A method and system for monitoring the operating state of a power module

Abstract

The application relates to the technical field of power monitoring, and discloses a power module operation state monitoring method and system, the method comprising the following steps: performing multi-frequency calibration on a high-frequency electrical signal of a power module to obtain a standardized multi-dimensional electrical signal; decomposing the standardized multi-dimensional electrical signal into a limited bandwidth modal component and performing Hilbert transformation to obtain a high-dimensional state characteristic vector; performing local linear embedding dimension reduction on the high-dimensional state characteristic vector to obtain a low-dimensional reference coordinate; obtaining a low-dimensional monitoring coordinate of a current high-dimensional state characteristic vector; constructing a decision graph of the low-dimensional reference coordinate, and performing cluster center spacing analysis on boundary points in the decision graph to obtain a dynamic adaptive threshold; performing state deviation degree analysis on the power module, and performing difference comparison between a state anomaly index after the analysis and the dynamic adaptive threshold to generate a state anomaly early warning signal of the power module; and the application can improve the efficiency of power module operation state monitoring.

Description

Technical Field

[0001] This invention relates to the field of power monitoring technology, and in particular to a method and system for monitoring the operating status of power modules. Background Technology

[0002] In existing power module operation status monitoring processes, the processing of high-frequency electrical signals lacks a standardized multi-frequency calibration procedure. The signal decomposition and feature extraction processes fail to fully extract the state information from the electrical signals, making it difficult to accurately capture subtle state changes during power module operation. The extracted feature vectors cannot comprehensively reflect the actual operating status of the module. Furthermore, the processing methods for high-dimensional features lack rationality, easily leading to redundancy or loss of feature information, affecting the effectiveness of subsequent state analysis.

[0003] The existing threshold setting method for power module monitoring lacks dynamic adjustment capability and cannot adaptively update according to the real-time operating conditions of the module. The analysis of the degree of deviation in the state only considers local or global deviation data, resulting in insufficient accuracy in judging the state anomaly index and low response efficiency of anomaly warning. It is difficult to predict power module failures in advance. Therefore, how to improve the accuracy and intelligence level of power module operating status monitoring has become an urgent problem to be solved. Summary of the Invention

[0004] This invention provides a method and system for monitoring the operating status of power modules to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a method for monitoring the operating status of a power module, comprising:

[0006] S1. Perform multi-frequency calibration on the high-frequency electrical signal of the power module to obtain a standardized multi-dimensional electrical signal of the high-frequency electrical signal;

[0007] S2. Perform variational mode decomposition on the standardized multidimensional electrical signal, and perform Hilbert transform on the decomposed finite bandwidth mode components to obtain the high-dimensional state feature vector of the power module.

[0008] S3. Perform local linear embedding dimensionality reduction on the high-dimensional state feature vector to obtain the low-dimensional reference coordinates of the high-dimensional state feature vector;

[0009] S4. Map the current high-dimensional state feature vector of the power module to the low-dimensional intrinsic space of the low-dimensional reference coordinates to obtain the low-dimensional monitoring coordinates of the current high-dimensional state feature vector.

[0010] S5. Based on the local density and relative distance of the low-dimensional reference coordinates, construct a decision graph of the low-dimensional reference coordinates, and perform clustering center spacing analysis on the boundary points in the decision graph to obtain the dynamic adaptive threshold of the power module.

[0011] S6. Based on the low-dimensional reference coordinates and the low-dimensional monitoring coordinates, perform a state deviation analysis on the power module, and compare the analyzed state anomaly index with the dynamic adaptive threshold to generate a state anomaly warning signal for the power module.

[0012] In a preferred embodiment, the step of performing multi-frequency calibration on the high-frequency electrical signal of the power module to obtain a standardized multi-dimensional electrical signal of the high-frequency electrical signal includes:

[0013] The high-frequency electrical signal of the power module is bandpass filtered to obtain the fundamental signal component and harmonic signal component of the high-frequency electrical signal;

[0014] The fundamental signal component and the harmonic signal component are subjected to amplitude normalization processing, and the normalized sub-signals are time-stamp aligned to obtain the multi-band signal of the high-frequency electrical signal.

[0015] According to the preset channel sequence, the multi-band signal is recombined into a multi-dimensional data to obtain a standardized multi-dimensional electrical signal of the high-frequency electrical signal.

[0016] In a preferred embodiment, the step of performing variational mode decomposition on the standardized multidimensional electrical signal and applying Hilbert transform to the decomposed finite-bandwidth mode components to obtain the high-dimensional state feature vector of the power module includes:

[0017] Based on the preset number of modal decompositions, the standardized multidimensional electrical signal is subjected to frequency domain iterative decomposition to obtain the initial modal components of the standardized multidimensional electrical signal.

[0018] The initial modal components are updated by convergence determination to obtain the finite bandwidth intrinsic mode functions of the standardized multidimensional electrical signal;

[0019] The instantaneous amplitude and instantaneous frequency of the finite bandwidth intrinsic mode function are analyzed, and the time-frequency integral mapping of the finite bandwidth intrinsic mode function is performed to obtain the Hilbert marginal spectrum of the finite bandwidth intrinsic mode function;

[0020] The energy distribution of the Hilbert marginal spectrum is analyzed to obtain the energy entropy and spectral centroid of the Hilbert marginal spectrum;

[0021] Multidimensional feature fusion is performed on the energy entropy and the spectral centroid to obtain the high-dimensional state feature vector of the power module.

[0022] In a preferred embodiment, the step of performing local linear embedding dimensionality reduction on the high-dimensional state feature vector to obtain the low-dimensional reference coordinates of the high-dimensional state feature vector includes:

[0023] Obtain the normal high-dimensional state feature vector within the historical normal operation period from the high-dimensional state feature vector;

[0024] Topological association analysis is performed on the normal high-dimensional state feature vectors to obtain the neighborhood relationships between the normal high-dimensional state feature vectors;

[0025] Based on the neighborhood relationship, the normal high-dimensional state feature vector is locally linearly reconstructed to obtain the weight coefficient matrix of the normal high-dimensional state feature vector;

[0026] Based on the weight coefficient matrix, the normal high-dimensional state feature vector is embedded and mapped to obtain the low-dimensional reference coordinates of the normal high-dimensional state feature vector.

[0027] In a preferred embodiment, the step of embedding and mapping the normal high-dimensional state feature vector based on the weight coefficient matrix to obtain the low-dimensional reference coordinates of the normal high-dimensional state feature vector includes:

[0028] The weight coefficient matrix is ​​constrained and encoded to obtain the cost matrix of the normal high-dimensional state feature vector;

[0029] Perform generalized eigenvalue decomposition on the cost matrix to obtain the eigenvectors of the normal high-dimensional state eigenvectors;

[0030] A basis system is constructed on the feature vectors to obtain the embedding coordinate matrix of the feature vectors;

[0031] Based on the embedded coordinate matrix, the low-dimensional reference coordinates of the normal high-dimensional state feature vector are confirmed.

[0032] In a preferred embodiment, constructing a decision graph of the low-dimensional reference coordinates based on the local density and relative distance of the low-dimensional reference coordinates includes:

[0033] The local density of the low-dimensional reference coordinates is obtained by performing a distribution density measurement on the low-dimensional reference coordinates.

[0034] The relative distance between the low-dimensional reference coordinates is obtained by performing spacing parameter analysis on the low-dimensional reference coordinates.

[0035] Using the local density as the horizontal axis parameter and the relative distance as the vertical axis parameter, the local density value and relative distance value of the low-dimensional reference coordinates are mapped to a two-dimensional coordinate system to obtain the initial decision map of the low-dimensional reference coordinates.

[0036] The initial decision graph is interpreted using graph structure interpretation to obtain the decision graph with the low-dimensional reference coordinates.

[0037] In a preferred embodiment, the step of performing cluster center spacing analysis on the boundary points in the decision graph to obtain the dynamic adaptive threshold of the power module includes:

[0038] Based on the distribution characteristics of the decision graph, the boundary points and cluster centers in the decision graph are selected.

[0039] Perform nearest neighbor analysis on the reference coordinates of the boundary points and the reference coordinates of the cluster centers to determine the coordinates of the cluster centers to which the boundary points belong.

[0040] Based on the correspondence between the reference coordinates of the boundary points and the coordinates of the assigned cluster centers, the Euclidean distance between the reference coordinates of the boundary points and the coordinates of the assigned cluster centers is calculated.

[0041] Within a preset sliding time window, the Euclidean distance is updated by extreme value tracking to confirm the dynamic adaptive threshold of the power module.

[0042] In a preferred embodiment, the Euclidean distance is calculated using the following formula:

[0043] ;

[0044] In the formula, For the first Each boundary point at the sampling time With the The Euclidean distance between the cluster centers to which each boundary point belongs. For the index identifier of the boundary point, This serves as a time identifier for the sampling time. For the first Each boundary point at the sampling time The low-dimensional reference coordinates of the first dimensional components, For the first The cluster center to which each boundary point belongs at the sampling time The low-dimensional reference coordinates of the first dimensional components, Let be the embedding dimension of the low-dimensional eigenspace. For summation operations.

[0045] In a preferred embodiment, the step of analyzing the state deviation of the power module based on the low-dimensional reference coordinates and the low-dimensional monitoring coordinates includes:

[0046] Based on the neighborhood topology of the low-dimensional reference coordinates, geometric deviation quantization is performed on the low-dimensional monitoring coordinates to obtain local structural deviation data of the low-dimensional monitoring coordinates.

[0047] Spatial distance measurement is performed between the low-dimensional monitoring coordinates and the low-dimensional reference coordinates to obtain global deviation data between the low-dimensional monitoring coordinates and the low-dimensional reference coordinates;

[0048] Based on the local structural deviation data and the global deviation data, a comprehensive status assessment of the power module is performed to obtain the status anomaly index of the power module.

[0049] To address the above problems, the present invention also provides a power module operation status monitoring system, the system comprising:

[0050] A multi-frequency calibration module is used to perform multi-frequency calibration on the high-frequency electrical signals of the power module to obtain standardized multi-dimensional electrical signals of the high-frequency electrical signals;

[0051] The feature extraction module is used to perform variational mode decomposition on the standardized multidimensional electrical signal and perform Hilbert transform on the decomposed finite bandwidth mode components to obtain the high-dimensional state feature vector of the power module.

[0052] The benchmark construction module is used to perform local linear embedding dimensionality reduction on the high-dimensional state feature vector to obtain the low-dimensional benchmark coordinates of the high-dimensional state feature vector.

[0053] The online mapping module is used to map the current high-dimensional state feature vector of the power module to the low-dimensional intrinsic space of the low-dimensional reference coordinates to obtain the low-dimensional monitoring coordinates of the current high-dimensional state feature vector.

[0054] The threshold learning module is used to construct a decision graph of the low-dimensional reference coordinates based on the local density and relative distance of the low-dimensional reference coordinates, and to perform cluster center spacing analysis on the boundary points in the decision graph to obtain the dynamic adaptive threshold of the power module.

[0055] The status diagnosis module is used to analyze the degree of status deviation of the power module based on the low-dimensional reference coordinates and the low-dimensional monitoring coordinates, and to compare the difference between the analyzed status anomaly index and the dynamic adaptive threshold to generate a status anomaly warning signal for the power module.

[0056] Compared with the prior art, the present invention has the following beneficial effects:

[0057] 1. This invention achieves standardized processing of high-frequency electrical signals through multi-frequency calibration, and completes accurate extraction of high-dimensional state feature vectors by combining variational mode decomposition and Hilbert transform. Then, it constructs low-dimensional reference coordinates through local linear embedding dimensionality reduction, which not only ensures the integrity and accuracy of state features, but also simplifies the data operation dimension, and greatly improves the efficiency and accuracy of power module state feature extraction.

[0058] 2. This invention obtains a dynamic adaptive threshold by constructing a decision graph, and combines local structure and global deviation data to comprehensively analyze the state anomaly index, thereby achieving a comprehensive judgment of the degree of deviation of the power module state. The dynamic threshold can be adapted to the real-time operating conditions of the module, improving the accuracy of state anomaly judgment. At the same time, it can quickly generate anomaly warning signals, improving the real-time performance and reliability of power module operating status monitoring, and effectively ensuring the reference value of monitoring results. Attached Figure Description

[0059] Figure 1 This is a flowchart illustrating a method for monitoring the operating status of a power module according to an embodiment of the present invention.

[0060] Figure 2 This is a functional block diagram of a power module operation status monitoring system provided in an embodiment of the present invention;

[0061] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0062] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0063] This application provides a method for monitoring the operating status of a power module. The executing entity of this method includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the method for monitoring the operating status of a power module can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0064] Reference Figure 1The diagram shown is a flowchart illustrating a method for monitoring the operating status of a power module according to an embodiment of the present invention. In this embodiment, the method for monitoring the operating status of a power module includes:

[0065] S1. Perform multi-frequency calibration on the high-frequency electrical signal of the power module to obtain a standardized multi-dimensional electrical signal of the high-frequency electrical signal;

[0066] In this embodiment of the invention, the step of performing multi-frequency calibration on the high-frequency electrical signal of the power module to obtain a standardized multi-dimensional electrical signal of the high-frequency electrical signal includes:

[0067] The high-frequency electrical signal of the power module is bandpass filtered to obtain the fundamental signal component and harmonic signal component of the high-frequency electrical signal;

[0068] The fundamental signal component and the harmonic signal component are subjected to amplitude normalization processing, and the normalized sub-signals are time-stamp aligned to obtain the multi-band signal of the high-frequency electrical signal.

[0069] According to the preset channel sequence, the multi-band signal is recombined into a multi-dimensional data to obtain a standardized multi-dimensional electrical signal of the high-frequency electrical signal.

[0070] A bandpass filter circuit with a fixed center frequency and passband range is built for the transmission frequency band of the high-frequency electrical signal of the power module. The high-frequency electrical signal of the power module is input into the bandpass filter circuit throughout its transmission. During the transmission of the signal in the circuit, all stray interference signals outside the preset frequency band are filtered out, and only the fundamental signal component and harmonic signal component that match the electrical characteristics of the power module are retained. Finally, the fundamental signal component and harmonic signal component corresponding to the high-frequency electrical signal of the power module are output.

[0071] The fundamental and harmonic signal components are connected to independent amplitude conditioning circuits. The impedance matching circuit within the conditioning circuit adjusts the input and output impedances to a consistent state. Then, the gain adjustment circuit within the circuit adjusts the amplitudes of the two signal components separately, bringing the amplitudes of both the fundamental and harmonic signal components within a preset amplitude reference range. This completes the amplitude normalization of the fundamental and harmonic signal components. Subsequently, the amplitude-normalized fundamental and harmonic signal sub-signals are connected to a time synchronization module. Using the sampling clock during the operation of the power module as a unified time reference, each data acquisition point on the two sub-signals is precisely matched to the corresponding sampling clock time node, achieving timestamp alignment of the normalized sub-signals. The two timestamp-aligned sub-signals are then integrated to finally obtain the multi-band high-frequency electrical signal of the power module.

[0072] A fixed channel order is pre-set based on the signal acquisition channel numbers of the power module from low to high. The processed fundamental and harmonic signals of the multi-frequency band signal are sequentially assigned to the designated channels of the multi-dimensional data storage according to the preset channel order. The signal data in each designated channel is arranged continuously and orderly. At the same time, the signal data of all channels are restructured into a structured whole, so that the restructured signal data forms a standardized signal with multi-dimensional characteristics and fully conforms to the preset data format requirements. Finally, the standardized multi-dimensional electrical signal of the high-frequency electrical signal of the power module is obtained.

[0073] The beneficial effects are as follows: By constructing a bandpass filter circuit with fixed frequency band parameters to specifically filter the high-frequency electrical signals of the power module, the fundamental and harmonic signal components can be accurately and thoroughly separated, effectively eliminating the adverse effects of stray interference signals on the original electrical signals and ensuring the purity and integrity of the extracted signal components. Amplitude normalization processing uses hardware circuits to unify the amplitude of different signal components to a preset amplitude reference range, eliminating amplitude differences between different signal components and providing a unified amplitude reference standard for signal components in different frequency bands. Timestamp alignment uses the power module's own sampling clock as a unified time reference, achieving precise time synchronization between the two sub-signal data acquisition points, avoiding deviations in the time dimension, and maintaining data consistency across multiple frequency bands. Multi-dimensional data reorganization, performed in a preset channel order, transforms multi-frequency band signals into structured and standardized multi-dimensional electrical signals, providing a unified and standardized signal data foundation for subsequent decomposition and feature extraction of electrical signals, significantly improving the orderliness and accuracy of subsequent signal processing. The entire multi-frequency calibration process is completed through hardware circuits and dedicated modules to process and convert signals without any black-box operation. Each step has clearly defined operating standards and benchmark requirements, ensuring the reproducibility of the entire process and the stability of actual implementation. This allows the final standardized multi-dimensional electrical signals to truly, comprehensively and accurately reflect the high-frequency electrical signal characteristics of the power module during actual operation, laying a solid and reliable signal data foundation for the subsequent accurate extraction of the operating status characteristics of the power module.

[0074] S2. Perform variational mode decomposition on the standardized multidimensional electrical signal, and perform Hilbert transform on the decomposed finite bandwidth mode components to obtain the high-dimensional state feature vector of the power module.

[0075] In this embodiment of the invention, the step of performing variational mode decomposition on the standardized multidimensional electrical signal and applying Hilbert transform to the decomposed finite-bandwidth mode components to obtain the high-dimensional state feature vector of the power module includes:

[0076] Based on the preset number of modal decompositions, the standardized multidimensional electrical signal is subjected to frequency domain iterative decomposition to obtain the initial modal components of the standardized multidimensional electrical signal.

[0077] The initial modal components are updated by convergence determination to obtain the finite bandwidth intrinsic mode functions of the standardized multidimensional electrical signal;

[0078] The instantaneous amplitude and instantaneous frequency of the finite bandwidth intrinsic mode function are analyzed, and the time-frequency integral mapping of the finite bandwidth intrinsic mode function is performed to obtain the Hilbert marginal spectrum of the finite bandwidth intrinsic mode function;

[0079] The energy distribution of the Hilbert marginal spectrum is analyzed to obtain the energy entropy and spectral centroid of the Hilbert marginal spectrum;

[0080] Multidimensional feature fusion is performed on the energy entropy and the spectral centroid to obtain the high-dimensional state feature vector of the power module.

[0081] By pre-setting a fixed number of mode decompositions based on the frequency band distribution range of the electrical signals of the power module and the signal characteristics of the standardized multi-dimensional electrical signals, the standardized multi-dimensional electrical signals are imported into the frequency domain decomposition system. Using this preset number as the decomposition target, the signal is subjected to hierarchical decomposition in the frequency domain. Each decomposition extracts the mode components within an independent frequency band of the signal. After completing one decomposition, the same frequency domain decomposition operation is performed on the remaining undecomposed signals until the number of extracted mode components is completely consistent with the preset number of mode decompositions. All the mode components extracted by the frequency domain decomposition are integrated as a whole to finally obtain the initial mode components of the standardized multi-dimensional electrical signals.

[0082] The initial modal components are imported into the signal convergence determination system. For each initial modal component, two consecutive full-dimensional signal feature detections are performed. The signal feature values ​​obtained from the two detections are matched and calculated to determine the signal feature overlap between adjacent detection results. This overlap is compared item by item with a preset convergence determination threshold. If the overlap does not reach the preset threshold, the initial modal component undergoes fine-tuning and updating of its signal features before the above detection and comparison operations are repeated. If the overlap reaches the preset threshold, the initial modal component is determined to have converged. After performing the above convergence determination and update operations on all initial modal components, all converged modal components are reconstructed using signal structure to ultimately obtain the finite-bandwidth intrinsic mode functions of the standardized multi-dimensional electrical signal. The preset convergence determination threshold is a signal feature overlap of 95% between two adjacent feature detections; a single fixed value within the 90%-99% range can be selected as this threshold based on the power module signal monitoring accuracy requirements. If the overlap does not reach the threshold, perform iterative updates on the modal components; if the overlap reaches the threshold, determine that the modal components have converged.

[0083] The finite-bandwidth intrinsic mode function (EMF) is input into a signal feature analysis system. The system's amplitude detection unit collects and records the signal amplitude of the function at each time point in real time, forming a continuous amplitude variation data sequence to obtain the instantaneous amplitude of the finite-bandwidth EMF. Similarly, the system's frequency detection unit collects and records the signal frequency of the function at each time point in real time, forming a continuous frequency variation data sequence to obtain the instantaneous frequency of the finite-bandwidth EMF. This finite-bandwidth EMF is then imported into a Hilbert transform processing system, where continuous time-frequency domain integration is performed across the entire domain. All time-frequency domain feature data of the function are precisely mapped to a pre-defined standardized time-frequency coordinate space. The mapped time-frequency domain feature data are then systematically integrated and professionally plotted to ultimately obtain the Hilbert marginal spectrum of the finite-bandwidth EMF.

[0084] The Hilbert marginal spectrum is imported into the energy distribution analysis system. The signal energy values ​​of each frequency band within the spectrum are comprehensively and completely collected and quantified. Based on the core calculation logic of information entropy, the overall distribution characteristics of all collected frequency band energy values ​​are analyzed to quantify the dispersion of energy distribution within the spectrum, thus obtaining the energy entropy of the Hilbert marginal spectrum. Then, the energy values ​​of all frequency bands within the spectrum are weighted and calculated with the corresponding frequency band position coordinates. The comprehensive result of the weighted calculation determines the core center position of the signal energy distribution within the spectrum. This center position is taken as the spectral centroid of the Hilbert marginal spectrum, and finally, the energy entropy and spectral centroid of the Hilbert marginal spectrum are obtained simultaneously.

[0085] The energy entropy and spectral centroid of the Hilbert marginal spectrum obtained from the analysis are used as two core feature quantities of the power module's operating state. A multidimensional feature space dimension and feature dimension arrangement rule that match the power module's state monitoring requirements are pre-defined. According to the rule, all feature data of the energy entropy and all feature data of the spectral centroid are respectively assigned to the specified independent dimensions of the multidimensional feature space. All data of the two feature quantities are fully structured and arranged in an ordered manner. The integrated multidimensional feature data is transformed and processed according to the standard form of vector expression to finally obtain the high-dimensional state feature vector of the power module.

[0086] The beneficial effects include: pre-setting the number of mode decompositions based on the signal characteristics of the power module, making the frequency domain iterative decomposition more targeted, accurately extracting the initial mode components related to the operating state of the power module, avoiding under- or over-decomposition, and ensuring the effectiveness of signal decomposition. By pre-setting a convergence threshold for updating the convergence judgment, the signal characteristics of the initial mode components tend to stabilize. The resulting finite-bandwidth intrinsic mode functions accurately reflect the intrinsic characteristics of standardized multi-dimensional electrical signals, laying a reliable foundation for subsequent feature analysis. Accurate analysis of the instantaneous amplitude and frequency of the finite-bandwidth intrinsic mode functions is performed, and a Hilbert marginal spectrum is plotted using time-frequency integral mapping, transforming the time-frequency characteristics of the signal into an intuitive spectral form, facilitating subsequent energy distribution analysis. The energy entropy and spectral centroid obtained by energy distribution analysis of the Hilbert marginal spectrum can quantitatively characterize the operating state features of the power module from different dimensions, realizing in-depth mining of signal features. Finally, the two core feature quantities are transformed into high-dimensional state feature vectors through multi-dimensional feature fusion, which fully preserves the key feature information of the power module's operating state. This provides comprehensive and accurate feature data support for subsequent dimensionality reduction processing and state monitoring. Each step of the entire process has set clear operating standards and judgment criteria to ensure the reproducibility and accuracy of feature extraction.

[0087] S3. Perform local linear embedding dimensionality reduction on the high-dimensional state feature vector to obtain the low-dimensional reference coordinates of the high-dimensional state feature vector;

[0088] In this embodiment of the invention, the step of performing local linear embedding dimensionality reduction on the high-dimensional state feature vector to obtain the low-dimensional reference coordinates of the high-dimensional state feature vector includes:

[0089] Obtain the normal high-dimensional state feature vector within the historical normal operation period from the high-dimensional state feature vector;

[0090] Topological association analysis is performed on the normal high-dimensional state feature vectors to obtain the neighborhood relationships between the normal high-dimensional state feature vectors;

[0091] Based on the neighborhood relationship, the normal high-dimensional state feature vector is locally linearly reconstructed to obtain the weight coefficient matrix of the normal high-dimensional state feature vector;

[0092] Based on the weight coefficient matrix, the normal high-dimensional state feature vector is embedded and mapped to obtain the low-dimensional reference coordinates of the normal high-dimensional state feature vector.

[0093] The step of embedding and mapping the normal high-dimensional state feature vector based on the weight coefficient matrix to obtain the low-dimensional reference coordinates of the normal high-dimensional state feature vector includes:

[0094] The weight coefficient matrix is ​​constrained and encoded to obtain the cost matrix of the normal high-dimensional state feature vector;

[0095] Perform generalized eigenvalue decomposition on the cost matrix to obtain the eigenvectors of the normal high-dimensional state eigenvectors;

[0096] A basis system is constructed on the feature vectors to obtain the embedding coordinate matrix of the feature vectors;

[0097] Based on the embedded coordinate matrix, the low-dimensional reference coordinates of the normal high-dimensional state feature vector are confirmed.

[0098] From the storage database of high-dimensional state feature vectors, high-dimensional state feature vectors of all historical operating periods of the power module are retrieved. The historical normal operating period is defined as the period in which all operating parameters of the power module are within the preset rated range and there are no fault alarm information. The retrieved vectors are precisely filtered according to the time interval of this period. Data integrity verification is performed on the filtered vectors, and invalid vectors with missing or disordered data are removed. Only valid vectors with complete data and standardized format are retained. Finally, the normal high-dimensional state feature vectors within the historical normal operating period are obtained from the high-dimensional state feature vectors.

[0099] Normal high-dimensional state feature vectors are imported into a topological association parsing system. A fixed number of neighborhood searches is preset. Each normal high-dimensional state feature vector is used as a core search point. The feature similarity between the core search point and all other normal high-dimensional state feature vectors is calculated. The remaining vectors are sorted in descending order of feature similarity. The vectors with the highest preset number of neighborhood searches in the sorted list are selected as the neighborhood vectors of the core search point. The core search and neighborhood vector selection operations are performed for each normal high-dimensional state feature vector one by one. All neighborhood vectors corresponding to each vector and the feature similarity association information between them are fully recorded to form a standardized topological association table. Finally, the neighborhood relationships between normal high-dimensional state feature vectors are obtained.

[0100] Based on the neighborhood relationships between the obtained normal high-dimensional state feature vectors, each normal high-dimensional state feature vector is taken as the target vector, and all neighborhood vectors corresponding to the target vector are taken as reconstruction basis vectors. A linear combination operation is performed on the reconstruction basis vectors, and the weight values ​​corresponding to each reconstruction basis vector are continuously adjusted during the combination process until the result of the linear combination operation completely matches all feature data of the target vector. The weight values ​​corresponding to each target vector when the accurate reconstruction is completed are recorded. According to the arrangement order of the normal high-dimensional state feature vectors, the weight values ​​corresponding to all target vectors are arranged in an orderly manner to form a standardized matrix structure in which rows and columns correspond to the normal high-dimensional state feature vectors, and finally the weight coefficient matrix of the normal high-dimensional state feature vectors is obtained.

[0101] The weight coefficient matrix of the normal high-dimensional state feature vector is imported into the matrix encoding system. Pre-defined matrix constraint rules, adapted to the dimensionality reduction requirements of local linear embedding, are established. These rules explicitly define the value range of matrix elements and the correlation constraints between matrix rows and columns. Each element in the weight coefficient matrix is ​​encoded and corrected according to these constraints. Elements that do not conform to the constraints are numerically adjusted to fully meet the preset constraints while preserving the original dimension and element relationships of the matrix. The encoded and corrected matrix is ​​then structurally organized to obtain the cost matrix of the normal high-dimensional state feature vector. The preset constraints are two fixed rules: the sum of all elements in each row of the weight coefficient matrix equals 1; and the value range of all elements in the weight coefficient matrix is ​​[0,1].

[0102] The cost matrix of the normal high-dimensional state feature vector is imported into the eigenvalue decomposition system. A global matrix feature analysis is performed on the cost matrix to extract all feature components with independent feature representation capabilities. The feature components are sorted in descending order of their contribution to the overall feature of the matrix. Each sorted feature component is then transformed into a vector form so that each feature component forms a feature vector that conforms to the vector expression standard. All transformed feature vectors are then integrated in an orderly manner according to the contribution ranking results to finally obtain the feature vector of the normal high-dimensional state feature vector.

[0103] The eigenvectors of the normal high-dimensional state feature vectors are imported into the basis system construction system. A low-dimensional embedding dimension matching the requirements of low-dimensional baseline coordinate construction is pre-defined. The top N eigenvectors by eigenvalue ranking are selected as the basis vectors for low-dimensional embedding, where N is the preset low-dimensional embedding dimension. A complete low-dimensional embedding basis system is constructed based on these basis vectors. All normal high-dimensional state feature vectors are projected onto the corresponding low-dimensional space of this basis system. The coordinate values ​​of each normal high-dimensional state feature vector in each dimension of the low-dimensional space are accurately calculated. Following the order of the normal high-dimensional state feature vectors, all coordinate values ​​are arranged in an ordered manner, forming a standardized matrix structure where rows correspond to normal high-dimensional state feature vectors and columns correspond to low-dimensional embedding dimensions. This yields the embedding coordinate matrix of the feature vectors. The preset low-dimensional embedding dimension N is determined using the cumulative eigenvalue contribution rate method. Specifically, all eigenvalues ​​obtained from the generalized eigenvalue decomposition of the cost matrix are sorted in descending order of value, denoted as λ1, λ2, ..., λ... m (m is the original dimension of the high-dimensional state feature vector); Calculate the cumulative eigenvalue contribution rate according to the formula: C(k)=(λ1+λ2+…+λ k )÷(λ1+λ2+…+λ m )×100%, select the smallest positive integer k that satisfies C(k)≥95%, and this value of k is the preset low-dimensional embedding dimension N.

[0104] The embedded coordinate matrix of the obtained feature vector is used as the core coordinate data. According to the row and column arrangement rules of the embedded coordinate matrix, the row coordinate value corresponding to each normal high-dimensional state feature vector is used as the independent coordinate of the vector in the low-dimensional space. The coordinate values ​​of each row in the matrix are transformed into standard coordinate form and spatially calibrated one by one. All the calibrated coordinates are integrated to form a complete set of low-dimensional coordinates, and finally the low-dimensional reference coordinates of the normal high-dimensional state feature vector are obtained.

[0105] The beneficial effects are as follows: By using clear historical normal operation period criteria to screen normal high-dimensional state feature vectors, invalid vectors are eliminated, ensuring the effectiveness and accuracy of the data source for subsequent dimensionality reduction processing. Topological association parsing determines neighborhood relationships by pre-setting a fixed number of neighborhood searches, making the topological associations between vectors more targeted and accurately reflecting the local distribution characteristics of normal high-dimensional state feature vectors. Local linear reconstruction based on neighborhood relationships ensures that the weight coefficient matrix truly reflects the linear relationships between normal high-dimensional state feature vectors, providing a reliable matrix foundation for subsequent dimensionality reduction. Constraint encoding with clear rules on the weight coefficient matrix yields the cost matrix, ensuring the matrix fully adapts to the dimensionality reduction requirements of local linear embedding and avoiding data bias during the dimensionality reduction process. Generalized eigenvalue decomposition of the cost matrix extracts feature vectors based on contribution, effectively preserving the core feature information of normal high-dimensional state feature vectors and eliminating redundant features. By constructing a basis spectrum and obtaining the embedding coordinate matrix through a preset low-dimensional embedding dimension, the accurate projection of the normal high-dimensional state feature vector to the low-dimensional space is achieved. The resulting low-dimensional reference coordinates significantly reduce the data dimensionality and simplify subsequent calculations, while fully preserving the core topology and feature distribution characteristics of the normal high-dimensional state feature vector. This provides a precise and standardized low-dimensional coordinate reference for the online mapping and anomaly detection of the current state of the power module. Each step of the entire dimensionality reduction process has set clear operating standards and preset rules, with no black-box operations, ensuring the reproducibility of the process and the stability of the results.

[0106] In this embodiment of the invention, S4, the current high-dimensional state feature vector of the power module is mapped to the low-dimensional intrinsic space of the low-dimensional reference coordinates to obtain the low-dimensional monitoring coordinates of the current high-dimensional state feature vector;

[0107] The low-dimensional intrinsic space corresponding to the low-dimensional reference coordinates of the constructed normal high-dimensional state feature vector is retrieved. This low-dimensional intrinsic space contains a preset low-dimensional embedding basis spectrum and standardized high-dimensional to low-dimensional embedding mapping rules. The current high-dimensional state feature vector of the power module is imported into the data verification system, and the integrity and format of the feature data of the vector are verified dimension by dimension to ensure that its feature dimensions, data arrangement specifications are completely consistent with the normal high-dimensional state feature vector during the historical normal operation period. Parts with missing data or inconsistent formats are corrected and standardized to obtain a current high-dimensional state feature vector with standardized format and complete data. Then, according to the preset embedding mapping rules of the low-dimensional intrinsic space, the corrected current high-dimensional state feature vector is projected dimension by dimension into the low-dimensional intrinsic space. Based on the feature correlation relationship of the low-dimensional embedding basis spectrum, the coordinate components of the vector in each embedding dimension of the low-dimensional intrinsic space are accurately calculated. Then, all the calculated coordinate components are integrated in an orderly manner according to the coordinate arrangement specifications of the low-dimensional reference coordinates. The integrated coordinate components are transformed into standardized coordinates that conform to the coordinate expression form of the low-dimensional intrinsic space, and finally the low-dimensional monitoring coordinates of the current high-dimensional state feature vector of the power module are obtained.

[0108] The beneficial effects include: retrieving a low-dimensional eigenspace that matches the low-dimensional reference coordinates and using its preset basis system and mapping rules, ensuring the uniformity of the coordinate reference during the mapping process of the current high-dimensional state feature vector, avoiding coordinate deviations caused by differences in mapping rules; conducting full-dimensional data verification and correction on the current high-dimensional state feature vector, ensuring that it is completely matched with the normal high-dimensional state feature vector in terms of format and dimension, providing a qualified data foundation for subsequent accurate projection; and projecting the current high-dimensional state feature vector dimension by dimension to the low-dimensional eigenspace according to standardized rules and accurately calculating the coordinate components of each dimension, thus realizing the mapping of high-dimensional feature vectors to low-dimensional eigenspaces. Precise spatial mapping reduces data dimensionality while fully preserving the core state features of the current high-dimensional state feature vector. Furthermore, the low-dimensional monitoring coordinates strictly adhere to the arrangement specifications of the low-dimensional reference coordinates, allowing subsequent state deviation analysis to be conducted within a unified low-dimensional coordinate system. This ensures the effectiveness and accuracy of coordinate comparison during the analysis process. The resulting low-dimensional monitoring coordinates accurately and realistically reflect the current operating status characteristics of the power module, providing reliable real-time coordinate data support for subsequent anomaly detection. The entire mapping process has clear operational standards, and the data processing steps are reproducible, ensuring the stability and consistency of the low-dimensional monitoring coordinates.

[0109] S5. Based on the local density and relative distance of the low-dimensional reference coordinates, construct a decision graph of the low-dimensional reference coordinates, and perform clustering center spacing analysis on the boundary points in the decision graph to obtain the dynamic adaptive threshold of the power module.

[0110] In this embodiment of the invention, constructing a decision graph of the low-dimensional reference coordinates based on the local density and relative distance of the low-dimensional reference coordinates includes:

[0111] The local density of the low-dimensional reference coordinates is obtained by performing a distribution density measurement on the low-dimensional reference coordinates.

[0112] The relative distance between the low-dimensional reference coordinates is obtained by performing spacing parameter analysis on the low-dimensional reference coordinates.

[0113] Using the local density as the horizontal axis parameter and the relative distance as the vertical axis parameter, the local density value and relative distance value of the low-dimensional reference coordinates are mapped to a two-dimensional coordinate system to obtain the initial decision map of the low-dimensional reference coordinates.

[0114] The initial decision graph is interpreted using graph structure interpretation to obtain the decision graph with the low-dimensional reference coordinates.

[0115] The step of performing clustering center spacing analysis on the boundary points in the decision graph to obtain the dynamic adaptive threshold of the power module includes:

[0116] Based on the distribution characteristics of the decision graph, the boundary points and cluster centers in the decision graph are selected.

[0117] Perform nearest neighbor analysis on the reference coordinates of the boundary points and the reference coordinates of the cluster centers to determine the coordinates of the cluster centers to which the boundary points belong.

[0118] Based on the correspondence between the reference coordinates of the boundary points and the coordinates of the assigned cluster centers, the Euclidean distance between the reference coordinates of the boundary points and the coordinates of the assigned cluster centers is calculated.

[0119] Within a preset sliding time window, the Euclidean distance is updated by extreme value tracking to confirm the dynamic adaptive threshold of the power module.

[0120] The formula for calculating the Euclidean distance is as follows:

[0121] ;

[0122] In the formula, For the first Each boundary point at the sampling time With the The Euclidean distance between the cluster centers to which each boundary point belongs. For the index identifier of the boundary point, This serves as a time identifier for the sampling time. For the first Each boundary point at the sampling time The low-dimensional reference coordinates of the first dimensional components, For the first The cluster center to which each boundary point belongs at the sampling time The low-dimensional reference coordinates of the first dimensional components, Let be the embedding dimension of the low-dimensional eigenspace. For summation operations.

[0123] A fixed density metric neighborhood is preset. The neighborhood is defined in the low-dimensional intrinsic space with each low-dimensional reference coordinate as the center. The number of other low-dimensional reference coordinates contained in the neighborhood is counted. This number is used as the local density value of the corresponding low-dimensional reference coordinate. This statistical operation is performed on all low-dimensional reference coordinates in sequence, and a unique local density value is matched for each low-dimensional reference coordinate. Finally, the local density of the low-dimensional reference coordinates is obtained.

[0124] A fixed relative distance measurement rule is preset. For each low-dimensional reference coordinate, the spatial distance between that reference point and all low-dimensional reference coordinates in the low-dimensional intrinsic space with local density values ​​greater than that reference point is calculated. The smallest spatial distance is selected as the relative distance value of that reference point. If no low-dimensional reference coordinate has a local density value greater than that reference point, the preset maximum distance value is used as the relative distance value of that reference point. This calculation and selection operation is performed sequentially for all low-dimensional reference coordinates, and a unique relative distance value is matched for each low-dimensional reference coordinate, ultimately obtaining the relative distance of the low-dimensional reference coordinates. The preset maximum distance value is 1.2 times the diagonal Euclidean distance of the low-dimensional intrinsic space; this value is a fixed calculated value and is not subject to fluctuation or adjustment.

[0125] A standardized two-dimensional Cartesian coordinate system is constructed, with the horizontal axis set as the local density axis and the vertical axis as the relative distance axis. Following a pre-defined coordinate calibration rule, scales matching the local density values ​​are established on the horizontal axis, and scales matching the relative distance values ​​are established on the vertical axis. The local density and relative distance values ​​corresponding to each low-dimensional reference coordinate are treated as a set of two-dimensional coordinate data. This set of data is precisely mapped to its corresponding position in the two-dimensional Cartesian coordinate system, and the data points are labeled. This mapping and labeling operation is performed sequentially for all low-dimensional reference coordinates, ensuring that all low-dimensional reference coordinates form corresponding labeled points in the two-dimensional coordinate system, ultimately yielding the initial decision map of the low-dimensional reference coordinates. The pre-defined coordinate calibration rule is a fixed standard: the horizontal axis is calibrated with a fixed step size of 1; the vertical axis is calibrated with a fixed step size of 0.1 times the pre-defined maximum spacing value; and the scale interval completely covers the value range of all data points.

[0126] The initial decision map is subjected to global graph feature detection to identify the overall distribution pattern of data points, the division of dense regions, and the positional characteristics of discrete points. According to the preset graph structure optimization rules, overlapping annotation points in the initial decision map are separated and calibrated, blurred annotation points are clarified, and the coordinate axis scales and annotations of the graph are standardized and corrected. At the same time, the original local density and relative distance information of all data points in the initial decision map are completely preserved. After completing all optimization and correction operations, a standardized graph structure is formed, and finally a decision map with low-dimensional reference coordinates is obtained.

[0127] The system pre-determines the selection criteria for cluster centers and boundary points. It identifies cluster centers as points with local density values ​​in a pre-determined high-density range and relative distance values ​​in a pre-determined high-distance range in the decision graph, and boundary points as points located at the edge of dense regions in the decision graph and with relative distance values ​​in a pre-determined medium-distance range. Based on these criteria, all points in the decision graph are screened one by one, and all identified cluster centers and boundary points are independently identified and fully recorded. Finally, the boundary points and cluster centers in the decision graph are selected.

[0128] Taking the low-dimensional reference coordinates corresponding to each boundary point as the analysis object, the spatial distance between the boundary point and the low-dimensional reference coordinates corresponding to all cluster centers in the decision graph is calculated. The low-dimensional reference coordinates corresponding to the cluster center with the smallest spatial distance are determined as the assigned cluster center coordinates of the boundary point. This calculation and determination operation is performed on all boundary points in sequence. A unique assigned cluster center coordinate is matched for each boundary point, and a one-to-one correspondence between the boundary point and the assigned cluster center coordinates is formed. Finally, the assigned cluster center coordinates of all boundary points are determined.

[0129] For each boundary point and its corresponding cluster center coordinates, at the sampling time, the components of each embedded dimension of the low-dimensional reference coordinates of the boundary point and the components of each embedded dimension of the coordinates of its corresponding cluster center are extracted. The difference between the components of the same dimension of the two is calculated for each dimension, and the difference of each dimension is squared. The squared results of all dimensions are accumulated to obtain the sum of squares. The square root of the sum of squares is the Euclidean distance between the boundary point and its corresponding cluster center at the sampling time. This calculation operation is performed for all boundary points in turn, and the corresponding Euclidean distance value is matched for each boundary point. Finally, the Euclidean distance between all boundary points and their corresponding cluster centers is obtained.

[0130] A fixed-duration sliding time window is preset and continuously advanced according to the sampling time. Within each sliding time window, the Euclidean distance values ​​corresponding to all boundary points within that time window are collected. The largest Euclidean distance value is selected as the extreme value within that sliding time window. As the sliding time window continues to advance, the extreme values ​​within each time window are tracked and updated in real time. The extreme value selected in the latest sliding time window is determined as the dynamic adaptive threshold of the power module, thus confirming the dynamic adaptive threshold of the power module.

[0131] The beneficial effects are as follows: By pre-setting fixed density metric neighborhood ranges and relative distance metric rules, a unique local density value and relative distance value are matched for each low-dimensional baseline coordinate, allowing for precise quantification of the distribution characteristics of the low-dimensional baseline coordinates. This provides unified and accurate quantitative data for decision graph construction, ensuring the validity of the foundational data for subsequent graph construction. A standardized two-dimensional coordinate system is built and data mapping is completed to obtain the initial decision graph. Then, the decision graph is obtained through standardized graph structure interpretation. While retaining all features of the original data, the standardization and clarity of the graph are achieved, providing an intuitive and accurate graph basis for the selection of boundary points and cluster centers. By pre-setting clear criteria such as high-density intervals and high-distance intervals, boundary points and cluster centers are selected, providing a unified basis for the selection operation and avoiding bias caused by subjective judgment. The selected points accurately reflect the graph distribution characteristics of the decision graph. Nearest neighbor analysis matches a unique cluster center coordinate for each boundary point, forming a clear one-to-one correspondence. This clarifies the fixed computational objects for Euclidean distance calculation, ensuring the targeted nature of the distance calculation. The calculation of Euclidean distance employs a step-by-step process of subtraction, squaring, summation, and square root extraction across dimensions. This precisely quantifies the actual spatial distance between boundary points and their respective cluster centers in the low-dimensional intrinsic space, providing crucial quantitative data support for confirming the dynamic adaptive threshold. By using a preset sliding time window to track and update the extreme values ​​of the Euclidean distance, the dynamic adaptive threshold can be updated in real-time following the sampling process. This ensures the threshold accurately adapts to the real-time operating conditions of the power modules. Each step in the process has a clearly defined preset range, rules, interval, or duration. All operations have specific and implementable methods, eliminating any black-box behavior and fully guaranteeing the reproducibility of the technical solution. The products generated in each step serve as the basis for subsequent steps, forming a logically coherent and data-interoperable complete operational flow. The resulting dynamic adaptive threshold possesses real-time performance, adaptability, and accuracy, providing a dynamic and reliable benchmark for comparing the differences in the power module state anomaly indices, effectively improving the accuracy and real-time performance of subsequent power module state anomaly determinations.

[0132] S6. Based on the low-dimensional reference coordinates and the low-dimensional monitoring coordinates, perform a state deviation analysis on the power module, and compare the analyzed state anomaly index with the dynamic adaptive threshold to generate a state anomaly warning signal for the power module.

[0133] In this embodiment of the invention, the step of analyzing the state deviation of the power module based on the low-dimensional reference coordinates and the low-dimensional monitoring coordinates includes:

[0134] Based on the neighborhood topology of the low-dimensional reference coordinates, geometric deviation quantization is performed on the low-dimensional monitoring coordinates to obtain local structural deviation data of the low-dimensional monitoring coordinates.

[0135] Spatial distance measurement is performed between the low-dimensional monitoring coordinates and the low-dimensional reference coordinates to obtain global deviation data between the low-dimensional monitoring coordinates and the low-dimensional reference coordinates;

[0136] Based on the local structural deviation data and the global deviation data, a comprehensive status assessment of the power module is performed to obtain the status anomaly index of the power module.

[0137] The process involves retrieving the established neighborhood topology of the low-dimensional reference coordinates, pre-setting geometric deviation quantization rules adapted to this topology, and precisely placing the low-dimensional monitoring coordinates within the corresponding low-dimensional intrinsic space of the low-dimensional reference coordinates. Based on the distribution characteristics of the neighborhood topology, a corresponding topological neighborhood range is matched for the low-dimensional monitoring coordinates. The geometric offset of the low-dimensional monitoring coordinates from each low-dimensional reference coordinate within this topological neighborhood range is calculated across each embedding dimension. Simultaneously, the fit between the low-dimensional monitoring coordinates and the overall geometric arrangement of the low-dimensional reference coordinates within the topological neighborhood is quantified. All embedding dimension offsets and geometric arrangement fit are converted into a single standardized value according to the pre-set quantization rules. This value represents the local structural deviation data of the low-dimensional monitoring coordinates. The pre-set quantization rule is a fixed calculation formula: Local structural deviation data = Σ(Geometric offset of each embedding dimension × 1 / N) + (1 - Geometric arrangement fit), where N is the pre-set low-dimensional embedding dimension, and 1 / N is a fixed weight coefficient for each dimension.

[0138] A unified spatial distance measurement rule is preset within the low-dimensional intrinsic space. Based on this rule, the spatial straight-line distance from the low-dimensional monitoring coordinate to each low-dimensional reference coordinate is calculated one by one in the low-dimensional intrinsic space. The average value of all calculated spatial straight-line distances is obtained by summing all the spatial straight-line distances to obtain the basic distance value of the low-dimensional monitoring coordinate relative to the low-dimensional reference coordinate. Then, according to the preset requirements of power module status monitoring, a fixed correction coefficient is matched to the basic distance value. The basic distance value and the correction coefficient are multiplied to obtain the standardized distance value. This standardized distance value is the global deviation data between the low-dimensional monitoring coordinate and the low-dimensional reference coordinate.

[0139] A preset weight allocation ratio is established to meet the monitoring requirements of power module operation status. This ratio clearly defines the proportion of local structural deviation data and global deviation data in the comprehensive status assessment. Local structural deviation data is matched with corresponding preset weight values ​​and multiplied to obtain a local deviation weight value. Similarly, global deviation data is matched with corresponding preset weight values ​​and multiplied to obtain a global deviation weight value. The local and global deviation weight values ​​are then summed to obtain a comprehensive assessment value. This comprehensive assessment value is then converted into a standardized value conforming to the power module status assessment standard according to a preset index conversion rule. This standardized value is the power module's status anomaly index. The preset index conversion rule is a fixed calculation method: Status Anomaly Index = Comprehensive Assessment Value × 100, and the converted status anomaly index ranges from [0, 100].

[0140] The system retrieves the confirmed dynamic adaptive threshold of the power module, sets up a comparison and judgment rule for abnormal status, and directly compares the obtained power module status abnormality index with the dynamic adaptive threshold. If the status abnormality index is less than the dynamic adaptive threshold, the power module is determined to be in normal operating condition, and a no-warning feedback signal is generated to indicate normal operating condition. If the status abnormality index is greater than or equal to the dynamic adaptive threshold, the power module is determined to be in abnormal operating condition. According to the preset warning signal generation format, the system integrates core information such as the status abnormality index value, local and global deviation data, and generates a power module status abnormality warning signal with complete abnormality characterization information.

[0141] The beneficial effects are as follows: Geometric deviation quantification of low-dimensional monitoring coordinates is carried out based on the existing neighborhood topology of low-dimensional reference coordinates. Combined with preset quantification rules, local structural deviation data is obtained, allowing local deviation analysis to closely match the topological distribution characteristics of the low-dimensional reference coordinates. This accurately captures subtle deviations in the local structure of the power module's operating status, ensuring the relevance and accuracy of the local deviation data. Global deviation data is calculated and corrected using preset unified spatial distance measurement rules, achieving quantification of the overall spatial offset of the low-dimensional monitoring coordinates relative to all low-dimensional reference coordinates. This comprehensively reflects the overall deviation degree of the power module's operating status, overcoming the limitations of local analysis. Based on preset weight allocation ratios, a weighted comprehensive evaluation of local and global deviation data is performed. The resulting state anomaly index takes into account both the local structural characteristics and overall spatial distribution characteristics of the power module's operating status, achieving a comprehensive and integrated quantification of the state deviation degree. This allows the anomaly index to truly and accurately reflect the actual operating status of the power module. The abnormal state index is compared with a standardized value using a dynamic adaptive threshold. Corresponding feedback signals or abnormal warning signals are generated according to preset rules, providing a clear and unified basis for anomaly determination and avoiding subjective bias. The generated abnormal warning signals integrate core abnormal information, providing a clear reference for the state management of power modules. Each step of the process is implemented based on preset rules, proportions, and standards. All operations have specific and feasible implementation methods, with no black-box behavior, fully ensuring the reproducibility of the technical solution. The products generated in each step serve as the basis for subsequent steps, forming a logically coherent and data-interoperable complete analysis process. This effectively improves the accuracy of power module state deviation analysis and the reliability of abnormal state determination, making the generation of abnormal warning signals more real-time and targeted. It provides solid technical support for the accurate monitoring of power module operating status and early fault prediction.

[0142] like Figure 2 The diagram shown is a functional block diagram of a power module operation status monitoring system provided in an embodiment of the present invention.

[0143] The power module operation status monitoring system 100 described in this invention can be installed in an electronic device. Depending on the functions implemented, the power module operation status monitoring system 100 may include a multi-frequency calibration module 101, a feature extraction module 102, a benchmark construction module 103, an online mapping module 104, a threshold learning module 105, and a status diagnosis module 106. The module described in this invention can also be referred to as a unit, which refers to a series of computer program segments that can be executed by the processor of an electronic device and can perform a fixed function, and which are stored in the memory of the electronic device.

[0144] In this embodiment, the functions of each module / unit are as follows:

[0145] The multi-frequency calibration module 101 is used to perform multi-frequency calibration on the high-frequency electrical signal of the power module to obtain a standardized multi-dimensional electrical signal of the high-frequency electrical signal;

[0146] The feature extraction module 102 is used to perform variational mode decomposition on the standardized multidimensional electrical signal and perform Hilbert transform on the decomposed finite bandwidth mode components to obtain the high-dimensional state feature vector of the power module.

[0147] The benchmark construction module 103 is used to perform local linear embedding dimensionality reduction on the high-dimensional state feature vector to obtain the low-dimensional benchmark coordinates of the high-dimensional state feature vector.

[0148] The online mapping module 104 is used to map the current high-dimensional state feature vector of the power module to the low-dimensional intrinsic space of the low-dimensional reference coordinates to obtain the low-dimensional monitoring coordinates of the current high-dimensional state feature vector.

[0149] The threshold learning module 105 is used to construct a decision graph of the low-dimensional reference coordinates based on the local density and relative distance of the low-dimensional reference coordinates, and to perform cluster center spacing analysis on the boundary points in the decision graph to obtain the dynamic adaptive threshold of the power module.

[0150] The state diagnosis module 106 is used to analyze the state deviation of the power module based on the low-dimensional reference coordinates and the low-dimensional monitoring coordinates, and to compare the analyzed state anomaly index with the dynamic adaptive threshold to generate a state anomaly warning signal for the power module.

[0151] In the several embodiments provided by this invention, it should be understood that the disclosed methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and other division methods may be used in actual implementation.

[0152] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0153] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.

[0154] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0155] This application embodiment can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0156] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for monitoring the operating status of a power module, characterized in that, The method includes: S1. Perform multi-frequency calibration on the high-frequency electrical signal of the power module to obtain a standardized multi-dimensional electrical signal of the high-frequency electrical signal; S2. Perform variational mode decomposition on the standardized multidimensional electrical signal, and perform Hilbert transform on the decomposed finite bandwidth mode components to obtain the high-dimensional state feature vector of the power module. S3. Perform local linear embedding dimensionality reduction on the high-dimensional state feature vector to obtain the low-dimensional reference coordinates of the high-dimensional state feature vector; S4. Map the current high-dimensional state feature vector of the power module to the low-dimensional intrinsic space of the low-dimensional reference coordinates to obtain the low-dimensional monitoring coordinates of the current high-dimensional state feature vector. S5. Based on the local density and relative distance of the low-dimensional reference coordinates, construct a decision graph for the low-dimensional reference coordinates, and perform clustering center spacing analysis on the boundary points in the decision graph to obtain the dynamic adaptive threshold of the power module, including: Based on the distribution characteristics of the decision graph, the boundary points and cluster centers in the decision graph are selected. Perform nearest neighbor analysis on the reference coordinates of the boundary points and the reference coordinates of the cluster centers to determine the coordinates of the cluster centers to which the boundary points belong. Based on the correspondence between the reference coordinates of the boundary points and the coordinates of the assigned cluster centers, the Euclidean distance between the reference coordinates of the boundary points and the coordinates of the assigned cluster centers is calculated. Within a preset sliding time window, the Euclidean distance is updated by extreme value tracking to confirm the dynamic adaptive threshold of the power module. S6. Based on the low-dimensional reference coordinates and the low-dimensional monitoring coordinates, perform a state deviation analysis on the power module, and compare the analyzed state anomaly index with the dynamic adaptive threshold to generate a state anomaly warning signal for the power module.

2. The method for monitoring the operating status of a power module as described in claim 1, characterized in that, The process of performing multi-frequency calibration on the high-frequency electrical signals of the power module to obtain standardized multi-dimensional electrical signals of the high-frequency electrical signals includes: The high-frequency electrical signal of the power module is bandpass filtered to obtain the fundamental signal component and harmonic signal component of the high-frequency electrical signal; The fundamental signal component and the harmonic signal component are subjected to amplitude normalization processing, and the normalized sub-signals are time-stamp aligned to obtain the multi-band signal of the high-frequency electrical signal. According to the preset channel sequence, the multi-band signal is recombined into a multi-dimensional data to obtain a standardized multi-dimensional electrical signal of the high-frequency electrical signal.

3. The method for monitoring the operating status of a power module as described in claim 1, characterized in that, The step of performing variational mode decomposition on the standardized multidimensional electrical signal and applying Hilbert transform to the decomposed finite-bandwidth mode components to obtain the high-dimensional state feature vector of the power module includes: Based on the preset number of modal decompositions, the standardized multidimensional electrical signal is subjected to frequency domain iterative decomposition to obtain the initial modal components of the standardized multidimensional electrical signal. The initial modal components are updated by convergence determination to obtain the finite bandwidth intrinsic mode functions of the standardized multidimensional electrical signal; The instantaneous amplitude and instantaneous frequency of the finite bandwidth intrinsic mode function are analyzed, and the time-frequency integral mapping of the finite bandwidth intrinsic mode function is performed to obtain the Hilbert marginal spectrum of the finite bandwidth intrinsic mode function; The energy distribution of the Hilbert marginal spectrum is analyzed to obtain the energy entropy and spectral centroid of the Hilbert marginal spectrum; Multidimensional feature fusion is performed on the energy entropy and the spectral centroid to obtain the high-dimensional state feature vector of the power module.

4. The method for monitoring the operating status of a power module as described in claim 1, characterized in that, The step of performing local linear embedding dimensionality reduction on the high-dimensional state feature vector to obtain the low-dimensional reference coordinates of the high-dimensional state feature vector includes: Obtain the normal high-dimensional state feature vector within the historical normal operation period from the high-dimensional state feature vector; Topological association analysis is performed on the normal high-dimensional state feature vectors to obtain the neighborhood relationships between the normal high-dimensional state feature vectors; Based on the neighborhood relationship, the normal high-dimensional state feature vector is locally linearly reconstructed to obtain the weight coefficient matrix of the normal high-dimensional state feature vector; Based on the weight coefficient matrix, the normal high-dimensional state feature vector is embedded and mapped to obtain the low-dimensional reference coordinates of the normal high-dimensional state feature vector.

5. The method for monitoring the operating status of a power module as described in claim 4, characterized in that, The step of embedding and mapping the normal high-dimensional state feature vector based on the weight coefficient matrix to obtain the low-dimensional reference coordinates of the normal high-dimensional state feature vector includes: The weight coefficient matrix is ​​constrained and encoded to obtain the cost matrix of the normal high-dimensional state feature vector; Perform generalized eigenvalue decomposition on the cost matrix to obtain the eigenvectors of the normal high-dimensional state eigenvectors; A basis system is constructed on the feature vectors to obtain the embedding coordinate matrix of the feature vectors; Based on the embedded coordinate matrix, the low-dimensional reference coordinates of the normal high-dimensional state feature vector are confirmed.

6. The method for monitoring the operating status of a power module as described in claim 1, characterized in that, The step of constructing a decision graph for the low-dimensional reference coordinates based on the local density and relative distance of the low-dimensional reference coordinates includes: The local density of the low-dimensional reference coordinates is obtained by performing a distribution density measurement on the low-dimensional reference coordinates. The relative distance between the low-dimensional reference coordinates is obtained by performing spacing parameter analysis on the low-dimensional reference coordinates. Using the local density as the horizontal axis parameter and the relative distance as the vertical axis parameter, the local density value and relative distance value of the low-dimensional reference coordinates are mapped to a two-dimensional coordinate system to obtain the initial decision map of the low-dimensional reference coordinates. The initial decision graph is interpreted using graph structure interpretation to obtain the decision graph with the low-dimensional reference coordinates.

7. The method for monitoring the operating status of a power module as described in claim 1, characterized in that, The formula for calculating the Euclidean distance is as follows: ； In the formula, For the first Each boundary point at the sampling time With the The Euclidean distance between the cluster centers to which each boundary point belongs. For the index identifier of the boundary point, This serves as a time identifier for the sampling time. For the first Each boundary point at the sampling time The low-dimensional reference coordinates of the first dimensional components, For the first The cluster center to which each boundary point belongs at the sampling time The low-dimensional reference coordinates of the first dimensional components, Let be the embedding dimension of the low-dimensional eigenspace. For summation operations.

8. The method for monitoring the operating status of a power module as described in claim 1, characterized in that, The analysis of the state deviation of the power module based on the low-dimensional reference coordinates and the low-dimensional monitoring coordinates includes: Based on the neighborhood topology of the low-dimensional reference coordinates, geometric deviation quantization is performed on the low-dimensional monitoring coordinates to obtain local structural deviation data of the low-dimensional monitoring coordinates. Spatial distance measurement is performed between the low-dimensional monitoring coordinates and the low-dimensional reference coordinates to obtain global deviation data between the low-dimensional monitoring coordinates and the low-dimensional reference coordinates; Based on the local structural deviation data and the global deviation data, a comprehensive status assessment of the power module is performed to obtain the status anomaly index of the power module.

9. A power module operation status monitoring system, characterized in that, The system for implementing the power module operation status monitoring method according to claim 1 includes: A multi-frequency calibration module is used to perform multi-frequency calibration on the high-frequency electrical signals of the power module to obtain standardized multi-dimensional electrical signals of the high-frequency electrical signals; The feature extraction module is used to perform variational mode decomposition on the standardized multidimensional electrical signal and perform Hilbert transform on the decomposed finite bandwidth mode components to obtain the high-dimensional state feature vector of the power module. The benchmark construction module is used to perform local linear embedding dimensionality reduction on the high-dimensional state feature vector to obtain the low-dimensional benchmark coordinates of the high-dimensional state feature vector. The online mapping module is used to map the current high-dimensional state feature vector of the power module to the low-dimensional intrinsic space of the low-dimensional reference coordinates to obtain the low-dimensional monitoring coordinates of the current high-dimensional state feature vector. A threshold learning module is used to construct a decision graph of the low-dimensional reference coordinates based on the local density and relative distance of the low-dimensional reference coordinates, and to perform clustering center spacing analysis on the boundary points in the decision graph to obtain the dynamic adaptive threshold of the power module, including: Based on the distribution characteristics of the decision graph, the boundary points and cluster centers in the decision graph are selected. Perform nearest neighbor analysis on the reference coordinates of the boundary points and the reference coordinates of the cluster centers to determine the coordinates of the cluster centers to which the boundary points belong. Based on the correspondence between the reference coordinates of the boundary points and the coordinates of the assigned cluster centers, the Euclidean distance between the reference coordinates of the boundary points and the coordinates of the assigned cluster centers is calculated. Within a preset sliding time window, the Euclidean distance is updated by extreme value tracking to confirm the dynamic adaptive threshold of the power module. The status diagnosis module is used to analyze the degree of status deviation of the power module based on the low-dimensional reference coordinates and the low-dimensional monitoring coordinates, and to compare the difference between the analyzed status anomaly index and the dynamic adaptive threshold to generate a status anomaly warning signal for the power module.

Images