Power equipment fault diagnosis method and system based on power large model

Abstract

The application relates to the technical field of power equipment fault detection, in particular to a power equipment fault diagnosis method and system based on a power large model, which comprises the following steps: converting multi-modal data of power cables during operation into a feature vector, inputting the feature vector into a pre-trained fault diagnosis model to obtain a preliminary diagnosis result and a confidence level; if the confidence level is not lower than a threshold, the preliminary diagnosis result is reserved; if the confidence level is lower than the threshold, the feature vector is taken as a current potential fault feature vector for secondary discrimination, which comprises the following steps: calculating the similarity and volatility of the current potential fault feature vector and a historical potential fault feature vector to obtain a significance index, analyzing a time change trend to obtain a cumulative trend index, comprehensively determining a potential fault index through the significance index and the cumulative trend index, and determining a final diagnosis result according to the potential fault index. The scheme improves the diagnosis accuracy of the fault diagnosis model for low-confidence potential faults and unknown faults.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power equipment fault detection. Specifically, it relates to a power equipment fault diagnosis method and system based on a power large model. BACKGROUND

[0002] The safe and stable operation of the power system is the cornerstone of social and economic development. As the lifeline of power transmission, the reliability of the operation state of the power cable is crucial. In recent years, with the development of deep learning technology, the use of power large models to process multi-modal data such as voltage, current, temperature, and partial discharge signals of power cables has become a leading technology to achieve accurate judgment and fault warning of power cables under different working conditions.

[0003] The prior art usually collects multi-modal data of power cables under different working conditions, trains a power equipment fault diagnosis model, and after training, a static mapping system from "features" to "known fault categories" has been formed inside the power equipment fault diagnosis model. Then, the multi-modal data of the power cable at a certain moment during operation is extracted, usually a feature vector is constructed, the feature vector is input into the power equipment fault diagnosis model, and the fault diagnosis result of the power cable at that moment is output, that is, various fault types and confidence levels of various fault types of the power cable at that moment, usually the fault type with the highest confidence level is selected as the fault at that moment, and those fault types with low confidence levels are no longer considered.

[0004] However, in the actual fault diagnosis process, for those completely new, never appeared in the training set fault mode or operating condition, the power equipment fault diagnosis model will produce a false judgment, for example, with the application of new materials, the introduction of new technology and the continuous change of the operating environment, the power cable may appear a completely new fault mechanism that has not been recorded in the historical data. When the multi-modal data corresponding to this new fault mode is input into the power equipment fault diagnosis model, since it cannot be matched to the known fault category, the fault diagnosis model will divide it into a similar known fault type with very low confidence, or directly exclude it as noise interference, resulting in inaccurate fault diagnosis results. SUMMARY

[0005] To solve the problem that the existing power equipment fault diagnosis model is prone to false judgment of new types of fault modes and new working conditions, and thus leads to inaccurate fault diagnosis results, the present application proposes a power equipment fault diagnosis method and system based on a power large model.

[0006] In a first aspect, the present application provides a power equipment fault diagnosis method based on a power large model, comprising:

[0007] convert the multi-modal data of the power cable at each moment in the operation process into a feature vector at the moment, and take the feature vector at any moment as a current feature vector;

[0008] input the current feature vector into a pre-trained power equipment fault diagnosis model, output a preliminary diagnosis result corresponding to the current feature vector and a confidence degree of the preliminary diagnosis result; if the confidence degree is greater than or equal to a preset confidence threshold, determine the preliminary diagnosis result as a fault diagnosis result of the power cable at the current moment; if the confidence degree is less than the preset confidence threshold, take the current feature vector as a current potential fault feature vector, and perform secondary screening, including:

[0009] determine a significance index of the current potential fault feature vector according to the similarity of the current potential fault feature vector and the historical potential fault feature vector and the volatility of the historical potential fault feature vector;

[0010] determine a cumulative trend index of the current potential fault feature vector by analyzing the change trend of the historical potential fault feature vector and the current fault feature vector in time;

[0011] take the product of the significance index and the cumulative trend index of the current potential fault feature vector as a potential fault index of the power cable at the current moment, and determine a fault diagnosis result of the power cable at the current moment based on a comparison result of the potential fault index and a preset potential fault index threshold.

[0012] The technical scheme converts multi-modal data into a feature vector, realizes quantitative integration of multi-dimensional information of the running state of the power cable, and provides comprehensive and calculable basic features for subsequent diagnosis; inputting the feature vector into a fault diagnosis model obtains a preliminary diagnosis result, which not only guarantees the efficiency of high-confidence fault diagnosis, but also avoids missing potential faults through a secondary screening mechanism for low-confidence samples; the construction of the significance index combines historical similar features and volatility analysis, effectively identifies the cluster characteristics of similar potential faults in the feature space, and reduces the interference of isolated noise; the cumulative trend index captures the evolution law of the fault feature over time through change analysis in the time dimension, especially suitable for the characteristics of progressive development of power equipment faults; the potential fault index is a comprehensive measure of aggregation and accumulation, which realizes precise quantitative evaluation of potential faults, and finally determines the diagnosis result through the index, thereby improving the accuracy of the diagnosis result of the fault diagnosis model for new fault modes and complex working conditions.

[0013] Preferably, the multi-modal data of the power cable at each moment in the operation process includes current, voltage, temperature, partial discharge signal, environmental humidity and environmental magnetic field strength.

[0014] Preferably, the significance index of the current potential fault feature vector is determined based on the following method: calculating the similarity between each historical potential fault feature vector and the current potential fault feature vector, as well as the volatility of all historical potential fault feature vectors; counting the number of historical potential fault feature vectors with similarity greater than or equal to a preset similarity threshold; and using the ratio of this number to the volatility as the significance index of the current potential fault feature vector.

[0015] This technical solution quantifies the similarity between the current potential fault feature vector and the historical potential fault feature vector, and performs comprehensive calculations by combining the discrete characteristics of the historical potential fault feature vector. This enables the effective capture of the spatial distribution pattern of fault features, which reduces the risk of isolated noise or random fluctuations being misjudged as faults, and enhances the ability to identify stable clusters of potential faults in the feature space.

[0016] Preferably, the volatility of all historical potential fault feature vectors is determined as follows: the centroid of all historical potential fault feature vectors is determined, and the Euclidean distance from each historical potential fault feature vector to the centroid is calculated; the mean, standard deviation, and median of the Euclidean distances from all historical potential fault feature vectors to the centroid are obtained, and the distribution pattern of the Euclidean distance is reflected by the ratio of the mean and the median; the volatility of all historical potential fault feature vectors is obtained by multiplying the ratio of the mean and the median by the standard deviation.

[0017] Preferably, the cumulative trend index of the current potential fault feature vector is determined as follows: historical potential fault feature vectors with similarity greater than a preset similarity threshold and the current potential fault feature vector are sorted chronologically, and the Euclidean distance between adjacent potential fault feature vectors is calculated to obtain an Euclidean distance sequence. The macro-cumulative trend index of the current potential fault feature vector is determined by performing linear regression analysis on the Euclidean distance sequence; the mean of all elements in the Euclidean distance sequence except the last element is calculated, and the ratio of the last element to the mean is determined as the recent aggravation index of the current potential fault feature vector; the product of the macro-cumulative trend index and the recent aggravation index is used as the cumulative trend index of the current potential fault feature vector.

[0018] This technical solution analyzes the characteristic evolution of potential failure modes with low confidence over time. The macro-cumulative index reflects the overall development trend of the failure, while the recent aggravation index quantifies the degree of abrupt change in the failure characteristics. The multiplication of the two achieves the integration of long-term trend stability and short-term change significance, so that the cumulative trend index not only conforms to the general law of gradual failure development, but also accurately responds to its possible accelerated deterioration characteristics.

[0019] Preferably, the macro-cumulative trend index of the current potential fault feature vector is determined as follows: the determination coefficient and fitted line of the Euclidean distance sequence are obtained through linear regression analysis; the normalized slope of the fitted line and the normalized mean of the Euclidean distance sequence are obtained; the normalized slope and the normalized mean are multiplied to obtain a trend component; the complementary value of the normalized mean and the determination coefficient is multiplied to obtain an amplitude component; the sum of the trend component and the amplitude component is used as the macro-cumulative trend index of the current potential fault feature vector; wherein, the complementary value of the determination coefficient is the difference between 1 and the determination coefficient.

[0020] Preferably, the fault diagnosis result of the power cable at the current moment is determined based on the following method: If the potential fault index is greater than or equal to the preset potential fault index threshold, it is determined that there is a potential fault at the current moment, and the following operations are performed: the historical potential fault feature vectors with similarity greater than the preset similarity threshold and the current potential fault feature vector are used to form a first vector cluster, and the feature vectors corresponding to each known fault type are used to form a second vector cluster. The spatial distribution characteristics of the first vector cluster and the second vector cluster are used to determine whether an unknown fault has occurred; if it has occurred, the potential fault at the current moment is an unknown fault; if it has not occurred, the potential fault at the current moment is a cumulative fault, and the fault type of the cumulative fault is determined by statistically analyzing the main fault types in the first vector cluster; if the potential fault index is less than the preset potential fault index threshold, it is determined that there is no potential fault at the current moment, and the preliminary diagnosis result is retained.

[0021] This technical solution combines hierarchical potential fault judgment logic with vector cluster space analysis to achieve accurate identification of potential faults in power cables. It not only improves the comprehensiveness and accuracy of fault diagnosis, but also enhances the ability to proactively identify unknown faults, providing a more detailed and reliable basis for the development of preventive maintenance, fault tracing, and handling strategies for power cables.

[0022] Preferably, determining whether an unknown fault has occurred based on the spatial distribution characteristics of the first and second vector clusters is done in the following way: calculating the probability of an unknown fault occurring at the current moment:

[0023]

[0024] In the formula, Given the possibility of unknown failures occurring at the current moment, Let be the minimum distance between the centroids of the first vector cluster and the centroids of each of the second vector clusters. The average centroid distance between all second vector clusters, when If the probability is greater than or equal to a preset probability threshold, an unknown fault is determined to have occurred at the current moment; when If the probability is less than the preset probability threshold, it is determined that no unknown fault has occurred at the current moment.

[0025] This technical solution provides a quantitative and rigorous basis for identifying unknown faults in power cables through dual criteria of centroid distance and inter-cluster distance. It accurately captures new fault types that differ significantly from known fault modes, and reduces the risk of missed or false judgments through dual verification. This provides a more accurate basis for early warning, type subdivision, and subsequent maintenance strategy formulation of power cable faults.

[0026] Preferably, if the potential fault at the current moment is an unknown fault, the following operations are also performed to adaptively evolve the power equipment fault diagnosis model: the specific fault type of the unknown fault is determined manually, and the current potential fault feature vector is marked according to the specific fault type. The marked current potential fault feature vector is added to the training dataset of the power equipment fault diagnosis model, and the adaptive evolution of the power equipment fault diagnosis model is achieved through incremental training.

[0027] This technical solution rapidly incorporates new fault features while retaining the existing diagnostic capabilities of the fault diagnosis model, enabling the model to continuously learn new fault modes, constantly enrich the fault identification library, and gradually improve the diagnostic accuracy and coverage of various newly added unknown fault modes. This forms a virtuous cycle from fault identification to manual calibration and then to model evolution, providing a dynamic optimization mechanism for the long-term reliability of power equipment fault diagnosis models and enhancing the adaptability of fault diagnosis models.

[0028] Secondly, the present invention also provides a power equipment fault diagnosis system based on a large power model. The power equipment fault diagnosis system includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of any power equipment fault diagnosis method.

[0029] The present invention has the following effects:

[0030] This invention constructs a preliminary diagnostic... Secondary screening The model evolution mechanism avoids the problem of misjudging new fault modes by traditional static fault diagnosis models. Secondary screening captures potential faults in low-confidence features from the dimensions of spatial correlation and temporal evolution, breaking through the limitations of traditional static mapping in identifying unknown modes. The adaptive evolution of the model enables it to dynamically adapt to the evolution of fault mechanisms and the migration of operating conditions, which not only improves the early identification accuracy of fault diagnosis models for new faults, but also improves the accuracy and reliability of fault diagnosis results for power equipment. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0032] Figure 2 This is a schematic diagram of the method flow for step S3 of the present invention. Detailed Implementation

[0033] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0034] The present invention provides a power equipment fault diagnosis method based on a large power model, such as... Figure 1 As shown, it includes:

[0035] S1: Acquire multimodal data of power cables during operation and convert it into feature vectors.

[0036] Fault characteristics of power cables are often manifested in the coordinated changes of multiple physical dimensions, and data from a single dimension are insufficient to fully reflect their operating status. For example, a sudden increase in partial discharge may be accompanied by an abnormal rise in temperature, while distortion of the environmental magnetic field may be related to insulation aging.

[0037] Therefore, this step integrates multimodal data, transforming scattered physical information into a unified feature vector, providing structured input for subsequent diagnostic models, and is the foundation for achieving accurate diagnosis.

[0038] The overall process of this step is as follows: first, collect multimodal data, then perform preprocessing, and then convert it into feature vectors.

[0039] Specifically, it includes:

[0040] S11: Multimodal data acquisition.

[0041] Real-time collection of six core data points of power cables during operation:

[0042] The first type is current. Hall effect sensors collect the real-time current in the cable core, reflecting the impact of load changes on the cable. Hall effect sensors operate based on the Hall effect; when current passes through the cable core, a ring-shaped magnetic field is generated around it. The strength of the magnetic field is proportional to the magnitude of the core current. Hall effect sensors can detect this magnetic field non-contactly and retrieve the real-time value of the core current through the law of electromagnetic induction. This non-invasive measurement method does not damage the cable insulation layer, making it suitable for online monitoring scenarios, especially meeting the safety monitoring requirements of power cables.

[0043] The second type is voltage, which uses a voltage divider sensor to obtain the voltage at both ends of the cable. The core principle of a voltage divider sensor (such as a resistive voltage divider or a capacitive voltage divider) is to use voltage dividing elements (resistors, capacitors, etc.) to indirectly measure high voltage by utilizing the voltage division law.

[0044] The third type is distributed temperature, which uses fiber optic grating sensor arrays to collect the temperature distribution of power cables and to capture local overheating phenomena. Fiber optic grating sensor arrays are one of the mainstream technologies for realizing distributed temperature monitoring of power cables. The principle is to etch multiple gratings with different center wavelengths on optical fibers. Each grating is an independent temperature sensing point, which can measure the temperature at different locations along the cable in real time, thereby forming a complete temperature distribution.

[0045] The fourth category consists of several statistical characteristics of partial discharge signals. Discharge pulse signals are detected using ultra-high frequency sensors. In the standard specifications for partial discharge detection and analysis, the following characteristics are most widely used due to their ease of acquisition, clear physical meaning, and strong correlation with defect states; they are the core analytical indicators for discharge pulse signals: First, the maximum discharge quantity, i.e., the maximum charge corresponding to a single discharge pulse, directly reflects the upper limit of the partial discharge intensity and is a core threshold indicator for assessing the insulation status of power equipment. Second, the discharge repetition rate, i.e., the number of discharge pulses per unit time, reflects the activity level of the discharge. A higher repetition rate indicates more frequent electric field disturbances at the insulation defect and a more unstable defect development trend. It is commonly used in the aging process of cable joints and insulation components; the repetition rate is an important early warning indicator. Third, the pulse peak value and average amplitude. As important time-domain characteristics of the signal, the pulse peak value is the maximum voltage / current amplitude of a single discharge; the average amplitude is the average value of all pulse peak values. The pulse peak value reflects the instantaneous intensity of a single discharge, while the average value reflects the overall average level of the discharge. Combining these two characteristics can distinguish between a few strong pulses and a majority of weak pulses. Fourth is the pulse width, which reflects the duration of the pulse from the rising edge to the falling edge. The half-peak width (WHM) is commonly used, i.e., the time interval during which the amplitude drops to 50% of its peak value. Different insulation defects have different discharge physical processes, resulting in significant differences in pulse width. Fifth is the signal morphological characteristics, including skewness and kurtosis. Skewness describes the asymmetry of the pulse amplitude distribution; kurtosis describes the steepness of the distribution. High kurtosis indicates the presence of numerous spike pulses, quantifying the distribution pattern of the partial discharge signal amplitude. It is a core feature for pattern recognition and is commonly used in the automatic identification of discharge patterns. For example, feature vectors are constructed using skewness and kurtosis to achieve automatic discharge pattern identification. These features cover the intensity, activity, waveform morphology, and distribution patterns of the discharge, and are easy to obtain (based on direct calculation or statistics of time-domain pulses), highly matching the technical solution of this invention. Therefore, they serve as key indicators in partial discharge signals.

[0046] The fifth category is ambient humidity, which is recorded using a humidity sensor to assess the risk of cable moisture absorption. Ambient humidity is one of the key environmental factors affecting the insulation performance of power cables. High humidity environments may cause the cable outer sheath and insulation layer to absorb moisture and become damp, accelerating insulation aging (such as water treeing aging of cross-linked polyethylene insulation under humid conditions), and even causing faults such as partial discharge and insulation breakdown. Therefore, recording ambient humidity through a humidity sensor is an important basis for assessing the risk of cable moisture absorption.

[0047] The sixth category is the ambient magnetic field strength. Magnetoresistive sensors measure the magnetic field distortion around the cable to reflect conductor structural anomalies. The current in a cable conductor generates a magnetic field surrounding it, the distribution of which is closely related to the conductor's structure. Under normal circumstances, the current distribution in a conductor (such as a round core or a multi-strand strand structure) is uniform, and the surrounding magnetic field exhibits a symmetrical ring distribution. If there are structural anomalies in the conductor (such as internal strand breaks, core eccentricity, or poor contact at joints leading to uneven current distribution), the symmetry of the magnetic field will be broken, causing magnetic field distortion (such as a sudden increase in local magnetic field strength, directional shift, or disruption of symmetry). Magnetoresistive sensors have high sensitivity to changes in the magnetic field and can detect subtle magnetic field distortions. By analyzing the degree, location, and morphology of the magnetic field distortion, it is possible to infer whether there are structural anomalies in the conductor.

[0048] S12: Data preprocessing.

[0049] At any given moment For the current moment, the 50 moments preceding the current moment (determined based on historical experience values) are set as all historical moments for the current moment. The six types of data collected at the current moment are filtered at a 50Hz power frequency to eliminate power grid interference, and then... The normalization formula standardizes various types of data to the same order of magnitude. Specifically, for the first... For data of type 3 (if it is type 3, the temperature of the power cable at each temperature sensing point is normalized separately; if it is type 4, each statistical feature of the partial discharge signal is normalized separately), the normalization formula is:

[0050]

[0051] In this formula, For the normalized first Class data, For the first Class data (raw values) The value range is from 1 to 6, corresponding to the 6 types of data in step S11. This is the mean of this type of data across all historical moments. This represents the standard deviation of this type of data across all historical moments.

[0052] S13: Construct feature vectors.

[0053] Will The time-normalized data from categories 1 to 6 are then concatenated sequentially to form... eigenvectors at time step , which serves as the current feature vector. for Current at any moment for Voltage at time, for The feature vector is formed by sequentially concatenating the temperatures at all temperature sensing points at any given time. for The feature vector is formed by sequentially concatenating all statistical features of the partial discharge signal at time 1. for The ambient humidity at all times, for The intensity of the ambient magnetic field at any given time.

[0054] Following these steps, the multimodal data of the power cable at each moment during operation can be transformed into a feature vector for that moment.

[0055] S2: Obtain preliminary diagnostic results and confidence levels of feature vectors through the fault diagnosis model.

[0056] A pre-trained power fault diagnosis model is acquired. This model, based on the Transformer architecture, has learned specific patterns for 12 known fault types, including short circuits and insulation aging. During fault diagnosis training, feature vectors are constructed from multimodal data of the power cable at each time step. These feature vectors are then labeled with their corresponding features under different fault types, enabling the large-scale power model to acquire fault diagnosis capabilities. The trained model has learned the feature patterns of known fault types and can quickly diagnose faults in the power cable's operating conditions at each time step.

[0057] The overall process of this step is as follows: First, input the current feature vector into the trained fault diagnosis model. Then, the fault diagnosis model outputs the preliminary diagnosis result and the confidence level of the preliminary diagnosis result. Finally, based on the confidence level of the preliminary diagnosis result, determine whether to start secondary screening. If secondary screening is started, then perform subsequent operations. If not started, then use the preliminary diagnosis result of the fault diagnosis model as the final diagnosis result, and the fault diagnosis process ends.

[0058] Specifically, it includes:

[0059] The current feature vector Input a pre-trained power equipment fault diagnosis model. This model captures the correlation between features through an attention mechanism and outputs a preliminary diagnosis result and its confidence level. The confidence level threshold is set to 0.8 (determined based on historical experience). The fault diagnosis model considers that when the confidence level of the preliminary diagnosis result is greater than or equal to 0.8, the preliminary diagnosis result is accurate and reliable. Conversely, when the confidence level of the preliminary diagnosis result is less than 0.8, the accuracy and reliability of the preliminary diagnosis result are relatively low.

[0060] In one example, the preliminary diagnostic results and confidence levels of the fault diagnosis model are as follows:

[0061] {"Fault Type":"Insulation Aging", "Confidence Level":0.92};

[0062] {"Fault Type":"Partial Discharge Exceeds Standard", "Confidence Level":0.88};

[0063] {"Fault Type":"Sheath Damage", "Confidence Level":0.85};

[0064] {"Fault Type":"Connector Overheating", "Confidence Level":0.81};

[0065] {"Fault Type":"Mechanical Damage", "Confidence Level":0.80};

[0066] In this example, the preliminary diagnostic results indicate that the power cable has a clear known fault type, reflecting that the power cable has multiple known faults at the same time (such as insulation aging accompanied by partial discharge). The confidence levels of various fault types included in the preliminary diagnostic results are all greater than or equal to the confidence threshold, which shows that the fault diagnosis model has achieved high reliability in identifying each fault type and can accurately capture the characteristics of different faults and their correlations.

[0067] This situation typically arises after long-term operation of power cables, as various components gradually age and become damaged, and faults may interact and accumulate. For example, insulation aging reduces insulation performance, making it prone to excessive partial discharge; sheath damage eliminates the protection of the cable's interior, potentially exacerbating mechanical damage and making joints more susceptible to overheating due to external environmental influences. Fault diagnosis models, through their internal attention mechanisms, can effectively uncover the potential connections between these fault characteristics, thus assigning a high confidence level to each fault. In terms of processing, since all fault confidence levels meet high reliability standards, "insulation aging," with the highest confidence level, can be used as the initial diagnostic result and prioritized for further processing.

[0068] Therefore, when the confidence levels of all types of faults included in the preliminary diagnosis results are greater than or equal to the confidence threshold, there is no need to initiate a secondary screening process. The fault type with the highest confidence level among all the fault types included in the preliminary diagnosis results is determined as the fault diagnosis result of the power cable at the current moment.

[0069] In one example, the preliminary diagnostic results and confidence levels of the fault diagnosis model are as follows:

[0070] {"Fault Type":"Insulation Aging", "Confidence Level":0.75};

[0071] {"Fault Type":"Moisture Deterioration", "Confidence Level":0.68};

[0072] {"Fault Type":"Electromagnetic Interference", "Confidence Level":0.52};

[0073] {"Fault Type":"Overload Operation", "Confidence Level":0.41};

[0074] {"Fault Type":"Outer Sheath Corrosion", "Confidence Level":0.33};

[0075] In this example, the preliminary diagnosis results indicate that the power cable may have a new type of unknown fault or a new complex operating condition interference. This shows that the fault diagnosis model cannot clearly match the known fault types. The confidence of each type of fault included in the preliminary diagnosis results is less than the confidence threshold, which means that the model's matching degree between the current feature vector and any known fault mode has not reached the reliable standard.

[0076] This situation typically corresponds to two scenarios: First, novel and unknown faults: the current fault mode has never appeared in the model training data (such as the unique degradation mode of new material cables), causing the fault diagnosis model to be unable to find a matching known type; this scenario is more likely. Second, complex operating conditions: sudden changes in the operating environment (such as strong surrounding electromagnetic interference or extreme weather) cause the data distribution to deviate from the training samples, making it difficult for the fault diagnosis model to accurately classify the fault; this scenario is less likely. In this case, directly ignoring low-confidence results may miss real fault risks. Further analysis and screening are needed to avoid omissions and misjudgments due to the model not having learned the fault mode.

[0077] The training data for any fault diagnosis model can only cover "known faults in known scenarios." However, equipment fault modes evolve continuously with technological iterations (such as new materials, new structures, and new operating logic). For example, the degradation of cables made of new materials may involve entirely new chemical / physical mechanisms that have never appeared in the fault samples of traditional cables, resulting in a lack of corresponding feature learning foundations for the model. This unknown is systematic; the feature vectors of new fault modes will be fundamentally different from those of known faults, and the model cannot generalize to identify them using existing knowledge, thus inevitably resulting in low-confidence results. From an engineering practice perspective, technological upgrades (such as material innovation and equipment iteration) are commonplace, and the emergence of new fault modes is a concomitant phenomenon of technological upgrades, occurring far more frequently than extreme operating conditions. Complex operating condition interference is the impact of temporary environmental mutations on data distribution, but its interference is accidental and limited. Furthermore, extreme operating conditions (such as strong electromagnetic interference and short-term extreme weather) are low-probability events. During normal operation, the working environment of equipment is usually within a relatively stable range (such as temperature, humidity, and electromagnetic compatibility thresholds), and the frequency of extreme mutations is low.

[0078] Therefore, when the confidence levels of all types of faults included in the preliminary diagnostic results are less than the confidence threshold, a secondary screening (spatial cluster analysis and time trend judgment) must be performed to use the current feature vector as the current potential fault feature vector. Potential fault signals are then mined from the historical correlation and dynamic evolution of the current potential fault feature vector to further confirm the fault nature of the current potential fault feature vector.

[0079] In one example, the preliminary diagnostic results and confidence levels of the fault diagnosis model are as follows:

[0080] {"Fault Type":"Connector Overheating", "Confidence Level":0.91};

[0081] {"Fault Type":"Partial Discharge Exceeds Standard", "Confidence Level":0.79};

[0082] {"Fault Type":"Insulation Aging", "Confidence Level":0.65};

[0083] {"Fault Type":"Mechanical Damage", "Confidence Level":0.58};

[0084] In this example, when the preliminary diagnosis indicates that the power cable has both a dominant fault and a minor fault, the confidence levels of each fault type output by the fault diagnosis model exhibit a mixed distribution. That is, some of the confidence levels for each fault type included in the preliminary diagnosis are above the confidence threshold, while others are below it. Although the confidence levels of each fault type output by the fault diagnosis model exhibit a mixed distribution, as long as there are preliminary diagnosis results with a confidence level greater than or equal to the confidence threshold, it indicates that the fault diagnosis model has a clear ability to identify fault modes. In this case, it is determined that the power cable has a fault type that meets the confidence requirement, and the dominant fault "joint overheating" is confirmed as the fault present in the power cable at the current moment.

[0085] This situation is essentially a derivative or accompanying phenomenon of a known fault, rather than an independent unknown fault mode. Identifying the dominant fault provides a clear basis for maintenance decisions; the dominant fault should be addressed first, while low-confidence types can be used as supplementary references to avoid wasting resources.

[0086] Therefore, when the confidence levels of each fault type in the output of the fault diagnosis model exhibit a mixed distribution, the fault type with the highest confidence level among all fault types included in the preliminary diagnosis results is identified as the fault diagnosis result of the power cable at the current moment. There is no need to initiate a secondary screening process. This processing logic not only ensures the accurate identification of core faults but also simplifies the analysis cost of non-critical information, which is in line with the principle of prioritizing the resolution of the main contradictions in engineering practice.

[0087] S3: Secondary screening is performed based on the significance index and cumulative trend index of the feature vector to obtain the final diagnostic result.

[0088] When the confidence level of the initial diagnosis is lower than the threshold, it is necessary to further explore potential fault signals through secondary screening. The core logic is that the essential characteristics of power cable faults are reflected in the dual dimensions of spatial clustering and temporal accumulation. Single-dimensional analysis is difficult to avoid noise interference or pattern misjudgment. The significance index and cumulative trend index of the feature vector precisely quantify the fault attributes from these two dimensions respectively. The combination of the two can achieve accurate judgment of potential faults.

[0089] The significance index is calculated by comparing the similarity between the current potential fault feature vector and the historical potential fault feature vector, as well as the volatility of the historical potential fault feature vector. Essentially, it measures whether the current feature forms a stable cluster with a sufficient number of historical potential fault features in a high-dimensional space. This is because the multimodal parameter combinations of the same type of fault (such as the correlation between partial discharge signal features and temperature distribution) have inherent consistency and exhibit spatial clustering characteristics. This index ensures that the identified fault features have repeating patterns by eliminating isolated noise from discrete distributions.

[0090] Among them, the cumulative trend index quantifies the gradual deterioration of faults over time by analyzing the evolution of similar feature vectors. Power equipment faults are mostly progressive (such as insulation aging accompanied by a continuous increase in partial discharge), and the differences in their feature vectors will show regular changes over time. This index avoids misjudging static parameter fluctuations as faults by capturing this dynamic evolution.

[0091] In summary, the combination of the two essentially achieves a dual verification of "spatial pattern verification" and "temporal evolution confirmation": the significance index ensures that the current feature belongs to a stable fault mode cluster, and the cumulative trend index verifies the real evolution of the pattern over time. Together, they eliminate the interference of isolated noise and random fluctuations, and finally accurately identify the real faults hidden in the low confidence results (including unknown faults and cumulative known faults), making up for the limitations of fault diagnosis models in identifying complex and novel fault modes.

[0092] Specifically, such as Figure 2 As shown, it includes:

[0093] S31: Determine significance indicators based on the similarity and volatility of historical potential fault feature vectors.

[0094] The significance index measures the degree of clustering between current potential fault feature vectors and historical potential fault feature vectors to determine whether a stable fault mode has formed. Specifically:

[0095] If a stable fault mode is formed, it indicates that the current potential fault feature vector and a sufficient number of historical potential fault feature vectors exhibit a significant clustering state in the feature space. That is, it has a high similarity with historical potential fault feature vectors, and the overall fluctuation of historical potential fault feature vectors is small. This shows that the current potential fault feature vector is not an isolated random disturbance, but a continuation of the same type of potential fault features in the time dimension. Its feature pattern has consistency and repeatability, which conforms to the objective law that fault features usually show a clustered distribution. The feature vectors caused by the same fault mechanism will not randomly discrete.

[0096] If a stable fault mode is not formed, it indicates that the current potential fault feature vector does not form an effective cluster with the historical potential fault feature vector in the feature space. This may be because the historical potential fault feature vector has high volatility, and even if similar historical potential fault feature vectors exist, they cannot form a stable cluster. The current potential fault feature vector has low similarity with all historical potential fault feature vectors, and its feature mode deviates significantly from the distribution range of existing potential faults. This may be due to occasional isolated anomalies caused by extreme operating conditions and sensor malfunctions, which cannot form a stable association with historical potential fault feature vectors.

[0097] The overall process of this step is as follows: first, obtain the historical potential fault feature vectors of the current potential fault feature vector; calculate the similarity between the historical potential fault feature vectors and the current potential fault feature vectors; analyze the volatility of the historical potential fault feature vectors; and finally determine the significance index of the current potential fault feature vectors.

[0098] Specifically, it includes:

[0099] S311: Obtain the feature vector of historical potential faults.

[0100] In one embodiment, the method for determining the historical potential fault feature vector is the same as the method for determining the current potential fault feature vector. That is, following the steps S12 and S13, first obtain the feature vectors corresponding to all historical moments at the current moment, and then obtain the preliminary diagnosis results of each of the feature vectors corresponding to all historical moments through step S2. If the confidence of all fault types included in the preliminary diagnosis result of the feature vector corresponding to a certain historical moment is less than the confidence threshold, the feature vector corresponding to that historical moment is taken as the historical potential fault feature vector, and then all historical potential fault feature vectors of the current potential fault feature vector are obtained.

[0101] S312: Calculate similarity and volatility.

[0102] In one embodiment, the similarity between each historical potential fault feature vector and the current potential fault feature vector is determined based on cosine similarity. That is, the cosine similarity between each historical potential fault feature vector and the current potential fault feature vector is calculated, and the cosine similarity is used as the similarity between the historical potential fault feature vector and the current potential fault feature vector.

[0103] In one embodiment, the volatility of all historical potential fault feature vectors is determined based on the following:

[0104] Determine the centroid of all historical potential fault feature vectors This represents the central location of all historical potential fault feature vectors. The total number of historical potential fault feature vectors. For the first A historical potential fault feature vector, This represents the index of the historical potential fault feature vector, with a value range of [value range missing]. .

[0105] Calculate the feature vector of each historical potential fault to The Euclidean distance, for example, the first Historical potential fault feature vectors to The Euclidean distance is .

[0106] Extracting distance statistics: Converting all historical potential fault feature vectors to... The mean of the Euclidean distance is denoted as All historical potential fault feature vectors to The standard deviation of the Euclidean distance is denoted as All historical potential fault feature vectors to The median of the Euclidean distance is denoted as .

[0107] Determine the volatility of all historical potential fault feature vectors:

[0108]

[0109] In the formula, For the volatility of all historical potential fault feature vectors, For all historical potential fault feature vectors to The mean of the Euclidean distance, For all historical potential fault feature vectors to The median of the Euclidean distance, For all historical potential fault feature vectors to The standard deviation of the Euclidean distance, The median is used to reflect the skewness of the Euclidean distance distribution. In a symmetrical distribution, the median and the mean are close, with a ratio of approximately 1. If there are many extreme values ​​(e.g., in a positively skewed Euclidean distance distribution, there are many large Euclidean distances), the mean will be significantly larger than the median, and the ratio will increase. This directly reflects the degree of asymmetry in the Euclidean distance distribution. The Euclidean distance directly measures the dispersion of the Euclidean distance; the larger the value, the looser the distribution of the historical potential fault feature vectors around the centroid. The parameter is used to prevent unreasonable calculations caused by a denominator of 0, and its value is [value missing]. To avoid the extreme case where all historical potential fault feature vectors are exactly the same, which would result in the mean and median of the Euclidean distance from all historical potential fault feature vectors to the centroid being 0, thus causing computational errors.

[0110] The result of multiplying the two is This combines the skewness and dispersion of the Euclidean distance distribution. When the distribution of historical potential fault feature vectors is concentrated and symmetrical, The smaller the value, the higher the consistency of the historical potential fault feature vector, and the smaller the fluctuation; when the distribution is significantly skewed and the dispersion is large, A larger value indicates significant differences in the historical potential fault feature vectors. The greater the volatility, the more likely it is to contain fault features of different types or severity.

[0111] This calculation process can capture the overall dispersion range of the data and reflect the impact of extreme values ​​on the distribution pattern, thus more accurately characterizing the volatility of historical potential fault feature vectors.

[0112] S313: Determine the significance index.

[0113] In one embodiment, the significance index of the current potential fault feature vector is determined as follows: a similarity threshold of 0.7 is set (based on historical experience values), and the number of historical potential fault feature vectors with a similarity greater than or equal to 0.7 is counted; the ratio of this number to the volatility of historical potential fault feature vectors is used as the significance index of the current potential fault feature vector.

[0114] The specific formula is as follows:

[0115]

[0116] In this formula, This serves as a significance index for the current potential fault feature vector. The number of historical potential fault feature vectors with a similarity greater than or equal to 0.7. The volatility of all historical potential fault feature vectors. The larger the value, the more significant the current potential fault feature vector is, indicating that it has formed a close cluster with a sufficient number of historical potential fault feature vectors. The parameter used to prevent the denominator from being zero has a value of . To avoid the extreme case where all historical potential fault feature vectors are exactly the same, thus preventing volatility... The result is a calculation error due to the value being 0.

[0117] When obtaining the significance index of the current potential fault feature vector, the volatility of all historical potential fault feature vectors is introduced. Its core function is to correct the dispersion characteristics of historical potential fault feature vectors. The influence of this allows the significance index to more accurately reflect the cluster significance of the current potential fault feature vector in historical data.

[0118] Specifically: It only measures the number of historical potential fault feature vectors similar to the current potential fault feature vector, but this number depends on the distribution characteristics of the historical potential fault feature vectors themselves. If the historical potential fault feature vectors are highly volatile, that is, the historical vectors are scattered and have low consistency, even... A larger fluctuation in the current potential fault feature vector could be because the historical potential fault feature vector itself is randomly and loosely distributed. Similar historical potential fault feature vectors are only a small subset of this random distribution, and therefore cannot be considered part of a stable fault cluster. Conversely, if the historical potential fault feature vector exhibits low volatility, meaning its distribution is concentrated and consistent, then... A larger value indicates that the current potential fault feature vector closely matches a stable fault cluster, making it more significant.

[0119] therefore, The essence is to Normalization to the volatile background of historical potential fault feature vectors avoids misjudgment of similarity due to the dispersion of historical potential fault feature vectors themselves, and ultimately enables the significance index of the current potential fault feature vector to more objectively measure whether the current potential fault feature vector belongs to a significant fault feature cluster.

[0120] S32: Determine cumulative trend indicators based on the time evolution trend of historical potential fault feature vectors.

[0121] Power cable faults often develop gradually (e.g., insulation aging worsens over time), and their dynamic evolution is difficult to capture solely through spatial characteristics. Cumulative trend indicators, by analyzing the time-series changes of eigenvectors, determine whether a fault exhibits a continuous worsening trend, thus overcoming the static limitations of spatial analysis.

[0122] The overall process of this step is as follows: first, sort the current potential fault feature vectors and similar historical potential fault feature vectors; then calculate the Euclidean distance sequence; next, construct macro trend indicators and recent aggravation indicators; and finally synthesize cumulative trend indicators.

[0123] Specifically, it includes:

[0124] S321: Sort by time.

[0125] The historical potential fault feature vectors with a similarity greater than or equal to 0.7 with the current potential fault feature vector are sorted in chronological order.

[0126] S322: Obtain the Euclidean distance sequence.

[0127] After sorting, the Euclidean distance between every two adjacent potential fault feature vectors is calculated. Since there are a total of historical potential fault feature vectors with a similarity greater than or equal to 0.7, the following conditions are met: There are [number] potential fault feature vectors, plus the current potential fault feature vectors, totaling [number]. Given a number of potential fault feature vectors, the following can be calculated: Euclidean distance, this Each Euclidean distance constitutes a sequence of Euclidean distances. .

[0128] The physical significance of the Euclidean distance sequence lies in quantifying the degree of difference in the potential fault feature vectors at adjacent time points (each dimension of the feature vector corresponds to the difference in the key operating parameters of the power cable). The larger the Euclidean distance, the more significant the difference in the multi-dimensional parameter combination between the two time points.

[0129] When the Euclidean distance shows a continuous increasing trend over time, it essentially reflects an accelerating dynamic evolution rate of potential fault characteristics. For power equipment, fault development often follows a gradual deterioration pattern; for example, insulation aging progresses from local defects to overall breakdown, and joint overheating progresses from slight temperature rise to the risk of melting. In the early stages of a fault, changes in relevant parameters (such as partial discharge quantity and temperature deviation) are usually relatively gradual, with small differences between adjacent time intervals. As the fault progresses, the magnitude and rate of parameter changes gradually increase (e.g., the partial discharge pulse frequency increases from 10 times per hour to 100 times per hour, and the temperature rises from ambient temperature +5℃ to +20℃). These drastic and coordinated changes in these multi-dimensional parameters lead to an increase in the Euclidean distance between adjacent potential fault feature vectors.

[0130] This increase in Euclidean distance is not a random fluctuation. Since the Euclidean distance sequence corresponds to potential fault feature vectors (with the current potential fault feature vector) that have a similarity greater than the similarity threshold, they belong to the same type of potential fault mode. The expansion of their feature differences does not come from mode switching, but rather from the progressive development of fault severity within the same fault mode. For example, for the same type of insulation defect fault, initially it only manifests as weak partial discharge (feature vector A). After 1 hour, due to the expansion of the defect, a significant temperature increase occurs (feature vector B), and the Euclidean distance between the two is small. After another hour, the partial discharge surges and the temperature exceeds the threshold (feature vector C). The Euclidean distance between B and C is significantly greater than the Euclidean distance between A and B. In this process, the increase in Euclidean distance directly corresponds to the evolution of the fault from a minor hidden danger to a significant fault.

[0131] Conversely, if the Euclidean distance remains stable or fluctuates slightly, it indicates that the potential fault characteristics are in a state of slow change or stable fluctuation, which may correspond to the fault latency period or that parameter disturbances during normal operation have not exceeded the fault evolution threshold. Therefore, the increasing trend of the Euclidean distance sequence can serve as the core basis for judging whether a potential fault has entered a rapid development stage, providing quantitative dynamic evolution evidence for subsequent cumulative trend analysis.

[0132] S323: Linear regression analysis to determine macro-cumulative trend indicators.

[0133] The macro-cumulative trend indicator achieves scientific quantification of long-term macroeconomic cumulative characteristics by dynamically integrating "trend directionality" and "overall magnitude of change." The core essence of macro-cumulative trend is a long-term, continuous process of change, analyzed primarily from two dimensions: first, whether the trend change has a stable direction; and second, whether the overall magnitude of the trend change is significant. The reliability of these two dimensions dynamically changes with data characteristics. When the trend is stable, trend directionality is key; when the trend is chaotic, the overall magnitude of change better reflects the cumulative effect of failures.

[0134] In one embodiment, the macro-cumulative trend index of the current potential fault feature vector is determined based on the following method:

[0135] The coefficient of determination and the fitted line of the Euclidean distance sequence were obtained through linear regression analysis. It is used to measure the goodness of fit of a fitted straight line. In regression analysis, The closer the value is to 1, the stronger the explanatory power of the fitted line for the Euclidean distance sequence; that is, the closer the elements of the Euclidean distance sequence are to the fitted line, the more stable the trend. , It is the slope of the fitted straight line, reflecting the long-term trend. When the difference between potential fault feature vectors at adjacent time points increases continuously over time, it reflects an accelerating trend of coordinated deterioration of feature parameters (such as partial discharge and temperature deviation) under the same fault mode, indicating that the fault is in a continuous and accumulating process; when This indicates that the difference between potential fault feature vectors at adjacent time points decreases over time, which may correspond to the mitigation of fault features (such as the dissipation of temporary disturbances) or a tendency to stabilize, and does not meet the trend characteristics of cumulative deterioration. It is the independent variable. This is the intercept term of the linear regression equation, and the mean of the Euclidean distance sequence is calculated. .

[0136] Use separately The function normalizes the slope of the fitted line and the mean of the Euclidean distance sequence: , ,in, For the normalized slope, For the normalized mean, exist The range of values ​​in the interval is ,exist The range of values ​​in the interval is ,pass Map the range to Within the interval, normalization is performed.

[0137] The macro-cumulative trend index of the current potential fault feature vector satisfies the following relationship:

[0138]

[0139] In this formula, This serves as a macro-cumulative trend indicator for the current potential fault feature vector. For the normalized slope, The coefficient of determination for the Euclidean distance sequence. The coefficient of determination is a fundamental statistic that can be directly obtained in linear regression analysis. It is a standard output indicator of linear regression analysis and can be obtained directly using conventional statistical tools or algorithms without the need for additional complex derivation or special processing. For the normalized mean, for The complementary values. The cumulative trend is the continuous change of the potential fault feature vector over time. The slope is related to the direction and strength of the cumulative trend, and the coefficient of determination is a reliability check of the changing trend. The larger the coefficient of determination, the more realistic the changing trend, and vice versa.

[0140] In this formula, As a trend component, it is used to reflect the trend of change. This is an amplitude component used to reflect the overall magnitude of characteristic changes. The two are fused together by addition to quantify the long-term macro trend.

[0141] when The larger the value, and the closer it is to 1, the more linearly the Euclidean distance sequence changes (e.g., the fault characteristics continue to worsen), and the more stable the trend of change. Give Higher weights, weights shift towards trend components tilt, Positive and negative become dominant The core: If at this time Trend components It is a positive value and the value is relatively large. Overall, the value is relatively high, directly quantifying the trend of stable and deteriorating fault accumulation (such as a continuous surge in partial discharge due to insulation aging), highlighting the contribution of long-term, continuously deteriorating faults to the accumulation trend; at this time, if Trend components If the value is negative and the absolute value is large, The overall value is relatively low, accurately reflecting the trend of stable and mitigated faults (such as improved heat dissipation conditions causing the joint temperature deviation to gradually decrease).

[0142] when The smaller the value, the closer it is to 0, indicating that the linearity of the Euclidean distance sequence is difficult to explain the changes, showing an unstable trend. Give Higher weights, weights shifted to magnitude components Inclined, through the mean of the Euclidean distance Reflecting the overall cumulative scale, even if the trend fluctuates, if the overall distance remains consistently large, it will lead to... The larger the value, the more cumulative the anomaly. This dynamic weighting design avoids the limitations of a single indicator.

[0143] In summary, by dynamically integrating "trend directionality" and "overall change magnitude", the macro-cumulative trend indicator was accurately quantified. The combination of positive slope and high determination coefficient is the quantitative characteristic of typical cumulative faults (such as a stable trend that continues to deteriorate); the combination of positive slope and low determination coefficient captures atypical cumulative characteristics that deteriorate during fluctuations (such as intermittent partial discharges but with increasing overall intensity); and negative slope, regardless of the level of determination coefficient, reduces the possibility of cumulative faults (such as the mitigation or stabilization of fault characteristics).

[0144] The final macro-cumulative trend indicator can not only scientifically quantify the strength of the long-term macro trend, but also resist noise interference, providing an interpretable and implementable quantitative basis for early fault warning and maintenance decision-making, and improving the accuracy and robustness of cumulative fault identification.

[0145] S324: Determine the recent aggravation index of the current potential fault feature vector.

[0146] The core function of the recent aggravation indicator is to quantify the degree of abrupt change in the characteristics of potential faults within the latest time window, and to compensate for the smoothing effect that macro trend analysis may have on short-term drastic changes. Its design logic is highly consistent with the dynamic evolution characteristics of power equipment faults.

[0147] In one embodiment, the recent aggravation index of the current potential fault feature vector is determined based on the following:

[0148] Calculate the mean of all elements in the Euclidean distance sequence except the last element, and determine the ratio of the last element to this mean as the recent aggravation index of the current potential fault feature vector. The essence of this ratio is to calibrate recent changes using historical normal levels to determine whether the latest changes deviate from the normal evolutionary pattern of this type of fault.

[0149] In a Euclidean distance sequence, the mean of all elements except the last element represents the typical range of change for this type of potential fault during its historical evolution. It serves as a statistical anchor point for the regular evolution rate of fault characteristics over a past period (from the earliest similar feature vector to the moment preceding the current vector). For example, the historical Euclidean distance mean for a certain type of insulation defect fault is 3.2, meaning that the characteristic differences between adjacent time points have typically remained at this level throughout its development. The last element of the Euclidean distance sequence corresponds to the difference between the current potential fault feature vector and the previous similar historical potential fault feature vector, directly reflecting the range of change in fault characteristics within the most recent time interval.

[0150] A higher recent aggravation index indicates that the magnitude of the most recent characteristic change significantly exceeds the historical average, meaning that the fault evolution rate has suddenly accelerated recently. For example, the historical Euclidean distance average for an overheating fault in a cable joint is 2.5, reflecting a normal temperature rise of 2 to 3°C per hour. However, the last element is 8.3, reflecting a temperature rise of 10°C in the most recent hour, and the recent aggravation index is 3.3, significantly exceeding the historical average. This indicates that the overheating of the joint has entered a stage of rapid deterioration from a slow rise, possibly accompanied by critical conditions such as a sudden increase in contact resistance, requiring urgent intervention.

[0151] When the recent aggravation index is closer to 1, it indicates that recent changes are consistent with historical patterns, and the fault is still in a stable evolution phase, requiring no additional emergency response, but continued monitoring of macroeconomic trends is necessary. When the aggravation index is smaller and closer to 0, it indicates that the recent changes are significantly lower than historical levels, which may mean that the fault has eased (e.g., a sudden drop in ambient temperature has suppressed local overheating).

[0152] S325: Composite cumulative trend indicator.

[0153] In one embodiment, the cumulative trend index of the current potential fault feature vector is synthesized by multiplying the macro-cumulative index with the recent aggravation index as the cumulative trend index of the current potential fault feature vector.

[0154] The specific calculation formula is as follows:

[0155]

[0156] In this formula, This is a cumulative trend indicator of the current potential fault feature vector. This serves as a macro-cumulative trend indicator for the current potential fault feature vector. This serves as an indicator of the recent aggravation of the current potential fault feature vector. The larger the value, the more significant the deterioration of the fault is in both the long-term trend and recent changes, and the stronger the time-cumulative effect.

[0157] The product synthesis method of the cumulative trend index aligns with the evolutionary nature of power equipment faults. It avoids misjudging short-term stagnation due to significant long-term trends, and also prevents ignoring isolated disturbances with no clear long-term direction due to short-term abrupt changes. It ensures that high index values ​​correspond only to the true cumulative fault state that has been continuously deteriorating in the long term and has recently accelerated in its evolution. It captures the typical process of power equipment faults (such as joint overheating and insulation aging) from quantitative to qualitative change. The long-term trend reflects the gradual development of the fault mechanism, while the recent aggravation reflects the parameter mutations in the critical stage. The coupling of the two forms a dual verification of "long-term trend + recent mutation", which enables the cumulative trend index to reflect both the overall evolution direction of the fault and capture short-term critical changes, thereby improving the diagnostic accuracy of potential faults.

[0158] S33: Comprehensively determine the potential fault indicators of the current potential fault feature vector.

[0159] A single indicator is insufficient to comprehensively determine potential faults; a fusion of indicators is necessary for comprehensive evaluation. Potential fault indicators combine significance and cumulativeity to achieve hierarchical judgments of "whether it is a fault" and "fault type," serving as a core decision-making step in secondary screening. The diagnostic results output in this step are directly linked to the subsequent model evolution mechanism, forming a complete diagnostic loop of "calculating potential fault indicators → determining whether a potential fault exists → distinguishing between unknown faults and cumulative faults."

Claims

1. A method for fault diagnosis of power equipment based on a large power model, characterized in that, include: The multimodal data of the power cable at each moment during operation is transformed into the feature vector at that moment, and the feature vector at any moment is used as the current feature vector. The current feature vector is input into the pre-trained power equipment fault diagnosis model, and the preliminary diagnosis result and the confidence level of the preliminary diagnosis result are output. If the confidence level is greater than or equal to the preset confidence level threshold, the preliminary diagnosis result is determined as the fault diagnosis result of the power cable at the current moment. If the confidence level is less than the preset confidence threshold, the current feature vector is used as the current potential fault feature vector, and secondary screening is performed, including: The significance index of the current potential fault feature vector is determined based on the similarity between the current potential fault feature vector and the historical potential fault feature vector, as well as the volatility of the historical potential fault feature vector. The cumulative trend index of the current potential fault feature vector is determined by analyzing the changing trends of historical potential fault feature vectors and current fault feature vectors over time. The product of the significance index and the cumulative trend index of the current potential fault feature vector is used as the potential fault index of the power cable at the current moment. Based on the comparison between the potential fault index and the preset potential fault index threshold, the fault diagnosis result of the power cable at the current moment is determined, including: If the potential fault index is greater than or equal to a preset potential fault index threshold, a potential fault is determined to exist at the current moment, and the following operations are performed: The historical potential fault feature vectors with a similarity greater than a preset similarity threshold are combined with the current potential fault feature vector to form a first vector cluster, and a second vector cluster is formed using all feature vectors corresponding to each known fault type. The spatial distribution characteristics of the first and second vector clusters are used to determine whether an unknown fault has occurred, including: calculating the probability of an unknown fault occurring at the current moment. ;In the formula, Given the possibility of unknown failures occurring at the current moment, Let be the minimum distance between the centroids of the first vector cluster and the centroids of each of the second vector clusters. The average centroid distance between all second vector clusters; when If the probability is greater than or equal to a preset probability threshold, an unknown fault is determined to have occurred at the current moment; when If the probability is less than a preset probability threshold, it is determined that no unknown fault has occurred at the current moment; If a potential fault occurs, the current potential fault is an unknown fault; if no potential fault occurs, the current potential fault is a cumulative fault, and the fault type of the cumulative fault is determined by statistically analyzing the main fault types within the first vector cluster; if the potential fault index is less than the preset potential fault index threshold, it is determined that there is no potential fault at the current time, and the preliminary diagnostic results are retained.

2. The power equipment fault diagnosis method according to claim 1, characterized in that, The multimodal data of power cables at each moment during operation include: current, voltage, temperature, partial discharge signal, ambient humidity, and ambient magnetic field strength.

3. The power equipment fault diagnosis method according to claim 1, characterized in that, The significance index of the current potential fault feature vector is determined based on the following method: Calculate the similarity between each historical potential fault feature vector and the current potential fault feature vector, as well as the volatility of all historical potential fault feature vectors; The number of historical potential fault feature vectors with a similarity greater than or equal to a preset similarity threshold is counted; the ratio of this number to the volatility is used as the significance index of the current potential fault feature vector.

4. The power equipment fault diagnosis method according to claim 3, characterized in that, The volatility of all historical potential fault feature vectors is determined based on the following method: Determine the centroid of all historical potential fault feature vectors, and calculate the Euclidean distance from each historical potential fault feature vector to that centroid; Obtain the mean, standard deviation, and median of the Euclidean distance from all historical potential fault feature vectors to the centroid, and reflect the distribution pattern of the Euclidean distance by the ratio of the mean to the median; Multiplying the ratio of the mean to the median by the standard deviation yields the volatility of all historical potential fault characteristic vectors.

5. The power equipment fault diagnosis method according to claim 1, characterized in that, The cumulative trend index of the current potential fault feature vector is determined based on the following method: Historical potential fault feature vectors and current potential fault feature vectors with similarity greater than a preset similarity threshold are sorted in chronological order, and the Euclidean distance between adjacent potential fault feature vectors is calculated to obtain an Euclidean distance sequence. Linear regression analysis is performed on the Euclidean distance sequence to determine the macro-cumulative trend index of the current potential fault feature vector. Calculate the mean of all elements in the Euclidean distance sequence except the last element, and determine the ratio of the last element to the mean as the recent aggravation index of the current potential fault feature vector; The product of the macro-cumulative trend indicator and the recent aggravation indicator is used as the cumulative trend indicator of the current potential fault feature vector.

6. The power equipment fault diagnosis method according to claim 5, characterized in that, The current macro-cumulative trend index of potential fault feature vectors is determined based on the following method: The determination coefficients and fitted lines of the Euclidean distance sequence were obtained through linear regression analysis. Obtain the normalized slope of the fitted line and the normalized mean of the Euclidean distance sequence; Multiplying the normalized slope and the normalized mean yields a trend component; Multiplying the normalized mean and the complementary value of the coefficient of determination yields an amplitude component; The sum of the trend component and the magnitude component is used as the macro-cumulative trend indicator of the current potential fault feature vector; where the complementary value of the coefficient of determination is the difference between 1 and the coefficient of determination.

7. The power equipment fault diagnosis method according to claim 1, characterized in that, If the potential fault at the current moment is an unknown fault, the following operations are also performed to adaptively evolve the power equipment fault diagnosis model: The specific fault type of the unknown fault is determined manually, and the current potential fault feature vector is labeled according to the specific fault type. The labeled current potential fault feature vector is added to the training dataset of the power equipment fault diagnosis model. The adaptive evolution of the power equipment fault diagnosis model is achieved through incremental training.

8. A power equipment fault diagnosis system based on a large power model, characterized in that, The power equipment fault diagnosis system includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the power equipment fault diagnosis method as described in any one of claims 1-7.

Images