Substation equipment state monitoring embedded edge analysis system

Abstract

The invention relates to the technical field of power system equipment monitoring, in particular to a substation equipment state monitoring embedded edge analysis system, which comprises a monitoring front-end module for acquiring and transmitting substation equipment operation state data; the edge processing module comprises a multi-scale wavelet transform unit, a topology continuity analysis unit and a manifold learning unit, wherein the multi-scale wavelet transform unit decomposes data and separates signals and noise through a self-adaptive threshold mechanism; the topological persistence analysis unit maps the de-noised signal to a high-dimensional feature space, calculates a persistent coherence group, and extracts a topological feature vector; and the manifold learning unit constructs a nonlinear manifold structure, calculates a geodesic distance, and executes discharge type identification and anomaly evaluation, and the FPGA and ARM heterogeneous calculation architecture is adopted, and through the synergistic effect of multi-scale wavelet transform, topology persistence analysis and manifold learning, high-precision extraction of weak discharge signals and accurate identification of fault modes are realized.

Description

Technical Field

[0001] This invention relates to the field of power system equipment monitoring technology, specifically to an embedded edge analysis system for substation equipment condition monitoring, which is applied to preventive maintenance and fault diagnosis of substation equipment. Background Technology

[0002] Substations are critical nodes in the power system, and their safe and stable operation is essential for the reliability of the entire power grid. Substation equipment, such as transformers, circuit breakers, and GIS equipment, are prone to problems like insulation aging and partial discharge during long-term operation. If these issues are not detected and addressed in a timely manner, they may lead to equipment damage or even power grid accidents.

[0003] Existing substation equipment monitoring systems are mainly based on periodic detection and centralized data analysis, which have the following shortcomings: First, the data acquisition frequency is low, making it impossible to capture transient abnormal signals; second, a large amount of raw data needs to be transmitted to a remote server for processing, causing network bandwidth pressure and processing delays; third, traditional analysis methods are difficult to effectively identify weak discharge signals, resulting in high false alarm and false negative rates; and finally, there is a lack of intelligent anomaly identification and fault early warning mechanisms, making it impossible to accurately assess the health status of equipment.

[0004] With the development of edge computing technology, moving data processing and analysis functions closer to the data source can effectively reduce transmission latency and improve real-time performance. However, existing edge analysis systems still face technical bottlenecks in weak signal feature extraction and complex discharge mode identification, especially in the complex electromagnetic environment of substations. How to accurately separate noise from effective signals and how to identify potential fault modes from multi-dimensional features remain key issues that urgently need to be addressed. Summary of the Invention

[0005] The purpose of this invention is to provide an embedded edge analysis system for substation equipment condition monitoring. Through innovative multi-scale wavelet transform, topological persistence analysis, and manifold learning techniques, it achieves high-precision extraction of partial discharge signals from substation equipment and accurate identification of fault modes, thereby improving the accuracy of equipment condition monitoring and the timeliness of early warning.

[0006] This invention proposes an embedded edge analysis system for substation equipment condition monitoring, comprising:

[0007] The monitoring front-end module is used to collect substation equipment operating status data and encode and transmit the substation equipment operating status data;

[0008] An edge processing module is connected to the monitoring front-end module via a dedicated power fiber optic line and is used to receive the substation equipment operation status data. The edge processing module includes a multi-scale wavelet transform unit, a topology persistence analysis unit, and a manifold learning unit.

[0009] The multi-scale wavelet transform unit is used to perform multi-scale decomposition on the substation equipment operating status data to form a wavelet coefficient matrix, and to separate the signal and noise through an adaptive threshold mechanism to generate a denoised partial discharge signal.

[0010] The topological persistence analysis unit is connected to the multi-scale wavelet transform unit and is used to receive the denoised partial discharge signal, map the denoised partial discharge signal to a high-dimensional feature space to form point cloud data, calculate the persistent homology group based on the point cloud data, generate a persistence map characterizing the topological features, and extract the topological feature vector from the persistence map.

[0011] The manifold learning unit is connected to the topology persistent analysis unit and is used to receive the topology feature vector, construct a nonlinear manifold structure, calculate the geodesic distance of the topology feature vector on the nonlinear manifold structure, perform discharge type identification and anomaly degree assessment based on the geodesic distance, and generate equipment health status diagnosis results.

[0012] The application service module, connected to the edge processing module, is used to receive the device health status diagnosis results and generate early warning information and maintenance suggestions.

[0013] Preferably, the monitoring front-end module includes:

[0014] The sensor unit is used to collect partial discharge signals, temperature signals, vibration signals, and sound signals from substation equipment;

[0015] A data acquisition unit, connected to the sensor unit, is used to perform high-precision sampling of the partial discharge signal, temperature signal, vibration signal, and sound signal. The data acquisition unit adopts a microsecond-level precise timing mechanism and supports millisecond-level monitoring data acquisition time intervals.

[0016] A data encoding unit, connected to the data acquisition unit, is used to encode and encapsulate the sampled data to form standardized data packets;

[0017] A communication interface unit, connected to the data encoding unit, is used to transmit the standardized data packets to the edge processing module via a dedicated power fiber optic line.

[0018] Preferably, the multi-scale wavelet transform unit includes:

[0019] A signal preprocessor is used to standardize and divide the operating status data of the substation equipment into time windows.

[0020] The wavelet decomposition engine, connected to the signal preprocessor, is used to automatically select the optimal wavelet basis function according to the signal characteristics, perform multi-scale decomposition on the substation equipment operating status data, and generate the wavelet coefficient matrix.

[0021] An adaptive threshold generator, connected to the wavelet decomposition engine, is used to estimate the background noise level based on the signal statistical characteristics and generate dynamic thresholds for different decomposition levels.

[0022] A signal reconstructor, connected to the wavelet decomposition engine and the adaptive threshold generator, is used to filter and retain meaningful wavelet coefficients according to the dynamic threshold, and generate the denoised partial discharge signal through inverse wavelet transform.

[0023] Preferably, the topology persistence analysis unit includes:

[0024] A feature space builder is used to map the denoised partial discharge signal to a high-dimensional feature space through a time delay embedding method to generate the point cloud data.

[0025] The filter function library stores various filter functions, including energy density functions, frequency distribution functions, waveform morphology functions, and phase relationship functions.

[0026] A continuous homology calculation engine, connected to the feature space builder and the filter function library, is used to construct a multi-scale simple complex sequence based on the point cloud data and the filter functions in the filter function library, and to calculate the continuous homology group in different dimensions.

[0027] The feature vectorizer, connected to the persistent cohomology computation engine, is used to extract persistent Barker codes from the persistent graph, compute topological invariants, and generate the topological feature vector.

[0028] Preferably, the manifold learning unit includes:

[0029] A manifold structure learner is used to construct a K-nearest neighbor graph based on the topological feature vector, analyze the geometric structure of the local neighborhood, optimize the global coordinate system, and generate coordinate representations of data points on the nonlinear manifold structure.

[0030] A geodesic distance calculator, connected to the manifold structure learner, is used to construct a weighted graph structure based on the K-nearest neighbor graph, execute the shortest path algorithm on the graph, and calculate the geodesic distance between any two points.

[0031] A dynamic template manager is used to maintain a reference template library for different discharge types and dynamically update the reference template library based on newly acquired data.

[0032] An anomaly detection evaluator, connected to the geodesic distance calculator and the dynamic template manager, is used to calculate the deviation of new data points from known manifolds, analyze the local density distribution around the data points, comprehensively evaluate the degree of anomaly, and generate the device health status diagnosis results.

[0033] Preferably, the edge processing module adopts a heterogeneous computing architecture of FPGA and ARM, wherein:

[0034] The FPGA implements the computationally intensive tasks in the multi-scale wavelet transform unit and the topology persistence analysis unit;

[0035] The ARM processor implements the manifold learning unit and system control tasks;

[0036] The edge processing module also includes a task scheduler for dynamically allocating computing tasks between the FPGA and the ARM processor based on priority and resource utilization efficiency.

[0037] Preferably, the persistent cohomology calculation engine in the topology persistence analysis unit performs the following processing:

[0038] Multiple filtering functions are determined to describe the features of the point cloud data from different perspectives;

[0039] Construct a simple complex sequence by changing the filter value;

[0040] Track the appearance and disappearance of topological features and record them as a persistent map;

[0041] Statistical information and topological invariants are extracted from the persistent graph to form the topological feature vector.

[0042] Preferably, the manifold learning unit includes an adaptive learning mechanism, which is used for:

[0043] The nonlinear manifold structure is continuously updated from new data through incremental learning.

[0044] Detect and adapt to characteristic changes caused by equipment aging;

[0045] The algorithm complexity is dynamically adjusted based on computing resource availability.

[0046] The model parameters were adjusted based on the diagnostic results confirmed by humans.

[0047] Preferably, the system further includes a data storage module, which is connected to the edge processing module and is used for:

[0048] Store raw monitoring data and process intermediate results;

[0049] Save historical diagnostic results and trends in equipment health status;

[0050] Maintain the discharge type reference template library;

[0051] Provides data query and retrospective analysis functions.

[0052] Preferably, the edge processing module further includes a resource adaptive management unit, which is used for:

[0053] Monitor system operating status and resource usage;

[0054] Dynamically allocate computing resources based on task priority and urgency;

[0055] Reduce feature dimensions and processing accuracy when computational resources are limited;

[0056] Adjust the monitoring frequency and data processing frequency according to the equipment status to achieve efficient utilization of system resources.

[0057] The present invention has the following beneficial effects:

[0058] 1. By employing multi-scale wavelet transform and an adaptive threshold mechanism, weak partial discharge signals can be effectively separated from background noise, improving signal detection sensitivity and reducing false alarm rate. In actual testing, compared to the traditional fixed threshold method, the detection sensitivity is improved by approximately 30%, and the detection rate of weak discharges (less than 5 pC) is significantly enhanced.

[0059] 2. The topological persistence analysis method is adopted to extract topological features from the discharge signal. These features are robust to changes in signal intensity and time scale, and can characterize the essential structural features of the discharge signal, providing a more reliable basis for fault identification.

[0060] 3. By introducing manifold learning technology, a nonlinear representation space for discharge modes is constructed. The similarity of discharge modes is measured by geodesic distance. Compared with the traditional Euclidean distance method, the accuracy of fault type identification is improved from 80% to over 95%, especially the ability to distinguish between discharge modes of similar types but different severity is significantly enhanced.

[0061] 4. By adopting a heterogeneous computing architecture of FPGA and ARM, the computing resources are utilized efficiently, and complex algorithms are calculated in real time at the edge with a processing latency of less than 200ms, which meets the real-time monitoring requirements of substation equipment status.

[0062] 5. It has an adaptive learning mechanism that can continuously learn from new data and optimize the model to adapt to feature changes during equipment aging and maintain the accuracy and effectiveness of long-term monitoring.

[0063] Through the above innovations, the system of this invention can detect anomalies in the early stages of fault development, providing strong support for preventive maintenance, extending equipment life by 15% to 30%, reducing unexpected power outages, and significantly improving the safety and reliability of power grid operation. Attached Figure Description

[0064] Figure 1 This is an overall architecture diagram of the embedded edge analysis system for substation equipment condition monitoring of the present invention;

[0065] Figure 2 This is a structural block diagram of the monitoring front-end module of the present invention;

[0066] Figure 3 This is a flowchart of the processing of the multi-scale wavelet transform unit of the present invention;

[0067] Figure 4 This is a flowchart of the topology persistence analysis unit of the present invention;

[0068] Figure 5 This is a flowchart of the processing of the manifold learning unit of the present invention. Detailed Implementation

[0069] Please refer to Figures 1-5 The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Those skilled in the art should understand that the following embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0070] like Figure 1 As shown, the substation equipment status monitoring embedded edge analysis system provided by the present invention mainly includes a monitoring front-end module 1, an edge processing module 2, an application service module 3, and a data storage module 4.

[0071] The monitoring front-end module 1 is deployed on the equipment side of the substation to collect equipment operating status data; the edge processing module 2 is connected to the monitoring front-end module 1 through a dedicated power fiber optic line to receive and process monitoring data and perform fault diagnosis; the application service module 3 receives the diagnostic results from the edge processing module 2 and generates early warning information and maintenance suggestions; the data storage module 4 is responsible for storing raw monitoring data, processing intermediate results and historical diagnostic results, and supports data query and retrospective analysis.

[0072] like Figure 2 As shown, the monitoring front-end module 1 includes a sensor unit 11, a data acquisition unit 12, a data encoding unit 13, and a communication interface unit 14.

[0073] Sensor unit 11 includes a partial discharge sensor, a temperature sensor, a vibration sensor, and a sound sensor. It is deployed on key equipment such as transformers, circuit breakers, and GIS equipment within the substation to collect partial discharge signals, temperature signals, vibration signals, and sound signals from the equipment. Preferably, the partial discharge sensor is an ultra-high frequency (UHF) sensor with an operating frequency range of 300MHz-3GHz, capable of effectively capturing discharge pulse signals generated inside the substation equipment.

[0074] The data acquisition unit 12 is connected to the sensor unit 11 and performs high-precision sampling, converting analog signals into digital signals. In this invention, the data acquisition unit 12 employs a microsecond-level precise timing mechanism, supporting millisecond-level monitoring data acquisition time intervals. For partial discharge signals, the sampling rate is set to 100 MS / s (megasamples / second) to capture nanosecond-level discharge pulses; for temperature signals, the sampling interval is 1 minute; for vibration and sound signals, the sampling rate is set to 20 kS / s (kilosamples / second).

[0075] The data encoding unit 13 receives the digital signals from the data acquisition unit 12, encodes and encapsulates them to form standardized data packets. In one embodiment of the present invention, the data encoding unit 13 uses differential encoding for some discharge signals to reduce the amount of data; and uses incremental encoding for slowly changing signals such as temperature and vibration. The data packet format includes: device ID (4 bytes), signal type (2 bytes), timestamp (8 bytes), data length (4 bytes), and data body (variable length).

[0076] The communication interface unit 14 is connected to the data encoding unit 13, and transmits standardized data packets to the edge processing module 2 via a dedicated power fiber optic line. This invention employs a power system-specific fiber optic communication protocol with a transmission rate of 10Gbps, ensuring real-time transmission and security of monitoring data.

[0077] Edge processing module 2 is the core component of this invention, including a multi-scale wavelet transform unit 21, a topology persistence analysis unit 22, a manifold learning unit 23, and a resource adaptive management unit 24. Edge processing module 2 adopts a heterogeneous computing architecture of FPGA and ARM, where the FPGA is responsible for implementing computationally intensive tasks, and the ARM processor is responsible for system control and some algorithm execution.

[0078] The multi-scale wavelet transform unit 21 includes a signal preprocessor 211, a wavelet decomposition engine 212, an adaptive threshold generator 213, and a signal reconstructor 214.

[0079] The signal preprocessor 211 performs standardization processing and time window division on the received substation equipment operating status data. The standardization processing adopts the zero-mean normalization method to adjust the signal amplitude to the range of [-1,1]; the time window division adopts the sliding window technique with a window length of 10ms and an overlap rate of 50% to ensure the continuous analysis of the signal.

[0080] The wavelet decomposition engine 212 automatically selects the optimal wavelet basis function based on signal characteristics, performs multi-scale decomposition, and generates a wavelet coefficient matrix. In this invention, considering the characteristics of some discharge signals, the Daubechies wavelet (db4) is preferably used as the basis function, and the decomposition level is set to 5-8 levels, adaptively adjusted according to the signal frequency characteristics. The wavelet decomposition process can be represented as follows:

[0081] ,

[0082] in: The wavelet coefficients represent the scale. and location The coefficient value at the location; The input signal represents the discrete time points. The amplitude at that point; For wavelet basis functions at scale and location The scaling and translation form is defined as follows: ; It is the wavelet mother function; Represents all discrete time points Summation; To decompose into levels, the value range is: to ; This is a position parameter, and its value range depends on the signal length.

[0083] The adaptive threshold generator 213 estimates the background noise level based on signal statistical characteristics and generates dynamic thresholds for different decomposition levels. In one embodiment of the invention, an improved Donoho thresholding method is used, and the threshold calculation formula is as follows:

[0084] ,

[0085] in: For the first The threshold of the layer; For the first The standard deviation of the layer wavelet coefficients is estimated using the median absolute deviation (MAD) method, and the calculation formula is as follows: ; For the first Layer wavelet coefficients The median absolute deviation; This is the signal length, i.e., the number of sampling points in the signal; To adjust the coefficients, the high-frequency layer takes a larger value (approximately). - The low-frequency layer has smaller values ​​(approximately). - ), to adapt to the characteristics of different frequency components; It is the natural logarithm; This represents the square root operation.

[0086] The signal reconstructor 214 selects meaningful wavelet coefficients based on a dynamic threshold and generates a denoised partial discharge signal through inverse wavelet transform. The coefficient selection uses a soft thresholding method.

[0087] ,

[0088] in: These are the filtered wavelet coefficients; This is a sign function that returns 1 when the input is positive, -1 when the input is negative, and zero. Represents wavelet coefficients The absolute value; For the first The threshold of the layer.

[0089] The signal reconstruction process is as follows:

[0090] ,

[0091] in: For the reconstructed denoised signal, let n represent the amplitude at discrete time point n. This represents summing over all decomposition levels j; This represents the summation over all positional parameters k; These are the filtered wavelet coefficients; These are wavelet basis functions.

[0092] The topology persistence analysis unit 22 is the key innovation of this invention, which includes a feature space builder 221, a filter function library 222, a persistence coherence calculation engine 223, and a feature vectorizer 224.

[0093] The feature space builder 221 maps the denoised partial discharge signal to a high-dimensional feature space using a time-delay embedding method, generating point cloud data. The mathematical expression of time-delay embedding is:

[0094] ,

[0095] in: Let be the embedded d-dimensional vector, representing the feature point at time i; The value of the denoised signal at time point i; d is the time delay parameter, representing the time interval between adjacent sampling points; d is the embedding dimension, representing the dimension of the constructed feature vector; i is the time index, ranging from 1 to... Where N is the total signal length. In this embodiment of the invention, the optimal time delay is determined using a mutual information method. The value typically ranges from 5 to 20 sampling points; the embedding dimension d is determined by the pseudo-nearest neighbor method, with a typical value of 10-30.

[0096] The filter function library 222 stores various filter functions for describing the characteristics of point cloud data from different perspectives. The following filter functions are included in this invention:

[0097] Energy density function: , describes the signal energy distribution; where Representing vectors The Euclidean norm, i.e. , For vectors The Each component.

[0098] Frequency distribution function: FFT The dominant frequency component describes the spectral characteristics; where FFT... Represents a vector Performing a Fast Fourier Transform (FFT), the dominant frequency component refers to the FFT. The frequency component with the largest mid-amplitude.

[0099] Waveform shape function: , describes the steepness of the waveform; where For vectors The maximum value in, For vectors The minimum value in, For vectors The standard deviation of is calculated using the following formula: For vectors The mean, calculated using the formula is: .

[0100] Phase relation function: , describes the phase relationship; where and They are vectors The The and the first One portion, For the arctangent function, it is usually chosen and Make the time difference between them be To capture the phase relationship of a fraction of a period.

[0101] The Continuous Homology Computation Engine 223 constructs multi-scale simple complex sequences based on point cloud data and filtering functions to compute continuous homology groups of different dimensions. Specific steps include:

[0102] (1) For a given filter function and threshold Construct the Vietoris-Rips complex:

[0103] ,

[0104] in: The Vietoris-Rips complex constructed at the threshold ε; Let k be a simplex with k+1 vertices. composition; For point and The Euclidean distance between them is calculated using the following formula: ,in Point The l-th coordinate component; Indicates all pairs and All meet the distance condition .

[0105] (2) As the threshold ε changes from small to large, a nested complex sequence is formed:

[0106] ,

[0107] in: This represents a series of increasing thresholds; Indicates at the threshold The complex constructed below the threshold is A subset of the complex constructed below, when hour.

[0108] (3) Calculate the homology group of each complex, track the appearance and disappearance of topological features (such as connected components, cycles, holes, etc.), and generate a persistence graph. Homology group calculation involves complex algebraic topology theory, and the specific implementation uses a standard computational topology algorithm library.

[0109] In embodiments of the present invention, continuous homology calculation typically sets a threshold range. With a step size of 0.05, calculate the continuous homology group in 0-2 dimensions. 0-dimensional homology group. Corresponding connected components, 1D homology group Corresponding to the ring structure (or "hole"), 2D homology group Corresponding to a cavity structure (or "cavity").

[0110] The feature vectorizer 224 extracts persistent Barker codes from the persistent graph, computes topological invariants, and generates topological feature vectors. The persistent Barker codes record the lifecycle of the topological features.

[0111] ,

[0112] in: For a persistent set of Barker codes; Indicates the first The birth and death points of each topological feature. This is the threshold (birth point) at which this feature appears. This is the threshold (death point) at which the feature disappears. This represents the total number of topological features.

[0113] Based on persistent Barker codes, compute the following topological invariants:

[0114] Betty number sequence: , representing the number of 0-dimensional, 1-dimensional, and 2-dimensional topological features, respectively; Betty number Defined as the first Conflict groups Rank.

[0115] Persistent entropy: ,in ; Indicates the first The proportion of the duration of each feature in the total duration; Represents the natural logarithm; Indicates all features Summation; Indicates all features Sum.

[0116] Lifespan Distribution: Statistics The distribution characteristics, such as mean, variance, and skewness. The formula for calculating the mean is: The formula for calculating variance is: The formula for calculating skewness is: .

[0117] The final generated topological feature vector is in the form of:

[0118] ,

[0119] in: These are topological feature vectors; Betti numbers for 0-dimensional, 1-dimensional, and 2-dimensional topological features; The persistent entropy represents the topological features of 0-dimensional, 1-dimensional, and 2-dimensional dimensions, respectively. These represent the mean, variance, and skewness of the lifetime, respectively. The vector dimension is typically 15-30, but can be adjusted depending on the specific application requirements.

[0120] The manifold learning unit 23 includes a manifold structure learner 231, a geodesic distance calculator 232, a dynamic template manager 233, and an anomaly detection evaluator 234.

[0121] The manifold structure learner 231 constructs a K-nearest neighbor graph based on topological feature vectors, analyzes the geometric structure of local neighborhoods, optimizes the global coordinate system, and generates coordinate representations of data points on the nonlinear manifold structure. In embodiments of the present invention, the locally linear embedding (LLE) algorithm is used to implement manifold learning, and the main steps include:

[0122] (1) Constructing a K-nearest neighbor graph: For each data point Find its K nearest neighbors. ,in This is a neighborhood index set. In this invention, the value of K is typically set to 10-20, adaptively adjusted according to data density. This represents the topological feature vector of the i-th data point; Representation and data points The set of indices of the K most recent points.

[0123] (2) Calculate local reconstruction weights: For each data point, calculate the weight coefficients for linear reconstruction of that point using its neighboring points by solving the following optimization problem:

[0124] ,

[0125] in: This is the weight matrix. The weighting coefficient represents the point. At the reconstruction point Contribution weight at any time; This indicates that for all data points Summation; Indicates a point All neighboring points Summation; Denotes the Euclidean norm; This indicates the search for the weight matrix that minimizes the objective function. Weighting coefficients The following constraints must be met: , and when hour, .

[0126] (3) Calculate the low-dimensional embedding coordinates: Based on the reconstruction weights, solve the following optimization problem:

[0127] ,

[0128] in: It is a low-dimensional embedded coordinate matrix; For data points The coordinate representation on a low-dimensional manifold is a A dimensional vector, where The dimension after dimensionality reduction is typically 2-5. The weighting coefficients are those calculated in step (2); This indicates that for all data points Summation; Indicates a point All neighboring points Summation; Represents the norm; This indicates the search for a low-dimensional coordinate matrix that minimizes the objective function. .

[0129] The geodesic distance calculator 232 constructs a weighted graph structure based on a K-nearest neighbor graph and executes the shortest path algorithm on the graph to calculate the geodesic distance between any two points. In this invention, Dijkstra's algorithm is used to calculate the geodesic distance, and the edge weights are defined as follows:

[0130] ,

[0131] in: For nodes and nodes The edge weights between them, i.e., the Euclidean distance; and They are nodes and nodes The corresponding topological feature vector; This represents the Euclidean norm.

[0132] For any two points and Its geodesic distance is defined as:

[0133] ,

[0134] in: Point and points Geodesic distance between them; For connection and The path on the manifold is a parameterized curve. ,in Describe a manifold that satisfies and ; Representing a path In parameters The tangent vector at the point; The norm of the tangent vector represents the instantaneous velocity of the path at that point. Indicates along the path The line integral; Indicates all connections and We search for the path that minimizes the integral among the possible paths. In the discrete case, we approximate the calculation using Dijkstra's shortest path algorithm.

[0135] The dynamic template manager 233 maintains a reference template library for different discharge types and dynamically updates the library based on newly acquired data. The reference templates include feature vectors and manifold coordinates for typical discharge types such as internal discharge, surface discharge, corona discharge, and levitation discharge. Template updates employ an incremental learning approach; for confirmed samples, they are incorporated into existing templates with weights ranging from 0.1 to 0.3.

[0136] ,

[0137] in: The updated template is a vector with the same dimensions as the feature vector; The original template is a vector with the same dimension as the feature vector; The new sample features are vectors with the same dimensions as the feature vectors. To update the weight, it is a scalar with a value ranging from 0 to 1, usually set to 0.1-0.3.

[0138] The anomaly detection evaluator 234 calculates the deviation of new data points from the known manifold, analyzes the local density distribution around the data points, comprehensively assesses the degree of anomaly, and generates equipment health status diagnostic results. The anomaly score calculation formula is:

[0139] ,

[0140] in Anomaly ratings typically range from 0 to 1, with higher values ​​indicating a greater degree of anomaly. The distance to the nearest template is a scalar representing the Euclidean distance between the new data point and the nearest template in the reference template library; Local density is a scalar representing the density of data points surrounding a new data point. The calculation formula is: ,in For the feature vector of the new data point, For the data points in the training set, The kernel width parameter is typically set to the average distance of the training data. The geodesic distance to the normal region manifold is a scalar representing the shortest path distance from the new data point to the manifold formed by the normal data points; The weighting coefficients are set to 0.3, 0.3, and 0.4 respectively, and satisfy the following conditions: .

[0141] The criteria for grading the degree of abnormality are as follows:

[0142] normal: ;

[0143] Minor abnormalities: ;

[0144] Moderate abnormality: ;

[0145] Serious abnormality: ;

[0146] The edge processing module 2 also includes a resource adaptive management unit 24, which monitors the system's operating status and resource usage, and dynamically allocates computing resources according to task priority and urgency to achieve efficient utilization of system resources.

[0147] Specifically, the resource adaptive management unit 24 performs the following functions:

[0148] Monitor system resource status such as CPU / FPGA load, memory usage, and storage space;

[0149] Dynamically allocate computing resources based on task priority and urgency;

[0150] Reduce feature dimensions and processing accuracy when computational resources are limited;

[0151] Adjust the monitoring frequency and data processing frequency according to the equipment status;

[0152] In an embodiment of the present invention, the system sets a three-level resource scheduling strategy:

[0153] Normal mode: Full-function operation, monitoring frequency is set to standard.

[0154] Energy-saving mode: Reduces the calculation accuracy of non-critical functions and decreases the monitoring frequency;

[0155] Emergency Mode: Prioritize the operation of core functions and increase the monitoring frequency of critical equipment.

[0156] Application service module 3 is connected to edge processing module 2, receives device health status diagnosis results, and generates early warning information and maintenance suggestions.

[0157] The main functions of Application Service Module 3 include:

[0158] Receive and parse the diagnostic results from edge processing module 2;

[0159] Different levels of early warning information are generated based on the diagnostic results;

[0160] Based on equipment characteristics and historical data, provide targeted maintenance recommendations;

[0161] The device status and trend analysis are displayed through a visual interface.

[0162] In embodiments of the present invention, the warning levels are divided into four levels:

[0163] Information level: The equipment is operating normally and requires no special attention;

[0164] Warning level: Minor abnormalities have occurred; please pay attention.

[0165] Warning level: Moderate abnormality has occurred; it is recommended to arrange for examination.

[0166] Alarm level: A serious anomaly has occurred; immediate inspection is recommended.

[0167] Maintenance recommendations are generated based on different types of discharge faults and equipment characteristics, including specific details such as inspection items, maintenance cycles, and spare parts preparation.

[0168] Data storage module 4 is connected to edge processing module 2 and is used to store raw monitoring data, process intermediate results, historical diagnostic results and equipment health status change trends, and provide data query and retrospective analysis functions.

[0169] The main functions of data storage module 4 include:

[0170] Store raw monitoring data and process intermediate results;

[0171] Save historical diagnostic results and trends in equipment health status;

[0172] Maintain the discharge type reference template library;

[0173] Provides data query and retrospective analysis functions;

[0174] In embodiments of the present invention, the data employs a hierarchical storage strategy:

[0175] Hot data (last month): stored in high-speed memory on edge nodes;

[0176] Warm data (last year): stored on the local hard drive of the edge node;

[0177] Cold data (over 1 year): stored on cloud servers.

[0178] The data is organized using a time-series database structure, supporting efficient time-series queries and analysis. Separate indexes are created for important events (such as equipment malfunctions and failures) to facilitate rapid retrieval and correlation analysis.

[0179] The system of this invention also includes an adaptive learning mechanism, enabling the system to continuously learn and optimize from new data, adapting to equipment aging and environmental changes. The adaptive learning mechanism is mainly embodied in the manifold learning unit 23, specifically including:

[0180] Incremental learning: continuously updating the nonlinear manifold structure and reference template from new data;

[0181] Concept drift adaptation: Detecting and adapting to characteristic changes caused by equipment aging;

[0182] Computational resource awareness: Dynamically adjust algorithm complexity based on computing resource availability;

[0183] Feedback learning: Adjust model parameters based on manually confirmed diagnostic results.

[0184] In embodiments of the present invention, the system performs a model update weekly, adjusting model parameters based on newly collected data and manual verification results. For emerging new discharge patterns, they are added to the reference template library after expert confirmation, and the relevant feature weights are adjusted.

[0185] The system of this invention has been deployed and tested in multiple substations. Taking the main transformer monitoring of a 220kV substation as an example, the system deployed 8 UHF sensors, with a monitoring frequency set to 10 minutes / time, and operated continuously for 6 months under normal load conditions.

[0186] Test results show that the system successfully identified three internal weak discharge events, two of which went undetected during routine periodic inspections. By adjusting operating parameters promptly and implementing preventative maintenance, potential equipment failures were avoided. Compared to traditional monitoring methods, the main performance advantages of this invention's system include:

[0187] The sensitivity of weak discharge detection is improved by about 30%, and the minimum detectable discharge quantity is reduced to 5pC;

[0188] The accuracy rate of discharge type identification has been improved to over 95%.

[0189] The false alarm rate decreased from 15% to below 3%;

[0190] Processing latency is reduced to less than 200ms, enabling real-time monitoring and early warning.

[0191] In addition, the system's adaptive learning mechanism enables monitoring performance to gradually improve as the running time increases, with the fault prediction accuracy after 6 months being about 15% higher than that at the time of initial deployment.

[0192] In summary, the embedded edge analysis system for substation equipment condition monitoring provided by this invention achieves high-precision extraction of partial discharge signals and accurate identification of fault modes of substation equipment through innovative multi-scale wavelet transform, topological persistence analysis and manifold learning technologies. This improves the accuracy of equipment condition monitoring and the timeliness of early warning, providing a strong guarantee for the safe and stable operation of substation equipment.

[0193] The embodiments described above are merely illustrative of specific implementations of the present invention, and while the descriptions are detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. An embedded edge analysis system for substation equipment condition monitoring, characterized in that, include: The monitoring front-end module is used to collect substation equipment operating status data and encode and transmit the substation equipment operating status data; An edge processing module is connected to the monitoring front-end module via a dedicated power fiber optic line and is used to receive the substation equipment operation status data. The edge processing module includes a multi-scale wavelet transform unit, a topology persistence analysis unit, and a manifold learning unit. The multi-scale wavelet transform unit is used to perform multi-scale decomposition on the substation equipment operating status data to form a wavelet coefficient matrix, and to separate the signal and noise through an adaptive threshold mechanism to generate a denoised partial discharge signal. The topological persistence analysis unit is connected to the multi-scale wavelet transform unit and is used to receive the denoised partial discharge signal, map the denoised partial discharge signal to a high-dimensional feature space to form point cloud data, calculate the persistent homology group based on the point cloud data, generate a persistence map characterizing the topological features, and extract the topological feature vector from the persistence map. The manifold learning unit is connected to the topology persistent analysis unit and is used to receive the topology feature vector, construct a nonlinear manifold structure, calculate the geodesic distance of the topology feature vector on the nonlinear manifold structure, perform discharge type identification and anomaly degree assessment based on the geodesic distance, and generate equipment health status diagnosis results. The application service module, connected to the edge processing module, is used to receive the device health status diagnosis results and generate early warning information and maintenance suggestions.

2. The embedded edge analysis system for substation equipment condition monitoring according to claim 1, characterized in that, The monitoring front-end module includes: The sensor unit is used to collect partial discharge signals, temperature signals, vibration signals, and sound signals from substation equipment; A data acquisition unit, connected to the sensor unit, is used to perform high-precision sampling of the partial discharge signal, temperature signal, vibration signal, and sound signal. The data acquisition unit adopts a microsecond-level precise timing mechanism and supports millisecond-level monitoring data acquisition time intervals. A data encoding unit, connected to the data acquisition unit, is used to encode and encapsulate the sampled data to form standardized data packets; A communication interface unit, connected to the data encoding unit, is used to transmit the standardized data packets to the edge processing module via a dedicated power fiber optic line.

3. The embedded edge analysis system for substation equipment condition monitoring according to claim 1, characterized in that, The multi-scale wavelet transform unit includes: A signal preprocessor is used to standardize and divide the operating status data of the substation equipment into time windows. The wavelet decomposition engine, connected to the signal preprocessor, is used to automatically select the optimal wavelet basis function according to the signal characteristics, perform multi-scale decomposition on the substation equipment operating status data, and generate the wavelet coefficient matrix. An adaptive threshold generator, connected to the wavelet decomposition engine, is used to estimate the background noise level based on the signal statistical characteristics and generate dynamic thresholds for different decomposition levels. A signal reconstructor, connected to the wavelet decomposition engine and the adaptive threshold generator, is used to filter and retain meaningful wavelet coefficients according to the dynamic threshold, and generate the denoised partial discharge signal through inverse wavelet transform.

4. The embedded edge analysis system for substation equipment condition monitoring according to claim 1, characterized in that, The topology persistence analysis unit includes: A feature space builder is used to map the denoised partial discharge signal to a high-dimensional feature space through a time delay embedding method to generate the point cloud data. The filter function library stores various filter functions, including energy density functions, frequency distribution functions, waveform morphology functions, and phase relationship functions. A continuous homology calculation engine, connected to the feature space builder and the filter function library, is used to construct a multi-scale simple complex sequence based on the point cloud data and the filter functions in the filter function library, and to calculate the continuous homology group in different dimensions. The feature vectorizer, connected to the persistent cohomology computation engine, is used to extract persistent Barker codes from the persistent graph, compute topological invariants, and generate the topological feature vector.

5. The embedded edge analysis system for substation equipment condition monitoring according to claim 1, characterized in that, The manifold learning unit includes: A manifold structure learner is used to construct a K-nearest neighbor graph based on the topological feature vector, analyze the geometric structure of the local neighborhood, optimize the global coordinate system, and generate coordinate representations of data points on the nonlinear manifold structure. A geodesic distance calculator, connected to the manifold structure learner, is used to construct a weighted graph structure based on the K-nearest neighbor graph, execute the shortest path algorithm on the graph, and calculate the geodesic distance between any two points. A dynamic template manager is used to maintain a reference template library for different discharge types and dynamically update the reference template library based on newly acquired data. An anomaly detection evaluator, connected to the geodesic distance calculator and the dynamic template manager, is used to calculate the deviation of new data points from known manifolds, analyze the local density distribution around the data points, comprehensively evaluate the degree of anomaly, and generate the device health status diagnosis results.

6. The embedded edge analysis system for substation equipment condition monitoring according to claim 1, characterized in that, The edge processing module adopts a heterogeneous computing architecture of FPGA and ARM, wherein: The FPGA implements the computationally intensive tasks in the multi-scale wavelet transform unit and the topology persistence analysis unit; The ARM processor implements the manifold learning unit and system control tasks; The edge processing module also includes a task scheduler for dynamically allocating computing tasks between the FPGA and the ARM processor based on priority and resource utilization efficiency.

7. The embedded edge analysis system for substation equipment condition monitoring according to claim 1, characterized in that, The persistent coherence computation engine in the topology persistence analysis unit performs the following processing: Multiple filtering functions are determined to describe the features of the point cloud data from different perspectives; Construct a simple complex sequence by changing the filter value; Track the appearance and disappearance of topological features and record them as a persistent map; Statistical information and topological invariants are extracted from the persistent graph to form the topological feature vector.

8. The embedded edge analysis system for substation equipment condition monitoring according to claim 1, characterized in that, The manifold learning unit includes an adaptive learning mechanism, which is used for: The nonlinear manifold structure is continuously updated from new data through incremental learning. Detect and adapt to characteristic changes caused by equipment aging; The algorithm complexity is dynamically adjusted based on computing resource availability. The model parameters were adjusted based on the diagnostic results confirmed by humans.

9. The embedded edge analysis system for substation equipment condition monitoring according to claim 1, characterized in that, The system further includes a data storage module, which is connected to the edge processing module and is used for: Store raw monitoring data and process intermediate results; Save historical diagnostic results and trends in equipment health status; Maintain the discharge type reference template library; Provides data query and retrospective analysis functions.

10. The embedded edge analysis system for substation equipment condition monitoring according to claim 1, characterized in that, The edge processing module further includes a resource adaptive management unit, which is used for: Monitor system operating status and resource usage; Dynamically allocate computing resources based on task priority and urgency; Reduce feature dimensions and processing accuracy when computational resources are limited; Adjust the monitoring frequency and data processing frequency according to the equipment status to achieve efficient utilization of system resources.

Images