A machine learning-based power equipment anomaly early warning method, device and medium

Abstract

This invention discloses a machine learning-based method, device, and medium for early warning of power equipment anomalies, relating to the field of machine learning technology. The method includes: identifying early, weak anomalies in an online health representation sequence; generating a set of candidate anomaly events; locating potential faulty components within the candidate anomaly event set and calculating the corresponding anomaly value index to form a set of high-value anomaly events; performing executability constraint processing on the high-value anomaly event set to generate a set of executable maintenance tasks and a resource constraint gap; calculating the maintenance priority index of the executable maintenance task set and classifying early warning levels based on the resource constraint gap, generating an early warning output record set. This invention, by constructing a machine learning-based health state learning model and performing executability constraint processing on high-value anomaly events, effectively distinguishes between early, weak anomalies and random fluctuations, improving the executability of early warning information and the efficiency of resource allocation.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and in particular to a method, device and medium for early warning of power equipment anomalies based on machine learning. Background Technology

[0002] With the continuous expansion of power equipment scale and the increasing complexity of its operation structure, power equipment operation status monitoring and fault early warning technologies are gradually shifting from traditional periodic maintenance modes to online monitoring and predictive maintenance modes. In recent years, the rapid development of multi-source sensing technology, data acquisition technology, and information and communication technology has enabled key power equipment such as transformers, switchgear, circuit breakers, instrument transformers, and line terminals to generate large-scale multi-source time-series data in real time, including electrical parameters, vibration signals, temperature information, partial discharge data, acoustic characteristics, oil chromatography parameters, and environmental operating status. Based on this, machine learning technology has shown good adaptability in complex nonlinear modeling, making data-driven health status assessment and anomaly detection an important research direction in the field of power equipment operation and maintenance management.

[0003] However, existing technologies still have certain shortcomings. Most existing anomaly detection methods focus on identifying whether the operating status of equipment deviates from the normal range, but they are insufficient in characterizing the stability characteristics, persistence characteristics, and consistency characteristics of multi-source data during the anomaly evolution process. This makes it difficult to effectively distinguish between early weak anomalies and short-term random disturbances, thus affecting the accuracy and reliability of early warning results. Existing technologies usually use anomaly identification results directly as the basis for early warning output, lacking a systematic evaluation mechanism for the feasibility of anomaly event operation and maintenance handling. For example, they do not fully consider actual operation and maintenance factors such as the availability of maintenance resources, power outage constraints, and maintenance implementation windows, resulting in a lack of effective connection between early warning results and actual operation and maintenance scheduling. This can easily lead to maintenance decisions that cannot be implemented or are out of priority, reducing the practical application value of predictive maintenance strategies. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a machine learning-based method for early warning of power equipment anomalies to address the problems of insufficient consistency analysis and improper integration with actual operation and maintenance conditions.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a machine learning-based method for early warning of power equipment anomalies, comprising: collecting equipment operating status data and performing data quality governance to generate an aligned time-series segment set; extracting multi-source health features from the aligned time-series segment set; constructing a health status learning model based on historical multi-source health features, inputting the multi-source health features into the health status learning model, performing nonlinear amplification mapping with a deviation sensitivity enhancement layer, and performing consistency and trend stability checks with a time-series stability discrimination layer to form an online health representation sequence; identifying early weak anomalies in the online health representation sequence, generating a candidate anomaly event set, locating potential faulty components in the candidate anomaly event set and calculating the corresponding anomaly value index to form a high-value anomaly event set; performing executability constraint processing on the high-value anomaly event set to generate an executable maintenance task set and a resource constraint gap; calculating the maintenance priority index of the executable maintenance task set, and classifying the early warning level according to the resource constraint gap, generating an early warning output record set.

[0007] As a preferred embodiment of the machine learning-based power equipment anomaly early warning method of the present invention, the steps of collecting equipment operating status data and performing data quality governance to generate an aligned time-series segment set, and extracting multi-source health features from the aligned time-series segment set are as follows. The equipment operation status data is processed by time-aligned resampling, missing value imputation, outlier truncation and unit normalization, and segmented into slices according to continuous sliding time windows to generate an aligned time series fragment set; Through time domain analysis, frequency domain analysis, and statistical analysis, the time domain features, frequency domain features, and statistical features of the aligned time series fragment set are extracted and spliced ​​in chronological order to form multi-source health features.

[0008] As a preferred embodiment of the machine learning-based power equipment anomaly early warning method of the present invention, the specific steps of constructing a health status learning model based on historical multi-source health features are as follows: An autoencoder is used to map historical multi-source health features to a low-dimensional latent space and perform a nonlinear transformation to construct a bias sensitivity enhancement layer. The deviation change rate is calculated based on historical multi-source health characteristics, and consistency judgment and trend stability test are performed to construct a time-series stability discrimination layer. A health state learning model is constructed by cross-layer fusion and hierarchical stacking of the bias sensitivity enhancement layer and the temporal stability discrimination layer.

[0009] As a preferred embodiment of the machine learning-based power equipment anomaly early warning method of the present invention, the specific steps for forming the online health representation sequence are as follows: The multi-source health features are input into the bias sensitivity enhancement layer and nonlinearly amplified and mapped to calculate the enhanced health bias. The enhanced health deviation is input into the temporal stability discrimination layer for consistency and trend stability testing, and the output is a health status stability score. Nonlinear mapping and feature fusion are performed on the enhanced health deviation quantity and health status stability score to form an online health representation sequence.

[0010] As a preferred embodiment of the machine learning-based power equipment anomaly early warning method of the present invention, the specific steps for generating the candidate anomaly event set are as follows: Anomaly metrics are calculated for online health characterization sequences using multi-scale anomaly metrics. Based on the anomaly metric, the occurrence time, duration, anomaly intensity, and fluctuation trend of each candidate anomaly event in the online health representation sequence are statistically analyzed to generate a set of candidate anomaly events.

[0011] As a preferred embodiment of the machine learning-based power equipment anomaly early warning method of the present invention, the specific steps for forming a set of high-value anomaly events are as follows: By using feature correlation analysis to obtain the multi-source feature contribution of the candidate abnormal event set, and by statistically analyzing the topological mapping relationship between equipment components, potential faulty components can be located. Calculate the maintenance impact, risk reduction potential, and resource consumption sensitivity of potential faulty components, and prioritize them to form a set of high-value anomalies.

[0012] As a preferred embodiment of the machine learning-based power equipment anomaly early warning method of the present invention, the specific steps for generating the set of executable maintenance tasks and resource constraint gaps are as follows: Analyze the spare parts availability, manpower availability, power outage permit status, and maintenance window availability status of each high-value anomaly event in the set of high-value anomalies, and generate the corresponding execution feasibility. Summarize high-value anomalies with an execution feasibility rating, and analyze their severity, impact range of faulty components, risk reduction potential, resource availability, and equipment health history to form a set of executable maintenance tasks; Calculate the required human resources, tools, equipment, and downtime for each executable high-value anomaly, and determine the resource constraint gap by assessing the availability of existing resources.

[0013] As a preferred embodiment of the machine learning-based power equipment anomaly early warning method of the present invention, the specific steps of calculating the maintenance priority index of the executable maintenance task set, classifying the early warning level according to the resource constraint gap, and generating an early warning output record set are as follows. The abnormal value index, equipment health status, task complexity, and resource requirements of each executable task in the executable maintenance task set are statistically analyzed to obtain the maintenance priority index. Based on the resource constraint gap and maintenance priority index, the executable tasks in the executable maintenance task set are divided into high warning level, medium warning level and low warning level. Extract the task ID, task priority, faulty component information, anomaly level, early warning lead time, resource gap information, and repair time from the set of executable maintenance tasks, and generate an early warning output record set.

[0014] In a second aspect, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the machine learning-based power equipment anomaly early warning method described in the first aspect of the present invention.

[0015] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the machine learning-based power equipment anomaly early warning method described in the first aspect of the present invention.

[0016] The beneficial effects of this invention are as follows: By constructing a health status learning model based on machine learning, it is possible to effectively distinguish between early weak anomalies and random fluctuations, thereby improving anomaly identification from instantaneous deviation judgment to continuous status judgment, thus enhancing the ability to identify early faults and reducing the false alarm rate; at the same time, by performing executability constraint processing on high-value anomaly events, it generates a set of executable maintenance tasks by comprehensively considering conditions such as spare parts, manpower, power outage permits, and maintenance windows, and calculates the resource constraint gap to achieve early warning classification, so that the anomaly assessment results are directly transformed into implementable operation and maintenance decisions, thereby improving the executability of early warning information and the efficiency of resource allocation, and realizing closed-loop optimization from anomaly detection to maintenance action decision-making. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a machine learning-based early warning method for power equipment anomalies.

[0019] Figure 2 This is a flowchart for data processing and multi-source health feature extraction.

[0020] Figure 3 The flowchart illustrates the formation of a health status learning model and an online health representation.

[0021] Figure 4 This is a flowchart of the abnormal early warning decision-making process. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides a machine learning-based method for early warning of power equipment anomalies, comprising the following steps: S1. Collect equipment operating status data and perform data quality management to generate an aligned time series fragment set and extract multi-source health features from the aligned time series fragment set.

[0026] S1.1 Perform time-aligned resampling, missing value imputation, outlier truncation and dimension normalization processing on the equipment operation status data, and segment it according to the continuous sliding time window to generate an aligned time sequence fragment set.

[0027] It should be noted that the equipment operating status data includes electrical parameter data, vibration signal data, temperature data, partial discharge data, acoustic signal data, oil chromatographic component data, as well as environmental and operating condition data; electrical parameter data includes voltage, current, active power, reactive power, and power factor; vibration signal data is collected by vibration sensors installed in the bearing housing or shell; temperature data is obtained by winding temperature sensors or infrared thermometers; partial discharge data is collected by ultrasonic or high-frequency sensors; acoustic signal data is obtained by acoustic arrays or single-point acoustic sensors; oil chromatographic component data is obtained through online oil chromatographic monitoring devices; environmental and operating condition data include ambient temperature, humidity, load rate, number of start-stop cycles, and operating mode indicators.

[0028] The equipment operating status data is sequenced into time intervals according to the data recording time. The distribution of time intervals between all adjacent records is statistically analyzed. The time interval with the highest frequency is selected as the reference sampling interval and used as a unified sampling interval for time alignment resampling processing, so that the equipment operating status data forms a continuous sequence under a unified time scale.

[0029] The missing locations in the equipment operation status data are identified according to a unified time scale; the length of continuous missing data is calculated based on the benchmark sampling interval, and a fixed multiple of the benchmark sampling interval is used as the upper limit of the missing data filling length (e.g., 20 to 30 times the benchmark sampling interval); time neighborhood interpolation is performed to fill missing segments that do not exceed the upper limit of the missing data filling length, and neighboring value retention is performed to fill missing segments that exceed the upper limit of the missing data filling length.

[0030] After completing the missing value imputation, a distribution statistics sliding window is constructed according to a fixed time length (e.g., a time span of 5 to 10 minutes), and the distribution statistics sliding window is moved in chronological order to locally extract the equipment operation status data; all values ​​within each distribution statistics sliding window are sorted, and the values ​​at the two ends of the sorting result are read as the distribution interval boundaries; outlier truncation is performed on the equipment operation status data within the window according to the distribution interval boundaries, and values ​​less than the lower boundary or greater than the upper boundary are replaced with the corresponding boundary values; the preset proportion is determined according to the historical fluctuation range of the equipment operation status data and is used to limit the effective range of the local value distribution; the preset proportion takes the small proportion range at both ends of the local value distribution (e.g., the range of the first 0.5% to the first 2% and the last 0.5% to the last 2%).

[0031] After outlier truncation, the equipment operating status data is normalized to map each dimension of the data to a uniform numerical range. After normalization, a time-segmented sliding window is constructed, with the length determined based on the statistical results of the equipment operating status change cycle (e.g., the median range of the interval with the highest proportion in the distribution of operating status change cycles). The equipment operating status data is continuously sliced ​​according to the time-segmented sliding window to form continuous time segments with a fixed time span. All time-segmented sliding window truncation results are numbered in chronological order and bound to time ranges to generate an aligned time-series segment set.

[0032] S1.2. Through time domain analysis, frequency domain analysis and statistical analysis, extract the time domain features, frequency domain features and statistical features of the aligned time series fragment set, and splice them in time order to form multi-source health features.

[0033] Perform time-domain analysis on the equipment operation status data in each time segment, statistically analyze the amplitude change, rate of change and fluctuation of the equipment operation status data in the current time segment, and record them according to the data source identifier to form the time-domain feature sequence of the corresponding time segment.

[0034] After completing the time domain analysis, frequency domain analysis is performed on the equipment operation status data in the same time segment. The frequency components of the equipment operation status data in the time segment are decomposed, and the energy distribution of the main frequency components (e.g., frequency components whose energy proportion is in the top 20% of the total energy) is statistically analyzed to form the frequency domain feature sequence of the corresponding time segment.

[0035] After completing the frequency domain analysis, statistical analysis is performed on the equipment operating status data within the same time segment. The average value, the difference between the maximum and minimum values, the median value, and the difference between the upper and lower quantile values ​​within the time segment are calculated (e.g., the difference between the first 25% and the last 25% of the value distribution). These values ​​are recorded according to the data source identifier to form a statistical feature sequence for the corresponding time segment. The time domain feature sequence, frequency domain feature sequence, and statistical feature sequence for the same time segment are then sequentially concatenated according to the time order corresponding to the time segment to generate multi-source health features that correspond one-to-one with the time segment. All multi-source health features corresponding to all time segments are arranged in chronological order to form a multi-source health feature sequence.

[0036] S2. Construct a health status learning model based on historical multi-source health features, and input the multi-source health features into the health status learning model. The bias sensitivity enhancement layer performs nonlinear amplification mapping, and the temporal stability discrimination layer performs consistency and trend stability tests to form an online health representation sequence.

[0037] S2.1. An autoencoder is used to map historical multi-source health features to a low-dimensional latent space and perform a nonlinear transformation to construct a bias sensitivity enhancement layer.

[0038] It should be noted that an autoencoder is an unsupervised learning method for learning the internal structure of data. The operation involves inputting the input data into the encoding structure record by record, compressing the input data layer by layer to obtain a low-dimensional representation, and then reconstructing the data layer by layer based on the low-dimensional representation. By continuously adjusting the internal parameters, the difference between the reconstructed data and the original input data is gradually reduced. When the reconstruction difference stabilizes, the intermediate compressed representation can reflect the typical change structure of historical multi-source health features and serve as a low-dimensional potential representation of historical multi-source health features.

[0039] Historical multi-source health features are divided into continuous data segments in chronological order. Each data segment has a fixed time length (e.g., a time span of 30 to 60 minutes). Each data segment is sequentially input into an autoencoder. Compressed representation generation and reconstruction recovery are performed on each data segment, and the internal parameters are continuously adjusted so that the difference between the reconstruction result and the input data gradually converges. After the reconstruction difference stabilizes, the compressed representation result is extracted as the low-dimensional latent representation of the corresponding data segment.

[0040] The low-dimensional latent representations corresponding to all data segments are summarized, the overall distribution range of the low-dimensional latent representations is statistically analyzed, and the main distribution interval of the low-dimensional latent representations is determined (e.g., the interval in the middle 90% range of the low-dimensional latent representation numerical distribution). The degree of deviation of each low-dimensional latent representation from the center position of the main distribution interval is calculated, and a nonlinear mapping relationship is established based on the degree of deviation, so that the magnitude of numerical change is amplified nonlinearly when the degree of deviation increases, while the numerical change within the main distribution interval remains unchanged. The low-dimensional latent representation generation method and the nonlinear mapping relationship are combined and defined to form a deviation sensitivity enhancement layer.

[0041] S2.2 Calculate the deviation change rate based on historical multi-source health characteristics, and perform consistency judgment and trend stability test to construct a time series stability discrimination layer.

[0042] It should be noted that the central location of the main distribution interval of the statistical low-dimensional latent representation (e.g., the average of the absolute differences of each dimension) is used as a reference benchmark; the numerical difference between the low-dimensional latent representation corresponding to each time segment and the reference benchmark is calculated dimension by dimension, and the absolute value of the difference of each dimension is taken and then the dimensions are summarized to obtain the health deviation sequence arranged in time order; the difference change of the health deviation sequence is calculated according to the adjacent time segments, and normalized by the time segment interval to obtain the deviation change rate sequence.

[0043] Within a range of consecutive adjacent time segments, the proportion of consistent change direction and the range of amplitude changes in the deviation change rate are statistically analyzed to form a consistency statistical result (e.g., the proportion of consistent change direction in 10 consecutive time segments is not less than 70%). A trend statistical time window is constructed over a longer continuous time span (e.g., a time span of 2 to 6 hours). Within the trend statistical time window, the number of reversals in the direction of the deviation change rate and the range of fluctuation amplitude are statistically analyzed to form a trend stability statistical result. The deviation change rate sequence, the consistency statistical result, and the trend stability statistical result are combined to form a time series stability discrimination layer.

[0044] S2.3. Perform cross-layer fusion and hierarchical stacking of the bias sensitivity enhancement layer and the temporal stability discrimination layer to construct a health state learning model.

[0045] It should be noted that a one-to-one mapping relationship is established between the output variable set of the bias sensitivity enhancement layer and the discriminant variable set of the temporal stability discrimination layer according to their time positions. A unified feature expression space is constructed based on this mapping relationship, ensuring that the information on the magnitude of bias changes and the information on temporal evolution stability remain synchronously corresponding in the same expression structure. After completing the construction of the unified feature expression space, the output path of the bias sensitivity enhancement layer is used as the pre-processing path, and the output path of the temporal stability discrimination layer is used as the post-discrimination path. A hierarchical connection relationship is established according to a fixed processing order, so that the output of the pre-processing path can be directly used as the input constraint variable of the post-processing path, forming a continuous processing structure. The unified feature expression space and the continuous processing structure are encapsulated as a whole to form a health state learning model.

[0046] Next, the health status learning model is trained. Historical multi-source health features are used as the sample set. A data loader (such as DataLoader) is used to perform batch loading of samples and feature diversification. Numerical standardization is performed through Z-score normalization to generate augmented samples. The augmented samples are input into the health status learning model, and gradient backpropagation is performed through an optimizer (such as Adam optimizer) to update the parameters of the health status learning model. The training loss is calculated based on the reconstruction deviation between the reconstruction result output by the health status learning model and the input samples. During the training process, a learning rate scheduler is applied to perform dynamic learning rate adjustment, and the current training loss is calculated after each training round. The augmented samples are divided into F data subsets and F-fold cross-validation is performed. The reconstruction loss of the health status learning model is calculated during each fold cross-validation, and the average reconstruction loss of all folds is taken to obtain the validation loss value. When the validation loss no longer decreases within a certain number of consecutive training rounds (such as 10 rounds), the training is terminated, and the trained health status learning model is output.

[0047] S2.4. Perform nonlinear amplification mapping on the multi-source health feature input deviation sensitivity enhancement layer to calculate the enhanced health deviation amount.

[0048] It should be noted that the multi-source health features are input into the deviation sensitivity enhancement layer in chronological order, and based on the low-dimensional latent space mapping relationship formed by the historical multi-source health features, the multi-source health features at each time moment are converted into the corresponding latent spatial location representation; the distribution location and discrete distribution degree of the historical multi-source health features in the low-dimensional latent space are read, and the spatial deviation distance between the current latent spatial location and the historical distribution location is calculated to obtain the original health deviation amount at the corresponding time location.

[0049] The original health deviation is input into the nonlinear amplification mapping relationship inside the deviation sensitivity enhancement layer, and nonlinear amplitude modulation is performed according to the relative position of the original health deviation in the discrete interval of the historical distribution. The higher the mapping amplification is when the deviation is closer to the outer region of the historical distribution, the enhanced health deviation is obtained. The enhanced health deviation corresponding to each time position is arranged in chronological order to form an enhanced health deviation sequence.

[0050] S2.5. Input the enhanced health deviation quantity into the time series stability discrimination layer for consistency test and trend stability test, and output the health status stability score.

[0051] It should be noted that the enhanced health deviation quantity is input into the temporal stability discrimination layer in chronological order, and the change amplitude and direction between adjacent time positions are calculated at continuous time positions to generate the enhanced health deviation quantity change sequence. The enhanced health deviation quantity change sequence is subjected to directional consistency statistical processing within continuous time segments, specifically by calculating the time proportion of consecutive occurrences of the same change direction, which is used as the change consistency characterization value.

[0052] After completing the directional consistency statistical processing, the difference between the maximum and minimum values ​​of the rate of change of the enhanced health deviation quantity change sequence within a continuous time interval is calculated, and the average absolute deviation between the rate of change and the average rate of change within the current continuous time interval is calculated. The difference between the maximum and minimum values ​​and the average absolute deviation are jointly mapped to obtain the change stability characterization value. The change consistency characterization value and the change stability characterization value are jointly mapped according to the time position to generate the health status stability score at the corresponding time position.

[0053] It should also be noted that during the joint mapping process, interval normalization is performed on each representation quantity involved in the mapping, so that the value range of each representation quantity falls uniformly between zero and one. After the normalization process is completed, the corresponding weight is determined according to the degree of dispersion of each representation quantity in the current continuous time interval, wherein the degree of dispersion is represented by the difference between the maximum and minimum values ​​of the current representation quantity sequence. The weighted combination process is performed on each representation quantity according to the determined weight to generate the joint mapping result.

[0054] When jointly mapping the difference between the maximum and minimum values ​​and the average absolute deviation, the difference between the maximum and minimum values ​​is defined as the rate of change fluctuation, and the average absolute deviation is used as the rate of change discrete quantity. Interval normalization is performed on both the rate of change fluctuation and the rate of change discrete quantity, mapping them uniformly to a value range between zero and one. After completing the interval normalization, the maximum and minimum values ​​of the rate of change fluctuation sequence within the current continuous time interval are counted, and the difference is calculated as the first discrete quantity. Simultaneously, the maximum and minimum values ​​of the rate of change discrete quantity sequence are counted, and the difference is calculated as the second discrete quantity. The sum of the first and second discrete quantities is used as the total discrete quantity, and the proportions of the first and second discrete quantities in the total discrete quantity are calculated as corresponding weights. A weighted combination is performed on the normalized rate of change fluctuation and normalized rate of change discrete quantity values ​​at the same time position according to the corresponding weights to obtain the stability characterization value for the corresponding time position.

[0055] S2.6. Perform nonlinear mapping and feature fusion on the enhanced health deviation quantity and health status stability score to form an online health representation sequence.

[0056] It should be noted that the enhanced health deviation sequence and the health status stability score sequence are read in chronological order, and a one-to-one correspondence is established at the same time position to form a time-aligned bivariate state pair. For the enhanced health deviation at each time position, a nonlinear mapping process is performed on the amplitude interval. The enhanced health deviation is divided into continuously changing segments according to the numerical range, and a nonlinear function transformation with different curvatures is performed on different segments. This makes the segments with smaller values ​​maintain a smooth mapping, and the segments with larger values ​​produce a higher mapping amplification, thus obtaining the nonlinear deviation characterization value. The health status stability score is subjected to a reverse modulation mapping process, so that when the stability score is low, it has an amplifying modulation effect on the nonlinear deviation characterization value, and when the stability score is high, it has an inhibitory modulation effect on the nonlinear deviation characterization value, thus obtaining the stability modulation deviation characterization value.

[0057] The stability modulation deviation characterization values ​​are processed into continuous trajectory encoding in chronological order. The change trend connection and trajectory continuity constraints are applied to the stability modulation deviation characterization values ​​of adjacent time positions to ensure that the state changes of adjacent time positions maintain a continuous path expression, generating a time-continuous health status trajectory representation. The health status trajectory representations corresponding to all time positions are arranged in chronological order to form an online health characterization sequence.

[0058] S3. Identify early weak anomalies in the online health representation sequence, generate a set of candidate anomaly events, locate potential faulty components in the candidate anomaly event set and calculate the corresponding anomaly value index to form a set of high-value anomaly events.

[0059] S3.1 Calculate the anomaly measurement value of the online health representation sequence through multi-scale anomaly measurement.

[0060] The expression for calculating the anomaly metric is: ; in, Indicates time position The corresponding anomaly metric, Indicates the time position of online health characterization sequences The value, Indicates the number of time spans involved in the anomaly metric calculation. Indicates the first A time span on a time scale Indicates the time position of online health characterization sequences Backtracking The value of each time interval. Indicates the first The baseline statistical length at each time scale Indicates the time position of online health characterization sequences The value, Indicates the time position of online health characterization sequences Backtracking The value of each time interval.

[0061] It should be noted that continuous historical time segments are selected from the online health representation sequence as statistical segments, and the number of time locations contained in the statistical segment is defined as the baseline statistical length.

[0062] S3.2. Based on the anomaly metric, calculate the occurrence time, duration, anomaly intensity, and fluctuation trend of each candidate anomaly event in the online health representation sequence, and generate a set of candidate anomaly events.

[0063] It should be noted that the abnormal measurement value sequence is read in chronological order, and the time positions in the abnormal measurement value sequence with the same direction of change and continuous change amplitude are grouped into the same continuous abnormal segment. The start time position corresponding to each continuous abnormal segment is read as the occurrence time of the candidate abnormal event, and the time span between the corresponding end time position and the start time position is taken as the duration of the candidate abnormal event.

[0064] The maximum and average values ​​of the abnormality metrics within each continuous abnormal segment are calculated, and the maximum and average values ​​of the abnormality metrics are jointly encoded to obtain the abnormality intensity field of the candidate abnormal event.

[0065] Calculate the differential change between adjacent time positions of the abnormal measurement value in each continuous abnormal segment in chronological order, and determine the fluctuation trend field of the abnormal measurement value based on the positive and negative changes of the differential change at continuous time positions (e.g., positive values ​​indicate an upward trend, negative values ​​indicate a downward trend, and alternating positive and negative values ​​indicate a fluctuation trend).

[0066] The occurrence time, duration, anomaly intensity, and fluctuation trend fields corresponding to each continuous abnormal segment are structured and summarized to generate a set of candidate abnormal events.

[0067] S3.3. Obtain the multi-source feature contribution of the candidate abnormal event set through feature correlation analysis, and statistically analyze the topological mapping relationship between equipment components to locate potential faulty components.

[0068] It should be noted that, for each candidate abnormal event in the candidate abnormal event set, the corresponding time segment covered by the occurrence time and duration is read, and the multi-source health feature sequence corresponding to the current candidate abnormal event is extracted from the same time segment to form a candidate abnormal event feature segment; the feature change amplitude of the multi-source health feature sequence in the candidate abnormal event feature segment is calculated according to the data source identifier, and the feature change amplitude corresponding to each data source identifier is aligned and correlated with the change amplitude of the abnormality metric value of the current candidate abnormal event to obtain the correlation contribution value corresponding to each data source identifier; the correlation contribution value corresponding to each data source identifier is normalized so that the sum of the contribution values ​​corresponding to each data source identifier is a preset total, and the normalized contribution value is determined as the multi-source feature contribution degree.

[0069] A set of associated paths for components is established based on the topological mapping relationship between equipment components. The contribution of multi-source features is mapped to the corresponding set of equipment components according to the data source identifier, so as to obtain the component contribution distribution of candidate abnormal events. The set of components with the highest contribution ratio is selected from the component contribution distribution and sorted from high to low contribution ratio to obtain the sorted list of potential faulty components corresponding to the candidate abnormal events. The sorted lists of potential faulty components of all candidate abnormal events in the candidate abnormal event set are summarized to generate the location result of potential faulty components corresponding to the candidate abnormal event set.

[0070] S3.4 Calculate the maintenance impact, risk reduction potential, and resource consumption sensitivity of potential faulty components, and prioritize them to form a set of high-value abnormal events.

[0071] The expression for calculating the influence of maintenance sorting is: ; in, This represents the index e of the candidate abnormal event relative to the index of the potential faulty component. The maintenance sorting impact quantity, This field represents the anomaly strength at index e of the candidate anomaly event. This indicates the duration field at index e of the candidate exception event. This represents the numerical encoding of the fluctuation trend field at candidate exception event index e. This represents the candidate abnormal event index e and the potential faulty component index. The contribution of components between them.

[0072] It should be noted that the component contribution distribution corresponding to the potentially faulty component is read, and the number of associated paths and path connection levels of the potentially faulty component in the equipment component topology mapping relationship are counted. The component contribution distribution and the number of associated paths and path connection levels are then jointly mapped to obtain the risk propagation range characterization value of the potentially faulty component, which serves as the risk reduction potential. The resource requirement record corresponding to each potentially faulty component is read, and the current human resource requirement, tool and equipment requirement, and downtime requirement of the potentially faulty component are counted. The proportion of each resource requirement in the total resource requirement of all potentially faulty components is calculated as the resource occupancy sensitivity.

[0073] It should also be noted that, in order to eliminate the influence of the time dimension of the duration field on the calculation result of the maintenance ranking influence, the duration field is subjected to interval normalization processing, so that the value of the duration field falls uniformly between zero and one, and the normalized duration value is substituted into the maintenance ranking influence calculation formula.

[0074] The impact of maintenance ranking, risk reduction potential, and resource consumption sensitivity are jointly mapped to obtain the anomaly value index. Candidate anomalies that rank at the top of the anomaly value index (such as the top 10%) are identified as high-value anomalies and then aggregated to form a set of high-value anomalies.

[0075] S4. Perform executability constraint processing on the set of high-value abnormal events to generate a set of executable maintenance tasks and resource constraint gaps.

[0076] S4.1 Analyze the spare parts availability, manpower availability, power outage permit status, and maintenance window availability status of each high-value anomaly event in the high-value anomaly event set, and generate the corresponding execution feasibility.

[0077] It should be noted that the availability of spare parts, manpower, power outage permit status, and maintenance window availability are derived from the operation and maintenance resource registration record. The operation and maintenance resource registration record is a resource status record formed and continuously updated during the equipment operation and maintenance management process. For each high-value anomaly in the high-value anomaly event set, the corresponding potential faulty component information is read, and the spare parts inventory status, manpower scheduling status, power outage permit approval status, and maintenance time arrangement status corresponding to the current potential faulty component are retrieved from the operation and maintenance resource registration record based on the potential faulty component identifier. The retrieved spare parts inventory status, manpower scheduling status, power outage permit approval status, and maintenance time arrangement status are processed item by item for status matching judgment. Statuses that meet the maintenance execution conditions are marked as executable statuses, and statuses that do not meet the maintenance execution conditions are marked as restricted statuses, generating a resource status matching result sequence. The number of executable statuses in the resource status matching result sequence is counted, and the proportion of the number of executable statuses to the total number of resource statuses is calculated as the execution feasibility of the current high-value anomaly event. The execution feasibility corresponding to all high-value anomalies is summarized according to the anomaly event identifier to form an execution feasibility record set.

[0078] S4.2 Summarize high-value anomalies with an execution feasibility rating of executable, and analyze their severity, impact range of faulty components, risk reduction potential, resource availability, and equipment health history to form a set of executable maintenance tasks.

[0079] It should be noted that the process involves reading all execution feasibility values ​​from the execution feasibility record set and sorting them according to their numerical values ​​to form an execution feasibility sorting sequence; calculating the change in the difference between adjacent sorting positions in the execution feasibility sorting sequence and determining the sorting position corresponding to the largest change in difference as the execution feasibility boundary position; identifying high-value anomalies above the execution feasibility boundary position in the execution feasibility sorting sequence as executable anomalies, and compiling all executable anomalies to form an executable anomaly list; and reading the anomaly intensity field and duration field corresponding to each high-value executable anomaly in the executable anomaly list and performing joint accumulation processing to obtain the anomaly severity characterization value of the current high-value executable anomaly.

[0080] Read the sorted list of potential faulty components corresponding to each executable high-value anomaly event, and count the number of related components that the potential faulty component can reach based on the topological mapping relationship between the equipment components, and use it as the fault impact range characterization value; combine the risk reduction potential and the fault impact range characterization value in sequence to form a risk disposal benefit characterization record; use the execution feasibility corresponding to each executable high-value anomaly event as the resource availability characterization value.

[0081] Read the historical health status change range of the online health representation sequence of the equipment corresponding to the executable high-value anomaly event before the anomaly occurred, and calculate the average value of the historical health status change range as the equipment health historical representation value; summarize the anomaly severity representation value, risk disposal benefit representation record, resource availability representation value, and equipment health historical representation value in a structured manner according to the anomaly event identifier, and sort the corresponding potential faulty components as maintenance target component information to form an executable maintenance task record; summarize the executable maintenance task records corresponding to all executable high-value anomaly events to generate an executable maintenance task set.

[0082] S4.3 Calculate the required human resources, tools and equipment, and downtime for each executable high-value anomaly event, and obtain the resource constraint gap by assessing the availability of existing resources.

[0083] It should be noted that the sorted list of potential faulty components corresponding to each executable maintenance task in the executable maintenance task set is read, and the maintenance resource requirement information corresponding to the current potential faulty component is retrieved from the operation and maintenance resource registration record. The maintenance resource requirement information includes the human resource requirement, tool and equipment requirement, and downtime requirement.

[0084] Read the available human resources, available tools and equipment, and available downtime length corresponding to the maintenance resource demand information in the operation and maintenance resource registration record; calculate the difference between demand and availability for each type of resource. When the demand is greater than the availability, record the difference between demand and availability as the resource constraint gap for that type of resource. When the demand is less than or equal to the availability, record the resource constraint gap for that type of resource as zero.

[0085] The resource constraint gaps for the same executable maintenance task, including the human resource constraint gap, tool and equipment constraint gap, and downtime constraint gap, are summarized and recorded to form the resource constraint gap corresponding to the current executable maintenance task. The resource constraint gaps of all executable maintenance tasks in the set of executable maintenance tasks are then structured according to task identifiers to form the resource constraint gap recording results.

[0086] S5. Calculate the maintenance priority index of the set of executable maintenance tasks, classify the early warning level according to the resource constraint gap, and generate an early warning output record set.

[0087] S5.1 Calculate the abnormal value index, equipment health status, task complexity and resource requirements of each executable task in the executable maintenance task set, and obtain the maintenance priority index.

[0088] It should be noted that the abnormal value index corresponding to each executable maintenance task is read, and the health status stability score of the associated time segment in the online health representation sequence corresponding to the current executable maintenance task is read. The abnormal value index and the health status stability score are normalized in the same direction to form a risk urgency representation value. The ranking position of each executable maintenance task in the corresponding potential faulty component ranking list is read, and the component impact ranking representation value is generated according to the relative position ratio of the ranking position in the total number of potential faulty components (for example, when the ranking is in the top 20%, it is recorded as a high impact ranking segment).

[0089] The resource constraint gap corresponding to the executable maintenance task is read, and the human resource constraint gap, tool and equipment gap, and downtime gap in the resource constraint gap are standardized to generate a resource occupancy pressure characterization value. The risk urgency characterization value, component impact ranking characterization value, and resource occupancy pressure characterization value are jointly mapped to obtain a comprehensive ranking characterization value. Based on the comprehensive ranking characterization value, all executable maintenance tasks in the executable maintenance task set are ranked. The ranking results are recorded in a structured manner according to the task identifier to generate a maintenance priority index corresponding to each executable maintenance task.

[0090] It should also be noted that the health status of the equipment is reflected by the health status stability score, and the task complexity is reflected in the resource occupancy pressure characterization value and the component impact ranking characterization value. The component impact ranking characterization value reflects the magnitude of the impact of potential faulty components on the equipment, and the resource occupancy pressure characterization value reflects the amount of resources required to execute the task. Resource requirements are measured by the resource constraint gap.

[0091] S5.2 Based on the resource constraint gap and maintenance priority index, the executable tasks in the executable maintenance task set are divided into high warning level, medium warning level and low warning level.

[0092] It should be noted that all maintenance priority indices are sorted from largest to smallest to form a maintenance priority ranking sequence, and a corresponding cumulative percentage sequence is generated according to the ranking position. The incremental difference between adjacent ranking positions in the cumulative percentage sequence is calculated to form a cumulative percentage increment sequence. The mean of the entire cumulative percentage increment sequence is calculated to obtain the cumulative percentage increment mean. The incremental difference at each position in the cumulative percentage increment sequence is compared with the cumulative percentage increment mean. Continuous ranking segments with incremental differences greater than the cumulative percentage increment mean are marked as high-risk urgent segments, continuous ranking segments with incremental differences less than the cumulative percentage increment mean are marked as low-risk urgent segments, and ranking segments at the boundary between the two types of segments are marked as medium-risk urgent segments. Read the ranking position of each executable maintenance task in the maintenance priority ranking sequence, and generate a risk urgency segment marker based on its ranking segment; sort the resource constraint gaps corresponding to all executable maintenance tasks from largest to smallest to form a resource pressure ranking sequence, and generate a corresponding cumulative percentage sequence based on the ranking position; calculate the incremental difference between adjacent ranking positions in the cumulative percentage sequence of resource pressure to form a cumulative percentage increment sequence of resource pressure; calculate the mean of the entire cumulative percentage increment sequence of resource pressure to obtain the mean of the cumulative percentage increment of resource pressure; compare the incremental difference at each position in the cumulative percentage increment sequence of resource pressure with the mean of the cumulative percentage increment of resource pressure, mark consecutive ranking segments with incremental differences greater than the mean of the cumulative percentage increment of resource pressure as high resource pressure segments, mark consecutive ranking segments with incremental differences less than the mean of the cumulative percentage increment of resource pressure as low resource pressure segments, and mark ranking segments at the boundary between the two types of segments as medium resource pressure segments.

[0093] Read the sorting position of each executable maintenance task in the resource pressure sorting sequence, and generate a resource pressure segment label based on its sorting segment; perform joint mapping processing on the risk urgency segment label and the resource pressure segment label to generate a warning level label corresponding to each executable maintenance task. Tasks that are simultaneously in the highest segment of both the risk urgency segment level and the resource pressure segment level are labeled as high warning level, tasks that are in only one segment of the highest segment or both segments of the middle segment are labeled as medium warning level, and the remaining tasks are labeled as low warning level.

[0094] S5.3 Extract the task ID, task priority, faulty component information, anomaly level, early warning lead time, resource gap information, and repair time from the executable maintenance task set, and generate an early warning output record set.

[0095] It should be noted that the task ID, maintenance priority index, potential faulty components, warning level marker, and resource constraint gap amount corresponding to the executable maintenance task are read and processed for field association to form a basic task record. At the same time, the abnormal occurrence time associated with the candidate abnormal event set corresponding to the executable maintenance task is read, and the execution time of the current operation is used as the warning generation time, and the time difference between the warning generation time and the abnormal occurrence time is used as the warning lead time to form a time difference record field. The downtime requirement obtained during the calculation of the resource constraint gap amount of the current executable maintenance task is read and used as the repair duration record field. The basic task record, time difference record field, and repair duration record field are structurally concatenated and then aggregated and uniformly encoded according to the task ID to generate a warning output record set.

[0096] This embodiment also provides a computer device applicable to the machine learning-based power equipment anomaly early warning method, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the machine learning-based power equipment anomaly early warning method proposed in the above embodiment.

[0097] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0098] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the machine learning-based power equipment anomaly early warning method proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0099] In summary, this invention improves early fault identification and reduces false alarm rates by constructing a health status learning model based on machine learning, effectively distinguishing between early weak anomalies and random fluctuations, and elevating anomaly identification from instantaneous deviation judgment to continuous state judgment. Simultaneously, by applying executability constraints to high-value anomaly events, it generates a set of executable maintenance tasks by integrating conditions such as spare parts, manpower, power outage permits, and maintenance windows, and calculates resource constraint gaps to achieve early warning classification. This allows anomaly assessment results to be directly transformed into implementable operation and maintenance decisions, thereby improving the executability of early warning information and resource allocation efficiency, and achieving closed-loop optimization from anomaly detection to maintenance action decision-making.

[0100] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for power equipment anomaly early warning based on machine learning, characterized in that: include, Collect equipment operating status data and perform data quality management to generate an aligned time series fragment set and extract multi-source health features from the aligned time series fragment set; A health status learning model is constructed based on historical multi-source health characteristics. The multi-source health characteristics are then input into the health status learning model. A bias sensitivity enhancement layer performs nonlinear amplification mapping, and a temporal stability discrimination layer performs consistency and trend stability tests to form an online health representation sequence. Identify early weak anomalies in the online health representation sequence, generate a set of candidate anomaly events, locate potential faulty components in the candidate anomaly event set and calculate the corresponding anomaly value index to form a set of high-value anomaly events. Perform executability constraint processing on the set of high-value anomalies to generate a set of executable maintenance tasks and resource constraint gaps; Calculate the maintenance priority index of the set of executable maintenance tasks, classify the early warning levels according to the resource constraint gap, and generate an early warning output record set. 2.The machine learning based power equipment anomaly early warning method of claim 1, wherein: The process involves collecting operational status data from the acquisition equipment, performing data quality management, generating an aligned time-series segment set, and extracting multi-source health features from the aligned time-series segment set. The specific steps are as follows. The equipment operation status data is processed by time-aligned resampling, missing value imputation, outlier truncation and unit normalization, and segmented into slices according to continuous sliding time windows to generate an aligned time series fragment set; Through time domain analysis, frequency domain analysis, and statistical analysis, the time domain features, frequency domain features, and statistical features of the aligned time series fragment set are extracted and spliced ​​in chronological order to form multi-source health features. 3.The machine learning based power equipment anomaly early warning method of claim 1, wherein: The specific steps for constructing a health status learning model based on historical multi-source health characteristics are as follows: An autoencoder is used to map historical multi-source health features to a low-dimensional latent space and perform a nonlinear transformation to construct a bias sensitivity enhancement layer. The deviation change rate is calculated based on historical multi-source health characteristics, and consistency judgment and trend stability test are performed to construct a time-series stability discrimination layer. A health state learning model is constructed by cross-layer fusion and hierarchical stacking of the bias sensitivity enhancement layer and the temporal stability discrimination layer. 4.The machine learning based power equipment anomaly early warning method of claim 1, wherein: The specific steps for forming the online health representation sequence are as follows: The multi-source health features are input into the bias sensitivity enhancement layer and nonlinearly amplified and mapped to calculate the enhanced health bias. The enhanced health deviation is input into the temporal stability discrimination layer for consistency and trend stability testing, and the output is a health status stability score. Nonlinear mapping and feature fusion are performed on the enhanced health deviation quantity and health status stability score to form an online health representation sequence. 5.The machine learning based power equipment anomaly early warning method of claim 4, wherein: The specific steps for generating the candidate set of abnormal events are as follows: Anomaly metrics are calculated for online health characterization sequences using multi-scale anomaly metrics. Based on the anomaly metric, the occurrence time, duration, anomaly intensity, and fluctuation trend of each candidate anomaly event in the online health representation sequence are statistically analyzed to generate a set of candidate anomaly events.

6. The machine learning-based early warning method for power equipment anomalies as described in claim 5, characterized in that: The specific steps for forming a set of high-value anomalous events are as follows. By using feature correlation analysis to obtain the multi-source feature contribution of the candidate abnormal event set, and by statistically analyzing the topological mapping relationship between equipment components, potential faulty components can be located. Calculate the maintenance impact, risk reduction potential, and resource consumption sensitivity of potential faulty components, and prioritize them to form a set of high-value anomalies. 7.The machine learning based power equipment anomaly early warning method of claim 6, wherein: The specific steps for generating the set of executable maintenance tasks and the resource constraint gap are as follows. Analyze the spare parts availability, manpower availability, power outage permit status, and maintenance window availability status of each high-value anomaly event in the set of high-value anomalies, and generate the corresponding execution feasibility. Summarize high-value anomalies with an execution feasibility rating, and analyze their severity, impact range of faulty components, risk reduction potential, resource availability, and equipment health history to form a set of executable maintenance tasks; Calculate the required human resources, tools, equipment, and downtime for each executable high-value anomaly, and determine the resource constraint gap by assessing the availability of existing resources. 8.The machine learning based power equipment anomaly early warning method of claim 1, wherein: The specific steps for calculating the maintenance priority index of the executable maintenance task set, classifying early warning levels based on resource constraint gaps, and generating an early warning output record set are as follows. The abnormal value index, equipment health status, task complexity, and resource requirements of each executable task in the executable maintenance task set are statistically analyzed to obtain the maintenance priority index. Based on the resource constraint gap and maintenance priority index, the executable tasks in the executable maintenance task set are divided into high warning level, medium warning level and low warning level. Extract the task ID, task priority, faulty component information, anomaly level, early warning lead time, resource gap information, and repair time from the set of executable maintenance tasks, and generate an early warning output record set. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is characterized in that: When the processor executes the computer program, it implements the steps of the machine learning-based power equipment anomaly early warning method according to any one of claims 1 to 8.

10. A computer readable storage medium having stored thereon a computer program, characterized in that: When the computer program is executed by the processor, it implements the steps of the machine learning-based power equipment anomaly early warning method according to any one of claims 1 to 8.

Images