Method and system for maintenance diagnosis of intelligent electrical equipment

Abstract

This invention relates to the field of electrical equipment anomaly diagnosis technology, and particularly to an intelligent electrical equipment maintenance and diagnosis method and system. It includes: acquiring a multi-source sequence comprising temperature, partial discharge quantity, effective value of vibration acceleration, and effective value of load current; preprocessing, normalizing, and standardizing the multi-source sequence to obtain a standardized multi-source sequence; constructing a time-varying correlation matrix to characterize the dynamic coupling relationship between parameters, extracting highly correlated synchronous change segments to form dynamic feature vectors, and establishing a health prediction model based on a long short-term memory network, identifying abnormal trends through residual analysis; further combining local outlier density and amplitude spectrum dominant frequency components for clustering and differentiation, accurately determining equipment degradation segments, and updating the historical degradation pattern library to generate a diagnostic report. This invention solves the problems of difficulty in characterizing the dynamic correlation of multi-source parameters and easy misjudgment of environmental disturbances, and achieves accurate identification of electrical equipment degradation trends.

Description

Technical Field

[0001] This invention relates to the field of electrical equipment anomaly diagnosis technology, and in particular to an intelligent electrical equipment maintenance and diagnosis method and system. Background Technology

[0002] With the continuous improvement of power systems and industrial automation, intelligent electrical equipment such as transformers, switchgear, cable systems, and motors are increasingly widely used in power grids and industrial production. These devices typically integrate multiple online monitoring systems, including temperature sensors, partial discharge monitoring devices, vibration monitoring units, and current acquisition modules, to reflect the equipment's operating status in real time through multi-source operational data. In actual operation, complex coupling relationships often exist between various operating parameters of electrical equipment. For example, changes in load current may cause temperature increases, which in turn may accelerate insulation aging and enhance partial discharge activity; loosening of mechanical structures may lead to changes in vibration amplitude, creating a linkage effect with current fluctuations. Simultaneously, time lag response and feedback amplification phenomena may exist between different parameters, and their correlation strength changes with the degree of equipment aging and operating conditions. Therefore, equipment degradation usually manifests as a gradual evolution of dynamic correlations between multiple parameters, rather than a sudden change in a single parameter.

[0003] However, existing intelligent electrical equipment maintenance and diagnostic methods mostly employ fixed threshold judgments or simple correlation analysis, independently monitoring each parameter and triggering an alarm when a parameter exceeds a set threshold. These methods fail to fully consider the time-varying correlation characteristics and dynamic dependencies between multi-source data, making it difficult to accurately characterize the dynamic coupling structure between parameters over time, easily leading to misjudgments or missed diagnoses. Furthermore, equipment operation is also affected by external factors such as changes in ambient temperature and load cycle fluctuations, generating periodic or phased disturbances. Existing methods struggle to effectively distinguish between periodic environmental fluctuations and the continuous cumulative degradation trend of equipment, easily resulting in frequent false alarms or delayed identification of degradation signals.

[0004] Therefore, there is an urgent need for an intelligent electrical equipment maintenance and diagnosis method and system that can fuse multi-source time-series data, dynamically analyze the correlation between parameters, identify trend anomalies that deviate from normal operating modes, and distinguish between periodic environmental disturbances and continuous degradation signals, thereby improving the accuracy and reliability of equipment maintenance and diagnosis. Summary of the Invention

[0005] This invention provides a method and system for the maintenance and diagnosis of intelligent electrical equipment, which enables refined monitoring of the entire process of intelligent electrical equipment operation status and identification of trend deterioration through multi-source data fusion and dynamic modeling mechanisms.

[0006] In a first aspect, the present invention provides a method for maintaining and diagnosing intelligent electrical equipment, the method comprising: Step S1: Obtain a multi-source sequence including temperature, partial discharge quantity, effective value of vibration acceleration, and effective value of load current; perform preprocessing, normalization, and standardization on the multi-source sequence to obtain a standardized multi-source sequence; Step S2: Calculate the correlation coefficients between parameters in the standardized multi-source sequence, filter the correlation coefficients based on time delay mutual information, and determine the time-varying correlation matrix based on the filtered high-correlation parameters; Step S3: Determine the parameter pairs that exceed the dynamic correlation threshold based on the time-varying correlation matrix; extract the synchronous change segments corresponding to the time periods of the parameter pairs; and concatenate them into a dynamic feature vector sequence; input the dynamic feature vector sequence into a long short-term memory network to obtain a prediction sequence; obtain the residual sequence based on the prediction sequence and the actual standardized multi-source sequence; Step S4: Determine the set of abnormal trend segments by comparing the residual sequence with the adaptive deviation threshold; perform cluster analysis on the set of abnormal trend segments based on the corresponding local outlier density and amplitude spectrum main frequency component to distinguish equipment degradation segments; Step S5: Update the historical degradation pattern library based on the equipment degradation fragments and generate a diagnostic report.

[0007] Secondly, the present invention also provides an intelligent electrical equipment maintenance and diagnostic system for implementing the above-described method, the system comprising: A multi-source sequence acquisition unit is used to acquire multi-source sequences including temperature, partial discharge quantity, effective value of vibration acceleration, and effective value of load current; and to preprocess, normalize, and standardize the multi-source sequences to obtain standardized multi-source sequences. The correlation analysis unit is used to calculate the correlation coefficients between parameters in the standardized multi-source sequence, filter the correlation coefficients by combining time delay mutual information, and determine the time-varying correlation matrix based on the filtered high correlation parameters. The residual calculation unit is used to determine the parameter pairs that exceed the dynamic correlation threshold based on the time-varying correlation matrix; extract the synchronous change segments corresponding to the time period of the parameter pairs; and concatenate them into a dynamic feature vector sequence; input the dynamic feature vector sequence into a long short-term memory network to obtain a prediction sequence; and obtain a residual sequence based on the prediction sequence and the actual standardized multi-source sequence. The degradation analysis unit is used to determine the set of abnormal trend segments by comparing the residual sequence with the adaptive deviation threshold; and to perform cluster analysis on the set of abnormal trend segments based on the corresponding local outlier density and amplitude spectrum main frequency component to distinguish the equipment degradation segments. The diagnostic result generation unit is used to update the historical degradation pattern library based on the deteriorated equipment segments and generate a diagnostic report.

[0008] Thirdly, the present invention also provides a computer-readable storage medium storing instructions that, when executed by a processor, implement the above-described method.

[0009] The beneficial effects of this invention are as follows: This invention constructs a stable and consistent standardized multi-source sequence by performing median filtering, anomaly removal, and double standardization on the effective values ​​of temperature, partial discharge, vibration acceleration, and load current. This provides a high-quality data foundation for subsequent analysis and reduces the interference of on-site noise and dimensional differences on the model. Through a rolling window correlation analysis and time-delay mutual information joint screening mechanism, a time-varying correlation matrix that can be dynamically updated over time is constructed. This accurately characterizes the linear and dynamic coupling relationships between multiple parameters and their evolution characteristics with changing operating conditions, enabling a structured expression of the complex internal correlation structure of the equipment. Based on the time-varying correlation matrix, highly correlated synchronous change segments are extracted to form a fused dynamic feature vector sequence. A health status prediction model is established using a long short-term memory network, and the residual quantization between predicted and actual values ​​is used for analysis. By assessing the degree of deviation, the system proactively captures potential abnormal trends. At the residual level, an adaptive threshold and continuity judgment mechanism are introduced to identify persistent abnormal trends. Combined with dynamic time warping distance and frequency domain dominant frequency characteristics, clustering is performed to effectively distinguish between periodic environmental disturbances and gradual internal equipment degradation, avoiding false alarms. Furthermore, through matching and self-learning updates of a historical degradation pattern library, diagnostic experience is continuously accumulated and optimized, generating early warning reports containing degradation type, degradation degree, and maintenance recommendations. This forms a closed-loop diagnostic process from data acquisition, correlation modeling, trend identification, degradation judgment to decision support. Through the synergy of these technical solutions, the accuracy of anomaly identification, trend judgment capability, and operational and maintenance decision-making of intelligent electrical equipment under complex operating conditions are significantly improved. Attached Figure Description

[0010] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a flowchart of a maintenance and diagnostic method for intelligent electrical equipment in one embodiment; Figure 2 This is a flowchart of the time-varying matrix construction method in the embodiment.

[0012] Figure 3 This is a structural diagram of an intelligent electrical equipment maintenance and diagnostic system as described in the embodiment. Detailed Implementation

[0013] This invention provides a method and system for the maintenance and diagnosis of intelligent electrical equipment. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0014] For ease of understanding, the specific process of the embodiments of the present invention will be described below, such as... Figure 1 As shown, an embodiment of the present invention provides a method for maintaining and diagnosing intelligent electrical equipment, comprising: Step S1: Obtain a multi-source sequence including temperature, partial discharge quantity, effective value of vibration acceleration, and effective value of load current; perform preprocessing, normalization, and standardization on the multi-source sequence to obtain a standardized multi-source sequence; specifically including: The multi-source sequence is smoothed by a sliding window midpoint filter; and the range of each parameter dimension is unified by a minimum-maximum normalization process; the normalized sequence is then determined as a standardized multi-source sequence.

[0015] Specifically, to achieve high-quality input of multi-source data in the maintenance and diagnosis of intelligent electrical equipment, multi-source sequences including temperature, partial discharge, effective value of vibration acceleration, and effective value of load current are acquired. These multi-source sequences are collected in real time by sensors deployed on intelligent electrical equipment such as transformers, switch cabinets, or motors, and stored in alignment according to a unified timestamp, thereby forming a multi-dimensional time-series data matrix indexed by time. Due to factors such as electromagnetic interference, mechanical shock, and sampling errors in the field operating environment, the original multi-source sequences often contain pulse-type spike noise and isolated anomalies. Therefore, the above multi-source sequences need to be preprocessed, normalized, and standardized to ensure the stability and comparability of the data.

[0016] During implementation, the above multi-source sequences are subjected to sliding window mid-range filtering. For the temperature sequence, partial discharge quantity, and load current sequence, a sliding window of size 5 can be selected. For the vibration acceleration effective value sequence, the window size can be appropriately expanded to 7 according to its periodic fluctuation characteristics. For each time point in the sequence, all data values ​​within its sliding window are extracted and sorted. The median value after sorting is selected as the smoothed value of that time point to replace the original value. Through the above processing method, the instantaneous spike noise caused by electromagnetic interference can be effectively suppressed, while maintaining the overall trend of the sequence without weakening, thereby ensuring the continuity of equipment operation trend information.

[0017] After median filtering, to eliminate dimensional differences and scale inconsistencies between different physical quantities, the multi-source sequences are subjected to min-max normalization. Specifically, the minimum and maximum values ​​are calculated for each parameter sequence, and each data point is processed according to... A mapping process is performed to uniformly convert the effective values ​​of temperature, partial discharge, vibration acceleration, and load current to a value range of 0 to 1, thereby avoiding the bias effect of dimensional differences on subsequent correlation coefficient calculation and dynamic modeling. The normalized sequence is still arranged in chronological order to maintain the time synchronization relationship between multiple parameters.

[0018] The normalized data is identified as a standardized multi-source sequence. This standardized multi-source sequence maintains a one-to-one correspondence with the original collected data in the time dimension and achieves scale uniformity and distribution normalization in the parameter dimension. This provides a stable and reliable data foundation for subsequent construction of time-varying correlation matrices, extraction of dynamic feature vector sequences, and maintenance and diagnosis of intelligent electrical equipment.

[0019] Step S2: Calculate the correlation coefficients among the parameters in the standardized multi-source sequence, filter the correlation coefficients based on time delay mutual information, and determine the time-varying correlation matrix according to the filtered high-correlation parameters; such as Figure 2 As shown, it specifically includes:

[0020] A rolling time window is used to calculate the Pearson correlation coefficients between parameters on the standardized multi-source sequence, and the correlation coefficient sequences of each parameter pair in different time periods are obtained. The mutual information value of each parameter pair under different time delay conditions is calculated within a preset time delay range. The Pearson correlation coefficients are filtered based on the time delay mutual information value to obtain the Pearson correlation coefficients corresponding to parameter pairs whose mutual information values ​​exceed a preset filtering threshold under the corresponding time delay conditions. Highly correlated parameter pairs are extracted from the filtering results, and directional labels are added according to the positive or negative signs of the corresponding correlation coefficients. A time-varying correlation matrix is ​​constructed based on the highly correlated parameter pairs with positive and negative directional labels.

[0021] Specifically, after preprocessing and standardization of the multi-source sequences, a standardized multi-source sequence is obtained, indexed by time and containing temperature, partial discharge, effective value of vibration acceleration, and effective value of load current. The standardized multi-source sequences are synchronized in the time dimension. To characterize the dynamic coupling relationship between various parameters during the operation of intelligent electrical equipment, the correlation coefficients between the parameters are further calculated, and a time-varying correlation matrix is ​​constructed. Specifically, a rolling time window is used to calculate the Pearson correlation coefficients between the parameters on the standardized multi-source sequences. The window length is set to a preset number of time points, such as 30 consecutive sampling points, and the window is slid along the time axis point by point. Within each window, temperature sequence segments, partial discharge sequence segments, effective value of vibration acceleration sequence segments, and effective value of load current sequence segments are extracted respectively. For any two parameters forming a parameter pair, the Pearson correlation coefficient is calculated to obtain a correlation coefficient sequence that changes with time. This ensures that each parameter pair does not only correspond to a single correlation coefficient, but forms a dynamic correlation coefficient curve that corresponds one-to-one with time, which is used to reflect the change in linear correlation strength under short-term operating conditions.

[0022] However, in the operation scenarios of intelligent electrical equipment, there are often dynamic dependencies and time delay responses between various parameters. Therefore, it is difficult to accurately characterize the real coupling relationship based solely on the Pearson correlation coefficient. To address this, within a preset time delay range, such as 0 to 10 time points, the mutual information value of each parameter pair under different time delay conditions is calculated. Specifically, when calculating the time delay mutual information of a parameter pair X and Y, X(t) and Y(t+τ) are matched, where τ is the time delay variable. The corresponding mutual information value is calculated using a statistical probability distribution estimation method, thereby obtaining the mutual information function curve of the parameter pair under various time delay conditions. If the mutual information value under a certain time delay condition reaches a peak and exceeds a preset screening threshold, such as exceeding 1.2 times the average mutual information value, then the parameter pair is considered to have a significant dependency relationship under that time delay condition. Based on this, the Pearson correlation coefficients are screened according to the aforementioned time delay mutual information values. Only the Pearson correlation coefficients corresponding to parameter pairs whose mutual information values ​​exceed the preset screening threshold under the corresponding time delay condition are retained, while parameter pairs whose mutual information is insufficient to support the dependency relationship are eliminated. Through the above screening process, the joint constraint of linear correlation and dynamic dependency is achieved, thereby reducing the risk of misjudgment caused by accidental correlation or noise correlation.

[0023] After obtaining stable correlation coefficients filtered by mutual information, parameter pairs with absolute values ​​exceeding the high correlation threshold are further extracted as high-correlation parameter pairs. Directional labels are added based on the positive or negative sign of the corresponding correlation coefficients. Positive correlation indicates that the parameters change in the same direction, such as an increase in temperature accompanied by an increase in partial discharge; negative correlation indicates an opposite change, such as an increase in load current accompanied by a decrease in vibration amplitude. Finally, a time-varying correlation matrix is ​​constructed based on the correlation coefficient values ​​corresponding to the parameter pairs with positive and negative directional labels. This matrix uses the parameters corresponding to the high-correlation parameter pairs as row and column indices, filling the corresponding correlation coefficients into the corresponding positions. Parameter pairs that fail the screening are filled with zero, i.e., a correlation coefficient of 0. The matrix element values ​​are updated in real time as the rolling time window progresses. Through this technical solution, the constructed time-varying correlation matrix can dynamically reflect the multi-parameter coupling structure of intelligent electrical equipment at different operating stages, providing structured correlation information support for subsequent dynamic feature extraction and degradation trend identification.

[0024] Step S3: Determine the parameter pairs that exceed the dynamic correlation threshold based on the time-varying correlation matrix; extract the synchronous change segments corresponding to the time periods of the parameter pairs; and concatenate them into a dynamic feature vector sequence; input the dynamic feature vector sequence into a long short-term memory network to obtain a prediction sequence; obtain the residual sequence based on the prediction sequence and the actual standardized multi-source sequence; In step S3, extracting the dynamic feature vector sequence based on the time-varying correlation matrix includes: Based on the parameters in the time-varying correlation matrix, the start and end time points that exceed the corresponding dynamic correlation threshold are used to form corresponding time period boundaries. Based on the time period boundaries, sequence segments corresponding to the time period boundaries are extracted from the standardized multi-source sequences to obtain synchronous change segments. The length of the synchronous change segments is checked. When the segment length is less than the preset minimum length, it is extended to the adjacent time point. The extracted synchronous change segments are spliced ​​together point by point in chronological order to form a fused dynamic feature vector sequence. The feature vector at each time point includes temperature, partial discharge, effective value of vibration acceleration, and effective value of load current.

[0025] Specifically, in this embodiment, after the time-varying correlation matrix is ​​constructed, in order to transform the dynamic coupling relationship between multiple parameters into a structured input that can be learned by the prediction model, a dynamic feature vector sequence is further extracted based on the time-varying correlation matrix. The time-varying correlation matrix is ​​a parameter correlation structure that is updated with the rolling time window. Its matrix elements are Pearson correlation coefficient values ​​that have been filtered by time delay mutual information and meet the stability verification conditions. Therefore, it can reflect the effective coupling strength and direction between multiple parameters of the device in different operating stages.

[0026] During implementation, based on the start and end times of parameter pairs in the aforementioned time-varying correlation matrix exceeding the aforementioned dynamic correlation threshold, the corresponding time period boundaries are determined. The aforementioned dynamic correlation threshold is preferably adaptively calculated based on the statistical characteristics of historical correlation coefficient distributions, for example, by calculating the upper bound of the Pearson correlation coefficient confidence interval or by doubling the standard deviation of the historical sample mean, thereby ensuring that only statistically significant high-correlation parameter pairs are retained. Subsequently, the corresponding elements of the time-varying correlation matrix are scanned on the rolling time axis. When the correlation coefficient of a parameter pair remains above the corresponding dynamic correlation threshold for multiple consecutive time windows, the time point when it first exceeds the threshold is recorded as the start time point, and when it falls below the threshold is recorded as the end time point, thus forming the effective coupling time period boundary for that parameter pair. After obtaining the aforementioned time period boundaries, the normalized multi-source sequence... The parameter sequence segments corresponding to the time period are extracted from the column to obtain synchronous change segments. Since the temperature, partial discharge, vibration acceleration RMS value, and load current RMS value in the standardized multi-source sequence are all aligned according to a unified timestamp, the extracted parameter subsequences correspond in the time dimension, thus ensuring the synchronization of data within the segment. For example, when the correlation coefficient between temperature and partial discharge exceeds the corresponding dynamic correlation threshold during the first time period, or when the correlation coefficient between vibration acceleration and load current exceeds the corresponding dynamic correlation threshold during the second time period, the corresponding temperature, partial discharge, vibration acceleration RMS value, and load current RMS value subsequences within the first or second time period are synchronously extracted. For parameter sequences with different sampling frequencies, the timestamps can be aligned using linear interpolation before extraction to avoid synchronization errors caused by time misalignment.

[0027] Furthermore, the length of the aforementioned synchronous change segments is validated. If the segment length is less than the preset minimum length, it is extended forward or backward to adjacent time points to ensure that the segment has sufficient data for subsequent feature modeling, thereby improving the representativeness of the dynamic feature vector and reducing redundant information. After segment extraction, the synchronous change segments are dimensionally aligned. If different parameters result in differences in the length of the corresponding time period, the segment length is made consistent by truncation. Then, the aligned segments are spliced ​​together point by point in chronological order to form a fused dynamic feature vector sequence. Specifically, at each time point, a sequence including temperature, partial discharge, vibration intensity, etc., is constructed. The four-dimensional feature vectors of the effective values ​​of velocity and load current are generated. For example, temperature, discharge segments, vibration, and current segments are synthesized into a vector sequence in chronological order, so that each time step simultaneously reflects the coupling state of multiple physical fields such as heat, electricity, and mechanics. Through the above technical solution, the resulting fused dynamic feature vector sequence not only retains the high correlation between parameters within a specific time period, but also continuously expresses the evolution process of the equipment's operating state in the time dimension. This provides structured and dynamic input data for subsequent prediction modeling of long short-term memory networks, which helps to improve the sensitivity and accuracy of abnormal trend identification in the maintenance and diagnosis of intelligent electrical equipment.

[0028] Further, in step S3, obtaining the residual sequence includes: The dynamic feature vector sequence is input into a long short-term memory network for single-step rolling prediction; a corresponding prediction sequence is generated, which represents the expected normal fluctuation when the device is in a healthy state; at each time point, the difference between the predicted value in the prediction sequence and the corresponding value in the standardized multi-source sequence is calculated to form a residual value point by point; the residual values ​​at each time point are arranged in chronological order to obtain the residual sequence.

[0029] Specifically, after obtaining the fused dynamic feature vector sequence, in order to obtain the expected change trajectory of the device in the healthy operation mode and quantify the deviation of the actual operation state from the normal mode, a residual sequence is further obtained. The residual sequence is used to reflect the difference between the standardized multi-source sequence and the prediction results based on the historical normal operation mode, and is the basis for abnormal trend identification.

[0030] In the implementation process, a Long Short-Term Memory (LSTM) network model is first constructed, and the aforementioned dynamic feature vector sequence is input into the LTM network model for single-step rolling prediction. The LTM network includes an input layer, at least one LTM layer, and a fully connected output layer. The input layer receives the dynamic feature vector sequence arranged in chronological order, and its input dimension is equal to the dimension of the dynamic feature vector. The LTM layer selectively memorizes and updates the state information in historical time steps through the synergistic effect of the input gate, forget gate, and output gate, thereby learning the evolution law of multi-parameter coupling relationship over time. The fully connected output layer maps the high-dimensional temporal features extracted by the LTM layer into a prediction output consistent with the dimension of the standardized multi-source sequence. The prediction output corresponds to four parameters: temperature, partial discharge, effective value of vibration acceleration, and effective value of load current.

[0031] During the model training phase, data from a historical period in which the equipment is in a stable and healthy state are selected as training samples. These training samples include standardized multi-source sequences and their corresponding dynamic feature vector sequences. The dynamic feature vector at the current time point is used as input, and the true value of the standardized multi-source sequence at the next time point is used as the training target. The error between the predicted value and the true value is measured by the mean squared error loss function. The backpropagation algorithm combined with the adaptive moment estimation optimizer is used to iteratively optimize the network parameters, so that the model gradually learns the expected normal fluctuation pattern of the equipment under normal operating conditions and environmental disturbances. The expected normal fluctuation refers to the normal range of change of the above four parameters when the equipment does not deteriorate or become abnormal.

[0032] After model training is complete, the actual monitoring phase begins. For the dynamic feature vector sequence generated within the monitoring period, a single-step rolling prediction method is used to generate prediction results. Specifically, at each time point t, the latest dynamic feature vector within the current time window is input into the trained Long Short-Term Memory (LSTM) network model to obtain the predicted value for time point t+1. Then, the time window is slid forward one time step, and the above prediction process is repeated to generate corresponding predicted values ​​for each time point within the monitoring period. All predicted values ​​are arranged in chronological order to form a prediction sequence corresponding to the input sequence. This prediction sequence represents the expected normal fluctuation sequence of each monitoring parameter under the assumption that the equipment is in a healthy state. After obtaining the prediction sequence, at each time point, the predicted values ​​in the prediction sequence are compared with the standard values ​​at the same time point. The difference between the actual values ​​of the multi-source sequence is calculated, and the residual values ​​are calculated for each parameter, namely temperature, partial discharge, effective value of vibration acceleration, and effective value of load current. The residual value is equal to the actual standardized parameter value minus the predicted value, which is used to quantify the deviation of the actual observation result from the prediction result of the normal mode. Subsequently, the residuals of each parameter calculated at all time points are arranged in chronological order to form a residual sequence. When the equipment is in normal operation, the residual sequence usually shows a random distribution with small fluctuations around zero. When the equipment shows potential deterioration or abnormal trends, the actual observed values ​​will continuously deviate from the predicted normal fluctuation range, which will be reflected in the residual sequence as an increase in amplitude or continuous shift. Through the above technical solution, the residual sequence provides a quantitative basis for subsequent identification of abnormal trend segments and determination of equipment deterioration.

[0033] Step S4: Determine the set of abnormal trend segments by comparing the residual sequence with the adaptive deviation threshold; perform cluster analysis on the set of abnormal trend segments based on the corresponding local outlier density and amplitude spectrum main frequency component to distinguish equipment degradation segments; In step S4, determining the abnormal trend segment based on the residual sequence includes: An adaptive deviation threshold is determined based on the historical residual distribution; it is then determined whether multiple consecutive points in the residual sequence exceed the adaptive deviation threshold; if multiple consecutive points exceed the adaptive deviation threshold, the corresponding consecutive time periods are marked as abnormal trend segments, and an abnormal trend set with timestamps is obtained based on the abnormal trend segments.

[0034] Specifically, after obtaining the predicted sequence output by the Long Short-Term Memory network and subtracting it point-by-point from the actual standardized multi-source sequence to obtain the residual sequence, in order to identify the continuous abnormal evolution process from the above residual sequence, abnormal trend segments are further determined based on the residual sequence. The above residual sequence corresponds one-to-one with the original sampling timestamp in the time dimension, and its value reflects the degree of deviation of the current equipment operating state from the healthy operating mode. Therefore, analyzing the distribution characteristics and continuous change pattern of the residual is the key to distinguishing between random fluctuations and potential deterioration trends. Specifically, based on the historical residual distribution, the corresponding components for each component are determined, namely the temperature residual component, the partial discharge residual component, the vibration acceleration effective value residual component, and the load. The adaptive deviation threshold for the residual components of the current RMS value is determined by selecting the residual sequence within a preset time window prior to the current monitoring time as the historical residual dataset. This time window covers multiple operating cycles of the equipment to ensure that the statistical results reflect the typical fluctuation characteristics of the equipment under normal or near-normal conditions. Subsequently, statistical analysis is performed on each component in the historical residual dataset, and the standard deviation of each component, calculated by adding or subtracting a set multiple, is used as the upper and lower boundary thresholds of the corresponding adaptive deviation threshold. After obtaining the adaptive deviation threshold for each component, the current residual sequence is sequentially scanned to determine whether there are multiple consecutive time points where the residual value of at least one component exceeds the corresponding adaptive deviation threshold. During the scanning process, the residual value of each component at each time point is compared with the upper and lower thresholds of the corresponding adaptive deviation threshold. When the residual value of any component is higher than the upper threshold or lower than the lower threshold, a significant deviation is considered to exist at that time point, and a continuous counting mechanism is activated. If a time point does not exceed the above threshold range, the count is reset. To avoid misjudging transient interference or isolated anomalies as trend faults, a minimum continuous point threshold is preset. Only when the number of time points where the residual value of at least one component continuously exceeds the corresponding adaptive deviation threshold range reaches or exceeds the minimum continuous point threshold is a valid abnormal trend signal detected. By introducing a continuous constraint, random noise-induced... The system distinguishes between short-term deviations and continuous deviations caused by internal structural changes in the equipment, thereby improving the reliability of trend identification. When the continuity condition is met, the time point when the threshold is first exceeded is recorded as the starting time point of the abnormal trend segment, and the scanning continues along the time axis until the residual values ​​of all components are restored to the threshold range. The last time point of deviation before restoration is recorded as the ending time point, thus forming a continuous abnormal time interval. For each identified abnormal time interval, its start and end timestamps, corresponding parameter identifiers, and residual sequence segments within the interval are encapsulated to form an abnormal trend segment record. All abnormal trend segment records are summarized in chronological order to form an abnormal trend set with timestamps.

[0035] The above-mentioned technical solution can not only dynamically adapt to the normal fluctuation level of the equipment, but also accurately extract abnormal trend segments that reflect the degradation evolution process, providing a structured input basis for subsequent clustering and classification based on density and frequency domain features, thereby improving the accuracy and stability of the overall intelligent electrical equipment anomaly diagnosis.

[0036] Further, in step S4, the deterioration segments of the equipment are identified, including: Extract residual sequence segments corresponding to each abnormal trend segment in the abnormal trend set, calculate the dynamic time warping distance between each residual sequence segment and other residual sequence segments in the abnormal trend set, determine its K nearest neighbors, and calculate the local outlier density based on the average dynamic time warping distance to the K nearest neighbors; normalize and weight the residual sequence segments to obtain a comprehensive residual sequence, window the comprehensive residual sequence and perform a fast Fourier transform to obtain the amplitude spectrum, extract the frequency component with the largest amplitude as the dominant frequency component of the amplitude spectrum; represent each abnormal trend segment as a two-dimensional feature vector containing the local outlier density and the dominant frequency component of the amplitude spectrum, and cluster it using the K-means clustering method; based on the distribution of the dominant frequency component and the statistical characteristics of the local outlier density of each cluster, classify the clustering results into environmental periodic fluctuation clusters and continuous cumulative degradation clusters; identify the abnormal trend segments belonging to the continuous cumulative degradation clusters as equipment degradation segments.

[0037] Specifically, in order to accurately distinguish the abnormal trend segments that truly reflect the gradual deterioration of the equipment from a large amount of abnormal data under complex operating conditions, after obtaining the abnormal trend set, for each abnormal trend segment in the above abnormal trend set, its corresponding residual sequence segment is extracted. The above residual sequence segment is a sequence of residual values ​​of multiple consecutive sampling points obtained by truncating the original residual time series according to the start and end time index of the abnormal trend segment. Since the duration of each abnormal trend segment may be different, the length of the above residual sequence segments is allowed to be different. Each residual sequence segment retains its original time order to maintain the dynamic change characteristics of the abnormal form.

[0038] After obtaining all residual sequence segments, the dynamic time warping distance between each residual sequence segment and all other residual sequence segments in the anomalous trend set is calculated. This dynamic time warping distance is obtained by constructing a cumulative distance matrix between two time series and finding a path with the minimum cumulative distance while allowing for nonlinear stretching and compression of the time axis. This yields the optimal alignment distance between two time series of potentially different lengths, eliminating alignment bias caused by varying anomalous durations and ensuring that the distance metric truly reflects the similarity between anomalous patterns. For the current residual sequence segment, its dynamic time warping distances with the remaining segments are sorted from smallest to largest, and the K segments with the smallest distances are selected as its K nearest neighbors. The value of K is adaptively determined based on the total number of anomalous trend sets, preferably the square root of the total number of sets, to ensure local statistical stability. Meanwhile, excessive smoothing is avoided. After determining the K nearest neighbors, the dynamic time warping distance from the current residual sequence segment to all its K nearest neighbors is further calculated. The obtained dynamic time warping distance value is divided by the longer segment length of the two residual sequence segments involved in the calculation to obtain the length-normalized dynamic time warping distance. The reciprocal of this normalized dynamic time warping distance is defined as the local outlier density of the abnormal trend segment. When the average distance between a residual sequence segment and its neighboring segments is small, its corresponding density value is large, indicating that its abnormal shape has a high consistency with the surrounding abnormal segments in terms of dynamic change trend, reflecting the characteristics of group fluctuation. Conversely, when the average distance is large, its density value is small, indicating that its abnormal shape is relatively isolated, which may correspond to a unique abnormal response caused by the development of local defects inside the equipment. Thus, the quantitative characterization of each abnormal trend segment in terms of morphological similarity is completed.

[0039] After obtaining the density of local outliers, each residual sequence segment is normalized and then weighted to obtain a comprehensive residual sequence. Its frequency domain features are then extracted to identify whether the anomaly has stable periodicity. Specifically, since each time point in the above residual sequence field corresponds to the temperature residual, partial discharge residual, vibration acceleration effective value residual, and load current effective value residual, in order to facilitate overall analysis, the residual components at each time point are normalized and weighted to obtain a comprehensive residual value after weighting. The comprehensive residual sequence is then obtained, and a Hanning window function is applied to the comprehensive residual sequence for windowing processing. By smoothing and attenuating the two ends of the sequence, the spectral leakage caused by truncation during frequency domain transformation is reduced, thereby improving the accuracy of frequency component identification. Subsequently, a Fast Fourier Transform is performed on the windowed composite residual sequence to map the time-domain signal to the frequency domain, obtaining the corresponding complex spectrum data. The modulo operation is then performed on the complex spectrum to calculate the amplitude value corresponding to each frequency point, forming an amplitude spectrum. The frequency component with the largest amplitude is identified in the amplitude spectrum, and this frequency value is extracted as the dominant frequency component of the amplitude spectrum of the abnormal trend segment, so that the abnormal trend segment obtains a scalar value in the frequency domain that can represent its main oscillation period characteristics. When the anomaly originates from external periodic factors such as the alternation of ambient temperature day and night or the periodic change of power grid load, its dominant frequency component is usually concentrated in a specific low-frequency range corresponding to the daily or annual cycle. However, when the anomaly originates from the gradual deterioration process such as internal insulation aging, enhanced partial discharge, or mechanical wear, its frequency distribution often lacks stable peaks.

[0040] After obtaining the local outlier density and amplitude spectrum dominant frequency component for each anomalous trend segment, each anomalous trend segment is represented as a two-dimensional feature vector, where the first dimension is the local outlier density value and the second dimension is the amplitude spectrum dominant frequency component value. All anomalous trend segments form a set of two-dimensional feature vectors. Based on this set of feature vectors, K-means clustering is used for clustering. The number of clusters is preferably set to 2 to correspond to the two potential sources of environmental periodic fluctuations and continuous cumulative degradation. During the clustering process, two cluster centers are first initialized, and then the following process is iteratively executed: each two-dimensional feature vector is assigned to the cluster center with the closest Euclidean distance, and the coordinates of the cluster centers are recalculated based on all feature vectors in each cluster until the cluster centers no longer change or the preset maximum number of iterations is reached, thereby obtaining two stable clusters.

[0041] After clustering, the statistical characteristics of each cluster are analyzed. Specifically, the average value and distribution of all eigenvectors within each cluster are calculated in the second dimension (the dominant frequency component of the amplitude spectrum), and combined with the average value in the first dimension (the density of local outliers) for comprehensive judgment. When the average value of the dominant frequency component of a cluster is concentrated in a frequency range that matches the known environmental cycle, and the average value of the local outlier density within the cluster is high, it indicates that the anomalous segments within the cluster are similar in morphology and have stable periodicity, and should be identified as an environmental periodic fluctuation cluster. Conversely, when the average value of the dominant frequency component of another cluster is low or its distribution is scattered with no significant peak, and its average value of the local outlier density is low, it indicates that the anomalous segments within the cluster lack consistency in both morphology and frequency domain. The abnormal trend segments belonging to the aforementioned continuous cumulative degradation clusters should be identified as equipment degradation segments. The corresponding time intervals and related characteristic parameters are then output as the basis for subsequent degradation pattern identification, health index assessment, and diagnostic report generation. Through this technical solution, the data forms a mutually corroborating dual-feature structure in both the time-domain similarity and frequency-domain periodicity dimensions. Automatic classification is achieved through clustering algorithms, effectively distinguishing between group-based periodic fluctuations caused by external environmental factors and non-periodic isolated anomalies caused by the gradual development of internal equipment defects in long-term operation monitoring scenarios for intelligent electrical equipment. This avoids misjudging environmental interference as equipment failure and improves the accuracy and reliability of maintenance diagnosis results.

[0042] Step S5: Update the historical degradation pattern library based on the aforementioned equipment degradation segments and generate a diagnostic report; specifically including: The timestamp information, residual sequence segments, local outlier density, and amplitude spectrum dominant frequency components corresponding to the equipment degradation segment are extracted to form degradation feature data. The degradation feature data is then matched with existing degradation patterns in the historical degradation pattern library. When the matching degree is higher than a preset similarity threshold, the parameters of the corresponding degradation pattern are updated. When the matching degree is lower than the similarity threshold, the degradation feature data is written as a new degradation pattern into the historical degradation pattern library. Based on the duration, residual amplitude, and associated parameter pairs of the equipment degradation segment, an early warning report is generated that includes degradation type determination results, degradation degree assessment results, and recommended maintenance measures.

[0043] Specifically, after identifying and classifying equipment degradation segments, the historical degradation pattern library is updated, and a diagnostic report is generated to achieve self-learning and closed-loop optimization of equipment health management. Specifically, for each abnormal time interval identified as a degradation segment, its corresponding timestamp information is extracted, including the degradation start time, end time, and duration. Simultaneously, a residual sequence segment corresponding to that time interval is extracted from the residual sequence, and the local outlier density value and amplitude spectrum dominant frequency component value calculated in step S4 are read. These time features, morphological features, and frequency domain features together constitute the degradation feature data of the degradation segment. The residual sequence segment is used to characterize the dynamic evolution of the degradation process, and the local outlier density is used to characterize the anomaly within the overall anomaly set. The degree of isolation is determined by the amplitude spectrum dominant frequency component, which reflects whether the data has stable periodicity characteristics, thus forming a structured degradation feature description vector. After the degradation feature data is formed, it is input into a historical degradation pattern library for similarity matching. The historical degradation pattern library stores multiple pattern records corresponding to confirmed degradation types. Each pattern record includes the residual statistical features, density features, frequency domain features, and corresponding fault type labels and maintenance records of historical degradation segments. During the matching process, a weighted Euclidean distance or a method based on the fusion of dynamic time warping distance and feature vector distance can be used to calculate the similarity value between the current degradation feature data and the degradation feature data corresponding to each existing degradation pattern in the library, and the pattern corresponding to the degradation feature data with the highest similarity is selected as the candidate matching object.

[0044] When the highest similarity value exceeds the preset similarity threshold, the current deteriorated segment is determined to belong to the same deterioration type as the existing deterioration pattern. The parameters of the existing deterioration pattern are then updated, specifically including incremental updates to the residual mean, residual standard deviation, main frequency component statistical interval, and local outlier density statistical interval. This allows the historical deterioration pattern library to absorb new operational data features while maintaining its original knowledge structure, achieving continuous optimization and adaptive evolution of the patterns. When the highest similarity value falls below the preset similarity threshold, the current deteriorated segment is determined to potentially correspond to a new deterioration type or a new deterioration stage. The aforementioned deterioration feature data, along with its time information, associated parameter pairs, and preliminary diagnostic labels, are written into the historical deterioration pattern library as new deterioration pattern entries and assigned a unique pattern number. This expands the coverage of the pattern library and improves the system's ability to identify new faults or early deterioration patterns. After completing pattern matching or adding new entries, further analysis is performed based on the duration of the equipment deterioration segment, the peak residual amplitude, and the residual mean deviation. The system generates diagnostic reports based on the degree of displacement and associated parameters. The degradation type determination result is derived from the pattern label after matching with historical patterns or the feature classification result of newly created patterns. The degradation degree assessment result can be graded according to the magnitude of residual amplitude and duration, such as classified as mild degradation, moderate degradation, or severe degradation. The recommended maintenance measures are automatically generated based on the historical maintenance experience database corresponding to the degradation type. For example, for cases where temperature and partial discharge coupling is enhanced and residuals continue to rise, insulation condition testing and load adjustment are recommended. For cases where vibration and current abnormal coupling is enhanced, mechanical fastening structures or bearing conditions are recommended. Finally, the degradation type determination result, degradation degree assessment result, abnormal occurrence time interval, related parameter pair information, and corresponding maintenance recommendations are integrated into a structured early warning report and output through the system interface or remote monitoring platform. This enables a visualized presentation of the operating status of intelligent electrical equipment and precise operation and maintenance support, thus forming a complete closed-loop maintenance and diagnostic process from anomaly identification and pattern updating to decision support.

[0045] The present invention also provides an intelligent electrical equipment maintenance and diagnostic system for implementing the above-described method, such as... Figure 3 As shown, the system includes: A multi-source sequence acquisition unit is used to acquire multi-source sequences including temperature, partial discharge quantity, effective value of vibration acceleration, and effective value of load current; and to preprocess, normalize, and standardize the multi-source sequences to obtain standardized multi-source sequences. The correlation analysis unit is used to calculate the correlation coefficients between parameters in the standardized multi-source sequence, filter the correlation coefficients by combining time delay mutual information, and determine the time-varying correlation matrix based on the filtered high correlation parameters. The residual calculation unit is used to determine the parameter pairs that exceed the dynamic correlation threshold based on the time-varying correlation matrix; extract the synchronous change segments corresponding to the time period of the parameter pairs; and concatenate them into a dynamic feature vector sequence; input the dynamic feature vector sequence into a long short-term memory network to obtain a prediction sequence; and obtain a residual sequence based on the prediction sequence and the actual standardized multi-source sequence. The degradation analysis unit is used to determine the set of abnormal trend segments by comparing the residual sequence with the adaptive deviation threshold; and to perform cluster analysis on the set of abnormal trend segments based on the corresponding local outlier density and amplitude spectrum main frequency component to distinguish the equipment degradation segments. The diagnostic result generation unit is used to update the historical degradation pattern library based on the deteriorated equipment segments and generate a diagnostic report.

[0046] The present invention also provides a computer-readable storage medium storing instructions that, when executed by a processor, implement the above-described method.

[0047] In summary, this invention constructs a stable and consistent standardized multi-source sequence by performing median filtering, anomaly removal, and double standardization on the effective values ​​of temperature, partial discharge, vibration acceleration, and load current. This provides a high-quality data foundation for subsequent analysis and reduces the interference of on-site noise and dimensional differences on the model. Through a rolling window correlation analysis and time-delay mutual information joint screening mechanism, a time-varying correlation matrix that can be dynamically updated over time is constructed. This accurately characterizes the linear and dynamic coupling relationships between multiple parameters and their evolution characteristics with changing operating conditions, enabling a structured expression of the complex internal correlation structure of the equipment. Based on the time-varying correlation matrix, highly correlated synchronous change segments are extracted to form a fused dynamic feature vector sequence. A health status prediction model is established using a long short-term memory network, and the residual between the predicted and actual values ​​is used to predict the health status. The system quantifies operational deviations to proactively capture potential anomalies. Adaptive thresholds and continuity assessment mechanisms are introduced at the residual level to identify persistent anomalies. Clustering is performed using dynamic time warping distance and frequency domain dominant frequency characteristics to effectively distinguish between periodic environmental disturbances and gradual internal equipment degradation, avoiding false alarms. Furthermore, a matching and self-learning update mechanism based on a historical degradation pattern library enables continuous accumulation and optimization of diagnostic experience, generating early warning reports containing degradation type, degradation level, and maintenance recommendations. This forms a closed-loop diagnostic process from data acquisition, correlation modeling, trend identification, degradation assessment to decision support. Through the synergy of these technical solutions, the accuracy of anomaly identification, trend judgment capability, and operational decision-making precision of intelligent electrical equipment under complex operating conditions are significantly improved.

[0048] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0049] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0050] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for maintaining and diagnosing intelligent electrical equipment, characterized in that, The method includes: Step S1: Obtain a multi-source sequence including temperature, partial discharge quantity, effective value of vibration acceleration, and effective value of load current; perform preprocessing, normalization, and standardization on the multi-source sequence to obtain a standardized multi-source sequence; Step S2: Calculate the correlation coefficients between parameters in the standardized multi-source sequence, filter the correlation coefficients based on time delay mutual information, and determine the time-varying correlation matrix based on the filtered high-correlation parameters; Step S3: Determine the parameter pairs that exceed the dynamic correlation threshold based on the time-varying correlation matrix; extract the synchronous change segments corresponding to the time periods of the parameter pairs; and concatenate them into a dynamic feature vector sequence; input the dynamic feature vector sequence into a long short-term memory network to obtain a prediction sequence; obtain the residual sequence based on the prediction sequence and the corresponding actual standardized multi-source sequence. Step S4: Determine the set of abnormal trend segments by comparing the residual sequence with the adaptive deviation threshold; perform cluster analysis on the set of abnormal trend segments based on the corresponding local outlier density and amplitude spectrum main frequency component to distinguish equipment degradation segments; Step S5: Update the historical degradation pattern library based on the equipment degradation fragments and generate a diagnostic report.

2. The method according to claim 1, characterized in that, In step S1, the standardized multi-source sequence is obtained, including: The multi-source sequence is smoothed by a sliding window midpoint filter; and the range of each parameter dimension is unified by a minimum-maximum normalization process; the normalized sequence is then determined as a standardized multi-source sequence.

3. The method according to claim 1, characterized in that, In step S2, determining the time-varying correlation matrix based on the standardized multi-source sequences includes: A rolling time window is used to calculate the Pearson correlation coefficients between parameters on the standardized multi-source sequence, and the correlation coefficient sequences of each parameter pair in different time periods are obtained. The mutual information value of each parameter pair under different time delay conditions is calculated within a preset time delay range. The Pearson correlation coefficients are filtered based on the time delay mutual information value to obtain the Pearson correlation coefficients corresponding to parameter pairs whose mutual information values ​​exceed a preset filtering threshold under the corresponding time delay conditions. Highly correlated parameter pairs are extracted from the filtering results, and directional labels are added according to the positive or negative signs of the corresponding correlation coefficients. A time-varying correlation matrix is ​​constructed based on the Pearson correlation coefficients corresponding to the highly correlated parameter pairs with positive and negative directional labels.

4. The method according to claim 1, characterized in that, In step S3, extracting the dynamic feature vector sequence based on the time-varying correlation matrix includes: Based on the parameters in the time-varying correlation matrix, the start and end time points that exceed the corresponding dynamic correlation threshold are used to form corresponding time period boundaries. Based on the time period boundaries, sequence segments corresponding to the time period boundaries are extracted from the standardized multi-source sequences to obtain synchronous change segments. The length of the synchronous change segments is checked. When the segment length is less than the preset minimum length, it is extended to the adjacent time point. The extracted synchronous change segments are spliced ​​together point by point in chronological order to form a fused dynamic feature vector sequence. The feature vector at each time point includes temperature, partial discharge, effective value of vibration acceleration, and effective value of load current.

5. The method according to claim 4, characterized in that, In step S3, the residual sequence is obtained, including: The dynamic feature vector sequence is input into a long short-term memory network for single-step rolling prediction; a corresponding prediction sequence is generated, which represents the expected normal fluctuation when the device is in a healthy state; at each time point, the difference between the predicted value in the prediction sequence and the corresponding value in the standardized multi-source sequence is calculated to form a residual value point by point; the residual values ​​at each time point are arranged in chronological order to obtain the residual sequence.

6. The method according to claim 1, characterized in that, In step S4, determining the abnormal trend segment based on the residual sequence includes: An adaptive deviation threshold is determined based on the historical residual distribution; it is then determined whether multiple consecutive points in the residual sequence exceed the adaptive deviation threshold; if multiple consecutive points exceed the adaptive deviation threshold, the corresponding consecutive time periods are marked as abnormal trend segments, and an abnormal trend set with timestamps is obtained based on the abnormal trend segments.

7. The method according to claim 6, characterized in that, In step S4, the deterioration segments of the equipment are identified, including: Extract residual sequence segments corresponding to each abnormal trend segment in the abnormal trend set, calculate the dynamic time warping distance between each residual sequence segment and other residual sequence segments in the abnormal trend set, determine its K nearest neighbors, and calculate the local outlier density based on the average dynamic time warping distance to the K nearest neighbors; normalize and weight the residual sequence segments to obtain a comprehensive residual sequence, window the comprehensive residual sequence and perform a fast Fourier transform to obtain the amplitude spectrum, extract the frequency component with the largest amplitude as the dominant frequency component of the amplitude spectrum; represent each abnormal trend segment as a two-dimensional feature vector containing the local outlier density and the dominant frequency component of the amplitude spectrum, and cluster it using the K-means clustering method; based on the distribution of the dominant frequency component and the statistical characteristics of the local outlier density of each cluster, classify the clustering results into environmental periodic fluctuation clusters and continuous cumulative degradation clusters; identify the abnormal trend segments belonging to the continuous cumulative degradation clusters as equipment degradation segments.

8. The method according to claim 1, characterized in that, In step S5, the historical degradation pattern library is updated, and an early warning report is generated, including: The timestamp information, residual sequence segments, local outlier density, and amplitude spectrum dominant frequency components corresponding to the equipment degradation segment are extracted to form degradation feature data. The degradation feature data is then matched with existing degradation patterns in the historical degradation pattern library. When the matching degree is higher than a preset similarity threshold, the parameters of the corresponding degradation pattern are updated. When the matching degree is lower than the similarity threshold, the degradation feature data is written as a new degradation pattern into the historical degradation pattern library. Based on the duration, residual amplitude, and associated parameter pairs of the equipment degradation segment, an early warning report is generated that includes degradation type determination results, degradation degree assessment results, and recommended maintenance measures.

9. An intelligent electrical equipment maintenance and diagnostic system, used to implement the method as described in any one of claims 1-8, characterized in that, The system includes: A multi-source sequence acquisition unit is used to acquire multi-source sequences including temperature, partial discharge quantity, effective value of vibration acceleration, and effective value of load current; and to preprocess, normalize, and standardize the multi-source sequences to obtain standardized multi-source sequences. The correlation analysis unit is used to calculate the correlation coefficients between parameters in the standardized multi-source sequence, filter the correlation coefficients by combining time delay mutual information, and determine the time-varying correlation matrix based on the filtered high correlation parameters. The residual calculation unit is used to determine the parameter pairs that exceed the dynamic correlation threshold based on the time-varying correlation matrix; extract the synchronous change segments corresponding to the time period of the parameter pairs; and concatenate them into a dynamic feature vector sequence; input the dynamic feature vector sequence into a long short-term memory network to obtain a prediction sequence; and obtain a residual sequence based on the prediction sequence and the actual standardized multi-source sequence. The degradation analysis unit is used to determine the set of abnormal trend segments by comparing the residual sequence with the adaptive deviation threshold; and to perform cluster analysis on the set of abnormal trend segments based on the corresponding local outlier density and amplitude spectrum main frequency component to distinguish the equipment degradation segments. The diagnostic result generation unit is used to update the historical degradation pattern library based on the deteriorated equipment segments and generate a diagnostic report.

10. A computer-readable storage medium storing instructions thereon, characterized in that, When the instructions are executed by the processor, they implement the method as described in any one of claims 1-8.

Images