Power equipment partial discharge data processing method and system based on spatio-temporal feature fusion

Abstract

The application discloses a power equipment partial discharge data processing method and system based on space-time feature fusion, relates to the technical field of power equipment state monitoring, and comprises the following steps: continuously sampling and extracting a basic window feature set by adopting a double buffering mechanism, and merging continuous abnormal basic windows into a meta event; when the meta event exists, the observation window boundary is adaptively adjusted; an extreme value is managed by using a red-black tree and time-varying weighted aggregation is carried out, and aggregated features are output; waveform similarity is calculated and it is checked whether the features are aligned; a feature fingerprint is generated by using a hash algorithm, a Merkle tree is constructed to form a feature data snapshot; the aggregated features are stored by classification according to a feature sensitivity index, and the data integrity is ensured through edge cloud collaborative verification; the application guarantees the space-time continuity of the discharge process, realizes high-precision, low-delay and traceable partial discharge data processing, and effectively reduces the resource consumption of an embedded terminal.

Description

Technical Field

[0001] This invention relates to the field of power equipment condition monitoring technology, and in particular to a method and system for processing partial discharge data of power equipment based on spatiotemporal feature fusion. Background Technology

[0002] Partial discharge (PD) monitoring technology has become a key supporting tool in the field of power equipment health assessment and fault early warning, playing an irreplaceable role in ensuring the safe operation of the power grid and extending the service life of equipment. However, existing partial discharge monitoring methods still face multiple technical bottlenecks in practical applications, which restricts their widespread adoption in embedded terminals and resource-constrained environments.

[0003] First, the problem of feature distortion severely affects the reliability of monitoring results. Traditional PD signal processing often uses minute-level statistical methods, directly aggregating and calculating millisecond-level high-frequency sampling data after discretizing it according to a fixed time window. This approach ignores the continuity and inherent correlation of PD signals in time and space, resulting in key physical parameters such as maximum, minimum, and peak-to-peak values ​​failing to accurately reflect the true evolution trajectory of the discharge process, thereby reducing the accuracy of fault identification and lifetime prediction.

[0004] Secondly, storage redundancy leads to low system resource utilization. Existing monitoring systems typically store raw waveform data and statistical summary features separately, lacking a unified feature tracing mechanism. This siloed data storage model not only occupies a large amount of limited embedded storage space, but also requires cross-database queries during subsequent data retrieval, auditing, and backtracking analysis, increasing system design complexity and maintenance costs.

[0005] Furthermore, computational efficiency bottlenecks cannot be ignored. As monitoring duration and sampling frequency increase, the amount of historical data grows exponentially. Traditional solutions often require multiple traversals of the entire dataset for secondary or multiple aggregation calculations, leading to a significant increase in processing latency. At the same time, this places stringent demands on the computing power and energy consumption of embedded terminals, making it difficult to meet the application requirements that balance real-time performance and low power consumption. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for processing partial discharge data of power equipment based on spatiotemporal feature fusion, thereby solving the above-mentioned problems.

[0007] To achieve the above objectives, this invention provides a method for processing partial discharge data of power equipment based on spatiotemporal feature fusion, comprising the following steps: Step 100: A double buffering mechanism is used to continuously sample the partial discharge signal, and the basic window features are extracted within a basic window of a preset duration. The eigenvalues ​​in the basic window features include extreme values, window mean, and pulse count; Step 200: Use a red-black tree data structure to dynamically manage the extreme values ​​in all basic window features; Step 300: Perform time-varying weighted aggregation on all basic window features within the observation window to generate aggregated features; Step 400: Calculate the waveform similarity between the basic window feature sequence and the aggregated feature sequence, and determine whether to trigger local recalculation based on the waveform similarity. The method for local recalculation is to extract the basic window feature set corresponding to the observation period from the edge and regenerate the corresponding aggregated features. Step 500: Generate feature fingerprints for the aggregated features of each observation period and construct a Merkle tree structure to form a snapshot of the feature data; Step 600: In real time, compare the newly generated feature fingerprint with the feature fingerprint stored in the Merkle tree. If they are inconsistent, trigger local recalculation and update the Merkle tree. This includes the following steps: Step 601: Obtain the newly generated feature fingerprint for the current observation period and extract the historical feature fingerprint corresponding to the observation period from the Merkle tree; Step 602: Use the exact equality comparison method to determine whether the feature fingerprint and the historical feature fingerprint are consistent. If they are consistent, the value is 1; if they are inconsistent, the value is 0. Step 603: If the value is 1, no operation is performed; if the value is 0, a local recalculation is triggered. Step 700: Perform hierarchical compression storage of aggregated features at the edge end based on the feature sensitivity index; Step 800: Synchronize the feature data snapshot to the cloud for full consistency verification.

[0008] Furthermore, the double buffering mechanism is specifically as follows: Two buffers are set up to work in parallel for partial discharge signal acquisition and basic window feature calculation. While one is performing calculation, the other continuously acquires partial discharge signals. After the calculation is completed, the two buffers switch their working contents.

[0009] Furthermore, step 100 also includes merging consecutive base windows into a meta-event when an abnormal signal is detected in multiple consecutive base windows, and recording the start and end times of the meta-event. The method for determining abnormal signals is as follows: the system monitors each feature value in each basic window feature in real time and compares it with the preset abnormal threshold of each feature value. If any feature value in the current basic window feature exceeds its corresponding abnormal threshold, it is determined that an abnormal signal has been detected. When meta-events exist, the steps required to adaptively adjust the observation window boundaries include: Step 101: Initialize the initial start time of the observation window. Initial and final times ; Step 102: Read the start and end timestamps of the meta-event. and Determine whether a meta-event crosses the observation window boundary: If and Then the meta-event crosses the window's initial boundary; if and If the meta-event crosses the window end boundary; Step 103, based on the window overlap coefficient Calculate the boundary spread. The calculation formula is: ; in, This is the window overlap coefficient. For window overlap rate, For the rate of change of the signal, The duration of the observation window; Step 104: If the meta-event crosses the window's initial boundary, then shift the observation window's start time forward. If the meta-event crosses the end boundary of the observation window, the end time of the observation window will be postponed. .

[0010] Furthermore, the steps for dynamically managing the extreme values ​​among all basic window features using a red-black tree data structure include: Step 201: Insert the extreme values ​​in each basic window feature into the maximum value red-black tree and the minimum value red-black tree, respectively; Step 202: Whenever a new basic window feature is generated, the extreme value is inserted into the red-black tree. After insertion, the balance property of the tree is maintained by the rotation and recoloring operations of the red-black tree. Step 203: Locate the global maximum and global minimum values ​​within the current observation period by backtracking the red-black tree: the global maximum value is obtained by traversing the rightmost path of the maximum value red-black tree; the global minimum value is obtained by traversing the leftmost path of the minimum value red-black tree.

[0011] Furthermore, the steps for generating aggregated features include: Step 301: Calculate the time decay factor for each base window within the current observation window. The calculation formula is as follows: ; in, This is the decay factor for the i-th base window within the current observation window; This is the start timestamp of the i-th basic window within the current observation window; Step 302: Calculate the time-varying weighted average based on the time decay factor. The calculation formula is as follows: ; in This is the time-varying weighted average of the current observation window; is the feature value in the i-th basic window feature of the current observation window; n is the number of basic windows in the current observation window; Step 303: Combine the global maximum value, global minimum value, time-varying weighted average value and the start timestamp of the current observation period to form the aggregated feature of the current observation period.

[0012] Further, the waveform similarity between the base window feature sequence and the aggregated feature sequence is calculated, and the determination of whether to trigger local recalculation based on the waveform similarity includes the following steps: Step 401, Define the basic window feature sequence ,in For the first observation period Feature values ​​of the basic window features Number of base windows; aggregated feature sequences ,in For the first Feature values ​​in the aggregated features of each observation period, Number of observation periods; Step 402: The dynamic time warping algorithm is used to calculate the dynamic time warping distance between the basic window feature sequence and the aggregated feature sequence to align the basic window feature sequence and the aggregated feature sequence. Step 403: Calculate the waveform similarity based on the aligned base window feature sequence and aggregated feature sequence. The calculation formula is as follows: ; in, and These are the average values ​​of the basic window feature sequence and the aggregated feature sequence, respectively. This represents the total number of aligned sampling points. This represents the waveform similarity between the basic window feature sequence and the aggregated feature sequence. The method for determining whether to trigger local recalculation based on waveform similarity is as follows: if the waveform similarity is less than the preset waveform similarity threshold, it is determined that the feature is misaligned and local recalculation is triggered; if the waveform similarity is not less than the preset waveform similarity threshold, it is determined that the feature alignment is successful and no local recalculation is required; the aggregated feature after local recalculation replaces the original output aggregated feature.

[0013] Furthermore, the steps of generating feature fingerprints from the aggregated features for each observation period and constructing a Merkle tree structure to form a snapshot of the feature data include: Step 501: Standardize and serialize the aggregated features of each observation period, and convert them into byte strings of each observation period in a fixed format. Step 502: The byte string of each observation period is used to calculate the hash value of each observation period of fixed length through a hash function, which is recorded as the feature fingerprint of each observation period. Step 503: Construct a Merkle tree based on the feature fingerprints of each observation period, specifically including the following steps: Step 5031: Use the feature fingerprint generated in each observation period as the leaf node of the tree; Step 5032: Concatenate the hash values ​​of two adjacent leaf nodes, and perform another hash operation on the concatenated result to obtain the hash value of the parent node; Step 5033, and so on recursively upwards until a unique Merkle root is calculated and a snapshot of the feature data with the current timestamp is formed.

[0014] Furthermore, the hierarchical compression and storage of aggregated features at the edge based on the feature sensitivity index includes the following steps: Step 701, calculate the feature sensitivity index. The calculation formula is as follows: ; ; in As a feature sensitivity index, For characteristic rate of change, For diagnostic weighting, The standard deviation of the aggregated eigenvalues. This is the time-varying weighted average of the current observation window; Step 702, if If the basic window feature set and aggregated features are fully preserved and stored at the edge, then... Then, a sliding window average block compression algorithm is used to generate a compressed feature summary from the aggregated features and store it at the edge; where, This is the feature sensitivity threshold.

[0015] Furthermore, synchronizing feature data snapshots to the cloud for full consistency verification includes the following steps: Step 801: The cloud periodically receives feature data snapshots from the edge. Step 802: The cloud verifies and corrects the received feature data snapshot using a preset model or algorithm; Step 803: Compare the Merkle root maintained independently in the cloud with the Merkle root synchronized from the edge: if they are equal, the data is consistent and the verification is successful; if they are not equal, the data is inconsistent, and a local recalculation is initiated. After the local recalculation, the new aggregated features will be used to update the local Merkle tree to generate a new Merkle root, and then synchronized to the cloud again for comparison and confirmation; if they are still not equal, an alarm will be issued.

[0016] A power equipment partial discharge data processing system based on spatiotemporal feature fusion, used to execute any of the above-described power equipment partial discharge data processing methods based on spatiotemporal feature fusion, includes: Signal acquisition and preprocessing module: used to sample partial discharge signals, extract basic window features, and generate meta-events; The spatiotemporal feature fusion calculation module is used to dynamically adjust the observation window boundary, dynamically manage extreme values, and perform multi-scale feature fusion calculations on the basic window feature set. Feature verification and version management module: used to verify the consistency between the basic window feature set and the aggregated features, and to maintain the data version history through a Merkle tree; Storage optimization and edge-cloud collaboration module: used to implement tiered storage and verification interaction with cloud servers.

[0017] Therefore, the present invention employs the above-mentioned method and system for processing partial discharge data of power equipment based on spatiotemporal feature fusion, which has the following beneficial effects: 1. By combining "event slicing" and "dynamic window", a complete and reversible feature mapping model from millisecond-level transients to minute-level trends was established, solving the problem of feature spatiotemporal continuity break in traditional methods.

[0018] 2. Red-black trees are applied to real-time extremum management in embedded environments, achieving efficient and accurate updating and querying of global extrema with constant-level memory increments.

[0019] 3. By integrating feature fingerprints, Merkle trees, and edge-cloud collaborative verification, a data integrity assurance mechanism with tamper-resistant, traceable, and efficient auditing capabilities has been built on resource-constrained terminals.

[0020] 4. By dynamically evaluating feature sensitivity, the lossless / compressed storage strategy is adaptively switched, and computationally intensive verification tasks are offloaded to the cloud, achieving an optimal balance between terminal resource consumption and computational accuracy.

[0021] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0022] Figure 1This is a schematic diagram of the power equipment partial discharge data processing system architecture based on spatiotemporal feature fusion according to the present invention; Figure 2 This is a schematic diagram of the DTW algorithm of the present invention; Figure 3 This is a flowchart illustrating the generation process of the feature fingerprint and Merkle tree in this invention. Figure 4 This is a flowchart illustrating the generation, storage, and comparison with historical fingerprints of the present invention. Figure 5 This is a diagram illustrating the Merkle tree update process of the present invention. Detailed Implementation

[0023] The following detailed description of embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely illustrates selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0024] Please see Figure 1-5 A method for processing partial discharge data of power equipment based on spatiotemporal feature fusion includes the following steps: Step 100: A double buffering mechanism is used to continuously sample the partial discharge signal, and the basic window features are extracted within a basic window of a preset duration. It is understood that the execution subject of this invention can be an embedded terminal or a small edge server, and no specific limitation is made here. This embodiment of the invention will be described using an embedded terminal as an example.

[0025] A single buffer cannot simultaneously receive new data during feature computation, leading to signal loss during the computation period. A dual-buffering mechanism decouples data acquisition and feature computation by having two buffers work alternately, ensuring the continuity and integrity of millisecond-level high-frequency sampling. Therefore, a dual-buffering mechanism is introduced at the sensor layer, setting up two buffers that operate in parallel: a first buffer and a second buffer. Wherein: First buffer: Data acquisition and basic window feature calculation are performed within a fixed basic window duration to extract basic window features such as extreme values, mean values, and pulse counts of the partial discharge signal. In an embodiment of the present invention, the basic window duration is set to 10ms, which can meet the high-frequency sampling requirements of the partial discharge signal.

[0026] Second buffer: During the calculation in the first buffer, partial discharge signals are continuously received to ensure the continuity of data acquisition.

[0027] This dual-buffering mechanism ensures that while one buffer is busy with calculations, the other buffer can still collect data without interruption, thus effectively preventing data loss. In this embodiment, when the first buffer is performing basic window feature calculations for the first 10ms window, the second buffer simultaneously collects partial discharge signal data for the second 10ms window; after the first window calculation is completed, the two buffers switch roles, and this process repeats continuously.

[0028] It should be noted that the methods for extracting basic window features include: (1) Extreme value extraction: Within each basic window, traverse all sampling points and record the maximum and minimum values ​​of the partial discharge signal amplitude.

[0029] (2) Mean calculation: The window mean is obtained by averaging all sampling points within the base window. .

[0030] (3) Pulse counting: Count the number of times the amplitude of the partial discharge signal exceeds the preset pulse threshold within the basic window. It is used to characterize the degree of discharge activity within the basic window.

[0031] The extreme values, mean values, and pulse counts extracted above together constitute the basic window features of this basic window. ;in and They represent the first The maximum and minimum values ​​of the partial discharge signal amplitude for each basic window. Indicates the first The window mean of the basic windows, Indicates the first The pulse count of each basic window.

[0032] Traditional methods process signals independently within fixed windows, artificially fragmenting continuous discharge events across windows and failing to fully reflect the spatiotemporal continuity of the discharge process. By constructing meta-events, abnormal signals spanning multiple basic windows are integrated into a complete event unit, providing event boundary references for subsequent dynamic window adjustments and ensuring that the observation window can fully encompass important discharge events. In embodiments of this invention, when abnormal signals are detected within multiple consecutive basic windows, the consecutive basic windows are merged into a meta-event, and the start and end timestamps of the meta-event are recorded.

[0033] Specifically, for partial discharge events that span multiple base windows, the system employs an event slicing method to automatically generate "meta-events" to ensure event continuity.

[0034] The specific method for determining abnormal signals is as follows: the system monitors the feature values ​​in each basic window feature in real time, including peak voltage, pulse amplitude, and energy. A set of abnormal thresholds is preset. If any feature value in the current basic window feature exceeds its corresponding abnormal threshold, it is considered that an abnormal signal has been detected in that basic window.

[0035] This processing method can effectively integrate discharge signals across windows, ensuring that subsequent data processing and analysis can reflect the continuity of the actual discharge event. In an embodiment of the present invention, if a complete partial discharge event lasts 35ms, the event will span four 10ms basic windows. The system will automatically merge the basic window features of these four windows into a single meta-event for overall analysis.

[0036] The generation process of meta-events is as follows: During data acquisition, when continuous... When abnormal signals are detected in all basic windows, the system records the start time of the first abnormal basic window. End time of the last abnormal base window And then this The basic window features of each basic window are merged into a single meta-event, represented as: ; in, Represents a meta-event. This represents the starting moment of the meta-event. The end point of the meta-event. This is the set of basic window features contained in this meta-event. For the first The basic window features of a basic window.

[0037] Fixed window boundaries may truncate or interrupt important discharge events across the window. Dynamically adjusting the window boundaries ensures that meta-events are fully contained within the observation window. When meta-events exist, the window overlap coefficient is dynamically calculated based on the signal change rate to adaptively adjust the observation window boundaries.

[0038] Specifically, after receiving a query or timed command, the system extracts all basic window feature data within the current observation window from the basic window feature set. In one embodiment of the present invention, if the observation window duration is... If set to 5 minutes, it contains 30,000 basic window feature data (5 minutes ÷ 10 ms = 30,000 windows).

[0039] First, basic window features that fall within the current observation window are selected based on the timestamp. Then, these basic window features are sorted in chronological order to form a complete sequence of basic window features.

[0040] The specific steps for dynamically adjusting window boundaries are as follows: Step 101: Initialize the initial start time of the observation window. Initial and final times ; Step 102: Read the start and end timestamps of the meta-event. and Determine whether a meta-event crosses the observation window boundary: If and Then the meta-event crosses the window's initial boundary; if and If the meta-event crosses the window end boundary; Step 103, based on the window overlap coefficient Calculate the boundary spread. The calculation formula is: ; in, This is the window overlap coefficient. For window overlap rate, For the rate of change of the signal, The duration of the observation window; Step 104: If the meta-event crosses the window's initial boundary, then shift the observation window's start time forward. If the meta-event crosses the end boundary of the observation window, the end time of the observation window will be postponed. .

[0041] Step 200: Use a red-black tree data structure to dynamically manage the extreme values ​​in the basic window feature set, and obtain the global maximum and global minimum values; Specifically, the system uses a red-black tree data structure to dynamically manage the extreme value data of each basic window feature in the basic window feature set. A red-black tree is a self-balancing binary search tree, and its construction follows these rules: (1) Each node is colored either red or black; (2) The root node is black; (3) All leaf nodes (NIL nodes) are black; (4) If a node is red, then its children must be black; (5) All paths from any node to each of its leaf nodes contain the same number of black nodes.

[0042] The specific methods for extreme value management are as follows: (1) Extract the extreme values ​​(maximum values) from each basic window feature. and minimum value Insert them as nodes into the corresponding red-black trees (maximum red-black tree and minimum red-black tree).

[0043] (2) Whenever a new basic window feature is generated, its extreme value is inserted into the red-black tree. After insertion, the balance of the tree is maintained by rotation and recoloring operations of the red-black tree.

[0044] (3) Locate the global maximum and global minimum values ​​within the current observation period by backtracking the red-black tree: ; ; in, The global maximum value within the current observation period is obtained by traversing the rightmost path of the maximum value red-black tree. The global minimum value within the current observation period is obtained by traversing the leftmost path of the minimum value red-black tree. This represents the base window number within the current observation period.

[0045] Leveraging the balancing property of red-black trees, the time complexity of the above extreme value query and update operations is O(n log n). Compared to traditional linear traversal methods (time complexity...), This significantly improves computational efficiency. In an embodiment of the invention, for a 5-minute observation period containing 30,000 basic window features, the red-black tree method requires only about 15 comparison operations to locate the global extremum. A linear traversal would require 30,000 comparisons.

[0046] Step 300: Analyze the basic window feature set within the observation window. Perform time-varying weighted aggregation and calculate the time-varying weighted mean. and with the global maximum value and global minimum Combining to form aggregate features As output; Traditional arithmetic averaging assigns equal weight to all data points, failing to reflect the differences in data importance over time, leading to an excessive influence of earlier data on the current state assessment. Time-varying weighted aggregation, through a time decay factor, gives higher weight to recent data, more accurately reflecting the current state of the device. Specifically, a time decay factor is applied to all basic window features in the basic window feature set for weighted averaging, thereby reducing the influence of earlier data on the current state and giving the latest data a higher weight in the overall assessment.

[0047] The formula for calculating the time decay factor is: ; in, The first one in the current observation window The time decay factor of a basic window; The first one in the current observation window The start timestamp of each basic window.

[0048] Time-varying weighted average The calculation formula is: ; in, This is the time-varying weighted average of the current observation window; For the first Feature values ​​(such as mean) in a basic window feature or extreme values ); This represents the number of base windows within the current observation window.

[0049] Finally, the global maximum value will be... Global minimum value and time-varying weighted average The aggregated features for this observation period are formed by combining the following expressions: ; in, Indicates aggregation features, This is the start timestamp of the observation period. This aggregation feature... As output.

[0050] Step 400: Calculate the waveform similarity between the basic window feature set and the aggregated features, and determine whether to trigger local recalculation based on the waveform similarity. The millisecond-level basic window feature sequence and the minute-level aggregated feature sequence differ in time scale, and traditional point-to-point comparison methods cannot effectively measure their consistency. The dynamic time warping algorithm, by finding the optimal alignment path, can establish feature mapping relationships across different time scales and accurately assess whether the aggregation process preserves the waveform characteristics of the original signal. Specifically, it defines the basic window feature sequence... ,in For the first observation period Feature values ​​of the basic window features (such as the mean) ), The number of base windows. Define the aggregated feature sequence. ,in For the first Aggregated features of observation periods Eigenvalues ​​(such as time-varying weighted average) ), This represents the number of observation periods.

[0051] The computation steps of the Dynamic Time Warping (DTW) algorithm are as follows: (1) Construct the distance matrix ,in Represents the basic window feature sequence The Feature values ​​and aggregated feature sequences of basic window features The The Euclidean distance between eigenvalues ​​in the aggregated features of a given observation period.

[0052] (2) Calculate the cumulative distance matrix Its recursive formula is: ; in, Indicates starting from the origin Time The minimum cumulative distance.

[0053] (3) Calculate the dynamic time warped distance using the following formula: ; in, Indicates all possible alignment paths, for and The minimum cumulative distance between the two sequences. After alignment, the total number of sampling points in the base window feature sequence and the aggregated feature sequence is the same, both being N.

[0054] (4) Calculate the waveform similarity (Pearson correlation coefficient) based on the aligned sequences. The calculation formula is as follows: ; in, and These are the basic window feature sequences. and aggregated feature sequences The average value; This represents the total number of aligned sampling points. Represents the basic window feature sequence and aggregated feature sequences The waveform similarity, with a value range of , where 1 represents a perfect positive correlation, -1 represents a perfect negative correlation, and 0 represents no linear correlation.

[0055] like This is then determined to be feature misalignment and triggers local recalculation, where The preset waveform similarity threshold is set to 0.85 in this embodiment. If... If the alignment is successful, it indicates that the feature alignment is successful and no calibration is required. The system will continue with subsequent feature fingerprint generation and feature data snapshot storage.

[0056] The method for local recalculation is to extract the basic window feature set corresponding to the observation period from the edge and regenerate the corresponding aggregated features.

[0057] The effects of the dynamic time warping algorithm are as follows: (1) Improved the robustness of feature extraction. The DTW algorithm can cope with the nonlinear changes of partial discharge signals in time, so that the extracted aggregated features can more accurately reflect the actual discharge mode.

[0058] (2) By aligning the similarity, the system can more accurately determine whether the current discharge state is consistent with the historical pattern, or whether new changes that need attention have occurred (e.g., changes in discharge type, increased severity, etc.).

[0059] (3) Precise alignment and similarity calculation reduce misjudgments caused by time misalignment, enabling local recalculation to play a more effective role, avoiding unnecessary waste of computing resources, and ensuring timely response when anomalies that really need attention occur.

[0060] (4) DTW ensures that only features that have been accurately verified and aligned can ultimately form a data snapshot, thereby improving the integrity and traceability of the data processed by the entire system.

[0061] Step 500: Calculate the aggregated features for each observation period to generate the feature fingerprint for the corresponding observation period, and construct a Merkle tree to form a feature data snapshot; Because aggregated feature data is large in volume, direct comparison is inefficient and incurs high storage costs; traditional storage methods cannot quickly detect whether data has been tampered with. Hash algorithms compress aggregated features of arbitrary length into fixed-length feature fingerprints, exhibiting unidirectionality and collision resistance; any data modification will result in completely different fingerprints. The Merkle tree structure further organizes multiple feature fingerprints into a tree hierarchy, allowing efficient verification of the integrity of large amounts of data by comparing the Merkle roots, with a verification complexity of O(logn). In the embodiments of this invention, the steps for generating corresponding feature fingerprints from aggregated features using the SHA-256 hash algorithm are as follows: (1) Perform routine standardization and serialization on the aggregated features of each observation period and convert them into byte strings of each observation period in a fixed format.

[0062] (2) Using the byte string as input, a hash value of fixed length (256 bits, usually represented as a 64-bit hexadecimal string) is calculated using a hash function (SHA-256 in this embodiment), which is the feature fingerprint. The calculation formula is as follows: ; in, For the first The characteristic fingerprint of each observation period For the first A serialized byte string of aggregated features for each observation period This refers to the SHA-256 hash function. The characteristic of the SHA-256 hash algorithm is that even small changes in the input data will lead to a significant change in the output hash value, thus effectively ensuring the uniqueness of the feature fingerprint and the sensitivity of data integrity verification.

[0063] The Merkle tree is constructed as follows: (1) Use the feature fingerprint generated in each observation period as the leaf node of the tree; (2) Concatenate the hash values ​​of two adjacent leaf nodes, and perform hash operation on the concatenated result again to obtain the hash value of the parent node; (3) Continue recursively upwards until a unique Merkle root is calculated. This creates a snapshot of the feature data with the current timestamp.

[0064] This tree structure ensures that any tampering with the aggregated feature data of any leaf node will cause changes in the hash values ​​of its parent node and all ancestor nodes, ultimately affecting the Merkle root, thus achieving data version immutability. Furthermore, by comparing the Merkle root, the integrity of large amounts of aggregated feature data can be efficiently verified.

[0065] Step 600: In real time, the newly generated feature fingerprint is compared with the historical feature fingerprint stored in the Merkle tree. If they are inconsistent, a local recalculation is triggered to regenerate new aggregated features and update the Merkle tree.

[0066] During long-term operation, aggregated features may become corrupted or abnormal due to hardware failures, transmission errors, or external interference. By comparing feature fingerprints in real time, abnormal data can be detected promptly. Utilizing the basic window feature set at the edge for local recalculation, aggregated features exhibiting abnormal cycles can be repaired without reprocessing all historical data, achieving efficient data self-healing capabilities.

[0067] Specifically, the method for feature fingerprint comparison is as follows: (1) Obtain the feature fingerprint to be compared and the historical feature fingerprint: Obtain the newly generated feature fingerprint for the current observation period; extract the historical feature fingerprint corresponding to the observation period from the Merkle tree (i.e. the feature fingerprint that was first generated and stored in the leaf nodes of the Merkle tree in this period).

[0068] (2) Perform bit-by-bit comparison of the feature fingerprint: Due to the nature of hash algorithms, even slight changes in the original aggregated features can result in completely different hash values. Therefore, an exact equality comparison method is used to determine whether the feature fingerprint and historical feature fingerprints are consistent; a value of 1 indicates consistency, while a value of 0 indicates inconsistency.

[0069] (3) Determine an anomaly and trigger the refactoring process: If the value is 1, no operation is performed; if the value is 0, it is determined that there is a data anomaly in the aggregated features of the current observation period (which may be caused by hardware failure, transmission error or data corruption), triggering local recalculation.

[0070] Local recalculation only applies to data from specific observation periods where anomalies are detected, rather than recalculating all historical data, thereby reducing computational resource consumption and processing latency.

[0071] The Merkle tree update method is as follows: A new feature fingerprint is generated from the newly aggregated features obtained through local recalculation using the SHA-256 hash algorithm. This new feature fingerprint is then used to replace the corresponding leaf node in the Merkle tree. Next, the hash values ​​of all parent nodes along the path from the leaf node to the Merkle root are recalculated from bottom to top until a new Merkle root is generated, forming a new feature data snapshot and version chain.

[0072] Version chain formation: The Merkle root generated at the current time point is linked to the Merkle root at the previous time point using a hash. This chain structure ensures that any tampering with historical aggregated feature data will result in a hash value mismatch in subsequent chains, thus being detected immediately and enabling auditing and tracing of data modifications.

[0073] Step 700: Perform hierarchical compression storage of aggregated features at the edge end according to the feature sensitivity index; Different aggregated features have varying degrees of importance for fault diagnosis, and adopting a uniform storage strategy can lead to loss of accuracy for important features or waste of storage space. By comprehensively considering the feature change rate (reflecting the degree of fluctuation) and the diagnostic weight (reflecting the importance of diagnosis) to calculate the feature sensitivity index, high-sensitivity features and low-sensitivity features can be scientifically distinguished, and lossless storage and lossy compression strategies can be adopted respectively.

[0074] Specifically, feature sensitivity index The calculation formula is: ; in, It is a characteristic sensitivity index, a comprehensive indicator used to assess the severity and impact of the current partial discharge state; The characteristic change rate reflects the degree of fluctuation of the aggregated characteristic over time, and can be obtained by statistical methods such as the difference or standard deviation of aggregated characteristic values ​​at adjacent time points; As a diagnostic weight, the empirical value or expert knowledge preset based on the importance of the aggregated feature to the diagnosis of partial discharge faults is used. For example, aggregated features that are strongly correlated with discharge energy and pulse amplitude can be assigned a higher diagnostic weight.

[0075] Characteristic change rate The calculation formula is: ; in, The standard deviation of the aggregated eigenvalues. This is the time-varying weighted mean of the current observation window, and this formula is used to measure the relative dispersion of aggregated feature data.

[0076] Storage policy determination criteria: like If the feature is identified as a high-sensitivity feature, the aggregated feature is stored losslessly, preserving the complete set of basic window features and the aggregated feature. like If the feature is low-sensitivity, the aggregated feature is classified as such and a lossy compression storage is performed using a sliding window average block compression algorithm to generate a compressed feature summary.

[0077] in, This is the feature sensitivity threshold, which is set according to the equipment type and monitoring requirements in practical applications.

[0078] The calculation method for generating a compressed feature summary using the sliding window average block compression algorithm is as follows: (1) Set the size of the sliding window to The sliding step size is .

[0079] (2) For aggregated feature sequences Calculate the first The average value of each compressed block is calculated using the following formula: ; in, For the first The average value of each compressed block; To adjust the sliding window size; This is the sliding step size; For the aggregated feature sequence, the first The feature values ​​in the aggregated features of each observation period.

[0080] (3) According to step length Slide the window and repeat the above steps until all data points have been processed.

[0081] (4) Calculate the average sequence of the compressed blocks It is stored as a compressed feature summary, in which the number of compressed data points is: ; in, This represents the number of feature values ​​in the aggregated feature sequence. This indicates the floor function.

[0082] This method can significantly reduce the amount of data while preserving the overall trend of aggregated features, generating a compressed feature summary.

[0083] Step 800: Data storage and full consistency verification are performed through an edge-cloud collaborative architecture.

[0084] Due to limited storage and computing resources at the edge, complete data cannot be stored and complex verifications cannot be performed. The edge refers to embedded monitoring devices (embedded terminals) or small servers deployed at power equipment sites, close to the data source. The cloud has sufficient computing and storage resources, but network latency affects real-time performance. Through an edge-cloud collaborative architecture, the edge only stores lightweight feature fingerprints and compressed feature summaries for fast local queries, while complete data snapshots are synchronized to the cloud for in-depth analysis and full verification, achieving an optimal balance between real-time performance and reliability.

[0085] Specifically, the architecture of edge-cloud collaborative verification is divided into two parts: the edge and the cloud. Edge storage strategy: Only the feature fingerprint and compressed feature summary are stored in the embedded terminal to minimize local storage space usage.

[0086] Cloud synchronization mechanism: Regularly synchronize feature data snapshots to the cloud server.

[0087] Specific methods for performing full consistency verification and analysis in the cloud include: (1) The cloud periodically receives feature data snapshots from the edge; (2) The cloud uses various preset models or algorithms to perform multi-dimensional and in-depth analysis and verification of the received aggregated feature data and correct erroneous data; (3) By comparing the Merkle root maintained independently in the cloud with the Merkle root synchronized from the edge: if the two are equal, the data is consistent and the verification is successful; if the two are not equal, it indicates that the aggregated feature data at the edge may be abnormal or tampered with. At this time, local recalculation will be initiated. After local recalculation, the new aggregated features will be used to update the local Merkle tree to generate a new Merkle root, and then synchronized to the cloud again for comparison and confirmation; if they are still not equal, an alarm will be issued.

[0088] The power equipment partial discharge data processing system based on spatiotemporal feature fusion is used to execute the above-mentioned power equipment partial discharge data processing method based on spatiotemporal feature fusion, including: Signal acquisition and preprocessing module: used to sample partial discharge signals, extract basic window features, and generate meta-events; The spatiotemporal feature fusion calculation module is used to dynamically adjust the observation window boundary, dynamically manage extreme values, and perform multi-scale feature fusion calculations on the basic window feature set. Feature verification and version management module: used to verify the consistency between the basic window feature set and the aggregated features, and to maintain the data version history through a Merkle tree; Storage optimization and edge-cloud collaboration module: used to implement tiered storage and verification interaction with cloud servers.

[0089] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. 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 still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for processing partial discharge data of power equipment based on spatiotemporal feature fusion, characterized in that, Includes the following steps: Step 100: A double buffering mechanism is used to continuously sample the partial discharge signal, and the basic window features are extracted within a basic window of a preset duration. The eigenvalues ​​in the basic window features include extreme values, window mean, and pulse count; Step 200: Use a red-black tree data structure to dynamically manage the extreme values ​​in all basic window features; Step 300: Perform time-varying weighted aggregation on all basic window features within the observation window to generate aggregated features; Step 400: Calculate the waveform similarity between the basic window feature sequence and the aggregated feature sequence, and determine whether to trigger local recalculation based on the waveform similarity. The method for local recalculation is to extract the basic window feature set corresponding to the observation period from the edge and regenerate the corresponding aggregated features. Step 500: Generate feature fingerprints for the aggregated features of each observation period and construct a Merkle tree structure to form a snapshot of the feature data; Step 600: In real time, compare the newly generated feature fingerprint with the feature fingerprint stored in the Merkle tree. If they are inconsistent, trigger local recalculation and update the Merkle tree. This includes the following steps: Step 601: Obtain the newly generated feature fingerprint for the current observation period and extract the historical feature fingerprint corresponding to the observation period from the Merkle tree; Step 602: Use the exact equality comparison method to determine whether the feature fingerprint and the historical feature fingerprint are consistent. If they are consistent, the value is 1; if they are inconsistent, the value is 0. Step 603: If the value is 1, no operation is performed; if the value is 0, a local recalculation is triggered. Step 700: Perform hierarchical compression storage of aggregated features at the edge end according to the feature sensitivity index; Step 800: Synchronize the feature data snapshot to the cloud for full consistency verification.

2. The method for processing partial discharge data of power equipment based on spatiotemporal feature fusion according to claim 1, characterized in that, The double buffering mechanism is as follows: Two buffers are set up to work in parallel for partial discharge signal acquisition and basic window feature calculation. While one is performing calculation, the other continuously acquires partial discharge signals. After the calculation is completed, the two buffers switch their working contents.

3. The method for processing partial discharge data of power equipment based on spatiotemporal feature fusion according to claim 2, characterized in that, Step 100 also includes merging consecutive base windows into a meta-event when an abnormal signal is detected in multiple consecutive base windows, and recording the start and end times of the meta-event. The method for determining abnormal signals is as follows: the system monitors each feature value in each basic window feature in real time and compares it with the preset abnormal threshold of each feature value. If any feature value in the current basic window feature exceeds its corresponding abnormal threshold, it is determined that an abnormal signal has been detected. When meta-events exist, the steps required to adaptively adjust the observation window boundaries include: Step 101: Initialize the initial start time of the observation window. Initial and final times ; Step 102: Read the start and end timestamps of the meta-event. and Determine whether a meta-event crosses the observation window boundary: If and Then the meta-event crosses the window's initial boundary; if and If the meta-event crosses the window end boundary; Step 103, based on the window overlap coefficient Calculate the boundary spread. The calculation formula is: ； in, This is the window overlap coefficient. For window overlap rate, For the rate of change of the signal, The duration of the observation window; Step 104: If the meta-event crosses the window's initial boundary, then shift the observation window's start time forward: If the meta-event crosses the end boundary of the observation window, the end time of the observation window will be postponed. .

4. The method for processing partial discharge data of power equipment based on spatiotemporal feature fusion according to claim 3, characterized in that, The steps for dynamically managing extreme values ​​in all basic window features using a red-black tree data structure include: Step 201: Insert the extreme values ​​in each basic window feature into the maximum value red-black tree and the minimum value red-black tree, respectively; Step 202: Whenever a new basic window feature is generated, the extreme value is inserted into the red-black tree. After insertion, the balance property of the tree is maintained by the rotation and recoloring operations of the red-black tree. Step 203: Locate the global maximum and global minimum values ​​within the current observation period by backtracking the red-black tree: the global maximum value is obtained by traversing the rightmost path of the maximum value red-black tree; the global minimum value is obtained by traversing the leftmost path of the minimum value red-black tree.

5. The method for processing partial discharge data of power equipment based on spatiotemporal feature fusion according to claim 4, characterized in that, The steps for generating aggregated features include: Step 301: Calculate the time decay factor for each base window within the current observation window. The calculation formula is as follows: ； in, This is the decay factor for the i-th base window within the current observation window; This is the start timestamp of the i-th basic window within the current observation window; Step 302: Calculate the time-varying weighted average based on the time decay factor. The calculation formula is as follows: ； in () represents the time-varying weighted average of the current observation window; is the feature value in the i-th basic window feature of the current observation window; n is the number of basic windows in the current observation window; Step 303: Combine the global maximum value, global minimum value, time-varying weighted average value and the start timestamp of the current observation period to form the aggregated feature of the current observation period.

6. The method for processing partial discharge data of power equipment based on spatiotemporal feature fusion according to claim 5, characterized in that, Calculating the waveform similarity between the base window feature sequence and the aggregated feature sequence, and determining whether to trigger local recalculation based on waveform similarity, includes the following steps: Step 401, Define the basic window feature sequence ,in For the first observation period Feature values ​​of the basic window features Number of base windows; aggregated feature sequences ,in For the first Feature values ​​in the aggregated features of each observation period, Number of observation periods; Step 402: The dynamic time warping algorithm is used to calculate the dynamic time warping distance between the basic window feature sequence and the aggregated feature sequence to align the basic window feature sequence and the aggregated feature sequence. Step 403: Calculate the waveform similarity based on the aligned base window feature sequence and aggregated feature sequence. The calculation formula is as follows: ； in, and These are the average values ​​of the basic window feature sequence and the aggregated feature sequence, respectively. This represents the total number of aligned sampling points. This represents the waveform similarity between the basic window feature sequence and the aggregated feature sequence. The method for determining whether to trigger local recalculation based on waveform similarity is as follows: if the waveform similarity is less than the preset waveform similarity threshold, it is determined that the feature is misaligned and local recalculation is triggered; if the waveform similarity is not less than the preset waveform similarity threshold, it is determined that the feature alignment is successful and no local recalculation is required; the aggregated feature after local recalculation replaces the original output aggregated feature.

7. The method for processing partial discharge data of power equipment based on spatiotemporal feature fusion according to claim 6, characterized in that, The steps for generating feature fingerprints from aggregated features for each observation period and constructing a Merkle tree structure to form a snapshot of feature data include: Step 501: Standardize and serialize the aggregated features of each observation period, and convert them into byte strings of each observation period in a fixed format. Step 502: The byte string of each observation period is used to calculate the hash value of each observation period of fixed length through a hash function, which is recorded as the feature fingerprint of each observation period. Step 503: Construct a Merkle tree based on the feature fingerprints of each observation period, specifically including the following steps: Step 5031: Use the feature fingerprint generated in each observation period as the leaf node of the tree; Step 5032: Concatenate the hash values ​​of two adjacent leaf nodes, and perform another hash operation on the concatenated result to obtain the hash value of the parent node; Step 5033, and so on recursively upwards until a unique Merkle root is calculated and a snapshot of the feature data with the current timestamp is formed.

8. The method for processing partial discharge data of power equipment based on spatiotemporal feature fusion according to claim 7, characterized in that, The hierarchical compression and storage of aggregated features at the edge based on feature sensitivity metrics includes the following steps: Step 701, calculate the feature sensitivity index. The calculation formula is as follows: ； ； in As a feature sensitivity index, For characteristic rate of change, For diagnostic weighting, The standard deviation of the aggregated eigenvalues. This is the time-varying weighted average of the current observation window; Step 702, if If the basic window feature set and aggregated features are fully preserved and stored at the edge, then... Then, a sliding window average block compression algorithm is used to generate a compressed feature summary from the aggregated features and store it at the edge; where, This is the feature sensitivity threshold.

9. The method for processing partial discharge data of power equipment based on spatiotemporal feature fusion according to claim 8, characterized in that, Synchronizing feature data snapshots to the cloud for full consistency verification includes the following steps: Step 801: The cloud periodically receives feature data snapshots from the edge. Step 802: The cloud verifies and corrects the received feature data snapshot using a preset model or algorithm; Step 803: Compare the Merkle root maintained independently in the cloud with the Merkle root synchronized from the edge: if they are equal, the data is consistent and the verification is successful; if they are not equal, the data is inconsistent, and a local recalculation is initiated. After the local recalculation, the new aggregated features will be used to update the local Merkle tree to generate a new Merkle root, and then synchronized to the cloud again for comparison and confirmation; if they are still not equal, an alarm will be issued.

10. A power equipment partial discharge data processing system based on spatiotemporal feature fusion, characterized in that, It is used to execute the power equipment partial discharge data processing method based on spatiotemporal feature fusion as described in any one of claims 1-9, comprising: Signal acquisition and preprocessing module: used to sample partial discharge signals, extract basic window features, and generate meta-events; The spatiotemporal feature fusion calculation module is used to dynamically adjust the observation window boundary, dynamically manage extreme values, and perform multi-scale feature fusion calculations on the basic window feature set. Feature verification and version management module: used to verify the consistency between the basic window feature set and the aggregated features, and to maintain the data version history through a Merkle tree; Storage optimization and edge-cloud collaboration module: used to implement tiered storage and verification interaction with cloud servers.

Images