A multi-parameter electrical system fault diagnosis method

Abstract

The application relates to the technical field of electrical fault diagnosis, and discloses a multi-parameter electrical system fault diagnosis method. The method collects multi-dimensional operation parameters of an electrical system, wherein the multi-dimensional operation parameters include voltage waveform data, current waveform data and temperature distribution data; time-frequency joint analysis is performed on the multi-dimensional operation parameters to extract multi-scale electrical characteristics, wherein the multi-scale electrical characteristics include steady-state characteristic components and transient-state characteristic components; based on a historical fault case library, pattern matching is performed on the multi-scale electrical characteristics to generate an initial fault type set; a dynamic weight distribution algorithm is used to evaluate the confidence of fault types in the initial fault type set, and high-confidence fault types are screened out; and according to the high-confidence fault types, a fault evolution path model is constructed, which is used to describe the time sequence of fault characteristics.

Description

Technical Field

[0001] This invention relates to the field of electrical fault diagnosis technology, specifically to a fault diagnosis method for multi-parameter electrical systems. Background Technology

[0002] As the core power source of modern industrial production and social life, the reliability of electrical systems is of paramount importance. Electrical system faults can range from minor issues like equipment downtime affecting production to major accidents such as fires. Traditional electrical system fault diagnosis often relies on threshold alarms for single parameters (such as the RMS value of current) or the tripping action of protective devices. While these methods are simple and direct, they have significant shortcomings. Threshold alarms typically trigger only after a fault has occurred or developed to a certain extent, making them reactive alarms lacking early warning capabilities. Furthermore, early signs of faults in electrical systems often manifest as subtle harmonics in voltage and current waveforms, transient impacts, or abnormal changes in localized temperature. This information cannot be reflected in a single RMS parameter, making it difficult to detect early faults in a timely manner.

[0003] With the development of sensor technology, acquiring multi-dimensional parameters of electrical systems during operation has become possible, such as high-sampling-rate voltage and current waveforms and temperature distributions at key nodes. These data contain rich information about the system's health status. However, effectively extracting fault-related features from this massive, heterogeneous data and accurately identifying fault types remains a challenge. Existing methods may focus on steady-state feature analysis, neglecting important information contained in transient processes; or, while capable of extracting multiple features, they lack effective feature fusion and filtering mechanisms, leading to diagnostic results that are highly susceptible to noise interference and have low confidence levels. Furthermore, electrical faults are a dynamic evolutionary process, with characteristics changing over time from initial weak symptoms to eventual severe failure. Current diagnostic methods mostly focus on a "snapshot" judgment of the current state, lacking the ability to analyze and predict fault evolution trends, and failing to provide sufficient time windows for preventative maintenance. Therefore, a diagnostic method is needed that can deeply integrate multi-dimensional parameters, intelligently filter key features, and track the fault evolution process to achieve early and accurate diagnosis and risk warning of electrical system faults. Summary of the Invention

[0004] The purpose of this invention is to provide a fault diagnosis method for multi-parameter electrical systems to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a fault diagnosis method for multi-parameter electrical systems, the method comprising:

[0006] Collect multi-dimensional operating parameters of the electrical system, including voltage waveform data, current waveform data, and temperature distribution data;

[0007] Time-frequency joint analysis is performed on the multi-dimensional operating parameters to extract multi-scale electrical features, which include steady-state feature components and transient feature components.

[0008] Based on the historical fault case library, pattern matching is performed on the multi-scale electrical features to generate an initial set of fault types.

[0009] A dynamic weight allocation algorithm is used to evaluate the confidence of the fault types in the initial fault type set, and high-confidence fault types are selected.

[0010] Based on the high-confidence fault type, a fault evolution path model is constructed, which is used to describe the temporal correlation of fault features;

[0011] By combining real-time monitoring data, the fault evolution path model is dynamically corrected to generate optimized fault diagnosis results.

[0012] Preferably, the step of performing time-frequency joint analysis on the multi-dimensional operating parameters to extract multi-scale electrical features includes:

[0013] An adaptive decomposition algorithm is used to separate the voltage waveform data and current waveform data to obtain the low-frequency fundamental component and the high-frequency harmonic component.

[0014] Energy density analysis is performed on the high-frequency harmonic components to determine the range of abnormal frequency bands;

[0015] Based on the abnormal frequency band range, the transient feature components are extracted, and the transient feature components include pulse amplitude and duration;

[0016] Phase offset detection is performed on the low-frequency fundamental component to extract the steady-state characteristic component, which includes amplitude fluctuation rate and phase consistency index.

[0017] Preferably, the step of performing pattern matching on the multi-scale electrical features based on a historical fault case library to generate an initial fault type set includes:

[0018] The steady-state and transient feature components are input into a pre-trained fault classification model to output candidate fault types.

[0019] A matching score is calculated based on the similarity between the candidate fault type and the fault features in the historical fault case library;

[0020] Candidate fault types with matching scores higher than a preset threshold are selected to form the initial fault type set.

[0021] Preferably, the step of using a dynamic weight allocation algorithm to evaluate the confidence level of the fault types in the initial fault type set includes:

[0022] The weighting coefficients are dynamically adjusted based on the contribution of the steady-state and transient characteristic components.

[0023] Based on the weighting coefficients, calculate the overall confidence level of each fault type in the initial fault type set;

[0024] Fault types with an overall confidence level below the threshold are removed, while the high-confidence fault types are retained.

[0025] Preferably, the construction of the fault evolution path model includes:

[0026] Extract the historical fault feature sequence corresponding to the high-confidence fault type;

[0027] Analyze the temporal variation patterns of the historical fault characteristic sequences and establish a fault characteristic transition matrix;

[0028] Based on the fault feature transition matrix, predict the next stage evolution trend of the current fault features.

[0029] Preferably, the dynamic correction of the fault evolution path model includes:

[0030] Real-time acquisition of updated operating parameters of the electrical system, and extraction of newly added electrical features;

[0031] The newly added electrical features are compared with the prediction results of the fault evolution path model, and the deviation is calculated.

[0032] If the deviation exceeds the allowable range, the parameters of the fault feature transfer matrix are readjusted.

[0033] Preferably, generating the optimized fault diagnosis result includes:

[0034] Based on the revised fault evolution path model, output the final fault type and evolution stage;

[0035] Based on the temperature distribution data, the physical rationality of the final fault type is verified.

[0036] Preferably, verifying the physical plausibility of the final failure type includes:

[0037] Establish a mapping relationship between the temperature distribution data and electrical loss characteristics;

[0038] Determine whether the theoretical temperature rise corresponding to the final fault type is consistent with the actual temperature rise data;

[0039] If there is a discrepancy, the dynamic weight allocation algorithm and the fault evolution path model correction steps shall be re-executed.

[0040] Preferably, the multi-dimensional operating parameters of the acquired electrical system include:

[0041] Simultaneously acquire three-phase voltage and current signals to ensure timestamp alignment;

[0042] Temperature sampling values ​​of key nodes are obtained through a distributed temperature sensor network;

[0043] The voltage waveform data, current waveform data, and temperature distribution data are time-synchronized and calibrated.

[0044] Preferably, before performing time-frequency joint analysis on the multi-dimensional operating parameters, the method further includes:

[0045] The original data is segmented using a sliding window mechanism;

[0046] Apply a windowing function to each data segment to suppress spectral leakage;

[0047] Interpolation algorithms are used to compensate for signal distortion caused by uneven sampling intervals.

[0048] Compared with the prior art, the beneficial effects of the present invention are:

[0049] This invention significantly improves the depth and reliability of electrical system fault diagnosis through multi-level information processing and dynamic analysis. By acquiring multi-dimensional parameters such as voltage waveforms, current waveforms, and temperature distribution, a more comprehensive information foundation is provided for diagnosis, enabling it to move beyond simple electrical quantities. Joint time-frequency analysis of these multi-dimensional parameters simultaneously captures the time and frequency domain characteristics of the signal, effectively extracting both steady-state features characteristic of long-term operation and transient features specific to the early stages of a fault. This enhances the detection capability for early or intermittent faults.

[0050] An initial set of fault types is generated through pattern matching based on a historical fault case database, leveraging existing knowledge and experience to provide preliminary direction for diagnosis. A key optimization step is the introduction of a dynamic weight allocation algorithm to evaluate the confidence level of fault types in the initial set. This algorithm dynamically adjusts the confidence level of different fault hypotheses based on factors such as the degree of matching between current features and historical cases, and the salience of the features themselves, thereby selecting high-confidence fault types and reducing the possibility of misjudgments and omissions. Constructing a fault evolution path model is another significant advantage of this method, extending the diagnostic perspective from static point-based judgment to dynamic process analysis. This model describes the possible sequences of fault characteristics changing over time, helping to understand the development stages of the fault and the next possible signs. Finally, the evolution path model is dynamically revised using real-time monitoring data, ensuring that the diagnostic results are continuously updated in accordance with changes in system state. This method not only identifies existing faults but also provides insights into potential fault development trends, offering valuable information for scheduling maintenance and preventing fault escalation. Attached Figure Description

[0051] Figure 1 This is a comprehensive analysis diagram for multi-parameter fault diagnosis;

[0052] Figure 2 A flowchart for multi-dimensional time-frequency joint analysis of operating parameters and multi-scale electrical feature extraction;

[0053] Figure 3 A flowchart for multi-scale electrical feature pattern matching and initial fault type set generation;

[0054] Figure 4 This is a diagram of the state transition probability matrix for fault characteristics. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] Please see Figure 1This invention provides a multi-parameter electrical system fault diagnosis method. The method includes collecting multi-dimensional operating parameters of the electrical system, including voltage waveform data, current waveform data, and temperature distribution data; performing time-frequency joint analysis on the multi-dimensional operating parameters to extract multi-scale electrical features, including steady-state feature components and transient feature components; performing pattern matching on the multi-scale electrical features based on a historical fault case library to generate an initial fault type set; using a dynamic weight allocation algorithm to evaluate the confidence of the fault types in the initial fault type set and selecting high-confidence fault types; constructing a fault evolution path model based on the high-confidence fault types, which describes the temporal correlation of fault features; and dynamically correcting the fault evolution path model using real-time monitoring data to generate an optimized fault diagnosis result.

[0057] Example 1: See Figure 2 The time-frequency joint analysis employs an adaptive decomposition algorithm to process the acquired voltage and current waveform data. This algorithm, based on wavelet transform principles, achieves signal separation. After preprocessing, the voltage and current waveform data are input into the adaptive decomposition algorithm, which decomposes the signal into a low-frequency fundamental component and a high-frequency harmonic component through multi-resolution analysis. The low-frequency fundamental component mainly contains steady-state information near the power system frequency, while the high-frequency harmonic component covers a wide range of harmonic and transient components. The specific execution of the adaptive decomposition algorithm relies on the filter bank structure of the discrete wavelet transform. The db4 wavelet is selected as the basis function to perform multi-level decomposition on the voltage and current waveform data. Each level of decomposition generates approximation coefficients and detail coefficients; the approximation coefficients constitute the low-frequency fundamental component, and the detail coefficients constitute the high-frequency harmonic component. The number of decomposition levels is determined based on the sampling frequency and the frequency band of interest, with six levels performed to cover the frequency range from DC components to several kilohertz. The signal separation process requires setting appropriate thresholds to filter the coefficients, eliminating noise interference and retaining effective fault characteristic information. The threshold setting adopts a general threshold rule, with the threshold value being the signal standard deviation multiplied by the logarithmic transform coefficient.

[0058] Energy density analysis is performed on high-frequency harmonic components, using the short-time Fourier transform (SFT) method to calculate the energy distribution of the signal in the time-frequency plane. The high-frequency harmonic components are divided into multiple time segments, and a Hanning window function followed by a fast Fourier transform is applied to each time segment to obtain the spectral information. The energy value of each frequency point on the time axis is calculated, forming a time-frequency energy matrix. By analyzing the regions of abnormal energy concentration in the time-frequency energy matrix, the abnormal frequency band range is determined. The identification of the abnormal frequency band range is based on an energy threshold comparison method, setting a dynamic energy threshold; frequency bands exceeding the threshold are marked as abnormal. The dynamic energy threshold is set to twice the average energy of the background noise. Transient characteristic components are extracted based on the abnormal frequency band range, including two main indicators: pulse amplitude and duration. The pulse amplitude is directly extracted from the time-domain waveform, and extreme points in the waveform are identified using a peak detection algorithm. The duration is calculated using zero-crossing detection combined with envelope analysis to determine the start and end points of the pulse waveform. The pulse amplitude reflects the intensity of the fault discharge, while the duration characterizes the degree of sustained impact of the fault event. The extraction of transient feature components should take into account waveform distortion, and cubic spline interpolation algorithm should be used to improve measurement accuracy.

[0059] Phase shift detection of the low-frequency fundamental component is performed using phase-locked loop (PLL) technology to track phase changes. A digital PLL model is constructed, using the system's nominal frequency as a reference, to compare the actual and ideal phases of the low-frequency fundamental component in real time. The phase shift detection outputs a phase difference sequence, and the phase consistency index is calculated through statistical analysis of the phase difference sequence. Amplitude fluctuation is extracted from the envelope of the low-frequency fundamental component, and the ratio of the standard deviation to the mean of the waveform amplitude is calculated. Steady-state characteristic components include amplitude fluctuation and phase consistency index. Amplitude fluctuation characterizes the stability of voltage or current, while the phase consistency index reflects the equilibrium state of the three-phase system.

[0060] The entire time-frequency joint analysis process employs a sliding window mechanism, setting a fixed-length data window that slides along the time axis. Data within each window undergoes independent adaptive decomposition, energy density analysis, and feature extraction. The length of the sliding window is dynamically adjusted based on signal characteristics, taking an integer multiple of the power frequency period to ensure waveform integrity. The Blackman window function is used for windowing to reduce the impact of spectral leakage on the analysis results. During signal reconstruction, interpolation algorithms are used to compensate for data distortion caused by non-uniform sampling, improving the accuracy of feature extraction. The integration of multi-scale electrical features combines steady-state and transient feature components into a feature vector. The steady-state feature components include numerical quantization results of amplitude fluctuation rate and phase consistency indices, while the transient feature components include measured data of pulse amplitude and duration. The feature vectors are arranged in chronological order to form a feature sequence, providing input data for subsequent pattern matching. The dimension of the feature vectors is determined according to actual needs, containing six main feature parameters to comprehensively describe the operating state of the electrical system.

[0061] The parameter configuration of the adaptive decomposition algorithm needs to be optimized according to the specific application scenario, and the choice of wavelet basis function affects the feature extraction effect. The db4 wavelet has tight support and approximate symmetry, making it suitable for analyzing transient signals in power systems. Determining the number of decomposition levels requires a balance between frequency resolution and computational complexity; a 6-level decomposition achieves this balance. The time-frequency resolution setting for energy density analysis requires a trade-off, with a window length of 256 sampling points for the short-time Fourier transform. An adaptive threshold algorithm is used to determine the abnormal frequency band range, with the threshold dynamically adjusted based on the background noise level. The initial threshold is set to 1.5 times the base noise energy, and then fine-tuned based on real-time signal characteristics. Pulse amplitude detection needs to avoid false positives caused by noise; a moving average filtering method combined with differential operations is used to improve detection reliability. The calculation of duration requires precise determination of the pulse start and end points, applying rising and falling edge detection algorithms combined with hysteresis control logic.

[0062] The phase-locked loop (PLL) design for phase offset detection employs a second-order loop filter structure to ensure the speed and stability of phase tracking. The phase consistency index is calculated based on variance analysis of the three-phase phase difference; a smaller variance indicates a more balanced system. The amplitude fluctuation rate is calculated using a moving window statistical method, with the window length corresponding to a multiple of the power frequency cycle to ensure statistical significance. The overlap rate setting of the sliding window mechanism affects the real-time performance and continuity of feature extraction; a 50% overlap rate is used to balance performance requirements. The data processing flow in the time-frequency joint analysis process strictly adheres to signal processing specifications. Voltage and current waveform data undergo standardization preprocessing, including detrending and normalization. Coefficients generated by adaptive decomposition require threshold denoising, using a soft threshold function to retain effective signal components. Energy density analysis results require smoothing, using a moving average filter to eliminate random fluctuations. The extracted feature data undergoes standardization transformation to ensure comparability of features with different dimensions.

[0063] The real-time performance of the time-frequency joint analysis algorithm is ensured through optimized computational processes. A fast wavelet transform algorithm is employed to reduce computational complexity, and frequency domain decomposition characteristics are utilized to improve operational efficiency. The short-time Fourier transform for energy density analysis uses the FFT algorithm to accelerate spectrum calculation. Fixed-point arithmetic replaces floating-point arithmetic in the feature extraction process, improving processing speed while maintaining accuracy. A sliding window mechanism employs a circular buffer to manage data, reducing data movement overhead. The extraction of multi-scale electrical features meets the real-time requirements of online fault diagnosis, providing timely and reliable feature data for subsequent processing stages. The signal preprocessing stage includes digital filtering operations, using a Butterworth low-pass filter to eliminate high-frequency noise. The filter's cutoff frequency is set to one-quarter of the sampling frequency, and the stopband attenuation reaches 40 dB. The preprocessed signal is normalized, adjusting the amplitude range to the [-1, 1] interval. Normalization improves the numerical stability of subsequent algorithms and avoids large numerical calculation errors.

[0064] The number of wavelet decomposition layers is selected based on the sampling frequency and signal characteristics. For a system with a sampling rate of 10kHz, a 6-layer decomposition can cover the fundamental frequency band of 0-312.5Hz and the harmonic frequency band of 312.5Hz-5kHz. After each layer of decomposition, the detail coefficients are thresholded to remove noise-related coefficient components. A soft thresholding function is used, and the threshold value is adaptively adjusted according to the noise level. When reconstructing the signal, only important coefficient components are retained to achieve signal denoising and feature enhancement. Energy density analysis uses a sliding window short-time Fourier transform with a window function length of 256 points and a sliding step size of 128 points. The power spectral density is calculated after applying a Hanning window to the signal within each window, and the power spectral density is estimated using the periodogram method. Anomaly detection is achieved by comparing the ratio of energy at each frequency point to the average energy; frequency bands with a ratio exceeding 3.0 are marked as anomalies. The boundaries of anomaly frequency bands are accurately located using the gradient descent method to improve the accuracy of frequency band division.

[0065] Transient feature extraction comprises two stages: pulse detection and parameter calculation. Pulse detection employs a dual-threshold comparison method, setting high and low thresholds to eliminate interfering pulses. The pulse amplitude is calculated as the difference between the waveform's extreme value and the baseline, obtained through moving average filtering. Pulse duration is calculated as the time interval from the rising edge exceeding the low threshold to the falling edge falling below it. Pulse feature statistics include multiple parameters such as maximum amplitude, average amplitude, duration, and repetition frequency. Steady-state feature analysis focuses on the amplitude and phase characteristics of the fundamental component. Amplitude volatility is calculated using a sliding window standard deviation, with a window length of 10 power frequency cycles. Phase consistency is measured by the standard deviation of the three-phase phase difference, calculated based on sampling points within one cycle. Phase offset detection utilizes an improved phase-locked loop (PLL) structure, comprising a phase detector, loop filter, and voltage-controlled oscillator (VCO). The PLL bandwidth is set to 10Hz to balance tracking speed and noise immunity. Feature vector construction combines transient and steady-state features by weighting, with weight coefficients allocated according to feature importance. The transient feature weight is 0.6, and the steady-state feature weight is 0.4. This weight allocation is based on the feature's sensitivity to faults. The feature vectors are standardized to ensure consistent dimensions. Standardization uses the z-score method, converting the feature values ​​into a distribution with a mean of 0 and a standard deviation of 1. The processed feature vectors are then input into the subsequent pattern recognition module to complete the fault diagnosis process.

[0066] Example 2: See Figure 3 The pattern matching process inputs steady-state and transient feature components into a pre-trained fault classification model, which employs a support vector machine (SVM) architecture to achieve multi-class classification. After standardization and preprocessing, the steady-state and transient feature components form a feature vector containing feature data across multiple dimensions, including amplitude volatility, phase consistency index, pulse amplitude, and duration. The SVM model is trained based on fault sample data accumulated in a historical fault case database, containing 3,000 fully labeled fault case samples. Each fault case sample includes a feature vector and a corresponding fault type label, covering typical electrical fault types such as short-circuit faults, open-circuit faults, grounding faults, and insulation aging faults. The SVM model uses a radial basis function as the kernel function and solves for the optimal classification hyperplane using a sequential minimum optimization algorithm. During model training, five-fold cross-validation is used to adjust hyperparameters, including the penalty coefficient and the kernel width.

[0067] The fault classification model receives real-time feature vectors and outputs candidate fault types, which are represented by probability vectors indicating the likelihood of each fault type. The decision function of the support vector machine (SVM) model calculates the distance from the feature vectors to each category hyperplane, and the distance value is converted into a probability value using a sigmoid function. Fault types with a probability value exceeding 0.5 are initially selected as candidate fault types, forming an initial list of candidate fault types. A matching score is calculated based on the similarity between the candidate fault types and fault features in the historical fault case library. The similarity calculation uses a cosine similarity algorithm to measure the cosine value of the angle between feature vectors. Each candidate fault type is compared with a standard feature template of the same type of fault in the historical fault case library. The standard feature template is obtained through cluster analysis of historical fault feature vectors of the same type. The matching score is defined as the product of the cosine similarity value and the probability value output by the SVM, and the matching score ranges from 0 to 1. Candidate fault types with a matching score higher than 0.7 are selected to form the initial fault type set. The 0.7 threshold is determined based on a balance between recall and precision of historical data.

[0068] A dynamic weight allocation algorithm is used to assess the confidence of fault types in the initial fault type set. This algorithm dynamically adjusts the weight coefficients based on feature importance analysis. Feature importance analysis employs a random forest algorithm to calculate the contribution of steady-state and transient feature components. The random forest algorithm contains one hundred decision trees, and feature importance is assessed using Gini impurity reduction. The contribution of steady-state and transient feature components quantifies the influence of each feature on fault classification. Transient feature components contribute more to sudden faults, while steady-state feature components are more discriminative for progressive faults. The dynamic weight allocation algorithm dynamically adjusts the weight coefficients based on the contribution of steady-state and transient feature components, with a positive correlation between the weight coefficients and feature contributions. The weight coefficients are normalized so that the sum of the weights of steady-state and transient feature components is 1. The normalization process uses the softmax function. The weight coefficient of the transient feature component is calculated as the ratio of its contribution to the total contribution, and the weight coefficient of the steady-state feature component is calculated similarly.

[0069] The overall confidence level of each fault type in the initial fault type set is calculated based on weighted coefficients using a weighted summation model. This model uses the probability values ​​output by the support vector machine as the base confidence level and linearly weights them with the weighted coefficients. The overall confidence level is calculated as the dot product of the base confidence level and the weighted coefficients, reflecting the overall reliability of the fault type. Fault types with an overall confidence level below 0.6 are removed, while high-confidence fault types are retained. The 0.6 threshold is set based on the false alarm rate requirement. Fault types with an overall confidence level below the threshold are considered unreliable diagnoses and removed from the initial fault type set. The retained high-confidence fault types are sorted from highest to lowest overall confidence level, forming an optimized fault type list.

[0070] The real-time performance of the dynamic weight allocation algorithm is ensured through an incremental learning mechanism, which periodically updates the feature importance evaluation results. After processing every 100 new samples, the random forest model recalculates the feature contribution and dynamically adjusts the weight coefficients. The incremental learning mechanism adapts to changes in the operating state of the electrical system, ensuring that the weight allocation matches the actual operating conditions. The collaborative work of pattern matching and confidence assessment forms a closed-loop optimization system. The support vector machine model provides preliminary classification results, and the dynamic weight allocation algorithm refines the results. The two modules exchange information through a standard data interface. The support vector machine model outputs candidate fault types and probability values, while the dynamic weight allocation algorithm inputs feature vectors and candidate fault types to calculate the comprehensive confidence score. The system outputs a list of high-confidence fault types and their corresponding confidence scores, providing input for subsequent fault evolution path construction. The maintenance of the historical fault case library adopts a rolling update strategy, adding newly occurring fault cases to the library after verification. The case library capacity is set to a maximum of 10,000 cases, and old cases are eliminated using a least recently used strategy. The case library index structure uses a B+ tree to achieve fast retrieval and supports multi-dimensional queries based on fault type and occurrence time. The fault feature similarity calculation optimizes the traditional nearest neighbor algorithm by introducing a distance-weighted mechanism to improve matching accuracy. Similarity calculation considers not only the spatial distance of feature vectors but also incorporates a time dimension weight, making recently occurring fault cases more valuable. The time weight coefficient is negatively correlated with the time of occurrence of the case; cases within six months have a time weight of 1, while the time weight of cases older than six months decreases linearly.

[0071] The online learning function of the Support Vector Machine (SVM) model is implemented through a hot update mechanism, with model parameters updated incrementally periodically without interrupting service. The hot update process uses a shadow mode switching, where the new model smoothly replaces the old model after validation in the test environment. Model version management records each update log, supporting rapid rollback to a stable version. The robustness of the dynamic weight allocation algorithm is ensured through multi-dimensional verification, with threshold limits set for weight coefficient changes to avoid drastic fluctuations. Consistency checks are performed before and after weight coefficient updates, and abnormal changes trigger a manual review process. Multiple verification points are set internally to monitor the rationality of feature contribution calculations. The collaborative optimization of the fault classification model and the weight allocation algorithm adopts an alternating iterative approach: first, the weight coefficients are fixed to optimize the classification model parameters, then the classification model is fixed to optimize the weight coefficients. This alternating iteration is performed for three rounds, stopping optimization after reaching the convergence condition. The convergence condition is that the change in the overall confidence score between two consecutive iterations is less than 0.01. The initial fault type set is generated using a multi-threshold screening strategy, with different matching score thresholds set for different fault types. The matching score threshold for severe fault types such as short-circuit faults is set to 0.6, and the threshold for general fault types is set to 0.7. The multi-threshold strategy improves diagnostic sensitivity while ensuring safety and avoids missing serious faults.

[0072] The interpretability of confidence assessment results is achieved through contribution decomposition, where the overall confidence of each fault type can be decomposed into contribution scores for each feature dimension. These contribution scores are visualized as histograms. The visualization interface also displays matching details of historical similar cases. System real-time performance is enhanced through a parallel computing architecture, with pattern matching and confidence assessment executed in parallel on two independent computing threads. The two threads exchange data via shared memory. Computing resources are automatically allocated based on system resource utilization. The reliability of fault diagnosis results is enhanced through a multi-model voting mechanism, deploying decision tree and neural network models as auxiliary classifiers in addition to support vector machine models. The outputs of the three models are integrated through a voting mechanism, with voting weights allocated based on the historical accuracy of each model. This voting mechanism reduces the risk of misjudgment by a single model. Quality control of the historical fault case database employs an automatic verification mechanism. Newly added cases undergo three layers of checks: format verification, range verification, and logical verification. Format verification ensures data format compliance, range verification checks if feature values ​​are within a reasonable range, and logical verification verifies the causal relationship between features and fault types. Cases that fail verification are marked as pending review and await manual processing.

[0073] The adaptive capability of the dynamic weight allocation algorithm is achieved through a sliding window mechanism, with weight coefficients calculated based on the most recent 30 days of operational data. The sliding window length is set to 30 days, and the window content is updated daily. Samples within the window undergo time-decay weighting, with recent samples receiving higher weights to ensure the algorithm's sensitivity to system changes. The pattern matching process is optimized using feature selection technology, removing redundant features through correlation analysis. Feature selection uses the maximum information coefficient to evaluate the correlation between features and fault types, retaining feature dimensions with high correlation to fault types. Feature selection is performed quarterly. The generalization capability of the fault classification model is ensured through regularization techniques; the support vector machine model uses L2 regularization to prevent overfitting. The regularization coefficient is selected using cross-validation to balance model complexity and fitting ability. An early stopping strategy is employed during model training, terminating training when validation set performance no longer improves. The entire pattern matching and confidence assessment system adopts a modular design, with components connected through standard interfaces, supporting independent upgrades and maintenance. The system's operational status is monitored in real time, and key indicators such as processing latency, accuracy, and resource utilization are visualized through a dashboard. Abnormal operating conditions trigger alarms, notifying operations and maintenance personnel to intervene promptly. The standardization preprocessing of feature vectors uses the z-score method to transform each feature dimension into a distribution with a mean of 0 and a standard deviation of 1. The standardization parameters are calculated from historical data, including the mean and standard deviation of each feature. Newly input feature vectors are standardized using these fixed parameters to ensure processing consistency.

[0074] The Support Vector Machine (SVM) model employs a one-to-one strategy for multi-class classification, training a binary classifier for every two fault types. For n fault types, a total of n(n-1) / 2 binary classifiers are trained. A voting mechanism is used during prediction, with each binary classifier voting for its supported class; the class with the most votes is the final output. The Random Forest algorithm calculates feature importance using a permutation importance evaluation method, observing model performance changes by randomly permuting feature values. The feature importance score is defined as the degree of performance degradation; a greater performance degradation indicates a more important feature. The permutation importance evaluation is repeated thirty times, and the average is taken as the final importance score. The dynamic weight allocation algorithm uses a smooth transition strategy for weight coefficient updates, performing a weighted average of the new and old weight coefficients. The smoothing factor is set to 0.3, with the new weight coefficient accounting for 30% and the old weight coefficient accounting for 70%. This smooth transition avoids fluctuations in diagnostic results caused by sudden weight changes. The fault diagnosis results are output in a standardized JSON format, including fault type, confidence level, timestamp, and feature details fields. The JSON format facilitates integration with other systems and supports RESTful API calls. The output data is simultaneously written to the database for persistent storage, for use in historical queries and analysis.

[0075] Example 3: Constructing a fault evolution path model requires extracting historical fault feature sequences corresponding to high-confidence fault types, retrieved chronologically from a historical fault case database. Each historical fault feature sequence contains feature data from multiple time points during the fault development process, including the numerical changes of steady-state and transient feature components. After extracting the historical fault feature sequences, the temporal variation patterns of the sequences are analyzed using state transition analysis. The fault development process is discretized into multiple states, each corresponding to a feature value interval. State division is based on the natural breakpoints of feature values, using the Jenks natural breakpoint algorithm to divide continuous feature values ​​into five state levels. State transition analysis counts the frequency of transitions from each state to the next state, forming a state transition frequency matrix. A fault feature transition matrix is ​​established based on the state transition frequency matrix. The fault feature transition matrix is ​​a square matrix, where each element represents the conditional probability of transitioning from the current state to the next state. Its mathematical expression is:

[0076]

[0077] in: Represents the fault characteristic transition matrix. Indicates from state Transition to state The probability, The matrix represents the total number of states, and satisfies the following conditions: For any Both are true. Probability value By statistically analyzing historical data from the state Transition to state The calculation is based on the proportion of the number of transfers to the total number of transfers, and the specific formula is as follows:

[0078]

[0079] in: Indicates from state Transition to state The probability of, and the corresponding probability value. Indicates from state Transition to state Number of times, Indicates from state The total number of times a fault is transferred out. The fault characteristic transfer matrix captures the statistical regularity of fault development and reflects the trend of fault characteristics evolving over time.

[0080] The evolution trend of the current fault feature in the next stage is predicted based on the fault feature transition matrix. The prediction process uses a Markov chain model, mapping the current fault feature to the corresponding state. Query the fault feature transition matrix for the first... row vector The state with the highest probability value is selected as the most likely next state, while the transition probabilities of other states are recorded as auxiliary references. The evolution trend prediction results include the most likely next state and its probability, as well as a reliability assessment of the state transition. The fault evolution path model is dynamically corrected by combining real-time monitoring data, which is continuously acquired through a data acquisition system. Updated operating parameters of the electrical system are collected in real time, including the latest voltage waveform data, current waveform data, and temperature distribution data. The newly added electrical features are compared with the prediction results of the fault evolution path model, and the deviation between the actual observed values ​​and the predicted values ​​is calculated during the comparison process. The deviation is calculated using Euclidean distance to measure the difference between the actual feature vector and the predicted feature vector. For each feature dimension, the absolute difference between the actual value and the predicted value is calculated, and the square root of the sum of the squares of the differences in all dimensions is used to obtain the overall deviation. The deviation calculation formula takes into account the dimensional differences of different features and the features are standardized beforehand.

[0081] If the deviation exceeds the allowable range, the parameters of the fault feature transition matrix are readjusted. The allowable range is set based on the statistical characteristics of historical data, with the 95th percentile of historical deviations used as the threshold. When the actual deviation exceeds this threshold, the fault feature transition matrix is ​​considered to need updating. Parameter adjustment employs a Bayesian update method, using new state transition observations as evidence to update the probability distribution of the fault feature transition matrix. The parameter adjustment process of the fault feature transition matrix uses Bayesian smoothing techniques, applying them to each observed state transition. The corresponding probability value The updated formula is:

[0082]

[0083] in: Indicates from state Transition to state The probability, This is a smoothing parameter, typically set to 1. This represents the total number of states. This smooth update method avoids the zero-probability problem while ensuring that the weights of the new observations on the fault feature transition matrix are reasonable.

[0084] The dynamic correction process employs a sliding window mechanism, considering only state transition data within the most recent timeframe. The sliding window length is set to thirty days, with the window content updated daily. State transition data within the window has a time-sensitive weight, with more recent data receiving a higher weight and older data receiving a decreasing weight. The weighting function uses an exponential decay form. ,in Indicates the age of the data. The attenuation coefficient is set to 0.05 to ensure the model's sensitivity to system changes. The fault evolution path model is validated through backtesting, which uses historical data to verify the model's predictive accuracy. Historical data is divided into training and test sets. The training set is used to construct the initial fault feature transition matrix, and the test set is used to evaluate the prediction performance. Evaluation metrics include state prediction accuracy, trend prediction consistency, and deviation distribution. Backtesting is performed monthly, and model parameters are adjusted based on the results. The model prediction results are visualized using a state transition diagram, which visually displays the transition probabilities between states. Nodes represent fault states, directed edges represent state transitions, and labels on the edges indicate the transition probabilities. The state transition diagram helps maintenance personnel understand fault development patterns and provides decision support for preventative maintenance. The visualization system also provides a prediction trajectory display, showing the most likely development path of fault characteristics.

[0085] The comparison results between real-time monitoring data and model predictions are displayed in real-time on a monitoring dashboard. The dashboard shows the current deviation, allowable range, and key indicators of model confidence. When the deviation approaches the allowable range, the system issues an alert, prompting operations personnel to pay attention to changes in model performance. Alert information is color-coded to distinguish severity: green indicates normal, yellow indicates attention, and red indicates intervention is needed. The fault evolution path model is stored using versioned management, saving a new version of the fault feature transition matrix after each parameter adjustment. Version records include update time, adjustment reason, and impact assessment information, supporting model rollback and comparative analysis. Version history data helps analyze the model evolution process. Model performance monitoring sets multiple quality indicators, including prediction accuracy, response time, and resource utilization. Quality indicators are calculated and recorded in real-time, triggering alarms when indicators are abnormal. The monitoring system automatically generates performance reports and sends them to the operations team for analysis periodically. Performance trend analysis helps identify signs of model degradation and allows for optimization adjustments.

[0086] The fault evolution path model is integrated with other modules through a standard interface. The model receives real-time feature data and outputs prediction results and confidence scores. The interface uses asynchronous communication via message queues to ensure system decoupling and scalability. Messages are serialized using Protocol Buffers to improve transmission efficiency and compatibility. The model update process enables hot deployment, updating fault feature transition matrix parameters without service interruption. Hot deployment is achieved through a shadow mode, where the new model is validated in the test environment and then gradually switched to the production environment. The switching process uses a blue-green deployment strategy to minimize service interruption risk. A version rollback mechanism ensures rapid recovery to a stable state in case of problems. Long-term optimization of the fault evolution path model is achieved through reinforcement learning. The reinforcement learning agent receives reward signals based on prediction accuracy. The reward function is designed based on prediction bias; the smaller the bias, the higher the reward value. The agent optimizes the fault feature transition matrix parameters through an exploration-utilization tradeoff, gradually improving prediction accuracy. The reinforcement learning algorithm uses Q-learning with a learning rate of 0.1 and a discount factor of 0.9.

[0087] Model interpretability enhancement is achieved through feature importance analysis, which analyzes the contribution of each feature dimension to state transition prediction. Feature importance is quantified using SHAP values, which measure the marginal contribution of each feature to the prediction result. The results of importance analysis help understand the key drivers of fault development and provide guidance for feature engineering. Application scenarios for the fault evolution path model include preventative maintenance, fault early warning, and maintenance decision support. In preventative maintenance, the model predicts fault development trends to help formulate maintenance plans. In fault early warning, the model identifies abnormal development patterns and issues timely warnings. In maintenance decision support, the model evaluates the effectiveness of different maintenance strategies and supports optimization decisions. The input data quality of the fault evolution path model is ensured through multi-level validation, including format validation, range validation, and logical validation. Format validation verifies that the data format conforms to specifications, range validation checks that feature values ​​are within a reasonable range, and logical validation verifies the correlation between features. Data that fails validation is marked as abnormal and enters the manual review process. Data quality reports are generated regularly to help improve data collection quality.

[0088] See Figure 4This chart displays the fault characteristic state transition probability matrix, a core component of the fault evolution path model. The chart uses grayscale levels to represent different probability values, with darker colors indicating higher transition probabilities. Rows in the matrix represent the current state, and columns represent the next state. The value in each cell precisely displays the conditional probability of transitioning from the current state to the next. The chart also marks a typical fault evolution path, connecting the complete evolution process from state 1 to state 5 with a red line, visually demonstrating the typical pattern of fault development. This state transition analysis provides a statistical basis for predicting fault development trends and supports the formulation of preventative maintenance decisions. The probability matrix is ​​built based on statistical analysis of historical fault data, effectively capturing the statistical patterns of fault development.

[0089] Example 4: Based on the modified fault evolution path model, the system outputs the final fault type and evolution stage. The final fault type is determined from high-confidence fault types, and the evolution stage indicates the current development status of the fault. The evolution stage is divided into three levels: initial stage, intermediate stage, and late stage, each corresponding to different characteristic value ranges and severity. The output format of the final fault type and evolution stage adopts a standardized data structure, including fault type code, evolution stage identifier, timestamp, and confidence score fields. The physical rationality of the final fault type is verified by combining temperature distribution data, which is obtained in real time from a distributed temperature sensor network. A mapping relationship between temperature distribution data and electrical loss characteristics is established, and the mapping relationship is derived based on the thermo-electric coupling model. Electrical loss characteristics are calculated through current and voltage data, including resistance loss, core loss, and stray loss. The thermo-electric coupling model considers material thermal conductivity, convective heat transfer coefficient, and radiative heat transfer coefficient parameters, converting electrical losses into a theoretical temperature rise distribution.

[0090] Referring to Table 1, the theoretical temperature rise calculation uses a finite element thermal analysis model. The electrical system is divided into mesh elements, and the heat balance equation for each element is established based on the law of conservation of energy. Tetrahedral elements are used for mesh generation, and the element size is adaptively adjusted according to the drastic change in the thermal gradient. The thermal analysis model solves the steady-state heat conduction equation, with boundary conditions including convective and radiative heat transfer, and the ambient temperature as a known input. The theoretical temperature rise distribution is output as a temperature field contour map, displaying the predicted temperature values ​​at various locations within the equipment.

[0091] Table 1: Mapping Table of Fault Types and Temperature Characteristics;

[0092]

[0093] The system determines whether the theoretical temperature rise corresponding to the final fault type is consistent with the actual temperature rise data, which is extracted from the temperature distribution data. A consistency check uses a temperature difference comparison algorithm to calculate the absolute difference between the theoretical and actual temperature rises. For each measurement point, the allowable temperature difference range is determined by querying a mapping table based on the fault type code; the actual temperature difference is compared with the allowable temperature difference. The overall consistency assessment uses a weighted average method, assigning higher weights to key measurement points, with weight coefficients determined based on sensitivity analysis. The consistency assessment result is expressed as a percentage, with a result higher than 90% considered valid. If the theoretical and actual temperature rise data are inconsistent, the dynamic weight allocation algorithm and fault evolution path model correction steps are re-executed. Inconsistencies trigger a diagnostic result reassessment process, which includes feature weight adjustment, model parameter optimization, and historical case re-matching. The reassessment process is iterative, with a maximum of three iterations or until the temperature difference is within the allowable range. Each iteration records an adjustment log, including changes in weight coefficients, model parameter updates, and matching score changes. During iteration, the search scope is gradually expanded, considering more possible fault type assumptions.

[0094] The real-time nature of the temperature verification process is ensured through a parallel computing architecture, with temperature data acquisition and electrical characteristic analysis performed in parallel. The temperature data processing pipeline includes data cleaning, outlier removal, and temperature field reconstruction. Data cleaning employs a median filtering algorithm to remove sensor noise, outlier removal is based on the three-standard-deviation rule, and temperature field reconstruction uses a Kriging interpolation algorithm to generate a continuous temperature distribution. The entire processing flow is completed within 100 milliseconds, meeting the real-time requirements of online diagnostics. The fault diagnosis results output integrates multi-source verification information, including electrical characteristic confidence levels, temperature verification results, and evolution stage assessments. The output results adopt a hierarchical structure: the first layer is the final fault type and confidence level, the second layer is the verification details, and the third layer is supporting data. This hierarchical structure facilitates understanding of the diagnostic results by different user groups; maintenance personnel focus on the top-level conclusions, while expert users can view the detailed analysis process. The output data format adopts the JSON standard, supporting both machine reading and human reading.

[0095] The spatiotemporal alignment of temperature distribution data with electrical characteristics is achieved through timestamp synchronization and coordinate mapping. Timestamp synchronization uses the IEEE 1588 precision clock protocol to ensure time consistency between electrical and temperature data. Coordinate mapping establishes the correspondence between the locations of electrical equipment and temperature sensors, mapping temperature to the corresponding electrical nodes through three-dimensional coordinate transformation. The spatiotemporal alignment error is controlled within 1 cm and 1 millisecond to ensure the accuracy of data fusion. The alignment process uses a calibration board for on-site calibration. The visualization of physical rationality verification adopts a multi-view linkage approach. The main view displays the electrical characteristic trend, and the auxiliary view displays the temperature distribution cloud map. A linkage relationship is established between views, automatically synchronizing and displaying the corresponding temperature distribution when a specific time point is selected. The visualization system supports zooming, panning, and rotation operations, facilitating observation of data relationships from different angles. Color coding uses a heat map, with red representing high-temperature areas and blue representing low-temperature areas. The visualization system is implemented based on WebGL technology and supports cross-platform access.

[0096] Uncertainty quantification of the validation results is achieved through Monte Carlo simulation, which considers measurement error, model error, and parameter uncertainty. 1000 random samples are performed, with each sample perturbing the input parameters within the error range, and the distribution characteristics of the output results are statistically analyzed. The uncertainty output includes confidence intervals, standard deviations, and sensitivity indices to help assess the reliability of the diagnostic results. Monte Carlo simulation employs variance reduction techniques to improve computational efficiency. The temperature validation model is calibrated using a standard heat source calibration method. The standard heat source provides a known power heat output, and the model's predicted temperature is compared with the actual measured temperature. The calibration process is performed at each sensor node, recording calibration coefficients and error curves. Regular calibration ensures temperature measurement accuracy, with a calibration cycle of three months or when environmental conditions change significantly. Calibration data is stored in a database for measurement value correction.

[0097] Continuous optimization of fault diagnosis results is achieved through a feedback learning mechanism, which collects the differences between actual maintenance results and diagnostic results. When maintenance results confirm a correct diagnosis, the weight of relevant features is strengthened; when a diagnosis is incorrect, the cause of the error is analyzed and model parameters are adjusted. The feedback learning mechanism establishes a diagnostic accuracy evaluation system, with evaluation indicators including false negative rate, false positive rate, accuracy, and timeliness. Evaluation reports are generated monthly to guide system optimization. The electrical feature analysis module outputs fault type hypotheses, the temperature verification module performs physical rationality checks, and the consistency evaluation module makes a comprehensive judgment. The modular design supports functional expansion and algorithm upgrades, and new verification methods can be easily integrated into the system. The module interface adopts the gRPC framework, supporting multi-language development.

[0098] Preprocessing of temperature distribution data includes ambient temperature compensation, sensor drift correction, and thermal conduction delay modeling. Ambient temperature compensation eliminates the impact of environmental changes on measurements, sensor drift correction overcomes sensor performance degradation, and thermal conduction delay modeling considers the time lag in heat transfer. Preprocessing algorithms improve the accuracy and usability of temperature data, providing a foundation for accurate verification. Preprocessing parameters are customized according to equipment characteristics. Correlation analysis is used to analyze the relationship between fault evolution stages and temperature characteristics, calculating the correlation coefficient between each evolution stage and temperature indicators. Correlation analysis is based on historical data statistics to identify the changing patterns of temperature characteristics as the fault develops. The analysis results are used to optimize the evolution stage judgment rules and improve the accuracy of stage identification. The correlation is updated monthly to keep it synchronized with the system status. The Pearson correlation coefficient is used for correlation analysis, with a threshold set at 0.7.

[0099] The system performance monitoring system features a dedicated dashboard that displays real-time diagnostic accuracy, response time, and resource utilization metrics. Monitoring data is automatically recorded and analyzed, with alerts issued when metrics are abnormal. Performance reports are generated regularly to help optimize system performance and resource allocation. Monitoring data is retained for one year, supporting long-term trend analysis and capacity planning. The monitoring system is built on Prometheus and Grafana. Fault diagnosis knowledge is accumulated through a case library update, where lessons learned from each diagnostic case are recorded and categorized. The case library uses an ontology-based approach to organize knowledge, establishing semantic relationships between fault types, characteristics, and causes. Knowledge retrieval supports intelligent queries and similar case recommendations to improve diagnostic efficiency. The case library is updated monthly to continuously enrich diagnostic knowledge. Built on a graph database, the case library supports complex relationship queries. Temperature verification model optimization is achieved through machine learning methods, using historical data to train the temperature prediction model. The machine learning algorithm employs a gradient boosting decision tree, with features including electrical parameters, environmental conditions, and equipment status. Model training is performed weekly to maintain prediction accuracy. Model performance is evaluated through cross-validation to ensure generalization ability. Training data is automatically labeled, reducing manual workload. The traceability of diagnostic results is ensured through a complete data log, which records the entire process from data collection to result output. The log includes raw data, intermediate results, decision-making logic, and modification history. Log data is retained for three years to support post-event analysis and accountability.

[0100] Example 5: A maintenance strategy matrix is ​​generated based on fault type and evolution stage. The matrix includes four dimensions: maintenance measures, urgency, resource requirements, and time window. Maintenance measures include three options: cleaning maintenance, component replacement, and overall repair. Urgency is categorized into three levels: emergency handling, planned handling, and observation / monitoring. Resource requirements cover personnel allocation, spare parts requirements, and equipment requirements. The time window specifies the recommended execution time range. The maintenance strategy matrix is ​​generated using a decision tree algorithm. The decision tree is trained based on historical maintenance cases, with each node representing a fault characteristic condition and leaf nodes corresponding to specific maintenance solutions. The CART algorithm is used, with a maximum depth of 10 and a minimum leaf node sample size of 20. For the mid-stage of insulation aging faults, the maintenance strategy matrix outputs the following solution: maintenance measures include replacing insulation components and reinforcing insulation wrapping; urgency is marked as planned handling; resource requirements include two high-voltage electricians, spare insulation materials and parts, and specialized tools; and the recommended time window is within seven days. The maintenance strategy matrix also provides alternative solutions, including temporary enhanced monitoring and planned power outage maintenance, each accompanied by risk assessments and cost estimates. Risk assessment employs failure mode and impact analysis, while cost estimation is based on a regression model using historical maintenance data.

[0101] The maintenance strategy matrix and equipment operation plan are co-optimized. The equipment operation plan is obtained from the production management system and includes equipment commissioning time, load plan, and power outage window information. Co-optimization employs a multi-objective programming method, with objective functions including maximizing reliability, minimizing cost, and minimizing impact. The optimization process considers equipment load rate, environmental conditions, and actual constraints of spare parts inventory, transforming the multi-objective problem into a single-objective solution through linear weighting. The optimization algorithm uses a genetic algorithm with a population size of 100, 500 iterations, a crossover probability of 0.8, and a mutation probability of 0.1. Specific maintenance work orders and operation instructions are generated. Maintenance work orders include fault location information, maintenance steps, safety precautions, and acceptance criteria. Operation instructions detail the technical requirements of each operation step, including tool usage methods, process standards, and quality checkpoints. Maintenance work orders are distributed to relevant personnel through a workflow engine. Mobile terminals receive work orders and support on-site check-in, progress updates, and result feedback. Operation instructions include standard operating procedure diagrams, demonstrating the technical points of key operation steps. Work order numbers use a date-month-year serial number rule to ensure uniqueness.

[0102] The execution of the maintenance strategy is monitored in real time through an IoT platform, which connects to personnel positioning devices, tool sensors, and video surveillance systems. Real-time monitoring data includes maintenance personnel location, tool usage status, and on-site environmental parameters, updated every second. Abnormal situations are automatically detected and alarms are triggered, such as personnel entering hazardous areas, improper tool use, or sudden changes in environmental conditions. A large screen in the monitoring center displays a panoramic view of the maintenance progress, with different colors indicating task status: green for normal progress, yellow for requiring attention, and red for requiring intervention. Alarm information is sent simultaneously via sound and light, SMS, and push notifications. Maintenance effectiveness is evaluated using a before-and-after comparative analysis method, comparing changes in fault characteristics before and after maintenance. Baseline data of fault characteristics, including partial discharge amplitude, dielectric loss factor, and temperature distribution, are recorded before maintenance. After maintenance, data is collected at the same measurement point to calculate the improvement in characteristic parameters. Effectiveness evaluation indicators include characteristic parameter recovery rate, stability improvement, and reliability growth; each indicator is weighted to calculate a comprehensive evaluation score. Evaluation data is stored in a maintenance case library to provide a reference for subsequent strategy optimization. The evaluation report uses a standard template and includes data tables, trend charts, and analytical conclusions.

[0103] The maintenance knowledge base is updated based on this maintenance practice. It uses a graph structure, where nodes represent maintenance objects, fault types, and maintenance measures, and edges represent the relationships between them. A complete record of this maintenance case has been added, including fault characteristics, diagnostic process, maintenance measures, and effectiveness evaluation data. The knowledge base is updated using an incremental learning approach, integrating new cases with existing knowledge and adjusting association weights and confidence levels. The knowledge retrieval interface supports semantic queries, allowing for the retrieval of similar cases based on natural language descriptions. The knowledge base is synchronized weekly to ensure data consistency. Maintenance resource scheduling considers the needs of multi-task collaboration. The resource scheduling algorithm is based on constrained programming theory, aiming to optimize resource utilization while meeting time windows. Constraints include personnel skill matching, tool availability, spare parts inventory, and safety procedures. The optimization objective is to minimize total cost and the longest completion time. The scheduling algorithm outputs a detailed resource allocation plan, including personnel scheduling plans, tool allocation plans, and spare parts delivery arrangements. The scheduling results are visualized as a Gantt chart, intuitively showing the time arrangement and resource usage of each task. Resource conflict detection uses a look-ahead algorithm to provide early warnings of resource shortages.

[0104] Safety management during maintenance utilizes a digital work order system. The digital work order includes a list of safety measures, hazard analysis, and emergency response plan. Maintenance personnel confirm the implementation of safety measures via a mobile application, and the system authorizes operation after verifying the completeness of the measures. Key steps have secondary confirmation steps, requiring review and signature from a supervising person. The safety management system monitors the work environment in real time, immediately alerting staff upon detecting abnormalities and remotely suspending work if necessary. Work order status is divided into five stages: preparation, review, approval, execution, and termination. Maintenance quality control employs a three-tiered acceptance system: self-inspection by the work supervisor, acceptance by professional management personnel, and confirmation by the equipment owner. Each level of acceptance has clear standards and record forms, and acceptance data is uploaded to the system in real time. Quality deviations trigger rectification processes, and serious quality issues initiate root cause analysis. Quality data statistics generate control charts for long-term tracking of maintenance quality trends. Quality records are maintained for ten years, supporting lifecycle management.

[0105] Continuous optimization of maintenance strategies is achieved through reinforcement learning. The reinforcement learning environment simulates equipment operating conditions, and the agent tries different maintenance strategies and receives reward signals. The reward function is designed based on multiple dimensions such as maintenance effectiveness, cost, and duration. The agent learns the optimal strategy through exploration-exploitation trade-offs. The trained strategy model is applied online, and parameters are fine-tuned based on actual feedback. The reinforcement learning cycle is one month to maintain the adaptability of the strategy. Strategy model version management supports A / B testing. Maintenance cost analysis adopts a full lifecycle cost model, which includes direct costs, indirect costs, and risk costs. Direct costs refer to the expenses directly incurred by maintenance activities, indirect costs include power outage losses and efficiency impacts, and risk costs measure the potential losses from untimely maintenance. Cost analysis results guide maintenance budget allocation, prioritizing maintenance projects with high cost-effectiveness. Cost data is visualized using Sankey diagrams, clearly showing cost flow and proportions. Cost prediction uses time series analysis. Maintenance personnel skills management is achieved through a competency matrix, recording each person's skills, qualifications, work experience, and training records. Maintenance tasks are automatically matched with personnel competency requirements, recommending the most suitable personnel combination. Skill gap analysis identifies team competency gaps and generates targeted training plans. Personnel performance is evaluated based on a comprehensive assessment of maintenance quality, efficiency, and security indicators. Communication is via an API gateway. Microservices include fault diagnosis, policy generation, resource scheduling, and monitoring services. The system's resilient design ensures that a single point of failure does not affect overall operation, and critical services have redundant backups. A service mesh enables load balancing and failover.

[0106] The long-term value of maintenance data is unlocked through trend analysis, which identifies patterns in equipment performance degradation, changes in maintenance effectiveness, and cost fluctuations. Analysis results guide adjustments to preventative maintenance cycles, optimization of spare parts inventory, and improvements in staffing. Data mining algorithms employ time series analysis, cluster analysis, and association rule learning to uncover potential causal relationships and optimization opportunities. Analysis reports are generated regularly to support management decisions. Data dashboards display key indicators in real time. The maintenance strategy generation system is deeply integrated with other enterprise information systems, including asset management, production management, and supply chain systems. Integration is achieved through an enterprise service bus, and data exchange uses international standard interfaces. Deep integration eliminates information silos, enabling data to be entered once and shared globally. Data consistency between systems is ensured through middleware, and conflict resolution employs version coordination mechanisms. Interface monitoring ensures reliable data transmission.

[0107] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0108] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for diagnosing a fault in a multi-parameter electrical system, characterized by, The method comprises the following steps: Collecting multi-dimensional operation parameters of an electrical system, including voltage waveform data, current waveform data and temperature distribution data; Performing time-frequency joint analysis on the multi-dimensional operation parameters to extract multi-scale electrical features, including steady-state feature components and transient feature components; Based on a historical fault case library, performing pattern matching on the multi-scale electrical features to generate an initial fault type set; Using a dynamic weight distribution algorithm to evaluate the confidence of fault types in the initial fault type set and screening out high-confidence fault types; Based on the high-confidence fault types, constructing a fault evolution path model to describe the time sequence correlation of fault features; Combining real-time monitoring data, dynamically correcting the fault evolution path model to generate an optimized fault diagnosis result, The time-frequency joint analysis on the multi-dimensional operation parameters to extract multi-scale electrical features comprises the following steps: Using an adaptive decomposition algorithm to separate the voltage waveform data and the current waveform data to obtain low-frequency fundamental components and high-frequency harmonic components; Performing energy density analysis on the high-frequency harmonic components to determine the abnormal frequency range; Based on the abnormal frequency range, extracting the transient feature components, including pulse amplitude and duration; Performing phase shift detection on the low-frequency fundamental components to extract the steady-state feature components, including amplitude fluctuation rate and phase consistency index; The pattern matching on the multi-scale electrical features based on the historical fault case library to generate an initial fault type set comprises the following steps: Inputting the steady-state feature components and the transient feature components into a pre-trained fault classification model to output candidate fault types; According to the similarity of the candidate fault types and the fault features in the historical fault case library, calculating the matching scores; Screening out candidate fault types with matching scores higher than a preset threshold to constitute the initial fault type set; The confidence evaluation of fault types in the initial fault type set using a dynamic weight distribution algorithm comprises the following steps: Dynamically adjusting the weight coefficients according to the contribution degrees of the steady-state feature components and the transient feature components; Based on the weight coefficients, calculating the comprehensive confidence of each fault type in the initial fault type set; Eliminating fault types with comprehensive confidence lower than a critical value and retaining the high-confidence fault types; The construction of the fault evolution path model comprises the following steps: Extracting the historical fault feature sequence corresponding to the high-confidence fault types; Analyzing the time sequence variation law of the historical fault feature sequence to establish a fault feature transition matrix; According to the fault feature transition matrix, predicting the next stage evolution trend of the current fault feature; The dynamic correction of the fault evolution path model comprises the following steps: Real-time collecting updated operation parameters of the electrical system to extract new electrical features; Comparing the new electrical features with the prediction results of the fault evolution path model to calculate the deviation; If the deviation exceeds the allowed range, readjusting the parameters of the fault feature transition matrix. The generated optimized fault diagnosis result comprises: According to the corrected fault evolution path model, output the final fault type and evolution stage; Combined with the temperature distribution data, verify the physical rationality of the final fault type; The verification of the physical rationality of the final fault type comprises: Establish the mapping relationship between the temperature distribution data and the electrical loss characteristics; Determine whether the theoretical temperature rise corresponding to the final fault type is consistent with the actual temperature rise data; If not, re-execute the dynamic weight distribution algorithm and the fault evolution path model correction step.

2. The method for diagnosing a fault of a multi-parameter electrical system according to claim 1, wherein The collection of the multi-dimensional running parameters of the electrical system comprises: Synchronously collect three-phase voltage and current signals to ensure time stamp alignment; Obtain temperature sampling values of key nodes through a distributed temperature sensor network; Time-synchronize and calibrate the voltage waveform data, current waveform data and temperature distribution data.

3. The method for diagnosing a fault of a multi-parameter electrical system according to claim 2, wherein Before the time-frequency joint analysis of the multi-dimensional running parameters, it further comprises: Segment the original data using a sliding window mechanism; Apply a window function to each segment of data to suppress spectral leakage; Compensate for signal distortion caused by uneven sampling interval through interpolation algorithm.

Images