Power grid equipment fault intelligent diagnosis and early warning method based on grey correlation analysis

Abstract

The application relates to the technical field of power grid equipment state monitoring and fault diagnosis, and specifically discloses a power grid equipment fault intelligent diagnosis and early warning method based on grey correlation analysis, which reconstructs phase space of state monitoring data to construct high-dimensional operation track, dynamically generates an adaptive resolution coefficient matrix through a meta-learning network, extracts multi-dimensional differences between a real-time track and a fault benchmark track on three geometric levels of position, tangent and curvature, injects an improved grey correlation degree model to complete multi-dimensional correlation calculation and weighted fusion, and realizes joint decision of fault diagnosis and early warning through threshold determination and sliding window trend fitting. In addition, a dimension isolation extreme difference extraction and contrast perception balanced correction mechanism is introduced, so that the problem of weak dimension distinguishing force collapse caused by dimensional difference of multi-dimensional characteristic quantities is solved, and the discrimination sensitivity to early weak defects is ensured.

Description

Technical Field

[0001] This application relates to the field of power grid equipment condition monitoring and fault diagnosis technology, and more specifically, to a method for intelligent fault diagnosis and early warning of power grid equipment based on grey relational analysis. Background Technology

[0002] As the physical foundation for the safe and stable operation of power systems, power grid equipment faces various degradation threats during long-term service, such as insulation aging, partial discharge initiation, and mechanical fatigue. Real-time and accurate perception of its operating status and proactive prediction of fault trends are crucial for ensuring power supply reliability and managing equipment throughout its entire lifecycle. Grey relational analysis, due to its natural adaptability to small-sample and information-poor systems, has been gradually introduced into the field of power grid equipment condition assessment. It is used to assist in fault type identification by calculating the correlation between real-time monitoring sequences and known fault mode sequences.

[0003] However, existing diagnostic methods based on grey relational analysis have structural limitations at the core formula level—their resolution coefficients are fixed as constants and cannot adaptively adjust to dynamic changes in on-site operating conditions. This directly leads to false alarms due to over-amplification of normal fluctuations in scenarios with strong electromagnetic interference or sudden load changes, and false alarms due to insufficient resolution in the early stages of equipment degradation. Meanwhile, existing methods only perform discrete point-to-point static comparisons at a single time cross-section, failing to construct the equipment's operating state as a continuous evolution trajectory in a high-dimensional phase space. Therefore, they cannot utilize the evolution rate and acceleration information carried by the first and second derivatives of the trajectory to effectively isolate normal deviations caused by sudden changes in operating conditions and accurately pinpoint abnormal distortions driven by physical defects. Furthermore, even if the correlation analysis is extended to a multidimensional geometric feature space encompassing location, tangent direction, and curvature, since these three types of features originate from different physical measurement dimensions and have vastly different dimensions, if they are mixed indiscriminately when extracting the range parameter, the location dimension, which has a dominant numerical magnitude, will dominate the range value. This causes the correlation of the curvature dimension, which has a small magnitude, to degenerate to a narrow range approaching the limit value, thus losing the sensitivity to identify early defects that only show anomalies in trajectory evolution acceleration, fundamentally weakening the system's early warning capability.

[0004] Therefore, an optimized intelligent diagnosis and early warning scheme for power grid equipment faults based on grey relational analysis is desired. Summary of the Invention

[0005] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide a method for intelligent diagnosis and early warning of power grid equipment faults based on grey relational analysis, comprising: Step 1: Reconstruct the phase space of the acquired power grid equipment status monitoring data to obtain the trajectory tensor, and normalize and encode the acquired power grid equipment operating condition data to obtain the operating condition tensor. Step 2: Based on the pre-trained meta-learning network, dynamic feature extraction and channel attention weighting are performed on the working condition tensor to obtain the adaptive resolution coefficient matrix. Step 3: Perform baseline matching and multidimensional trajectory geometric feature extraction on the trajectory tensor and each baseline trajectory in the preset baseline trajectory set to obtain the geometric distance set. The geometric distance set includes the Euclidean space position difference, the difference of the angle between the first derivative tangent directions, and the sum of the two values. Step 4: Based on the adaptive resolution coefficient matrix, perform adaptive multidimensional dynamic grey relational degree calculation on the geometric distance set to obtain the comprehensive relational vector, which includes positional relational degree, tangent relational degree and curvature relational degree. Step 5: Perform fault diagnosis decision-making and trend warning triggering on the comprehensive correlation vector to obtain diagnosis and warning results.

[0006] Compared with existing technologies, this application proposes an intelligent fault diagnosis and early warning method for power grid equipment based on grey relational analysis. It constructs a high-dimensional continuous operating trajectory by reconstructing the phase space of condition monitoring data. A meta-learning network is used to perceive real-time operating conditions and dynamically generate an adaptive resolution coefficient matrix to replace traditional fixed constants. Then, multi-dimensional differences between the real-time trajectory and each fault baseline trajectory are extracted at three geometric levels: position, tangent direction, and curvature. The adaptive resolution coefficient matrix is ​​injected into an improved grey relational model for multi-dimensional correlation calculation and weighted fusion. Finally, joint decision-making for fault confirmation and early warning is achieved through static threshold judgment and sliding window trend fitting. Simultaneously, to address the problem of weak dimension discrimination collapse caused by dimensional differences between multi-dimensional geometric features, a dimensional isolation range extraction and contrast perception equilibrium correction mechanism is introduced to ensure that the correlation of each dimension is fully expanded within the complete value range, restoring the sensitivity to early and subtle defects and the ability to provide early warning. Attached Figure Description

[0007] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0008] Figure 1 This is a flowchart of a method for intelligent diagnosis and early warning of power grid equipment faults based on grey relational analysis, according to an embodiment of this application. Figure 2This is a data flow diagram illustrating an intelligent fault diagnosis and early warning method for power grid equipment based on grey relational analysis according to an embodiment of this application; Figure 3 The flowchart illustrates the process of dynamically extracting features from the operating condition tensor and applying channel attention weighting to obtain an adaptive resolution coefficient matrix using a pre-trained meta-learning network based on grey relational analysis, according to an embodiment of this application. Figure 4 This is a flowchart illustrating a method for intelligent diagnosis and early warning of power grid equipment faults based on grey relational analysis according to an embodiment of this application. The method involves benchmark matching and multidimensional trajectory geometric feature extraction between the trajectory tensor and each benchmark trajectory in a preset benchmark trajectory set to obtain a geometric distance set, which includes Euclidean space position differences, first derivative tangent direction angle differences, and so on. Figure 5 This is a flowchart illustrating an intelligent diagnosis and early warning method for power grid equipment faults based on grey relational analysis, according to an embodiment of this application. The method involves adaptively calculating the multidimensional dynamic grey relational degree of the geometric distance set using an adaptive resolution coefficient matrix to obtain a comprehensive relational vector. Figure 6 This is a flowchart illustrating the multidimensional dynamic grey relational degree calculation for balancing the geometric distance set based on dimensional isolation range extraction and comparative perception cross-dimensional equilibrium factor calculation, according to an embodiment of this application, to obtain a comprehensive correlation vector. Detailed Implementation

[0009] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0010] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0011] While this application makes various references to certain modules of the systems according to embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The modules described are merely illustrative, and different aspects of the systems and methods may use different modules.

[0012] Flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0013] Most existing fault diagnosis technologies for power grid equipment still rely on fixed resolution coefficients and static cross-sectional grey relational analysis. Under conditions of load fluctuations, ambient temperature changes, and strong noise interference, these methods are prone to misjudging normal operating disturbances as faults or masking early, subtle degradation features, leading to missed alarms. Especially when performing joint analysis of multi-dimensional geometric features such as location, tangent, and curvature, traditional methods are susceptible to collapse in their ability to distinguish weak features due to scale differences between different dimensions, making it difficult for the system to promptly identify the evolution trend of fault nascent stages. Therefore, this application proposes an intelligent fault diagnosis and early warning method for power grid equipment based on grey relational analysis. This method first synchronizes, filters, fuses, and decouples the collected power grid equipment condition monitoring data and operating condition data. Information representing equipment degradation is reconstructed into a phase space trajectory tensor, and information representing the external environment and operating load is encoded into an operating condition tensor, thus establishing the foundation for state evolution representation and environmental perception, respectively. Next, a pre-trained meta-learning network is used to dynamically extract features from the operating condition tensor, and an adaptive resolution coefficient matrix matching the current operating environment is generated using a channel attention mechanism. This enables the grey relational model to automatically adjust its resolution capability as operating conditions change. Finally, the real-time trajectory is compared with various fault baseline trajectories. The system performs trace-by-trace matching to extract multi-dimensional geometric features such as positional differences, tangent direction differences, and curvature variation differences. It also performs dynamic grey relational degree calculations using adaptive resolution coefficients. When necessary, it further avoids mutual suppression between features of different dimensions by using dimensional isolation range extraction and cross-dimensional balance correction. Finally, it performs weighted fusion of the obtained positional relational degree, tangent relational degree, and curvature relational degree to form a comprehensive relational vector. Combined with static threshold judgment and trend analysis under the sliding time window, it outputs diagnostic results containing fault type, severity, and early warning information, thereby achieving accurate identification and early warning of early deterioration of power grid equipment.

[0014] Figure 1 This is a flowchart of a method for intelligent diagnosis and early warning of power grid equipment faults based on grey relational analysis, according to an embodiment of this application. Figure 2 This is a data flow diagram illustrating a method for intelligent diagnosis and early warning of power grid equipment faults based on grey relational analysis, according to an embodiment of this application. Figure 1 and Figure 2As shown, an intelligent fault diagnosis and early warning method for power grid equipment based on grey relational analysis according to an embodiment of this application includes: S1, reconstructing the phase space of the acquired power grid equipment status monitoring data to obtain a trajectory tensor, and normalizing and encoding the acquired power grid equipment operating condition data to obtain an operating condition tensor; S2, performing dynamic feature extraction and channel attention weighting on the operating condition tensor based on a pre-trained meta-learning network to obtain an adaptive resolution coefficient matrix; S3, performing benchmark matching and multi-dimensional trajectory geometric feature extraction on the trajectory tensor and each benchmark trajectory in a preset benchmark trajectory set to obtain a geometric distance set, the geometric distance set including Euclidean space position difference, first derivative tangent direction angle difference, and so on; S4, performing adaptive multi-dimensional dynamic grey relational degree calculation on the geometric distance set based on the adaptive resolution coefficient matrix to obtain a comprehensive correlation vector, the comprehensive correlation vector including position correlation degree, tangent correlation degree, and curvature correlation degree; S5, performing fault diagnosis decision and trend early warning triggering on the comprehensive correlation vector to obtain a diagnosis and early warning result.

[0015] Specifically, in step S1, the acquired power grid equipment condition monitoring data is reconstructed in phase space to obtain a trajectory tensor, and the acquired power grid equipment operating condition data is normalized and encoded to obtain an operating condition tensor. It should be noted that, due to differences in sampling time, noise level, and physical meaning between the acquired power grid equipment condition monitoring data and operating condition data, directly entering the original sequence into grey relational calculations would introduce operating condition disturbances and weaken the ability to characterize the continuous evolution of equipment degradation. Based on this, the technical solution of this application first reconstructs the phase space of the condition monitoring data to obtain a trajectory tensor, and then normalizes and encodes the operating condition data to obtain an operating condition tensor. Through the above processing, the equipment degradation information and external load information can be respectively structured and represented, thereby effectively providing a unified data foundation for subsequent adaptive resolution coefficient calculation and multi-dimensional dynamic grey relational degree solution.

[0016] More specifically, in a concrete example of this application, the acquired condition monitoring data and operating condition data are first synchronized and aligned under a unified time reference, and then subjected to adaptive filtering. Specifically, multi-source monitoring sequences within the same sampling window are matched according to timestamps, missing sampling positions are interpolated to fill in gaps, abnormal jump values ​​are corrected, and the aligned sequences are filtered to reduce noise, thereby mitigating the effects of electromagnetic interference, sensor sampling jitter, and short-term disturbances, thus obtaining synchronized fused data. After this processing, the condition monitoring data and operating condition data form a consistent data foundation in the time dimension, facilitating subsequent joint analysis.

[0017] After obtaining the synchronously fused data, modal decoupling and feature separation processing are performed. Specifically, the synchronously fused data is first constructed into a feature matrix in chronological order. Then, based on principal component analysis, the feature matrix is ​​centered, analyzed for covariance, and principal components are extracted. Combining the load distribution of each principal component on different physical quantities, information reflecting insulation aging, partial discharge, and equipment degradation is extracted to form an equipment state sequence, and information reflecting changes in ambient temperature, load disturbance, operating voltage fluctuations, and humidity is extracted to form an external operating condition sequence. Subsequently, the equipment state sequence is reconstructed and mapped in phase space to form a trajectory tensor representing the continuous evolution of the equipment's operating state. Specifically, the delay time is first determined based on the temporal correlation of the equipment state sequence, and the embedding dimension is determined based on the signal expansion requirements in high-dimensional space. Then, the one-dimensional equipment state sequence is reconstructed into a high-dimensional state vector according to the delay coordinate method, and arranged in chronological order to form a trajectory tensor. The phase space reconstruction can be expressed as: in, Indicates the first There are several reconstructed state vectors, which, when arranged along the time dimension, form a trajectory tensor. Indicates the device state sequence at the 1st The sampled value at each moment. Indicates the delay time. This represents the embedding dimension. After this mapping, the original discrete state monitoring sequence is converted into high-dimensional tensor data that can characterize the state evolution trajectory, providing a basis for subsequent extraction of positional differences, tangent differences, and curvature differences. Finally, the external operating condition sequence is subjected to operating condition feature standardization and encoding to obtain the operating condition tensor. Specifically, the external operating condition sequence is first dimensionlessized according to the value range of each operating condition quantity within the statistical window to eliminate the dimensional differences between ambient temperature, load current, operating voltage, and humidity. Then, the dimensionless operating condition sequence is temporally arranged and tensorized and encoded according to a preset time window and feature dimension to form the operating condition tensor. The standardization process can be expressed as: in, Indicates the first Standard chemical condition characteristic values ​​at a given time point Indicates the first The original operating condition characteristic values ​​at each time point, This represents the minimum value of this operating condition characteristic within the statistical window. This represents the maximum value of the characteristic of this operating condition within the statistical window. After standardization and encoding, operating condition data from different sources and with different amplitude ranges are uniformly converted into a comparable tensor representation for subsequent operating condition sensing and adaptive resolution coefficient calculation.

[0018] Specifically, in step S2, based on a pre-trained meta-learning network, dynamic feature extraction and channel attention weighting are performed on the operating condition tensor to obtain an adaptive resolution coefficient matrix. It should be noted that, given that the signal-to-noise ratio and fault characterization intensity of power grid equipment monitoring signals dynamically change under conditions of load fluctuations, ambient temperature variations, humidity disturbances, and strong electromagnetic interference, traditional grey relational analysis using fixed resolution coefficients struggles to simultaneously differentiate between normal operating condition fluctuations and early subtle degradation features. This can easily lead to misjudging normal deviations caused by sudden load changes as faults, or masking early abnormal information in scenarios such as transformer insulation thermal aging, partial discharge between winding turns, GIS micro-water leakage, and localized overheating of high-voltage cable joints. Therefore, the technical solution of this application further utilizes a pre-trained meta-learning network to perform dynamic feature extraction and channel attention weighting on the operating condition tensor to obtain an adaptive resolution coefficient matrix. This maps the influence of operating condition factors such as ambient temperature, load current, operating voltage, and humidity on resolution capability into matrix parameters that vary with time and feature channels, transforming the traditional fixed constant resolution coefficients into an adaptive resolution coefficient expression that matches the current operating condition state. Through the above processing, the resolution of gray relational calculation can be dynamically adjusted according to the working conditions under different operating environments, thereby effectively suppressing false alarms caused by environmental disturbances and noise fluctuations, and improving the ability to identify early deterioration trends and weak fault symptoms of equipment.

[0019] Figure 3 This document describes a flowchart illustrating the process of dynamically extracting features from the operating condition tensor and applying channel attention weighting to obtain an adaptive resolution coefficient matrix using a pre-trained meta-learning network based on grey relational analysis, according to an embodiment of this application. Figure 3 As shown, step S2 includes: S21, capturing local working condition mutations in the working condition tensor based on a pre-trained meta-learning network to obtain an environmental feature map; S22, performing channel attention weighted evaluation on the environmental feature map to obtain weighted environmental features; S23, performing nonlinear mapping and matrix reshaping on the weighted environmental features to obtain an adaptive resolution coefficient matrix.

[0020] In step S21, based on a pre-trained meta-learning network, local operating condition mutations are captured in the operating condition tensor to obtain an environmental feature map. It should be noted that, due to the combined effects of load current fluctuations, ambient temperature changes, humidity disturbances, and operating voltage fluctuations during power grid equipment operation, operating condition information exhibits characteristics of local mutations, short-term transitions, and slow drifts in the time dimension. The subsequent generation of adaptive resolution coefficients needs to accurately reflect the actual impact of such operating condition changes on the signal-to-noise ratio and fault separability of the monitoring signal. If the original operating condition tensor is directly used in subsequent calculations, it is difficult to effectively distinguish the key change segments corresponding to normal operating condition fluctuations and abnormal disturbances. Therefore, the technical solution of this application further uses a pre-trained meta-learning network to capture local operating condition mutations in the operating condition tensor to obtain an environmental feature map. Through the above processing, local dynamic features closely related to changes in the equipment operating environment can be extracted from the operating condition tensor, thereby effectively providing a feature foundation with temporal identification capabilities for subsequent channel attention weighting and adaptive resolution coefficient matrix generation.

[0021] More specifically, in a concrete example of this application, the operating condition tensor, composed of ambient temperature, load current, operating voltage, and humidity, is first input into a pre-trained meta-learning network according to a preset time window, maintaining the correspondence between the features of each operating condition in the time dimension to ensure that the network extraction results can reflect the combined effect of external load and environmental conditions on the equipment's operating state within the same time period. Subsequently, a local receptive field scan is performed along the time axis of the operating condition tensor, and convolutional extraction is performed on the amplitude changes, rate of change, and short-term abrupt changes in adjacent time slices, so that the local patterns corresponding to load surges, temperature changes, abnormal humidity increases, and voltage fluctuations form separable representations in the latent space. The mapping from the operating condition tensor to the environmental feature map can be expressed as: in, This represents an environmental feature map, used to characterize the local environmental changes extracted dynamically from the operating condition tensor. This represents the input operating condition tensor, reflecting the combined distribution of ambient temperature, load current, operating voltage, and humidity at various times. Represents the convolution weights, used to extract local change patterns within a time window; symbol This represents the convolution operation. This represents the bias term, used to correct the baseline level of the convolution response. This represents a nonlinear activation function used to preserve effective responses and suppress ineffective negative decay. After this mapping, the scattered local variation information in the original operating condition sequence is transformed into an environmental feature map that characterizes the intensity and location of environmental disturbances. Following the capture of these local operating condition abrupt changes, the feature responses extracted at different time scales are further aggregated along the feature channels, ensuring that both short-period sudden changes and longer-period gradual trends are simultaneously preserved in the environmental feature map. The short-period features reflect transient operating condition disturbances, while the longer-period features reflect continuous load changes and gradual environmental changes.

[0022] Specifically, the pre-trained meta-learning network is trained offline. During training, multiple sets of operating condition tasks covering different load fluctuation levels, ambient temperature ranges, humidity changes, voltage disturbance intensity, and noise background are first constructed based on historical power grid equipment operation data. The operating condition data in each task are organized into an operating condition tensor, and the corresponding fault type, early warning result, or correlation optimization objective is used as supervision information. Subsequently, within each set of operating condition tasks, the operating condition tensor is used to generate resolution coefficient output, and the corresponding diagnostic error, early warning error, or correlation calculation error is combined to update the parameters within the task. Then, cross-task parameter aggregation and update are performed between multiple sets of operating condition tasks, so that the network gradually learns the mapping relationship between resolution coefficient and diagnostic effect under different operating conditions. After iterative training, the initial parameters of the network that can maintain rapid adaptation to unseen operating conditions are obtained, and these parameters are used as the pre-trained meta-learning network in step S21, so that the subsequent input operating condition tensor can be directly mapped to a feature expression that matches the current operating environment.

[0023] In step S22, channel attention weighting evaluation is performed on the environmental feature map to obtain weighted environmental features. It should be noted that since the types and intensities of operating condition changes corresponding to different channels in the environmental feature map, and their impact on the separability of monitoring signals, are not consistent, the roles of load current surges, ambient temperature jumps, operating voltage fluctuations, and humidity changes differ in the early stages of fault evolution. If each channel feature is equally weighted in the subsequent resolution coefficient mapping, redundant responses with low diagnostic relevance will be retained, and the effect of key operating condition features on adjusting the grey relational resolution capability will be weakened. Based on this, the technical solution of this application further performs channel attention weighting evaluation on the environmental feature map to obtain weighted environmental features. Through the above processing, features with a high degree of correlation with the current operating condition and signal-to-noise ratio changes can obtain higher weights, thereby effectively improving the matching ability of the subsequent adaptive resolution coefficient matrix to complex operating condition changes.

[0024] More specifically, in a concrete example of this application, the environmental feature map is first statistically aggregated along the channel dimension, compressing the overall response of each channel within the current time window into a channel description vector to obtain the global representation of each type of working condition feature in the current sample. Subsequently, this channel description vector is sequentially subjected to linear dimensionality reduction, nonlinear activation, and linear dimensionality increase, allowing the relative importance of each channel to be compared in a low-dimensional space. Then, a channel weight vector with values ​​between 0 and 1 is formed through normalization mapping, and this weight vector is mapped one-to-one with the channel dimension of the environmental feature map. This process can be represented as: in, This represents the weighted environmental characteristics, used to characterize the operating condition results after channel attention recalibration. This represents an environmental feature map, used to characterize the local operating condition abrupt change features extracted in the preceding steps. This indicates a global average aggregation operation, used to compress the timing response of each channel into channel description values. This represents the dimensionality reduction mapping parameters, used to compress the channel description vectors and extract the main differences. This represents a non-linear activation function used to preserve positive valid responses. This represents the dimensionality-upgrading mapping parameter, used to restore the low-dimensional comparison results to the original number of channels. This represents the normalization function, used to generate the weight coefficients for each channel. This represents an element-wise weighted operation, used to apply channel weights to the corresponding channels of the environmental feature map. After obtaining the channel weight vector, a channel-wise weighting process is performed on it and the environmental feature map. This process preserves highly correlated channels corresponding to sudden load increases, abnormal changes in ambient temperature, continuous increases in humidity, and unstable operating voltage, while suppressing the responses of stable background segments and channels with low correlation, thereby obtaining the weighted environmental features.

[0025] In step S23, the weighted environmental features are nonlinearly mapped and reshaped to obtain an adaptive resolution coefficient matrix. It should be noted that since the noise level, state fluctuation amplitude, and fault feature exposure of power grid equipment vary under different operating conditions, although the environmental features after dynamic feature extraction and channel attention weighting reflect the dominant information of operating condition changes, these features cannot be directly used in the grey relational formula calculation and need to be converted into a resolution parameter expression form that matches the relational calculation process. Based on this, the technical solution of this application further performs nonlinear mapping and matrix reshaping on the weighted environmental features to obtain an adaptive resolution coefficient matrix. Through the above processing, the influence of operating condition changes on the relational resolution capability can be transformed into a parameter matrix that varies with time and feature dimensions, thereby effectively suppressing misjudgments and missed judgments caused by fixed resolution coefficients under load disturbances, ambient temperature rises, humidity changes, and voltage fluctuations.

[0026] More specifically, in a specific example of this application, the weighted environmental features are first expanded into continuous feature vectors along a predetermined feature order. Then, based on the parameter dimension requirements of subsequent grey relational analysis, the target length of the output parameters is determined, establishing a one-to-one mapping relationship between the weighted environmental features and the target resolution parameters. The purpose of this process is to compress the condition-sensitive features obtained in the preceding steps into computable, constrainable, and rearrangeable parameter inputs, so as to generate a resolution coefficient expression adapted to the trajectory tensor. After completing the above feature processing, a nonlinear mapping process is performed on the weighted environmental features. Specifically, the weighted environmental features are input into a linear mapping unit to obtain an intermediate response corresponding to the target parameter length. Then, a nonlinear constraint is applied to this intermediate response to ensure that the output value stably falls within the resolution coefficient range required for grey relational analysis, thereby avoiding interference from parameter out-of-bounds errors in subsequent relational degree calculations. This nonlinear mapping can be expressed as: in, This represents the resolution coefficient vector obtained after nonlinear mapping, with each element taking values ​​between 0 and 1. It serves as the direct input for subsequent reshaping processing. This represents weighted environmental characteristics, used to characterize operational information after the channel importance recalibration has been completed. This represents the mapping weights, used to establish a linear correspondence between operating condition characteristics and resolution parameters. This represents the bias term, used to correct the baseline level of the mapping result. This represents the base of the natural logarithm. After obtaining the resolution coefficient vector, it is further reshaped according to the time and spatial embedding dimensions of the trajectory tensor, so that the resolution parameters structurally correspond to the subsequent geometric distance data. This reshaping process can be represented as: in, This represents the adaptive resolution coefficient matrix, used to apply differential adjustments to the grey relational calculations at different time steps and different feature dimensions. This indicates a reshaping operation. Indicates the length of the time dimension, corresponding to the number of time steps in the trajectory tensor. The spatial embedding dimension represents the feature dimension after phase space reconstruction. After this reshaping process, the originally one-dimensional arrangement of resolution parameters is transformed into a matrix form that matches the trajectory tensor and geometric distance set. This makes the resolution coefficients no longer a single fixed constant, but an adaptive resolution coefficient matrix that can reflect changes in current operating conditions, signal-to-noise ratio level, and fault separability, providing a parameter basis for the subsequent dynamic calculation of position correlation, tangent correlation, and curvature correlation.

[0027] Specifically, in step S3, the trajectory tensor is matched with each reference trajectory in a preset set of reference trajectories, and multidimensional trajectory geometric features are extracted to obtain a geometric distance set. This geometric distance set includes the Euclidean space position difference, the difference in the angle between the first derivative tangent directions, and the absolute deviation of curvature. It should be noted that, given the continuous evolution of power grid equipment degradation, simply performing grey relational comparison based on static features at a specific moment or section is insufficient to distinguish between normal trajectory deviations caused by load fluctuations and environmental changes, and abnormal evolution trajectories caused by insulation aging, partial discharge, micro-water leakage, and local overheating. This is especially true for scenarios such as the initial thermal aging of transformer insulation paper, the initiation period of inter-turn defects in windings, and abnormal temperature rise at high-voltage cable joints. Early anomalies often manifest first in changes in trajectory evolution direction and curvature, rather than significant shifts in position coordinates. Therefore, the technical solution of this application further performs benchmark matching and multidimensional trajectory geometric feature extraction on the trajectory tensor and each reference trajectory in a preset set of reference trajectories to obtain a geometric distance set. This geometric distance set includes the Euclidean space position difference, the difference in the angle between the first derivative tangent directions, and the absolute deviation of curvature. Through the above processing, the differences in position, evolution trend, and acceleration change between the real-time status trajectory and the baseline trajectory of different fault modes can be uniformly quantified, thereby effectively improving the ability to identify early deterioration characteristics and fault evolution trends of equipment, and providing a geometric basis for subsequent adaptive multidimensional dynamic grey relational degree calculation.

[0028] Figure 4 This document describes a method for intelligent diagnosis and early warning of power grid equipment faults based on grey relational analysis, according to an embodiment of this application. The method involves benchmark matching and multi-dimensional trajectory geometric feature extraction between the trajectory tensor and each benchmark trajectory in a preset benchmark trajectory set to obtain a geometric distance set. The geometric distance set includes Euclidean space position differences, differences in the angles between the first-order derivative tangent directions, and so on. (See flowchart for example.) Figure 4As shown, step S3 includes: S31, obtaining trajectory pairs by pairing the trajectory tensor with each reference trajectory in the reference trajectory set one by one in the time dimension, and calculating the absolute position difference of each pair of trajectory pairs in high-dimensional Euclidean space to obtain the position distance matrix; S32, differentiating the tangent vectors along the time axis for the real-time trajectory and the reference trajectory in the trajectory pair, and calculating the angle difference of the tangent directions at the same time step to obtain the tangent distance matrix; S33, differentiating the curvature scalar for the real-time trajectory and the reference trajectory in the trajectory pair, and calculating the absolute curvature difference at the same time step to obtain the curvature distance matrix; S34, performing multi-dimensional encapsulation aggregation on the position distance matrix, the tangent distance matrix, and the curvature distance matrix to obtain the geometric distance set.

[0029] In step S31, trajectory pairs are obtained by pairing the trajectory tensor with each reference trajectory in the reference trajectory set one by one in the time dimension, and the absolute position difference of each pair of trajectory pairs in high-dimensional Euclidean space is calculated to obtain the position distance matrix. It should be noted that since the trajectory tensor after phase space reconstruction represents the continuous evolution position of the power grid equipment state in high-dimensional state space, and the reference trajectories corresponding to different fault types have different spatial distribution characteristics at the same time scale, it is difficult to accurately characterize the trajectory deviation under conditions such as transformer insulation aging, partial discharge, switchgear insulation defects, and cable joint overheating without first establishing the time-by-time correspondence between the real-time trajectory and various reference trajectories. Based on this, the technical solution of this application further obtains trajectory pairs by pairing the trajectory tensor with each reference trajectory in the reference trajectory set one by one in the time dimension, and calculates the absolute position difference of each pair of trajectory pairs in high-dimensional Euclidean space to obtain the position distance matrix. Through the above processing, the spatial deviation of the real-time running trajectory from various known state modes can be uniformly quantified, thereby effectively providing the basic input of the position dimension for subsequent extraction of tangent difference, curvature difference and gray relational degree calculation.

[0030] More specifically, in a concrete example of this application, the trajectory tensor and the baseline trajectory set are first processed to be consistent in the time dimension, so that the real-time trajectory and each baseline trajectory maintain a correspondence in sampling length, time step sequence, and embedding dimension. For data with inconsistent sampling starting points or different time lengths, they are first truncated and aligned according to a unified time window, and then organized into a comparable time-series trajectory sequence based on the same sampling interval. The baseline trajectory set pre-includes trajectories under normal operating conditions, natural aging trajectories, and typical trajectories formed under different fault modes, thereby ensuring that the subsequent comparison objects are consistent with the actual diagnostic scenario.

[0031] After time-consistency processing, the trajectory tensor is paired with each reference trajectory in the reference trajectory set one by one according to time steps, forming multiple trajectory pairs. Each trajectory pair consists of a real-time trajectory within the same time window and a corresponding reference trajectory, maintaining a one-to-one correspondence of coordinate points at each time step. Through this pairing process, the real-time operating state is projected onto the reference frames of normal mode, aging mode, and fault mode, respectively, transforming the subsequent position difference calculation from an isolated single-trajectory analysis into a comparative analysis process oriented towards multiple known state modes. Subsequently, absolute position difference is calculated for each trajectory pair in high-dimensional Euclidean space to quantify the spatial deviation between the real-time trajectory and the reference trajectory at each time step. The position difference calculation can be expressed as: in, Indicates the first The baseline trajectory and the real-time trajectory are in the first... The absolute position difference at each time step reflects the degree of spatial offset between the two trajectories in the high-dimensional state space at that moment. Indicates the real-time trajectory at the 1st The trajectory coordinate vector corresponding to each time step. Indicates the first The baseline trajectory is in the first The trajectory coordinate vector corresponding to each time step, symbol The L2 norm is used to calculate the Euclidean distance between two high-dimensional coordinate vectors. This calculation directly yields the spatial position residuals of the real-time state relative to each reference state; the smaller the position difference, the closer the current operating state is to the corresponding baseline pattern. After obtaining the hourly position differences for each pair of trajectory pairs, all difference results are sequentially arranged according to the baseline trajectory index and time index to form a position distance matrix. This position distance matrix corresponds to different baseline trajectories in the row direction and different time steps in the column direction. Each element in the matrix represents the position deviation of the real-time trajectory relative to a certain baseline trajectory at a certain moment.

[0032] In step S32, the tangent vectors of the real-time trajectory and the reference trajectory in the trajectory pair are differentiated along the time axis, and the angle difference between the tangent directions at the same time step is calculated to obtain the tangent distance matrix. It should be noted that since the positional offset in high-dimensional phase space can only reflect the proximity of the real-time trajectory and the reference trajectory in the coordinate plane, it is difficult to reveal the directional change of the equipment state over time. Processes such as the initiation of thermal aging of transformer insulation, the development of inter-turn discharge in windings, the deterioration of switchgear insulation, and local overheating of cable joints often first manifest as a change in the trajectory evolution direction, and then further as the accumulation of positional offsets. Based on this, the technical solution of this application further differentiates the tangent vectors of the real-time trajectory and the reference trajectory in the trajectory pair along the time axis, and calculates the angle difference between the tangent directions at the same time step to obtain the tangent distance matrix. Through the above processing, the directional differences between the equipment state evolution trends can be quantified, thereby effectively improving the ability to distinguish between early deterioration processes and adjacent fault modes.

[0033] More specifically, in a concrete example of this application, the temporal continuity of the formed trajectory pairs is first verified, and the real-time trajectory and the reference trajectory are arranged sequentially according to a uniform sampling step size to ensure that the same trajectory pair has a differentiable relationship between adjacent time steps. For each trajectory pair, continuous high-dimensional trajectory coordinates are extracted using the same time index, and a discrete tangent vector is constructed from the coordinate difference between the previous time step and the current time step, so that the state evolution direction at each time step can be uniformly described. The real-time trajectory and the reference trajectory... The baseline trajectory is in the first The tangent vectors at each time step are represented as follows: in, Indicates the real-time trajectory at the 1st The tangent vector at each time step is used to characterize the instantaneous direction of the current state evolution. Indicates the first The baseline trajectory is in the first The tangent vector at each time step is used to characterize the evolution direction of the corresponding reference mode at that time. Indicates the real-time trajectory at the 1st High-dimensional coordinate vector at each time step Indicates the first The baseline trajectory is in the first High-dimensional coordinate vector at each time step This represents the time step between adjacent sampling moments. After obtaining the tangent vectors of the real-time trajectory and the reference trajectory, the directions of the two tangent vectors at the same time step are further compared to extract the angle difference between the trajectory evolution trends. The angle difference is calculated using the inner product relationship of the tangent vectors and is used as the directional deviation at that time step. Its calculation formula is: in, Indicates the first The baseline trajectory and the real-time trajectory are in the first... The difference in the tangent direction angle at each time step is used to quantify the difference in the evolution direction of the two trajectories at that moment. This represents the inner product of two tangent vectors. and Let L1 and L2 represent the real-time trajectory tangent vector and the reference trajectory tangent vector, respectively. This represents the regularization constant, used to prevent computational instability caused by a zero denominator. The inverse cosine operation is used to convert directional similarity into angular difference. This formula allows for the explicit representation of trend deviations that are difficult to identify solely based on positional differences. A smaller angular difference indicates a more consistent evolution direction between the real-time trajectory and the corresponding baseline trajectory; a larger angular difference indicates a more significant deviation in their respective development trends. Subsequently, the angular differences between each pair of trajectories across all time steps are arranged according to the baseline trajectory index and the time index, forming a tangent distance matrix. Each row of this matrix corresponds to a baseline trajectory, and each column corresponds to a time step. The matrix elements represent the difference in tangent direction between the real-time trajectory and the corresponding baseline trajectory at the given time.

[0034] In step S33, the curvature scalar derivatives of the real-time trajectory and the reference trajectory in the trajectory pair are calculated, and the absolute curvature deviation at the same time step is calculated to obtain the curvature distance matrix. It should be noted that, since the state trajectory of power grid equipment does not necessarily manifest as a significant shift in position coordinates in the early stages of a fault, but rather as a change in evolutionary acceleration, especially in the early stages of thermal aging of transformer insulation paper, the initiation stage of partial discharge between winding turns, micro-water leakage in switchgear, and local overheating of cable joints, abnormalities in the degree of trajectory bending often appear earlier than differences in position and tangential direction. If only positional distance and tangential distance are used for discrimination, it is difficult to fully identify these early weak fault characteristics. Based on this, the technical solution of this application further calculates the curvature scalar derivatives of the real-time trajectory and the reference trajectory in the trajectory pair, and calculates the absolute curvature deviation at the same time step to obtain the curvature distance matrix. Through the above processing, the difference between the real-time trajectory and the reference trajectory at the level of evolutionary acceleration can be quantified, thereby effectively improving the ability to identify early signs of equipment degradation and complex fault evolution processes.

[0035] More specifically, in a concrete example of this application, the real-time trajectory and the reference trajectory in the trajectory pair are first arranged sequentially according to a uniform time step, and the trajectory coordinates of three consecutive time steps are used as the basic data for curvature differentiation, so that the degree of trajectory curvature at each time step can be represented in discrete form. For the beginning and end positions of the time series, the boundary processing is completed using a one-sided difference method; for the middle position, the second-order change is calculated using a central difference method to ensure that the curvature representation remains continuous throughout the entire time window. Based on this, the real-time trajectory and the reference trajectory are respectively... The curvature scalar is differentiated using the baseline trajectory. The curvature scalar is represented by the second norm of the second-order discrete derivative of the high-dimensional trajectory coordinates with respect to time, and its calculation formulas are as follows: in, Indicates the real-time trajectory at the 1st The curvature scalar at each time step is used to reflect the bending intensity of the trajectory at that moment. Indicates the first The baseline trajectory is in the first The curvature scalar at each time step is used to reflect the bending strength of the corresponding reference mode at that moment. Indicates the real-time trajectory at the 1st High-dimensional coordinate vector at each time step Indicates the first The baseline trajectory is in the first High-dimensional coordinate vector at each time step This indicates the time step between adjacent sampling times. The L2 norm is used to convert high-dimensional second-order variations into scalar form. Through this calculation, the bending changes caused by insulation degradation, enhanced discharge, or thermal defect accumulation during trajectory evolution are extracted as directly comparable curvature features. Subsequently, the absolute deviation between the real-time trajectory curvature scalar and the baseline trajectory curvature scalar at the same time step is calculated to quantify the degree of deviation between the two at the trajectory acceleration level. The absolute curvature deviation can be expressed as: in, Indicates the first The baseline trajectory and the real-time trajectory are in the first... The absolute deviation of curvature at each time step reflects the difference in the degree of curvature between the two trajectories at that moment, denoted by [symbol missing]. This represents the absolute value operation. The smaller the deviation value, the closer the real-time trajectory is to the corresponding reference trajectory in terms of evolutionary acceleration; the larger the deviation value, the more significant the deviation between the current operating state and the corresponding reference mode in terms of deep dynamic characteristics. For early defects where anomalies only occur in trajectory curvature, this deviation can form a recognizable response before the position difference. Subsequently, the absolute curvature deviations corresponding to all reference trajectories are arranged according to the reference trajectory index and time index to form a curvature distance matrix. The rows of this matrix correspond to different reference trajectories, and the columns correspond to different time steps. The matrix elements are used to characterize the curvature difference of the real-time trajectory relative to each reference trajectory at each time step. Thus, the discrete curvature comparison results are organized into a unified temporal matrix structure and have a data interface consistent with the position distance matrix and tangent distance matrix, providing second-order dynamic feature input for subsequent construction of multi-dimensional geometric distance sets and grey relational degree calculation.

[0036] In step S34, the position distance matrix, tangent distance matrix, and curvature distance matrix are multidimensionally encapsulated and aggregated to obtain a geometric distance set. It should be noted that since the position distance matrix, tangent distance matrix, and curvature distance matrix reflect the differences in trajectory at three levels—spatial location, evolution direction, and evolution acceleration—each respectively, if these three are separately entered into subsequent correlation calculations, it is difficult to complete unified range extraction, resolution parameter injection, and multidimensional grey relational degree calculation under the same data structure. It is also not conducive to maintaining the correspondence of difference information between different fault modes at the same time step. Based on this, the technical solution of this application further performs multidimensional encapsulation and aggregation of the position distance matrix, tangent distance matrix, and curvature distance matrix to obtain a geometric distance set. Through the above processing, the multidimensional trajectory geometric differences can be organized into a unified structured input, thereby effectively providing a consistent data interface and joint analysis basis for subsequent adaptive multidimensional dynamic grey relational degree calculations.

[0037] More specifically, in a concrete example of this application, the position distance matrix, tangent distance matrix, and curvature distance matrix are first subjected to structural consistency verification to ensure a one-to-one correspondence between the three types of matrices in terms of baseline trajectory index, time step index, and data arrangement order. For data with inconsistent index orders, they are first rearranged according to a unified baseline trajectory numbering order and time series order, so that the positional differences, orientation differences, and curvature differences of the same baseline trajectory at the same time step can correspond to the same data coordinate position. After this processing, the three types of matrices can maintain a strict data alignment relationship during subsequent aggregation, avoiding misalignment of difference information at a certain time step or a certain baseline trajectory. Subsequently, multi-dimensional encapsulation aggregation processing is performed on the aligned three types of matrices, combining the position distance matrix, tangent distance matrix, and curvature distance matrix along a preset feature dimension, so that the three types of geometric differences corresponding to the same baseline trajectory and the same time step are encapsulated into a unified feature unit. This aggregation relationship can be expressed as: in, Represents the geometric distance set in the th order. The baseline trajectory, the first The feature vectors corresponding to each time step are used to uniformly characterize the positional difference, tangent direction angle difference, and absolute curvature deviation of the real-time trajectory relative to the corresponding reference trajectory at that moment. Represents the position distance matrix of the first position. The baseline trajectory is in the first Position difference at each time step Represents the tangent distance matrix of the th The baseline trajectory is in the first The difference in the tangent direction angle at each time step Represents the curvature distance matrix of the th The baseline trajectory is in the first The absolute curvature deviation at each time step. Thus, the three originally separate matrices are integrated into a joint geometric feature representation for the same diagnostic object.

[0038] Finally, all feature vectors are arranged in order according to the baseline trajectory dimension and the time dimension to form a geometric distance set, so that each baseline trajectory corresponds to a continuous multidimensional geometric difference sequence throughout the entire time window. This geometric distance set preserves the differences in spatial offset at the state modes such as transformer insulation aging, partial discharge, switchgear insulation defects, and abnormal temperature rise of cable joints, as well as the differences at the evolution direction and bending degree levels, and maintains consistency with the data organization form required for subsequent grey relational degree calculations. In this way, unified range extraction, correlation degree calculation, and comprehensive fusion of the three types of geometric differences can be directly performed in subsequent processing.

[0039] Specifically, in step S4, based on the adaptive resolution coefficient matrix, an adaptive multidimensional dynamic grey relational degree calculation is performed on the geometric distance set to obtain a comprehensive relational vector. The comprehensive relational vector includes positional relational degree, tangent relational degree, and curvature relational degree. It should be noted that, given that the positional differences, tangent direction differences, and curvature differences in the geometric distance set of power grid equipment will dynamically change with the operating conditions and fault evolution stages under load fluctuations, ambient temperature changes, humidity disturbances, and electromagnetic interference, the traditional grey relational calculation method using fixed resolution coefficients and a single global range is difficult to simultaneously take into account the dynamic resolution requirements of multidimensional features. Moreover, when the three heterogeneous geometric features of position, tangent, and curvature participate in the calculation together, the problem of strong dimensions dominating and weak dimensions suppressing their distinguishing ability is prone to occur. This makes it difficult to reliably identify weak anomalies in scenarios such as early thermal aging of transformer insulation, the initiation period of partial discharge between winding turns, micro-water leakage in switchgear, and local overheating of cable joints. Based on this, the technical solution of this application further performs adaptive multidimensional dynamic grey relational degree calculation on the geometric distance set based on the adaptive resolution coefficient matrix to obtain a comprehensive relational vector. This allows the resolution parameters obtained from the working condition perception to participate in the grey relational degree solution along with the position distance, tangent distance, and curvature distance. When necessary, it combines dimensional isolation range extraction and cross-dimensional balance correction to form a comprehensive relational result for each fault baseline trajectory. Through the above processing, the closeness of the real-time operating trajectory to different fault modes can be uniformly quantified, thereby effectively improving the adaptability of grey relational calculation under complex working conditions, the balance of inter-dimensional expression, and the ability to distinguish early degradation trends of equipment.

[0040] Figure 5 This is a flowchart illustrating an intelligent fault diagnosis and early warning method for power grid equipment based on grey relational analysis, according to an embodiment of this application. The method involves adaptively calculating multidimensional dynamic grey relational degree on a geometric distance set using an adaptive resolution coefficient matrix to obtain a comprehensive correlation vector. (See flowchart for example.) Figure 5 As shown, step S4 includes: S411, extracting global range parameters from the geometric distance set to obtain the minimum range and maximum range; S412, constructing a dynamic penalty term based on the adaptive resolution coefficient matrix and the maximum range, and substituting it, along with the minimum range and the geometric distance set, into the improved grey relational model to perform adaptive relational calculations to obtain the positional relational degree, tangent relational degree, and curvature relational degree; S413, performing weighted linear fusion of the positional relational degree, tangent relational degree, and curvature relational degree based on the preset geometric weight distribution coefficient to obtain a comprehensive relational vector.

[0041] In step S411, global range parameters are extracted from the geometric distance set to obtain the minimum and maximum ranges. It should be noted that since the positional differences, tangent direction differences, and curvature differences in the geometric distance set are discretely distributed across different reference trajectories and different time steps, if unified range boundary parameters are not extracted first, the resolution terms in subsequent grey relational analysis cannot establish a common reference scale. Based on this, the technical solution of this application further extracts global range parameters from the geometric distance set to obtain the minimum and maximum ranges. Through the above processing, unified upper and lower bounds of the difference can be provided for subsequent adaptive grey relational analysis, thereby effectively supporting the comparison of correlation between different fault modes.

[0042] More specifically, in a concrete example of this application, the position distance matrix, tangent distance matrix, and curvature distance matrix contained in the geometric distance set are first expanded sequentially, and a unified traversal sequence is established according to the baseline trajectory index and time step index, so that the three types of geometric differences participate in the extreme value search within the same retrieval range. This processing allows the geometric differences corresponding to different state modes such as transformer insulation aging, partial discharge, switchgear insulation defects, and abnormal temperature rise of cable joints to be processed in the same statistical space, so that the subsequent grey relational formula can use a unified range parameter. After completing the unified traversal, a global minimum value search and a global maximum value search are performed on all geometric difference elements to obtain the minimum range and the maximum range, respectively. The calculation relationship can be expressed as: in, This represents the minimum range, used to characterize the smallest boundary value among all difference elements within the geometric distance set. Represents the maximum range, used to characterize the largest boundary value among all difference elements within the geometric distance set. Indicates the first The baseline trajectory is in the first Position difference at each time step Indicates the first The baseline trajectory is in the first The difference in the tangent direction angle at each time step Indicates the first The baseline trajectory is in the first Absolute curvature deviation at each time step Indicates the baseline trajectory index. This represents the time step index. Thus, the differences originally scattered across the three types of matrices are summarized into unified upper and lower bound parameters. Subsequently, the obtained minimum and maximum ranges are propagated in correspondence with the original geometric distance set, enabling subsequent correlation calculations to handle positional features, tangent features, and curvature features under the same range reference. After this processing, all types of differences in the geometric distance set obtain a unified scale benchmark, allowing for the subsequent construction of a dynamic penalty term using the adaptive resolution coefficient matrix and completion of multidimensional grey correlation solution.

[0043] In step S412, a dynamic penalty term is constructed based on the adaptive resolution coefficient matrix and the maximum range. This term, along with the minimum range and the geometric distance set, is then substituted into the improved grey relational model for adaptive relational calculation to obtain the positional relational degree, tangent relational degree, and curvature relational degree. It should be noted that since the fluctuation ranges of the positional difference, the tangent direction angle difference, and the absolute curvature deviation are not consistent under different operating conditions, if a fixed resolution coefficient is still used for grey relational calculation, it will be difficult to simultaneously consider the dynamic resolution requirements under load fluctuations, ambient temperature rises, humidity changes, and voltage disturbances. Furthermore, it may weaken the subtle geometric differences corresponding to early faults such as transformer insulation thermal aging, partial discharge initiation, switchgear insulation deterioration, and local overheating of cable joints. Therefore, the technical solution of this application further constructs a dynamic penalty term based on the adaptive resolution coefficient matrix and the maximum range, and substitutes this term, along with the minimum range and the geometric distance set, into the improved grey relational model for adaptive relational calculation to obtain the positional relational degree, tangent relational degree, and curvature relational degree. Through the above processing, the impact of changes in operating conditions on the gray relational discrimination capability can be synchronously injected into the multidimensional geometric difference calculation process, thereby effectively improving the ability to distinguish the association of different fault modes under complex operating conditions.

[0044] More specifically, in a concrete example of this application, the adaptive resolution coefficient matrix, maximum range, minimum range, and geometric distance set are first read, and the various types of data are mapped according to the baseline trajectory index and time step index, so that the resolution coefficient, range parameter, and geometric difference remain consistent at the same time position. At this point, the position difference, tangent direction angle difference, and absolute curvature deviation in the geometric distance set are used as inputs for the three types of correlation calculations, respectively, and the corresponding elements in the adaptive resolution coefficient matrix are used as the resolution parameter inputs for the current time moment, thus establishing subsequent correlation calculations on a unified data coordinate system. Subsequently, a dynamic penalty term is constructed based on the adaptive resolution coefficient matrix and the maximum range, so that the resolution part in the grey relational formula no longer uses a fixed constant but changes with the working conditions. The dynamic penalty term can be expressed as: in, Indicates the first The dynamic penalty term at each time step is used to adjust the resolution of the grey relational formula at the current time. The adaptive resolution coefficient matrix is ​​represented in the th... The resolution coefficient element corresponding to each time step is extracted from the features of the preceding operating conditions and is used to reflect the resolution requirements under the current load, temperature, voltage, and humidity conditions. The maximum range in the geometric distance set is used as the scaling benchmark for the penalty term. This method increases the penalty term when the operating conditions are strongly disturbed and decreases when the operating conditions are relatively stable, making grey relational calculations adaptable to environmental changes. Based on this, the minimum range, the dynamic penalty term, and the three types of differences in the geometric distance set are substituted into the improved grey relational degree model to obtain the positional relational degree, tangent relational degree, and curvature relational degree. The corresponding calculation formulas are as follows: in, Indicates the first The baseline trajectory is in the first The positional correlation at each time step is used to characterize the degree of spatial proximity between the real-time trajectory and the corresponding reference trajectory. Indicates the first The baseline trajectory is in the first Tangential correlation at each time step is used to characterize the degree of closeness between two trajectories in terms of evolutionary direction. Indicates the first The baseline trajectory is in the first The curvature correlation at each time step is used to characterize the closeness of two trajectories at the level of evolutionary acceleration. This represents the minimum range within the geometric distance set, and is used as a lower bound reference for the grey relational formula. Indicates the first The baseline trajectory is in the first Position difference at each time step Indicates the first The baseline trajectory is in the first The difference in the tangent direction angle at each time step Indicates the first The baseline trajectory is in the first The absolute curvature deviation at each time step. Finally, the positional correlation, tangent correlation, and curvature correlation corresponding to all time steps and all reference trajectories are arranged in the original index order to form three types of correlation result matrices, which serve as the direct input for subsequent weighted linear fusion, providing a unified, continuous, and comparable correlation basis for the construction of subsequent comprehensive correlation vectors.

[0045] In step S413, based on preset geometric weight distribution coefficients, the position correlation, tangent correlation, and curvature correlation are weighted and linearly fused to obtain a comprehensive correlation vector. It should be noted that since position correlation, tangent correlation, and curvature correlation reflect the proximity of trajectories in spatial location, evolution direction, and evolution acceleration, respectively, using any single correlation coefficient alone is insufficient to fully characterize the overall proximity relationship between the current state of the power grid equipment and each benchmark fault mode. Therefore, the technical solution of this application further uses preset geometric weight distribution coefficients to weightedly and linearly fuse the position correlation, tangent correlation, and curvature correlation to obtain a comprehensive correlation vector. Through the above processing, the multi-dimensional correlation results can be uniformly mapped to a comprehensive representation for each benchmark trajectory, thereby effectively supporting subsequent fault diagnosis and trend warning determination.

[0046] More specifically, in a concrete example of this application, the positional correlation, tangent correlation, and curvature correlation are first read and aligned according to the same baseline trajectory index and time step index, ensuring that the three types of correlation results remain corresponding at the same time and under the same baseline trajectory. Subsequently, the three types of correlation are linearly weighted and summed according to a preset geometric weight distribution coefficient to obtain the fusion correlation value of each baseline trajectory at each time step, calculated as follows: in, Indicates the first The baseline trajectory is in the first The fusion correlation value at each time step Indicates the degree of positional relevance. Indicates the degree of tangent correlation. Indicates curvature correlation. , and These represent the weight coefficients corresponding to position, tangent, and curvature, respectively, and the sum of the weights is 1. This processing allows position offset, direction change, and curvature change to participate in the state evaluation in a predetermined proportion. Subsequently, the fused correlation values ​​at each time step are aggregated along the time dimension to obtain the comprehensive scalar correlation results for each reference trajectory, and arranged in the order of the reference trajectories to form a comprehensive correlation vector. Its calculation formula is: in, Represents the comprehensive correlation vector. to These represent the comprehensive correlation results relative to each baseline trajectory. This represents the total number of reference trajectories. Thus, the three types of related information—position, tangent, and curvature—are uniformly transformed into vector expressions that can be directly used for diagnostic judgment, providing a basis for subsequent identification of conditions such as transformer insulation aging, partial discharge, switchgear insulation defects, and abnormal temperature rise in cable joints.

[0047] In Example 1 above, the global range parameter is extracted from the geometric distance set and combined with the adaptive resolution coefficient matrix to complete the multidimensional dynamic grey relational degree calculation, which can already realize fault diagnosis and trend early warning of the operating status of power grid equipment. Further considering that the position difference, tangent direction angle difference, and absolute curvature deviation differ on a numerical scale, directly using a unified global range for relational degree calculation may affect the distinguishing ability of some geometric features. Therefore, Example 2 is proposed. While maintaining the overall technical process and upstream and downstream data interfaces unchanged, the range extraction and relational degree calculation method in step four is further improved. By introducing dimensional isolation range extraction, cross-dimensional balance correction, and balanced multidimensional dynamic relational degree calculation, the coordination and comparability of multidimensional geometric features in comprehensive diagnosis are improved.

[0048] Figure 6 This document presents a flowchart illustrating the multidimensional dynamic grey relational degree calculation for a power grid equipment fault intelligent diagnosis and early warning method based on grey relational analysis, according to an embodiment of this application. The calculation involves dimensional isolation range extraction and comparative perception cross-dimensional equilibrium factor calculation to balance and correct the geometric distance set, thereby obtaining a comprehensive correlation vector. (See flowchart for example.) Figure 6 As shown, step S4 includes: S421, extracting the dimensional range of the geometric distance set to obtain the dimensional range matrix; S422, calculating the cross-dimensional equilibrium factor of the dimensional range matrix to obtain the equilibrium range matrix; S423, calculating the multi-dimensional dynamic correlation degree of the equilibrium range matrix, the adaptive resolution coefficient matrix, and the geometric distance set to obtain the positional correlation degree, tangent correlation degree, and curvature correlation degree; S424, performing weighted linear fusion of the positional correlation degree, tangent correlation degree, and curvature correlation degree based on the preset geometric weight distribution coefficient to obtain the comprehensive correlation vector.

[0049] In step S421, the geometric distance set is subjected to dimension-isolated range extraction to obtain a dimensional range matrix. It should be noted that, given that the first embodiment mixes the position distance matrix, tangent distance matrix, and curvature distance matrix before uniformly searching for extrema, scale-dominated bias can easily occur, leading to the position dimension dominating the maximum range value, while the distinguishing ability of weaker features such as curvature is compressed. This, in turn, affects the early anomaly identification in scenarios such as the initial stage of transformer insulation thermal aging, the initiation stage of inter-turn partial discharge in windings, GIS micro-water leakage, and local overheating of cable joints. Therefore, the technical solution of this application further performs dimension-isolated range extraction on the geometric distance set to obtain a dimensional range matrix. Through the above processing, the position, tangent, and curvature dimensions can each establish their own range boundaries, thereby effectively avoiding scale suppression of weaker dimensions caused by cross-dimensional mixing for extrema.

[0050] More specifically, in a concrete example of this application, the geometric distance set is first dimensionally split, and the positional distance data, tangent distance data, and curvature distance data are arranged independently according to the baseline trajectory index and time step index, forming three isolated statistical spaces. After completing the dimensional split, the entire baseline trajectory and all time steps are traversed within the positional dimension, tangent dimension, and curvature dimension, and the minimum absolute difference and maximum absolute difference within each dimension are extracted independently. For the positional dimension, the isolation range is calculated as follows: in, Represents the geometric distance set of the first The set of trajectories in the first The absolute difference of position at each time step This represents the minimum range within the location dimension. This represents the maximum range within the location dimension. Indicates the baseline trajectory index. This represents the time step index. For the tangent dimension, the isolation range is calculated as follows: in, Represents the geometric distance set of the first The set of trajectories in the first The difference in the tangent direction angle at each time step This represents the minimum range within the tangent dimension. This represents the maximum range within the tangent dimension. For the curvature dimension, the isolation range is calculated as follows: in, Represents the geometric distance set of the first The set of trajectories in the first The absolute deviation of the curvature scalar at each time step This represents the minimum range within the curvature dimension. This represents the maximum range within the curvature dimension. The resulting six scalar parameters are encapsulated in dimensional order to form a dimensional range matrix, which is used for subsequent cross-dimensional equilibrium correction and multidimensional dynamic correlation calculation.

[0051] Subsequently, the ranges of the position dimension, tangent dimension, and curvature dimension are organized according to a unified data structure, ensuring that each dimension retains its true statistical distribution information. The six scalar parameters are then encapsulated into a structured dimensional range matrix for output. After dimensional isolation, the range parameters of the curvature dimension and the position dimension exist independently, fundamentally severing the scale suppression channel exerted on the curvature dimension by the position dimension through global extremum search. The range of each dimension faithfully reflects only its own statistical distribution characteristics.

[0052] For example, when the thermal aging of transformer insulation paper is in its early stages, the trajectory position deviation is not yet obvious, but the curvature change has already shown early anomalies. If the global hybrid range search method in the first embodiment is continued, a large number of range values ​​in the position dimension will dominate the maximum range, causing the range parameter corresponding to the curvature dimension to be compressed. The subsequent curvature correlation will be concentrated in a close numerical range, making it difficult to differentiate between different fault modes. After adopting this step, the curvature difference value only searches for the minimum and maximum ranges within the curvature dimension. The position dimension, tangent dimension, and curvature dimension each retain independent scale boundaries, thereby enabling the curvature dimension to maintain its due ability to identify trajectory bending anomalies caused by early thermal aging, weak discharge, or micro-water leakage, and providing accurate input for subsequent balanced range correction.

[0053] In step S422, a comparative perception cross-dimensional equilibrium factor calculation is performed on the dimensional range matrix to obtain an equilibrium range matrix. It should be noted that, given that after the dimensional isolation range extraction is completed, although the position dimension, tangent dimension, and curvature dimension have established their respective range boundaries, there are still orders of magnitude differences between them. If the significantly different dimension-specific ranges are directly substituted into the grey relational formula, the correlation degree values ​​output by each dimension will be unbalanced. The curvature dimension may still be concentrated in a narrow range close to 1.0, thus affecting the early anomaly identification in scenarios such as early thermal aging of transformer insulation paper, partial discharge initiation between winding turns, micro-water leakage in GIS, and local overheating of cable joints. Based on this, the technical solution of this application further performs a comparative perception cross-dimensional equilibrium factor calculation on the dimensional range matrix to obtain an equilibrium range matrix. Through the above processing, the cross-dimensional scale imbalance can be further reduced while maintaining the independence of each dimension, thereby effectively improving the scale fairness and distinguishing ability of subsequent multi-dimensional dynamic grey relational degree calculations.

[0054] More specifically, in a concrete example of this application, the minimum and maximum ranges of the position dimension, the tangent dimension, and the curvature dimension in the dimensional range matrix are first read, and the contrast of the distance distribution within each dimension is calculated to characterize the strength of the original distinguishing ability of that dimension. The contrast calculation formula is: in, Indicates the first The contrast of each geometric feature dimension reflects the breadth or narrowness of the dynamic range of distance values ​​within that dimension. Indicates the first The maximum range of each dimension Indicates the first The minimum range in each dimension, This represents the regularization constant, used to prevent the denominator from degenerating to zero. Represents a dimension index, where Indicates the location dimension. Indicates the tangent dimension. This represents the curvature dimension. This calculation transforms the dispersion of each dimension into a comparable scalar, providing a basis for subsequent equilibrium adjustments. Subsequently, an arithmetic mean is calculated on the contrast of the three dimensions to establish a cross-dimensional mean anchor point, serving as a neutral reference benchmark for subsequent equilibrium adjustments. The calculation formula is as follows: in, Anchor points representing the contrast mean of three-dimensional geometric features. Indicates contrast in the position dimension. Indicates the contrast of the tangent dimension. This represents the contrast of the curvature dimension. Using this mean anchor point, the deviation of each dimension from the overall average discriminative power can be determined, thus establishing a unified reference for gaining weaker dimensions and suppressing stronger dimensions. After obtaining the mean anchor point, the deviation of each dimension from the mean anchor point is mapped to a bounded equilibrium factor. The equilibrium factor is calculated as follows: in, Indicates the first A balance factor in each dimension, with values ​​between 0 and 1. This represents the steepness control parameter of the saturation function, used to adjust the sensitivity of the equalization correction. Indicates the cross-dimensional mean anchor point. This indicates the contrast of the current dimension. Represents the regularization constant. This represents the base of the natural logarithm. When the contrast of the current dimension is lower than the mean anchor point, the equalization factor increases to enhance the effectiveness of subsequent ranges in that dimension; when the contrast of the current dimension is higher than the mean anchor point, the equalization factor decreases to weaken the dominant role of that dimension in the overall correlation calculation. A saturation function is used for mapping to keep the equalization correction bounded, avoiding both the continuous compression of weak dimensions and excessive amplification of underlying noise. After obtaining the equalization factor, the ranges of each dimension in the dimensional range matrix are adjusted by dimension-wise scaling to form the equalization range matrix. Its adjustment form is: in, Indicates the first after equilibrium correction Minimum range in each dimension Indicates the first after equilibrium correction Maximum range in each dimension This represents the equilibrium factor for the current dimension. and These represent the dimensional isolation ranges before correction. This unifies the ranges of the position, tangent, and curvature dimensions to a scale range more suitable for joint correlation calculations, and encapsulates them in dimensional order into a balanced range matrix for direct use in subsequent balanced corrections of multidimensional dynamic correlation calculations. For example, when the contrast corresponding to the curvature dimension is lower than that of the position dimension, the curvature dimension receives a relatively larger balance factor, while the position dimension receives a relatively smaller balance factor. This results in a moderate amplification of the curvature range and a moderate compression of the position range, thus providing a more balanced parameter basis for subsequent multidimensional correlation calculations.

[0055] In step S423, a multidimensional dynamic correlation degree calculation is performed on the balanced range matrix, adaptive resolution coefficient matrix, and geometric distance set to obtain the position correlation degree, tangent correlation degree, and curvature correlation degree. It should be noted that, given that the three dimensions share the same set of global range parameters in the first embodiment, the position dimension tends to dominate the penalty term calculation, while the correlation degree of weaker features such as the curvature dimension is compressed, affecting the identification of anomalies such as early thermal aging, weak discharge, and micro-water leakage. Based on this, the technical solution of this application further performs a multidimensional dynamic correlation degree calculation on the balanced range matrix, adaptive resolution coefficient matrix, and geometric distance set to obtain the position correlation degree, tangent correlation degree, and curvature correlation degree. Through the above processing, the scale coordination of the correlation calculation of each dimension can be improved while maintaining the adaptive capability of the operating conditions.

[0056] More specifically, in a concrete example of this application, the equilibrium range matrix, adaptive resolution coefficient matrix, and geometric distance set are first aligned according to the baseline trajectory index and time step index, so that the position difference, tangent difference, and curvature difference are consistent with the corrected range of the corresponding dimension and the adaptive resolution coefficient at the current time, respectively. Then, the corrected range, adaptive resolution coefficient, and original difference of each dimension are substituted into the corresponding grey relational degree formula to obtain the positional relational degree, tangent relational degree, and curvature relational degree. The positional relational degree is calculated as follows: Tangent correlation is calculated as follows: The curvature correlation degree is calculated as follows: in, , and They represent the first The set of trajectories in the first Positional correlation, tangent correlation, and curvature correlation at each time step , This represents the corrected range for the position dimension. , This represents the corrected range in the tangent dimension. , The corrected range represents the curvature dimension. Indicates the first Adaptive resolution coefficients at each time step , and These represent the position difference, the difference in the tangent direction angle, and the absolute deviation of curvature, respectively. The three types of correlations obtained are then output in their original index order for subsequent weighted fusion. The penalty term in the curvature formula is on the same order of magnitude as the original curvature difference, allowing the correlation to expand across the entire numerical range, effectively amplifying the differences in curvature correlations between different fault modes.

[0057] In a specific scenario, when the device is in the early thermal aging stage, the position difference changes little while the curvature difference becomes abnormal first. After adopting this step, the curvature dimension calls the curvature-specific equilibrium correction range to participate in the solution, thereby preserving the response to early anomalies, while the position dimension and tangent dimension still maintain their ability to represent state offset and trend changes, respectively.

[0058] In step S424, based on preset geometric weight distribution coefficients, the positional correlation, tangent correlation, and curvature correlation are weighted and linearly fused to obtain a comprehensive correlation vector. It should be noted that since positional correlation, tangent correlation, and curvature correlation reflect the proximity of trajectories in spatial location, evolution direction, and evolution acceleration, respectively, without unified fusion, it is difficult to form an overall judgment result for each fault baseline. Furthermore, the enhanced second embodiment has already resolved the upstream scale imbalance problem; if a reasonable fusion method is still lacking, the early degradation information carried by the curvature dimension may still be difficult to stably enter the final diagnostic result. Therefore, the technical solution of this application further uses preset geometric weight distribution coefficients to weightedly and linearly fuse the positional correlation, tangent correlation, and curvature correlation to obtain a comprehensive correlation vector. Through the above processing, the multi-dimensional correlation results can be uniformly compressed into a comprehensive representation for each baseline trajectory, thereby effectively improving the predictability of subsequent fault diagnosis and trend warning.

[0059] More specifically, in a specific example of this application, the positional correlation, tangent correlation, and curvature correlation obtained after equalization correction are first aligned according to the same baseline trajectory index and time step index, and a preset geometric weight distribution coefficient is read so that the three types of correlation participate in the fusion calculation at the same time position. After completing the equalization and independent solution of the three-dimensional correlation, the positional correlation, tangent correlation, and curvature correlation at the same time step are linearly weighted and summed based on the preset geometric weight distribution coefficient to obtain the mixed correlation coefficient of the trajectory at a single time sampling point. Among them, the position dimension is used to characterize the state shift, the tangent dimension is used to characterize the evolution trend change, and the curvature dimension is used to characterize the evolution acceleration anomaly. The three dimensions participate in the state evaluation together according to a preset ratio to avoid a single dimension directly determining the final result. Subsequently, the three-dimensional correlation of each baseline trajectory at each time step is weighted and linearly fused, and mean dimensionality reduction is performed along the sliding time window direction to obtain the comprehensive scalar correlation value for each baseline trajectory, the calculation formula of which is: in, Indicates the first The comprehensive scalar correlation value corresponding to the fault benchmark. This represents the total number of time steps within the sliding time window. , and Let represent the geometric weight coefficients for the position dimension, tangent dimension, and curvature dimension, respectively, and satisfy . , Indicates the first The baseline trajectory is in the first Location correlation at each time step Indicates the first The baseline trajectory is in the first Tangent correlation at each time step Indicates the first The baseline trajectory is in the first Curvature correlation at each time step Indicates the time step index. This represents the baseline trajectory index. This formula allows for the further aggregation of correlation results at a single time step into stable evaluation values ​​for the entire time window. Finally, the comprehensive scalar correlation values ​​corresponding to all baseline trajectories are ordered and encapsulated into a comprehensive correlation vector, expressed as: in, This represents the final output composite correlation vector. to These represent the composite scalar correlation values ​​corresponding to each baseline trajectory. Represents the total number of categories of the baseline trajectory, symbol This indicates transposition. Finally, all benchmark scalar correlation values ​​are sequentially assembled and encapsulated into a comprehensive correlation vector. After this processing, multidimensional geometric features are uniformly transformed into vector expressions that can be directly used for fault diagnosis and judgment, and maintain a data interface consistent with downstream static threshold comparison and trend early warning analysis.

[0060] For example, when a transformer is in the early thermal aging stage of its insulation paper, the positional offset is not yet prominent, but the curvature correlation has already shown an abnormal response. In the second embodiment, since the upstream curvature dimension has been expanded within its effective value range after equalization correction, the early warning information carried by the curvature dimension can enter the comprehensive correlation vector according to the preset weights during the weighted fusion process in this step, without being unilaterally masked by the position dimension. Because the correlations of the three upstream dimensions are fully expanded within their respective complete value ranges after equalization correction, the weighted fusion here is no longer unilaterally dominated by the high discriminative power of the position dimension, and the early deterioration warning information carried by the curvature dimension can contribute to the comprehensive correlation vector in a reasonable proportion.

[0061] Specifically, in step S5, fault diagnosis decisions and trend warning triggers are performed on the comprehensive correlation vector to obtain diagnostic warning results. It should be noted that since the comprehensive correlation vector only represents the degree of proximity between the current operating state and each fault baseline trajectory, without further combining static threshold judgment and time-series trend analysis, it is difficult to distinguish between established deterministic faults and early degradation processes still in the evolution stage, thus affecting the timing of maintenance response in scenarios such as transformer insulation thermal aging, partial discharge, GIS insulation defects, and cable joint overheating. Based on this, the technical solution of this application further performs fault diagnosis decisions and trend warning triggers on the comprehensive correlation vector to obtain diagnostic warning results. Through the above processing, the current fault state, degradation trend intensity, and corresponding warning level can be given simultaneously, thereby effectively supporting on-site inspection and preventive maintenance.

[0062] More specifically, in a concrete example of this application, a static diagnostic judgment is first performed on the comprehensive correlation vector, then a trend fitting is performed on its historical change process, and finally, the static diagnostic results and trend analysis results are jointly used for decision-making to form a diagnostic warning result that includes fault type, severity level, and operation and maintenance guidance. First, the correlation confidence level corresponding to each fault type in the comprehensive correlation vector is compared with a preset static threshold value one by one, and the maximum fault correlation value is used as the static judgment basis at the current moment. The static diagnostic state can be represented as: in, This indicates the static diagnostic status. A value of 1 indicates that a deterministic fault has been established at the current moment, while a value of 0 indicates that a deterministic fault has not been established at the current moment. Represents the first in the comprehensive correlation vector The comprehensive correlation value corresponding to the fault class benchmark, Represents the fault baseline set, This represents the static threshold. This process allows for the early identification of fault states that have reached alarm conditions. After completing the static determination, the comprehensive correlation vector at each time point is sequentially pushed into a fixed-length first-in-first-out sliding time queue, and linear regression fitting is performed on the historical correlation degree sequence within the queue to extract the correlation degree change rate. For the ... The rate of change of a fault-like baseline can be expressed as: in, Indicates the first Rate of change in correlation of fault-like benchmarks Indicates the length of the sliding time queue. This represents the time index in the queue. Indicates the first The corresponding time of the first moment The comprehensive correlation value for fault types. This rate of change characterizes the growth rate of correlation over time; a continuous increase indicates that the corresponding equipment is in a deterioration process. After obtaining the static diagnostic status and the correlation rate of change, the two are combined for conditional screening and hierarchical fusion judgment. When the static diagnostic status indicates a deterministic fault, the corresponding fault type and severity alarm level are directly output; when the static diagnostic status does not reach the fault threshold but the correlation rate of change exceeds the trend threshold, a trend warning result is output to indicate that the equipment is in an early deterioration stage; when neither of the two conditions is met, a normal operation result is output. This combines the current status judgment with the future development trend, avoiding missed alarms or delayed warnings caused by relying solely on the correlation value at a single moment.

[0063] As described above, the intelligent diagnosis and early warning method for power grid equipment faults based on grey relational analysis according to the embodiments of this application can be implemented in various power grid equipment status monitoring and fault early warning scenarios, such as transformer online monitoring scenarios, GIS switchgear insulation defect monitoring scenarios, or high-voltage cable joint overheating early warning scenarios. In one possible implementation, this method can be integrated into a substation integrated online monitoring platform, power grid equipment operation and maintenance management platform, or dispatch master station intelligent analysis platform as a fault diagnosis and dynamic early warning engine based on multi-dimensional time-series grey relational analysis. For example, this method can be an independent diagnostic and early warning application running on a station-side edge computing gateway, or it can be a condition perception and adaptive relational calculation processing component in an existing equipment status monitoring device, or it can be an intelligent diagnostic analysis middleware service deployed on the master station side and receiving station-side monitoring data through a communication link; of course, the phase space trajectory reconstruction, adaptive resolution coefficient generation, multi-dimensional geometric distance extraction, and grey relational degree fusion judgment processing in this method can also run in online monitoring terminals, edge analysis units, or station control layer processing devices deployed at the substation site, serving as the underlying real-time analysis execution basis of the intelligent diagnosis and trend early warning system for power grid equipment faults.

[0064] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for intelligent diagnosis and early warning of power grid equipment faults based on grey relational analysis, characterized in that, include: Step 1: Reconstruct the phase space of the acquired power grid equipment status monitoring data to obtain the trajectory tensor, and normalize and encode the acquired power grid equipment operating condition data to obtain the operating condition tensor. Step 2: Based on the pre-trained meta-learning network, dynamic feature extraction and channel attention weighting are performed on the working condition tensor to obtain the adaptive resolution coefficient matrix. Step 3: Perform baseline matching and multidimensional trajectory geometric feature extraction on the trajectory tensor and each baseline trajectory in the preset baseline trajectory set to obtain the geometric distance set. The geometric distance set includes the Euclidean space position difference, the difference of the angle between the first derivative tangent directions, and the sum of the two values. Step 4: Based on the adaptive resolution coefficient matrix, perform adaptive multidimensional dynamic grey relational degree calculation on the geometric distance set to obtain the comprehensive relational vector, which includes positional relational degree, tangent relational degree and curvature relational degree. Step 5: Perform fault diagnosis decision-making and trend warning triggering on the comprehensive correlation vector to obtain diagnosis and warning results.

2. The intelligent diagnosis and early warning method for power grid equipment faults based on grey relational analysis according to claim 1, characterized in that, Step one includes: Step 1.1: Perform data synchronization alignment and adaptive filtering on the condition monitoring data and operating condition data to obtain synchronized fused data; Step 1.2: Based on principal component analysis, modal decoupling and feature separation are performed on the synchronously fused data to obtain the equipment state sequence characterizing the degradation of the equipment body and the external operating condition sequence characterizing the environmental load. Step 1.3: Perform phase space reconstruction mapping on the device state sequence to obtain the trajectory tensor; Step 1.4: Standardize and encode the external operating condition sequence to obtain the operating condition tensor.

3. The intelligent diagnosis and early warning method for power grid equipment faults based on grey relational analysis according to claim 1, characterized in that, Step two includes: Step 2.1: Based on the pre-trained meta-learning network, local abrupt changes in the working condition tensor are captured to obtain the environmental feature map. Step 2.2: Perform channel attention weighted evaluation on the environmental feature map to obtain weighted environmental features; Step 2.3: Perform nonlinear mapping and matrix reshaping on the weighted environmental features to obtain the adaptive resolution coefficient matrix.

4. The intelligent diagnosis and early warning method for power grid equipment faults based on grey relational analysis according to claim 1, characterized in that, Step three includes: Step 3.1: Pair the trajectory tensor with each reference trajectory in the reference trajectory set one by one in the time dimension to obtain trajectory pairs, and calculate the absolute position difference of each pair of trajectory pairs in the high-dimensional Euclidean space to obtain the position distance matrix. Step 3.2: Differentiate the tangent vectors along the time axis for the real-time trajectory and the reference trajectory in the trajectory pair, and calculate the angle difference between the tangent directions at the same time step to obtain the tangent distance matrix. Step 3.3: Perform curvature scalar differentiation on the real-time trajectory and the reference trajectory in the trajectory pair respectively, and calculate the absolute curvature difference at the same time step to obtain the curvature distance matrix; Step 3.4: Perform multidimensional encapsulation aggregation on the position distance matrix, tangent distance matrix, and curvature distance matrix to obtain the geometric distance set.

5. The intelligent diagnosis and early warning method for power grid equipment faults based on grey relational analysis according to claim 1, characterized in that, Step four includes: Step 4.1: Extract the global range parameters from the geometric distance set to obtain the minimum and maximum ranges; Step 4.2: Construct a dynamic penalty term based on the adaptive resolution coefficient matrix and the maximum range, and substitute it together with the minimum range and the geometric distance set into the improved grey relational degree model to perform adaptive relational solution to obtain the positional relational degree, tangent relational degree and curvature relational degree. Step 4.3: Based on the preset geometric weight distribution coefficients, perform weighted linear fusion of positional correlation, tangent correlation, and curvature correlation to obtain a comprehensive correlation vector.

6. The intelligent diagnosis and early warning method for power grid equipment faults based on grey relational analysis according to claim 1, characterized in that, Step five includes: Step 5.1: Using the indicator function structure, the association confidence of each fault type in the comprehensive association vector is compared with the preset static threshold value one by one to determine whether a deterministic fault is constituted at the current moment and to generate a static diagnostic state. Step 5.2: Push the comprehensive correlation vector into a fixed-length first-in-first-out sliding time queue, and use the least squares method to perform linear regression fitting on the historical correlation degree sequence in the queue and extract the slope parameter to obtain the correlation degree change rate. Step 5.3: Through the decision logic tree, the static diagnostic status and correlation change rate are jointly screened and graded to obtain diagnostic warning results that include fault type, severity level and operation and maintenance guidance.

7. The intelligent diagnosis and early warning method for power grid equipment faults based on grey relational analysis according to claim 4, characterized in that, Step four includes: Dimensional isolation range extraction is performed on the geometric distance set to obtain the dimensional range matrix; Comparative perception of the dimensional range matrix is ​​performed to calculate the cross-dimensional equilibrium factor in order to obtain the equilibrium range matrix; Multidimensional dynamic correlation degree solution is obtained by performing equilibrium correction on the equilibrium range matrix, adaptive resolution coefficient matrix and geometric distance set to obtain positional correlation degree, tangent correlation degree and curvature correlation degree; Based on the preset geometric weight distribution coefficients, the positional correlation, tangent correlation and curvature correlation are weighted and linearly fused to obtain a comprehensive correlation vector.

Images