. i The abnormality level L is assigned as follows: i :

[0130]

[0131] S63, transfer the corresponding number i and timestamp t i and abnormality level L i Combined construction of structured fault early warning record triples (i,t) i ,L i );

[0132] S64. Sort all generated fault warning record triplets in ascending order of number to generate a fault warning information set containing timestamp, observation subsequence number and anomaly level.

[0133] In this embodiment, S7 specifically includes:

[0134] S71. Configure a local exception cache structure inside the edge computing node, and record a triplet (i,t) for each fault warning. i ,L i The system receives data in real time and writes it as a key-value pair (t) with key i as the number and value as the value. i ,L i In the hash cache table structure;

[0135] S72, Read the current system local timestamp t sys The device unique identifier code e in the edge node configuration file id As global identifier metadata, it is appended to each cache record to construct a 5-tuple (i,t). i ,L i ,t sys ,e id );

[0136] S73. Set the task cycle of the timer scheduler to ΔT. In each scheduling cycle, extract all five-tuple records from the hash cache table in batches and sort them in ascending order according to the number i to generate a structured output queue of abnormal events.

[0137] S74. Encode each sorted quintuple into a data frame of a uniform format. The data frame structure includes a frame header identifier field, a timestamp field, a subsequence number field, an anomaly level field, and a node identifier field. Each field has a fixed length and is encapsulated into a binary transmission unit using byte alignment.

[0138] S75. Call the physical communication interface controller configured in the edge computing node, write the generated data frames one by one into the transmission buffer register, and transmit them frame by frame through the preset serial port or Ethernet port. The target address includes the upper-level monitoring platform and the local control module. The transmission method is unicast or multicast.

[0139] Example 1:

[0140] To verify the feasibility and practical application value of this invention, it was deployed for field testing at a 110kV urban substation in a certain province. This substation serves multiple surrounding residential areas and industrial enterprises. The equipment operates in an environment with strong electromagnetic interference and frequent load fluctuations, making it a typical medium-voltage, high-density power transmission and distribution node. The substation's equipment includes two main transformers, three voltage transformers, seven high-voltage switch units, and multiple busbar connection devices. Accurate monitoring of its operating status directly affects the stability of the power supply to the entire area.

[0141] The station's original condition monitoring system primarily employed a centralized architecture, requiring data to be aggregated to a central platform via fiber optic cables. Anomalies were then identified manually using thresholds or simple rules. This approach suffered from data lag, slow response, and weak trend identification capabilities, particularly performing poorly in handling gradual or combined anomalies. To improve the real-time performance and intelligence of the monitoring system, the station decided to introduce the edge computing-based power equipment fault prediction and early warning method proposed in this invention.

[0142] During the transformation process, edge acquisition terminals were deployed on key node devices. These terminals can simultaneously acquire four types of physical quantities: voltage, current, temperature, and frequency. The sampling frequency is 1Hz, and a local median filter module is configured to filter out high-frequency noise and abrupt changes. After linear normalization, the data is uniformly formed into multi-channel standardized time-series data. The system deploys the processing flow proposed in this invention on local edge nodes, including observation subsequence construction, Hearst exponent anomaly detection, construction of an improved random cut forest model, generation of fused prediction scores, and structured early warning output.

[0143] The system first divides standardized data into sliding segments with a fixed window size to construct a set of observation subsequences. Three-phase voltage and three-phase current data are extracted from each subsequence, and the Hearst exponent for each channel is calculated through unequal scale division and range ratio analysis. The resulting Hearst exponent is used to quantify the degree of trend change in the sequence and further constitutes a trend disturbance value index to reflect the fluctuation of equipment status. For example, from March 5th to 9th, 2025, the Hearst exponent of the B-phase current of main transformer No. 2 continuously decreased from 0.67 to 0.42, while the trend disturbance value increased from 0.09 to 0.25, indicating a significant deviation in the internal current fluctuation trend and preliminarily characterizing a potential local fault trend.

[0144] Subsequently, the system inputs the observed subsequences and corresponding trend perturbation values ​​into the improved random cut forest model. During tree construction, the system prioritizes splitting samples with high perturbation values, constructing trend-driven cutting paths. Each cutting tree selects a cutting axis in the multidimensional feature space based on the perturbation-weighted variance, forming a left-propagation structure, thereby improving the model's sensitivity to trend anomalies. The model outputs the isolation score of each subsequence, which is then fused with the trend perturbation value to calculate the predicted score. If the fused predicted score exceeds a set threshold, the system automatically identifies it as an anomaly and classifies it into mild, moderate, or severe anomalies according to a set level, outputting a structured number, timestamp, and level information.

[0145] To quantify the technical advantages of this invention, the substation analyzed and compared equipment operation data for a one-month period in March 2025. The technical team selected 500 randomly chosen anomaly warning records from March 1st to 31st and compared the identification effectiveness of traditional methods with that of this invention. The following are the statistical results of the comparison data:

[0146] Table 1 Comparison of Anomaly Detection and Response Performance

[0147] index Traditional methods Method of the present invention Average early warning response time (minutes) 26.8 3.9 Anomaly detection accuracy (manually verified) 76.30％ 93.40％ Trend-based anomaly detection rate (slowly changing anomalies) 48.70％ 91.10％ False positive rate (false alarms) 14.20％ 4.60％ Anomaly level refinement capability (can it output at multiple levels?) no yes Time (seconds) for local encapsulation and output of abnormal data Not supported 1.1 Structured data format compatibility (SCADA interface) No standard output 100％

[0148] As can be seen from the data in Table 1, this invention significantly outperforms traditional methods in terms of trend anomaly identification capabilities, especially in accurately identifying "gradually changing faults." In the early morning of March 9th, the system successfully detected a gradual increase in the trend disturbance value of the temperature channel of switchgear No. 1 (from 0.06 to 0.19) without any obvious instantaneous anomalies, ultimately triggering a moderate warning when the predicted score reached 0.87. The following day, maintenance personnel discovered during an inspection that the cable joint of the equipment had poor contact and was accompanied by abnormal heating. If not addressed promptly, this could potentially cause a tripping fault during high-load operation.

[0149] The system also supports the packaging and encapsulation of localized structured early warning information, using a unified field format (frame header, timestamp, device ID, level label, prediction score, etc.) for output. It can be directly connected to the SCADA platform via the station control system or Ethernet interface. In contrast, traditional solutions can only output binary status or text logs, which is not conducive to the integrated analysis of intelligent operation and maintenance platforms.

[0150] Furthermore, the system's operational stability was also verified in this pilot test. A total of 744 hours of operation were completed throughout the month, processing approximately 26.7 million real-time data entries. The average latency for edge node data processing remained within 1.3 seconds, with minimal error fluctuations and no memory overflows or system crashes. This demonstrates that the invention exhibits excellent engineering adaptability under high-frequency operation and complex multi-source input conditions.

[0151] Table 2 Summary of Actual Equipment Risk Identification

[0152]

[0153]

[0154] In summary, this embodiment fully demonstrates the applicability and technical superiority of the present invention in real-world, complex power system environments. By integrating multiple technical features such as Hearst exponent anomaly detection, a disturbance-driven randomized forest model, and a hierarchical early warning mechanism deployed at the edge, the system achieves efficient identification, rapid response, and intelligent encapsulation of potential faults in power equipment. This significantly outperforms traditional manual inspection and centralized analysis methods, not only improving the safety and continuity of system operation but also providing a practical solution for the intelligent operation and maintenance of large-scale power equipment.

[0155] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting and warning power equipment faults based on edge computing, characterized in that, Includes the following steps: S1. Collect voltage, current, temperature and frequency parameters, perform filtering and normalization processing, and generate a continuous time series data sequence; S2. Perform a sliding window segmentation operation on the continuous time series data sequence to form a set of observation subsequences of fixed length; S3. Input the set of observed subsequences into the Hearst exponent anomaly detection algorithm, calculate the Hearst exponent for each observed subsequence, and output the trend perturbation value that represents the degree of trend change. S4. Input the set of observed subsequences and the corresponding trend perturbation values ​​into the improved robust random cut forest model, adjust the cutting priority of data points according to the trend perturbation values, construct a density-weighted cutting tree, and output the isolation score of each observed subsequence. S5. Extract the trend disturbance value and isolation score, perform normalization processing on each, perform weighted fusion according to the set weight, and output the prediction score used to characterize potential faults. S6. Based on the comparison between the predicted score and the preset threshold, generate fault warning information including timestamp, observation subsequence number and anomaly level; S7. In the edge computing node, the fault warning information is encapsulated and processed, and then output to the external system or local device through the communication interface.

2. The method for power equipment fault prediction and early warning based on edge computing according to claim 1, characterized in that, S1 specifically includes: S11. Set up edge acquisition terminals at the power equipment operation site to synchronously acquire four types of physical quantities, namely three-phase voltage, three-phase current, ambient temperature and power grid frequency, through wired or wireless means. Collect continuously at a fixed sampling frequency and generate the original parameter data sequence by arranging them in ascending order of timestamp. S12. Divide the original parameter data sequence into multi-channel sub-sequences according to parameter type, and perform median filtering on each channel to filter out abrupt changes and high-frequency discrete points in each sampling window, forming a filtered and smoothed parameter channel sequence. S13. For each channel data in the filter smoothing parameter channel sequence, calculate the minimum and maximum values ​​of the channel within a preset observation time window, perform linear normalization according to the proportional conversion rule, compress all channel values ​​to a uniform numerical scale range, and form a normalized channel data sequence set. S14. Using each sampling time as a reference, concatenate all the normalized channel data corresponding to the sampling time into a feature vector, arrange all feature vectors in chronological order, and construct a standardized continuous time series data sequence.

3. The method for power equipment fault prediction and early warning based on edge computing according to claim 1, characterized in that, S2 specifically includes: S21. Extract the parameter feature vector sequence from the standardized continuous time series data sequence in chronological order, set a fixed window length and step size parameter, and determine the starting index sequence of the sliding window; S22. Based on the starting index of each sliding window, select a corresponding number of feature vectors to form an observation subsequence. All observation subsequences have the same length and the parameter arrangement order is consistent, forming a set of observation subsequences with temporal continuity. S23. During each slide, advance the starting index of the previous window by one step, and repeat the feature vector combination operation until all standardized continuous time series data are traversed to form a complete set of observation subsequences, with some overlap between the subsequences in the set. S24. Number each observation subsequence according to the sampling order and mark its starting time point in the original time series data to generate a set of observation subsequences.

4. The method for power equipment fault prediction and early warning based on edge computing according to claim 1, characterized in that, S3 specifically includes: S31. Extract the parameter sequences from the three-phase voltage and three-phase current channels for each observation subsequence in the observation subsequence set to form a trend analysis input sequence set. Each sequence in the trend analysis input sequence set maintains its original time order. S32. For each parameter sequence in the trend analysis input sequence set, set three sets of unequally spaced increasing scale segments, perform segmented calculations in sequence, determine the sampling length of each scale segment, and form a scale segment set {n1,n2,n3}. S33. Within each scale segment, divide the sample into segments, calculate the range R and standard deviation S within each segment, and obtain their ratio B = R / S. Take the logarithm of the scale length and the ratio to construct a point set (log(n), log(B)). Perform least-squares linear fitting on all point sets, and define the slope as the single-channel Hurst exponent H of the parameter sequence. j ; S34. For each observed subsequence, calculate the set of Hearst exponents {H} of all its trend analysis input sequences. j } Calculate the channel maximum value, channel average value, and mean square deviation value D. i Construct the trend disturbance value T i The calculation formula is as follows: T i =(max(H j )-mean(H j ))+D i ; S35. Match the observation subsequence number i with the corresponding trend disturbance value T. i Composition of trend disturbance label pairs (i,T) i All trend disturbance identifiers constitute a set of trend disturbance values ​​in chronological order.

5. The method for power equipment fault prediction and early warning based on edge computing according to claim 1, characterized in that, S4 specifically includes: S41. Construct each observation subsequence into a fixed-length feature vector X containing all parameter channels. i and its corresponding trend disturbance value T i Added as a perturbation factor, generating perturbation samples (X) i ,T i This forms a set of perturbed samples; S42. Arrange the disturbance sample set according to the trend disturbance value T. i The samples are sorted from largest to smallest. The sorted samples are used to determine the processing order when building the tree. Samples with larger perturbation values ​​are processed first and assigned to the left sub-branch first, forming a trend-driven left-side splitting path. S43. Select samples sequentially from the sorted perturbation sample set to construct a single cutting tree. The construction process of the single cutting tree includes the following operations: S431. Extract all samples to be split in the current tree node, set all parameter channels as candidate cutting dimensions, and calculate the perturbation-weighted variance V for each dimension. d The definition is as follows: Among them, X i,d Let be the value of sample i in the d-th dimension. This represents the weighted mean along the d-dimensional axis; S432. Select the dimension with the largest perturbation weighted variance as the current cutting axis, and randomly generate a cutting point p between the maximum and minimum values ​​of the dimension. d ; S433, All samples with cutting axis values ​​less than or equal to p d The sample value is assigned to the left child node, and the remaining samples are assigned to the right child node; when the sample value is within the tolerance range near the cut point [p d -∈,p d When the value of the trend disturbance is within the range of +∈], it is preferentially assigned to the left child node according to the magnitude of the trend disturbance value, forming a trend-dominated leftward propagation structure; S434. Repeat the feature vector extraction, perturbation weighted variance calculation, cutting axis selection, cutting point generation, and sample partitioning operations for the left and right child nodes respectively, until the leaf node contains only one sample, thus completing the construction of a single cutting tree. S44. Based on the same perturbation sample set, repeat the pruning tree construction process to generate k pruning trees with perturbation sorting and left propagation mechanisms, and combine them to form an improved robust random pruning forest model. S45. In the robust random cut forest model, record the path length traversed by each observed subsequence in all cut trees, denoted as . Where j is the pruning tree number, the average path length is calculated as follows: S46. Calculate the isolation score S of the observed subsequence. i The definition is as follows: Where n represents the total number of samples, and log2(n) is the logarithmic baseline of the theoretical expected path length; S47. Number each observation subsequence and associate it with the corresponding isolation score S. i The isolation degree output pairs are combined to form the isolation degree score set.

6. The method for power equipment fault prediction and early warning based on edge computing according to claim 1, characterized in that, S5 specifically includes: S51. Extract the trend perturbation value and isolation score of each observation subsequence, arrange them in order according to the observation subsequence number, and form the trend index sequence and the structure index sequence respectively. S52. Perform normalization processing on the trend indicator sequence and the structure indicator sequence respectively to generate normalized trend components and normalized structure components. S53. Set the weight coefficients of the trend component and the weight coefficients of the structure component, and the sum of the two is 1. Multiply them by the normalized trend component and the normalized structure component respectively, perform weighted summation, and generate the fusion prediction score corresponding to each observation subsequence. S54. Set a prediction score threshold for judging whether the observed subsequence is abnormal, compare each fusion prediction score with the threshold, and if the fusion prediction score is greater than or equal to the prediction score threshold, the observed subsequence is marked as abnormal; otherwise, it is marked as normal. S55. Output the fused prediction score as a numerical result to characterize the potential fault state.

7. The method for power equipment fault prediction and early warning based on edge computing according to claim 1, characterized in that, S6 specifically includes: S61. Extract the fusion prediction score p for each observation subsequence. i Subsequence number i and its start timestamp t in the standardized continuous time series data sequence i This constitutes the set of fusion prediction mapping triples (i,t) i ,p i ); S62. Set three level thresholds θ1, θ2, and θ3 locally on the edge computing node, satisfying θ1 > θ2 > θ3, and apply these thresholds to each fusion prediction score p. i The abnormality level L is assigned as follows: i : S63, transfer the corresponding number i and timestamp t i and abnormality level L i Combined construction of structured fault early warning record triples (i,t) i ,L i ); S64. Sort all generated fault warning record triplets in ascending order of number to generate a fault warning information set containing timestamp, observation subsequence number and anomaly level.

8. The method for power equipment fault prediction and early warning based on edge computing according to claim 1, characterized in that, Specifically, S7 includes: S71. Configure a local exception cache structure inside the edge computing node, and record a triplet (i,t) for each fault warning. i ,L i The system receives data in real time and writes it as a key-value pair (t) with key i as the number and value as the value. i ,L i In the hash cache table structure; S72, Read the current system local timestamp t sys The device unique identifier code e in the edge node configuration file id As global identifier metadata, it is appended to each cache record to construct a 5-tuple (i,t). i ,L i ,t sys ,e id ); S73. Set the task cycle of the timer scheduler to ΔT. In each scheduling cycle, extract all five-tuple records from the hash cache table in batches and sort them in ascending order according to the number i to generate a structured output queue of abnormal events. S74. Encode each sorted quintuple into a data frame of a uniform format. The data frame structure includes a frame header identifier field, a timestamp field, a subsequence number field, an anomaly level field, and a node identifier field. Each field has a fixed length and is encapsulated into a binary transmission unit using byte alignment. S75. Call the physical communication interface controller configured in the edge computing node, write the generated data frames one by one into the transmission buffer register, and transmit them frame by frame through the preset serial port or Ethernet port. The target address includes the upper-level monitoring platform and the local control module. The transmission method is unicast or multicast.

Images