A method, system, device, and medium for online diagnosing partial discharge of electrical equipment

Abstract

The application provides a method, system, device and medium for online diagnosis of partial discharge of electrical equipment. The method pre-processes the collected initial data set, adopts common sub-tree multiplexing, operator merging and precondition pruning to further reduce repeated calculation, performs probability calibration on the initial matching result to obtain a reliable feedback label of the initial matching result, performs historical back testing and symbol distillation on the initial matching result based on historical ultrahigh frequency data and the reliable feedback label to obtain an optimal symbol mechanism and a candidate symbol mechanism. This process can convert the partial discharge expert knowledge from natural language or scattered program rules into a symbol mechanism with type constraints and determined execution semantics, so that the mechanism can be checked, back tested, combined, multiplexed and versioned, and is beneficial to forming a maintainable diagnostic knowledge asset across manufacturers and across equipment types.

Description

Technical Field

[0001] This invention belongs to the field of high-voltage electrical equipment condition monitoring technology, specifically relating to a method, system, equipment, and medium for online diagnosis of partial discharge in electrical equipment. Background Technology

[0002] Partial discharge may occur in high-voltage and ultra-high-voltage electrical equipment due to manufacturing defects, insulation aging, metal particles, conductor tips, floating potential, insulation gaps, or surface defects. The steep pulses generated by partial discharge can form electromagnetic wave signals in the ultra-high frequency band. By collecting these signals using one or more ultra-high frequency sensors installed on GIS, GIL, transformers, or switchgear, waveform, spectrum, phase, and pulse statistics reflecting the discharge state can be obtained.

[0003] Existing online partial discharge diagnosis methods still have the following shortcomings:

[0004] 1. Criteria for partial discharge are usually stored in natural language, tables, program branches, or scattered expert rules, lacking unified input / output types and execution semantics, making them difficult to combine, reuse, audit, and migrate.

[0005] 2. UHF waveforms have a high sampling rate and contain a large number of sampling points within a single time window. When multiple sensors, multiple windows, and multiple rules are running simultaneously, it is difficult to meet real-time requirements by interpreting and executing them sequentially window by window and rule by rule.

[0006] 3. Although traditional black-box classification models can output the category of partial discharge, they are difficult to reliably provide the features, thresholds, phase intervals, durations and logical paths corresponding to each diagnostic conclusion, which is not convenient for maintenance personnel to verify.

[0007] 4. Using a fixed threshold to determine partial discharge is difficult to adapt to different equipment structures, sensor installation locations, propagation paths, noise environments, and operating conditions. Directly modifying the rules online can easily lead to an increase in false alarm or false negative rates.

[0008] 5. Existing solutions lack a complete closed loop of "candidate mechanism formation - legality verification - tensor quantification and compilation - historical backtesting - distillation screening - feedback update - gray-scale deployment", and cannot form a traceable and continuously updated judgment mechanism. Summary of the Invention

[0009] This invention addresses at least one of the technical problems existing in the prior art. To this end, embodiments of this invention provide a method, system, device, and medium for online diagnosis of partial discharge in electrical equipment. This method enables verification, backtesting, combination, reuse, and version management, and can form maintainable diagnostic knowledge assets across manufacturers and equipment types.

[0010] To achieve the above objectives, the present invention provides the following technical solution:

[0011] In a first aspect, embodiments of this disclosure provide a method for online diagnosis of partial discharge in electrical equipment, comprising the following steps:

[0012] Step S1: Collect the UHF waveform of at least one sensor in the high-voltage electrical equipment and form an initial dataset. Preprocess the initial dataset to obtain data batches for parallel processing.

[0013] Step S2: Calculate the partial discharge characteristics of each waveform window in the data batch. The partial discharge characteristics include time domain characteristics, frequency domain characteristics, time-frequency characteristics, phase characteristics, pulse statistical characteristics, and multi-channel spatial characteristics.

[0014] Step S3: Organize the partial discharge features and the pre-constructed partial discharge-specific symbol system into an input tensor, and generate a tensor computation graph according to the dependency relationship, node type and topology level of the symbol tree. Through batch processing, merging of similar operators, reuse of common subtrees, precondition pruning and heterogeneous parallelism, the initial matching result is obtained.

[0015] Step S4: Perform probability calibration on the initial matching result to obtain a reliable feedback label for the initial matching result. Based on historical high frequency data and the reliable feedback label, perform historical backtesting and symbol distillation on the initial matching result to obtain the preferred symbol mechanism and candidate symbol mechanism.

[0016] Step S5: Construct a formal online diagnostic environment based on the preferred symbol mechanism that is in a formally enabled state, and output the online diagnostic results.

[0017] Furthermore, the initial dataset in step S1 include:

[0018] ;

[0019] In the formula, It is a waveform sampling sequence; For the time of data collection; The sampling frequency; For sensor or channel identification; It is the power frequency phase; For equipment identification;

[0020] During the preprocessing, the UHF waveform is subjected to preprocessing including DC component removal, bandpass filtering, power frequency and narrowband interference suppression, background noise estimation, pulse triggering, baseline correction, amplitude normalization and / or abnormal saturation identification, and the data batch is obtained.

[0021] For multi-channel data, time alignment is performed based on a hardware clock, synchronization pulse, or correction signal, and waveform windows are divided according to a fixed length, pulse trigger, or sliding window method to obtain the data batch.

[0022] Furthermore, obtaining the partial discharge characteristics in step S2 includes the following steps:

[0023] Assuming the waveform window is , for the shape window After performing a short-time Fourier transform or wavelet transform, a coefficient matrix that varies with time and frequency is obtained, and the amplitude, energy and energy ratio of each time-frequency unit are used as the time-frequency characteristics of the waveform window.

[0024] The time-domain characteristics of the waveform window include the root mean square value. and window energy , wherein the root mean square value include:

[0025] ;

[0026] The window energy include:

[0027] ;

[0028] The frequency domain characteristics of the waveform window are represented by the spectral proportion. This indicates the percentage of the spectrum within the specified frequency band [f1, f2]. include:

[0029] ;

[0030] In the formula, For the spectrum of ultra-high frequency waveforms, To prevent positive numbers from being divided by zero;

[0031] The multi-channel spatial characteristics of the waveform window are represented by the arrival time difference of the multi-channel pulses. This indicates the arrival time difference of the multi-channel pulses. include:

[0032] ;

[0033] In the formula, for Arrival time of channel pulse; for Arrival time of channel pulse;

[0034] The pulse statistical features include pulse count, peak value, rise time, pulse width, pulse interval, and power frequency phase clustering, as well as the statistical values ​​of the pulse statistical features within a continuous time window;

[0035] Based on channel and channels The window energy is used to obtain the inter-channel energy attenuation ratio. :

[0036] ;

[0037] Calculate the channel within the preset time delay search range. and channels The normalized cross-correlation of the waveforms yields the cross-correlation peak value, which is then used as the correlation coefficient. The time delay corresponding to the peak cross-correlation value is used as the arrival time difference. ;

[0038] Based on the arrival time difference Align the multi-channel waveforms and calculate the correlation coefficient of any channel pair. and normalized residuals And calculate propagation consistency ,include:

[0039] ;

[0040] Among them, normalized residuals For channel The waveform and the amplitude-scaled channel The L2 norm of the difference between aligned waveforms divided by the L2 norm of channel i waveform and sum; The value range is [0, 1];

[0041] The power frequency phase, amplitude, and occurrence period of the pulse are mapped to a PRPD / PRPS spectrum, and the phase distribution and power frequency phase clustering are extracted from the PRPD / PRPS spectrum as the phase features.

[0042] Furthermore, the partial discharge-specific symbol system in step S3 includes waveform nodes, sensor nodes, time window nodes, time domain operation nodes, frequency domain operation nodes, phase nodes, statistical nodes, spatial correlation nodes, condition nodes, logic nodes, timing nodes, and result nodes. Each type of node is predefined with node name, number of parameters, input type, output type, allowed sub-nodes, numerical range, and execution function.

[0043] It also includes the initial symbol mechanism, which refers to the partial discharge diagnostic expression composed of registered nodes, parameters, dependencies and execution semantics. The partial discharge diagnostic expression includes at least feature calculation, condition judgment and result output. Each intermediate node of the partial discharge diagnostic expression stores the node value, threshold, input window and execution state.

[0044] The steps involved in generating the tensor computation graph are as follows:

[0045] Step A1: Generate candidate symbol mechanisms based on the initial symbol mechanism, including:

[0046] By selecting or searching for each configurable parameter from the allowed range using a preset parameterized rule template, one or more candidate symbol mechanisms are generated. The preset parameterized rule template is a symbolic expression skeleton that represents the verified partial discharge diagnostic rules as having a fixed node type and logical structure, and is configurable with threshold, frequency band, time window and applicable device range.

[0047] Alternatively, select a subtree with the same input and output types from historical effective symbol mechanisms, perform subtree cross-combination at type-compatible node positions, and mutate at least one of the threshold, frequency band, time window, comparison operator, or logical connection relationship to generate candidate symbol mechanisms;

[0048] Alternatively, a trained symbol mechanism generation model can encode the device type, target partial discharge type, available channels, feature vocabulary, upper limit of expression tree depth, and evaluation target as task conditions. Expressions are generated symbol by symbol using preorder traversal, postorder traversal, or parenthesized function forms. Syntax masks and type constraints are used during generation until a result node is generated and all parameters to be supplemented are closed, yielding candidate symbol mechanisms. During symbol generation, nodes matching the output type are selected from the registered symbol vocabulary based on the data type required by the input slot to be filled in the current parent node. Nodes with more parameters than the remaining number of generateable symbols, nodes whose minimum depth to complete exceeds the remaining tree depth, and nodes that violate node safety rules or create circular dependencies are removed. The remaining nodes are then used as the allowed set for the next symbol, and other nodes are masked as unselectable.

[0049] Step A2: Verify the syntax, type, and security of the candidate symbol mechanisms; normalize the expressions of the candidate symbol mechanisms that pass the verification; review the executability of the normalized expressions; and generate mechanism hash values ​​based on the normalized expressions that pass the review.

[0050] The expression normalization includes:

[0051] Calculate and replace the subtree containing only constants with the result;

[0052] Sort the operands of the commutative operation according to the hash values ​​of their child nodes;

[0053] Replace two consecutive logical NOTs with the original node;

[0054] For branches where the condition is always true or always false, only the reachable branches are retained;

[0055] In addition, map synonym node names to standard names in the registry;

[0056] The verification is used to identify always true, always false, and / or ineffective branches in the candidate symbol mechanism, and to process them as objects for expression normalization. The verification process includes:

[0057] Check whether the symbols of the candidate symbol mechanism belong to the registered symbol vocabulary;

[0058] Check whether the brackets and expression tree structure of the candidate symbol mechanism are complete;

[0059] Check whether the number of node parameters for the candidate symbol mechanism is correct;

[0060] Check whether the node input type and output type of the candidate symbol mechanism match;

[0061] Check whether the candidate symbol mechanism has division by zero, empty window and / or invalid frequency band;

[0062] Check whether the candidate symbol mechanism has circular references and / or excessive expression tree depth;

[0063] Check whether the candidate symbol mechanism invokes unregistered and / or disallowed operators;

[0064] Step A3: Organize the B data batches, C sensor channels, W time windows, and F partial discharge features into the input tensor. :

[0065] ;

[0066] Where X[b, c, w, f] represents the f-th normalized feature of the b-th batch, c-th channel, and w-th window, and a valid value mask with the same shape as X is set to identify missing or abnormal data;

[0067] Step A4: Using the mechanism hash value as the retrieval key, group the candidate symbol mechanisms that have passed the verification. Retain one duplicate candidate symbol mechanism with the same mechanism hash value and consistent normalized expression. Organize the retained M candidate symbol mechanism descriptions into a mechanism parameter tensor. :

[0068] ;

[0069] Each candidate symbol mechanism contains at most N nodes and nodes in D dimensions; each mechanism parameter tensor It should include at least node type encoding, child node index, threshold or frequency band parameter, time window parameter, and valid node mask;

[0070] Step A5: Based on the input tensor and the mechanism parameter tensor The initial matching result is obtained, wherein the input tensor Provide the feature values ​​and mechanistic parameter tensors of the sample to be diagnosed. Provides node structures and parameters for performing calculations and determinations on the aforementioned feature values;

[0071] The node types and dependency indices in the mechanism parameter tensor P are read according to the topological hierarchy of the symbolic expression tree, and nodes of the same type or with the same execution semantics are compiled into batch tensor operators; the batch tensor operators are derived from the input tensor. Alternatively, the intermediate results from the previous topology layer can be used as input, and the node values ​​can be calculated by combining the node parameters in the mechanism parameter tensor P. The nodes are then connected according to the dependency relationship to form a directed acyclic tensor computation graph, and finally, the initial matching results corresponding to B batches of data and M candidate symbol mechanisms are output.

[0072] Furthermore, in step S3, when the tensor computation graph is executed on the parallel computing unit to obtain the initial matching result, the parallel computing unit executes:

[0073] When executing the tensor computation graph, the dimension-corresponding data batches, monitored devices, sensor channels, waveform windows, partial discharge characteristics and / or candidate symbol mechanisms are processed in batches and heterogeneous parallel processing.

[0074] By implementing precondition pruning, nodes whose computational cost is lower than a preset cost threshold and whose Boolean output is a necessary condition for the execution of subsequent branches are executed first. The preconditions include pulse count, amplitude, data quality, and / or device applicability conditions. For batches, channels, or window samples that do not meet the preconditions, as well as subsequent subtree branches of the corresponding candidate symbol mechanism, Boolean masks are set, and frequency domain, time-frequency, and / or multi-channel correlation calculations are skipped.

[0075] For nodes of the same type, operator merging is performed through merging of similar operators and reuse of common subtrees; intermediate results are reused for nodes of the same common subtree.

[0076] In addition, the initial matching result, intermediate node value, execution delay, computational resource consumption and / or abnormal status of each candidate symbol mechanism are recorded.

[0077] Furthermore, in step S4, the first The degree of matching between the candidate symbol mechanism described in the previous section and the current partial discharge data is used as the original mechanism matching degree. And based on the original mechanism matching degree Get the trigger determination result ;

[0078] The original mechanism matching degree include:

[0079] ;

[0080] The trigger determination result include:

[0081] ;

[0082] in, Indicates the first The degree of matching between the candidate symbol mechanism described in the article and the current partial discharge data; This represents the minimum matching degree required for the candidate symbol mechanism to produce a positive determination, and serves as the trigger threshold.

[0083] When similar equipment and similar discharge tasks have no less than a preset number of labeled and calibrated samples, and both positive and negative samples reach a preset proportion, and pass the sample distribution consistency check, a logistic regression calibration, or order-preserving regression, or temperature scaling method is used to fit the probability mapping function. , obtained the Diagnostic probability of candidate symbol mechanism :

[0084] ;

[0085] Diagnostic probability is calculated using resampling methods or calibrated residual quantiles. When the confidence interval does not meet the calibration conditions, the output mechanism matching degree is... Risk Score and the "Uncalibrated" label;

[0086] Based on mechanism matching degree Diagnostic probability Data quality coefficient Severity coefficient Calculate risk score ,include:

[0087] ;

[0088] in, These are the non-negative weights corresponding to mechanism matching degree, diagnostic probability, data quality, and severity, respectively. This serves as a probability calibration status indicator; it indicates when the diagnostic probability has been calibrated. ,otherwise ; The risk score The value range is 0 to 100;

[0089] When the ratio between the preferred symbol mechanism and the candidate symbol mechanism does not meet the preferred threshold, some parameters of the candidate symbol mechanism are adjusted, historical backtesting and symbol distillation are performed again, and the preferred symbol mechanism and the candidate symbol mechanism are updated. The target mechanism library is then updated based on the updated preferred symbol mechanism.

[0090] When adjusting some parameters of the candidate symbol mechanism, including: feature threshold, feature frequency band, time window length, node parameters, logical connection relationship, mechanism fusion weight, applicable device range and / or enabled status.

[0091] Furthermore, in step S5, the preferred symbol mechanism that is in the officially enabled state is compiled and loaded into the official online diagnostic environment; the candidate symbol mechanism, the verification mechanism and the gray-scale operation mechanism are executed in a bypass environment isolated from the official alarm, and only evaluation records with verification identifiers are generated.

[0092] The target mechanism library includes: mechanism identifier, symbolic expression tree, tensor computation graph, applicable device type, applicable sensor type, target partial discharge type, feature threshold and time window, data version, historical backtesting index, expression complexity, mechanism creation source, mechanism version number and / or mechanism activation status;

[0093] The target mechanism library has at least candidate mechanisms, mechanisms under verification, gray-scale operation mechanisms, and states of official activation, deactivation, and rollback. When a gray-scale operation mechanism meets the conditions and is reviewed, it is switched to the official activation state.

[0094] When performing online partial discharge diagnosis based on the aforementioned formal online diagnostic environment, the following steps are included:

[0095] The mechanism of parallel execution of multiple optimal symbols on the target UHF data; for the first The optimal symbol mechanism is used to verify the comprehensive performance of the set. Equipment compatibility Input data quality and target data coverage The product is normalized to obtain the fusion weight. :

[0096] ;

[0097] Based on the fusion weight Calculate the fusion score for each partial discharge type k. :

[0098] ;

[0099] In the formula, For the first The optimal symbol mechanism determines the diagnostic probability or normalized matching degree of type k;

[0100] Preset weight threshold, select fusion weight The high-weighted selection mechanism, which prioritizes symbols above a preset weight threshold, is used in the fusion scoring. The maximum value is selected as the highest fusion score. When the partial discharge type determinations of multiple high-weight preferred symbol mechanisms are consistent, and the partial discharge type is the same as the partial discharge type corresponding to the highest fusion score, and the highest fusion score reaches the alarm threshold, the alarm conclusion of the partial discharge type is output. Otherwise, the status of pending review is output, and the conflict mechanism and evidence are displayed. The evidence includes mechanism identifier, input channel and time window, key feature value, judgment threshold corresponding to key feature, mechanism trigger threshold and alarm threshold, original mechanism matching degree, calibration status, diagnostic probability output when the probability calibration condition is met, and complete judgment path.

[0101] The online diagnostic results are generated as follows: the type corresponding to the highest fusion score is taken as the partial discharge type; the partial discharge status is determined by the highest fusion score and the alarm threshold; based on the risk score... The severity of partial discharge is determined by the interval; the suspected discharge location or region is determined based on the arrival time difference of multiple channels, installation location, energy attenuation ratio between channels, and propagation consistency; the mechanism and characteristics that contribute to the fusion score to the preset contribution threshold are respectively determined as the trigger symbol mechanism and key characteristics, and the node judgment path is output.

[0102] The online diagnostic results also include the discharge development trend: based on the number of pulses, window energy, characteristic frequency band energy and phase concentration of K consecutive time windows, and the matching degree of the original mechanism output by the preferred symbol mechanism in the corresponding window when it is in the officially activated state, linear regression or robust regression is performed with the window corresponding time as the independent variable, the regression slope is used as the growth rate, the difference between adjacent regression slopes divided by the corresponding time interval is used as the growth acceleration, and the growth rate, growth acceleration and the number of continuous windows in the same direction are used as trend features, where K is not less than 3;

[0103] Provided that the data quality and sample coverage reach the corresponding thresholds, if at least two preferred symbol mechanisms in the officially enabled state support the same trend, or if a high-weight preferred symbol mechanism continuously supports the trend, then the discharge development trend is determined according to the characteristics of the trend: when the confidence interval of the growth rate falls into the preset zero value neighborhood and the number of continuous windows reaches the stability determination threshold, a stable discharge development trend is output; when the stability determination condition is not met, and at least L windows in K windows change in the same direction and the confidence interval of the corresponding growth rate does not contain zero, the discharge development trend of rising, rapidly rising, or falling is output according to the positive or negative sign of the growth rate and the comparison result of the growth rate, growth acceleration, and preset rapid rise threshold; when the corresponding determination condition is not met, an uncertain discharge development trend is output, where 1≤L≤K;

[0104] Using a preferred symbol mechanism that includes time-series nodes, trend characteristics are extrapolated according to the current growth rate and growth acceleration within a preset prediction time H after the current moment. The extrapolated mechanism fusion score, normalized growth rate, growth acceleration and duration are weighted and mapped to 0 to 100 to obtain the trend risk score. [t, t+H] is used as the prediction time window, and the moment when the trend risk score first reaches the preset risk threshold is taken as the expected trigger time. The uncertainty range of the expected trigger time is given based on the regression residual or resampling results.

[0105] In a second aspect, embodiments of this disclosure provide a system for online diagnosis of partial discharge in electrical equipment, comprising:

[0106] The initialization unit is configured to acquire the ultra-high frequency waveform of at least one sensor in the high-voltage electrical equipment and form an initial dataset. The initial dataset is then preprocessed to obtain data batches for parallel processing.

[0107] The feature unit is configured to calculate the partial discharge features of each waveform window in the data batch, the partial discharge features including time-domain features, frequency-domain features, time-frequency features, phase features, pulse statistics features, and multi-channel spatial features;

[0108] The matching unit is configured to organize the partial discharge features and the pre-constructed partial discharge-specific symbol system into an input tensor, and generate a tensor computation graph based on the dependency relationship, node type and topological level of the symbol tree. The initial matching result is obtained through batch processing, merging of similar operators, reuse of common subtrees, precondition pruning and heterogeneous parallelism.

[0109] The distillation unit is configured to perform probability calibration on the initial matching result to obtain a reliable feedback label for the initial matching result, and to perform historical backtesting and symbol distillation on the initial matching result based on historical UHF data and the reliable feedback label to obtain the preferred symbol mechanism and the candidate symbol mechanism.

[0110] The output unit is configured to construct a formal online diagnostic environment based on the preferred symbol mechanism in a formally enabled state, and output the online diagnostic results.

[0111] In a third aspect, embodiments of this disclosure provide an electronic device, characterized in that the electronic device comprises:

[0112] At least one processor; and,

[0113] The memory is communicatively connected to the at least one processor; wherein,

[0114] The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method for online diagnosis of partial discharge in electrical equipment.

[0115] In a fourth aspect, embodiments of this disclosure provide a non-transitory computer-readable storage medium, characterized in that the non-transitory computer-readable storage medium stores computer instructions for causing the computer to perform the method for online diagnosis of partial discharge of electrical equipment.

[0116] Compared with the prior art, the present invention has the following beneficial technical effects:

[0117] This invention provides a method, system, device, and medium for online diagnosis of partial discharge in electrical equipment. The method preprocesses the initial dataset and transforms it into data batches capable of parallel processing. It calculates the partial discharge characteristics of each waveform window in the data batch, organizes these characteristics and a pre-constructed partial discharge-specific symbol system into an input tensor, generates a tensor computation graph based on the symbol tree's dependencies, node types, and topological hierarchy, and performs batch execution on heterogeneous computing resources. This enables simultaneous processing of multiple devices, channels, windows, and mechanisms, making it suitable for online applications with high sampling rates and ultra-high frequency data. It employs common subtree reuse, operator merging, and preconditioning. Pruning can further reduce redundant computation, providing a technical foundation for low-latency deployment and large-scale centralized analysis at the station level. Then, the initial matching results are probabilistically calibrated to obtain a reliable feedback label for the initial matching results. Based on historical UHF data and reliable feedback labels, the initial matching results are backtested and symbol distilled to obtain the preferred symbol mechanism and candidate symbol mechanism. This process can convert partial discharge expert knowledge from natural language or decentralized program rules into symbol mechanisms with type constraints and deterministic execution semantics, enabling the mechanism to be verified, backtested, combined, reused, and versioned. This is conducive to forming maintainable diagnostic knowledge assets across manufacturers and equipment types. Attached Figure Description

[0118] Figure 1 A flowchart illustrating an online diagnostic method for partial discharge in electrical equipment according to an embodiment of this disclosure is shown.

[0119] Figure 2 A schematic diagram of partial discharge feature extraction according to an embodiment of this disclosure is shown;

[0120] Figure 3 A characteristic amplitude distribution diagram of the "pre-slope" time interval during the partial discharge development stage of an embodiment of this disclosure is shown;

[0121] Figure 4 A characteristic diagram of amplitude distribution during the "middle of the slope" time interval in the partial discharge development stage of an embodiment of this disclosure is shown;

[0122] Figure 5 A characteristic diagram of amplitude distribution in the "post-slope" time interval during the partial discharge development stage of an embodiment of this disclosure is shown.

[0123] Figure 6 A diagram of an apparatus for online diagnosis of partial discharge in electrical equipment, according to an embodiment of the present disclosure, is shown. Detailed Implementation

[0124] The embodiments of this disclosure will now be described in detail with reference to the accompanying drawings.

[0125] The following specific examples illustrate the implementation of this disclosure. Those skilled in the art can easily understand other advantages and effects of this disclosure from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. This disclosure can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this disclosure. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0126] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this disclosure, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0127] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this disclosure. The drawings only show the components related to this disclosure and are not drawn according to the number, shape and size of the components in actual implementation. In actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0128] Furthermore, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.

[0129] Figure 1 A flowchart 100 of a method for online diagnosis of partial discharge in electrical equipment according to the first embodiment of this application is shown, as follows: Figure 1 As shown, it includes the following steps:

[0130] In step S101, the ultra-high frequency waveform of at least one sensor in the high-voltage electrical equipment is acquired and an initial dataset is formed. The initial dataset is preprocessed to obtain a data batch for parallel processing.

[0131] Specifically, ultra-high frequency waveforms are acquired by one or more ultra-high frequency sensors installed on the monitored high-voltage electrical equipment, forming raw data records which serve as the initial dataset in step S101. include:

[0132] ;

[0133] In the formula, It is a waveform sampling sequence; For the time of data collection; The sampling frequency; For sensor or channel identification; It is the power frequency phase; For equipment identification;

[0134] It should be noted that the raw data record may also include triggering mode, range, gain, sensor location, propagation path, operating voltage, load, switching operation, external radio interference, and maintenance event identifiers. The frequency response range, detection sensitivity, and time synchronization accuracy are retained for different acquisition devices. Specifically, UHF signals can be acquired by built-in, external, window-type, or portable sensors; continuous acquisition, threshold triggering, periodic acquisition, event triggering, or multi-level triggering can be used. There can be one or more sensors, and multiple channels can be synchronized using hardware synchronization, software time correction, or reference pulse correction. In addition to UHF signals, high-frequency current, ultrasound, transient ground voltage, equipment vibration, temperature, operating voltage, and switching operation information can also be synchronously input as auxiliary evidence, but the main input for partial discharge diagnosis still includes UHF waveforms or their derived characteristics.

[0135] During the preprocessing, the UHF waveform is subjected to preprocessing including DC component removal, bandpass filtering, power frequency and narrowband interference suppression, background noise estimation, pulse triggering, baseline correction, amplitude normalization and / or abnormal saturation identification, and the data batch is obtained.

[0136] For multi-channel data, time alignment is performed based on a hardware clock, synchronization pulse, or correction signal, and waveform windows are divided according to a fixed length, pulse trigger, or sliding window method to obtain the data batch.

[0137] It should be further explained that, when preprocessing the aforementioned UHF waveform, finite impulse response filtering, infinite impulse response filtering, wavelet thresholding denoising, adaptive noise cancellation, spectral subtraction, narrowband notch filtering, independent component analysis, or interference cancellation based on background templates can be employed. The waveform window can be a fixed length, a sliding length, an adaptive pulse center length, or a statistical window spanning the power frequency cycle. Features can consist of manually defined time-domain, frequency-domain, time-frequency, phase, and statistical features, or can include low-dimensional embeddings output by convolutional networks, autoencoders, or other feature extractors; low-dimensional embedding serves as a mechanism for inputting symbols to data nodes with defined dimensions and value ranges.

[0138] Next, proceed to step S102.

[0139] In step S102, the partial discharge characteristics of each waveform window in the data batch are calculated. The partial discharge characteristics include time domain characteristics, frequency domain characteristics, time-frequency characteristics, phase characteristics, pulse statistical characteristics, and multi-channel spatial characteristics.

[0140] Specifically, obtaining the partial discharge characteristics in step S102 includes the following steps:

[0141] Assuming the waveform window is , for the shape window After performing a short-time Fourier transform or wavelet transform, a coefficient matrix that varies with time and frequency is obtained, and the amplitude, energy and energy ratio of each time-frequency unit are used as the time-frequency characteristics of the waveform window.

[0142] The time-domain characteristics of the waveform window include the root mean square value. and window energy , wherein the root mean square value include:

[0143] ;

[0144] The window energy include:

[0145] ;

[0146] The frequency domain characteristics of the waveform window are represented by the spectral proportion. This indicates the percentage of the spectrum within the specified frequency band [f1, f2]. include:

[0147] ;

[0148] In the formula, The spectrum of the ultra-high frequency waveform. To prevent positive numbers from being divided by zero;

[0149] The multi-channel spatial characteristics of the waveform window are represented by the arrival time difference of the multi-channel pulses. This indicates the arrival time difference of the multi-channel pulses. include:

[0150] ;

[0151] In the formula, for Arrival time of channel pulse; for Arrival time of channel pulse;

[0152] The pulse statistical features include pulse count, peak value, rise time, pulse width, pulse interval, and power frequency phase clustering, as well as the statistical values ​​of the pulse statistical features within a continuous time window;

[0153] Based on channel and channels The window energy is used to obtain the inter-channel energy attenuation ratio. :

[0154] ;

[0155] Calculate the channel within the preset time delay search range. and channels The normalized cross-correlation of the waveforms is used to obtain the cross-correlation peak value, which is then used as the correlation coefficient. The time delay corresponding to the peak cross-correlation value is used as the arrival time difference. ;

[0156] Based on the arrival time difference Align the multi-channel waveforms and calculate the correlation coefficient of any channel pair. and normalized residuals And calculate propagation consistency ,include:

[0157] ;

[0158] Among them, normalized residuals For channel The waveform and the amplitude-scaled channel The L2 norm of the difference between aligned waveforms divided by the L2 norm of channel i waveform and sum; The value range is [0, 1];

[0159] The power frequency phase, amplitude, and occurrence period of the pulse are mapped to a PRPD / PRPS spectrum, and the phase distribution and power frequency phase clustering in the PRPD / PRPS spectrum are extracted as the phase features.

[0160] Next, proceed to step S103.

[0161] In step S103, the partial discharge features and the pre-constructed partial discharge-specific symbol system are organized into an input tensor, and a tensor computation graph is generated according to the dependency relationship of the symbol tree, node type and topological level. The initial matching result is obtained through batch processing, merging of similar operators, reuse of common subtrees, precondition pruning and heterogeneous parallelism.

[0162] Specifically, the partial discharge special symbol system in step S103 includes waveform nodes, sensor nodes, time window nodes, time domain operation nodes, frequency domain operation nodes, phase nodes, statistical nodes, spatial correlation nodes, condition nodes, logic nodes, timing nodes, and result nodes. Each type of node is predefined with node name, number of parameters, input type, output type, allowed sub-nodes, numerical range, and execution function.

[0163] It also includes the initial symbol mechanism, which refers to the partial discharge diagnostic expression composed of registered nodes, parameters, dependencies and execution semantics. The partial discharge diagnostic expression includes at least feature calculation, condition judgment and result output. Each intermediate node of the partial discharge diagnostic expression stores the node value, threshold, input window and execution state.

[0164] In this embodiment, an example expression of the initial symbol mechanism is shown:

[0165] Result("Floating potential discharge", "Alarm")

[0166] Continuous

[0167] And(

[0168] Greater(BandRatio(Wave(c,w), f1, f2), θ1),

[0169] Greater(PulseCount(Wave(c,w), θp), θ2),

[0170] Greater(PhaseConcentration(PRPD(c,w), Φ), θ3)

[0171] ), L, K

[0172] ), evidence=[BandRatio, PulseCount, PhaseConcentration])

[0173] Where f1, f2, θ1, θ2, θ3, θp, L, and K are fitted from the historical training set or given by experts, and are fixed on an independent validation set before entering the test. Each intermediate node of the expression stores the node value, threshold, input window, and execution state to form a traceable decision path.

[0174] The steps involved in generating the tensor computation graph are as follows:

[0175] Step A1: Generate candidate symbol mechanisms based on the initial symbol mechanism, including:

[0176] By selecting or searching for each configurable parameter from the allowed range using a preset parameterized rule template, one or more candidate symbol mechanisms are generated. The preset parameterized rule template is a symbolic expression skeleton that represents the verified partial discharge diagnostic rules as having a fixed node type and logical structure, and is configurable with threshold, frequency band, time window and applicable device range.

[0177] Alternatively, select a subtree with the same input and output types from historical effective symbol mechanisms, perform subtree cross-combination at type-compatible node positions, and mutate at least one of the threshold, frequency band, time window, comparison operator, or logical connection relationship to generate candidate symbol mechanisms;

[0178] Alternatively, a trained symbol mechanism generation model can encode the device type, target partial discharge type, available channels, feature vocabulary, upper limit of expression tree depth, and evaluation target as task conditions. Expressions are generated symbol by symbol using preorder traversal, postorder traversal, or parenthesized function forms. Syntax masks and type constraints are used during generation until a result node is generated and all parameters to be supplemented are closed, yielding candidate symbol mechanisms. During symbol generation, nodes matching the output type are selected from the registered symbol vocabulary based on the data type required by the input slot to be filled in the current parent node. Nodes with more parameters than the remaining number of generateable symbols, nodes whose minimum depth to complete exceeds the remaining tree depth, and nodes that violate node safety rules or create circular dependencies are removed. The remaining nodes are then used as the allowed set for the next symbol, and other nodes are masked as unselectable.

[0179] Step A2: Verify the syntax, type, and security of the candidate symbol mechanisms; normalize the expressions of the candidate symbol mechanisms that pass the verification; review the executability of the normalized expressions; and generate mechanism hash values ​​based on the normalized expressions that pass the review.

[0180] The expression normalization includes:

[0181] Calculate and replace the subtree containing only constants with the result;

[0182] Sort the operands of the commutative operation according to the hash values ​​of their child nodes;

[0183] Replace two consecutive NOT statements with the original node;

[0184] For branches where the condition is always true or always false, only the reachable branches are retained;

[0185] In addition, map synonym node names to standard names in the registry;

[0186] The verification is used to identify always true, always false, and / or ineffective branches in the candidate symbol mechanism, and to process them as objects for expression normalization. The verification process includes:

[0187] Check whether the symbols of the candidate symbol mechanism belong to the registered symbol vocabulary;

[0188] Check whether the brackets and expression tree structure of the candidate symbol mechanism are complete;

[0189] Check whether the number of node parameters for the candidate symbol mechanism is correct;

[0190] Check whether the node input type and output type of the candidate symbol mechanism match;

[0191] Check whether the candidate symbol mechanism has division by zero, empty window and / or invalid frequency band;

[0192] Check whether the candidate symbol mechanism has circular references and / or excessive expression tree depth;

[0193] Check whether the candidate symbol mechanism invokes unregistered and / or disallowed operators;

[0194] It should be noted that candidate symbol mechanisms that fail the verification or review in step A2 must not be included in the historical backtesting process or the online mechanism library. Expression normalization should maintain the execution semantics defined by the registered nodes; for commutative operands with the same child node hash value, the sorting is determined according to the normalized expression of the complete subtree. The deduplication check in step A4 is performed under the same symbol registry and execution semantic version. The complete normalized expression includes node type, parameters, dependencies, and result node information; the mechanism hash value is only used to narrow the search scope, and whether it is a duplicate is determined by the comparison result of the complete normalized expression.

[0195] For candidate symbol mechanisms with different structures after normalization, mechanism pairs with a common applicable input domain are first screened according to the applicable equipment range, input type, required channels, time window conditions, and output type. Pre-processed preset test waveforms or subsets of historical data are input into the mechanism pairs. The input samples cover typical partial discharge, non-discharge interference, missing or abnormal data states, and situations near the judgment threshold. The Boolean judgments, partial discharge types, and effective states of the outputs are compared item by item, and the numerical outputs are also compared. When all the non-numerical outputs corresponding to all test inputs are consistent, and the differences between the numerical outputs do not exceed the preset tolerance, the mechanism pair is marked as potentially equivalent. If any test input does not meet the comparison conditions, that input is recorded as a counterexample of inequivalence. Test consistency on a limited sample is only used to screen mechanism pairs to be further detected and is not directly used as the basis for merging mechanisms.

[0196] For mechanism pairs that can be fully converted into logical and arithmetic constraints, the common applicable input domain, parameter range, missing data handling rules, and time window conditions are encoded as constraints. A query is performed to check if there are inputs that cause the Boolean decisions, categories, or valid states of the two mechanisms to differ, or whose numerical output differences exceed a preset tolerance. If the query result is satisfactory, the obtained input is recorded as an example of inequivalence; if the query result is unsatisfactory, the outputs of the two mechanisms are confirmed to be equivalent within the input domain defined by the constraints; if the encoding cannot be fully completed, the solution times out, or the result is unknown, they are retained as different mechanisms. Equivalent mechanism merging is only performed when the equivalence conclusion covers the entire intended application range of both mechanisms and retains the associated records of their respective mechanism identifiers, versions, and evidence paths; otherwise, only the detection conclusion is recorded.

[0197] Step A3: Organize the B data batches, C sensor channels, W time windows, and F partial discharge features into the input tensor. :

[0198] ;

[0199] Where X[b, c, w, f] represents the f-th normalized feature of the b-th batch, c-th channel, and w-th window, and a valid value mask of the same shape as X is set to identify missing or abnormal data;

[0200] Step A4: Using the mechanism hash value as the retrieval key, group the candidate symbol mechanisms that have passed the verification. For duplicate candidate symbol mechanisms with the same mechanism hash value and consistent normalized expression, retain only one. Organize the retained M candidate symbol mechanism descriptions into a mechanism parameter tensor. :

[0201] ;

[0202] Each candidate symbol mechanism contains at most N nodes and nodes in D dimensions; each mechanism parameter tensor It should include at least node type encoding, child node index, threshold or frequency band parameter, time window parameter, and valid node mask;

[0203] Step A5: Based on the input tensor and the mechanism parameter tensor The initial matching result is obtained, wherein the input tensor Provide the feature values ​​and mechanistic parameter tensors of the sample to be diagnosed. Provides node structures and parameters for performing calculations and determinations on the aforementioned feature values;

[0204] The node types and dependency indices in the mechanism parameter tensor P are read according to the topological hierarchy of the symbolic expression tree, and nodes of the same type or with the same execution semantics are compiled into batch tensor operators; the batch tensor operators are derived from the input tensor. Alternatively, the intermediate results from the previous topology layer can be used as input, and the node values ​​can be calculated by combining the node parameters in the mechanism parameter tensor P. The nodes are then connected according to the dependency relationship to form a directed acyclic tensor computation graph, and finally, the initial matching results corresponding to B batches of data and M candidate symbol mechanisms are output.

[0205] It should be noted that during the compilation of the tensor computation graph:

[0206] 1. Assign shared intermediate results to common subtrees in different mechanisms;

[0207] 2. Generate Boolean masks for condition nodes and logical branches;

[0208] 3. Cache fixed thresholds, frequency band indexes, and time window indexes;

[0209] 4. Merge nodes of the same type and with the same parameter configuration into batch operators;

[0210] 5. Divide and schedule the computation graph according to the operator types supported by the target heterogeneous computing units.

[0211] Condition nodes can be grouped by topology level, node type, feature type, device, channel, window, or mechanism number. Execution plans can be generated using static graphs, dynamic graphs, just-in-time compilation, or pre-compilation. Common subtree reuse can occur within a single mechanism, or across mechanisms, device templates, or time windows.

[0212] More specifically, in step S103, when the tensor computation graph is executed on the parallel computing unit to obtain the initial matching result, the parallel computing unit executes:

[0213] When executing the tensor computation graph, the dimension-corresponding data batches, monitored devices, sensor channels, waveform windows, partial discharge characteristics and / or candidate symbol mechanisms are processed in batches and heterogeneous parallel processing.

[0214] Precondition pruning prioritizes nodes whose computational cost is below a preset cost threshold and whose Boolean output is a necessary condition for subsequent branch execution. These preconditions include pulse count, amplitude, data quality, and / or device applicability. For batches, channels, or window samples that do not meet the preconditions, as well as subsequent subtree branches of the corresponding candidate symbol mechanism, Boolean masks are set, and frequency domain, time-frequency, and / or multi-channel correlation calculations are skipped. It should be noted that precondition pruning can be implemented using Boolean masks, sparse indexes, conditional operators, and / or task queues.

[0215] For nodes of the same type, operator merging is performed through merging of similar operators and reuse of common subtrees; intermediate results are reused for nodes of the same common subtree.

[0216] In addition, the initial matching result, intermediate node value, execution delay, computational resource consumption and / or abnormal status of each candidate symbol mechanism are recorded.

[0217] In this embodiment, a tensor computation graph is executed on a CPU, GPU, NPU, FPGA or other parallel computing unit. At the same time, the matching results, intermediate node values, execution latency, computing resource usage and abnormal states of each mechanism are recorded to ensure that the diagnostic process is repeatable, interpretable and traceable.

[0218] Next, proceed to step S104.

[0219] In step S104, the initial matching result is probabilistically calibrated to obtain a reliable feedback label for the initial matching result. Based on historical UHF data and the reliable feedback label, the initial matching result is subjected to historical backtesting and symbol distillation to obtain the preferred symbol mechanism and candidate symbol mechanism.

[0220] Specifically, in step S104, the first The degree of matching between the candidate symbol mechanism described in the previous section and the current partial discharge data is used as the original mechanism matching degree. And based on the original mechanism matching degree Get the trigger determination result ;

[0221] The original mechanism matching degree include:

[0222] ;

[0223] The trigger determination result include:

[0224] ;

[0225] in, Indicates the first The degree of matching between the candidate symbol mechanism described in the article and the current partial discharge data; This represents the minimum matching degree required for the candidate symbol mechanism to produce a positive determination, and serves as the trigger threshold.

[0226] When similar equipment and similar discharge tasks have no less than a preset number of labeled and calibrated samples, and both positive and negative samples reach a preset proportion, and pass the sample distribution consistency check, a logistic regression calibration, or order-preserving regression, or temperature scaling method is used to fit the probability mapping function. , obtained the Diagnostic probability of candidate symbol mechanism :

[0227] ;

[0228] Diagnostic probability is calculated using resampling methods or calibrated residual quantiles. When the confidence interval does not meet the calibration conditions, the output mechanism matching degree is... Risk Score and the "Uncalibrated" label;

[0229] Based on mechanism matching degree Diagnostic probability Data quality coefficient Severity coefficient Calculate risk score ,include:

[0230] ;

[0231] in, These are the non-negative weights corresponding to mechanism matching degree, diagnostic probability, data quality, and severity, respectively. This serves as a probability calibration status indicator; it indicates when the diagnostic probability has been calibrated. ,otherwise ; The risk score The value range is 0 to 100;

[0232] When the ratio between the preferred symbol mechanism and the candidate symbol mechanism does not meet the preferred threshold, some parameters of the candidate symbol mechanism are adjusted, historical backtesting and symbol distillation are performed again, and the preferred symbol mechanism and the candidate symbol mechanism are updated. The target mechanism library is updated based on the updated preferred symbol mechanism. It should be noted that the mechanism library can be updated according to period, cumulative sample count, performance degradation event or manual instruction.

[0233] When adjusting some parameters of the candidate symbol mechanism, including: feature threshold, feature frequency band, time window length, node parameters, logical connection relationship, mechanism fusion weight, applicable device range and / or enabled status.

[0234] It should be noted that symbol distillation refers to the process of selecting a set of target symbol mechanisms from a large set of candidate symbol mechanisms that meet performance requirements, have controlled expression size, are tensorably quantizable, and have a clear scope of application. In some embodiments, target mechanism selection may employ weighted comprehensive scoring, Pareto front, hierarchical thresholding, or device type-based grouping. Multi-mechanism fusion may employ weighted voting, maximum confidence, evidence theory, Bayesian update, or a learning-based fusion engine. Regardless of the method used, individual mechanism results, intermediate node values, weight sources, and conflict resolution results are retained to support retrospective analysis.

[0235] It should be further noted that any node type, parameters, dependencies, and execution semantics that can be recovered can be used as the preferred symbol mechanism in this embodiment. The tree-like symbol mechanism can be stored using preorder sequences, postorder sequences, bracketed functions, abstract syntax trees, directed acyclic graphs with shared subtrees, bytecode, or domain-specific languages. Conditional nodes can use hard threshold comparisons, or comparisons with hysteresis intervals, fuzzy membership degrees, or piecewise functions to reduce frequent jumps near the threshold.

[0236] Candidate symbol mechanisms can be generated directly by experts, through template parameterization, genetic programming, program synthesis, historical mechanism combination mutation, or generative models. When using generative models, decoder-only, encoder-decoder, or other sequence generation structures can be used; model optimization can employ supervised fine-tuning, preference optimization, or reinforcement learning. If generative models are not used, a target mechanism library can still be formed through expert templates, historical backtesting, and parameter search, without affecting the implementation of the tensor quantization execution scheme.

[0237] Next, proceed to step S105.

[0238] In step S105, a formal online diagnostic environment is constructed based on the preferred symbol mechanism that is in a formally enabled state, and the online diagnostic results are output.

[0239] Specifically, in step S105, the preferred symbol mechanism that is in the officially enabled state is compiled and loaded into the official online diagnostic environment; the candidate symbol mechanism, the verification mechanism and the gray-scale operation mechanism are executed in a bypass environment isolated from the official alarm, and only evaluation records with verification identifiers are generated.

[0240] The target mechanism library includes: mechanism identifier, symbolic expression tree, tensor computation graph, applicable device type, applicable sensor type, target partial discharge type, feature threshold and time window, data version, historical backtesting index, expression complexity, mechanism creation source, mechanism version number and / or mechanism activation status;

[0241] The target mechanism library has at least candidate mechanisms, mechanisms under verification, gray-scale operation mechanisms, and states of official activation, deactivation, and rollback. When a gray-scale operation mechanism meets the conditions and is reviewed, it is switched to the official activation state.

[0242] When performing online partial discharge diagnosis based on the aforementioned formal online diagnostic environment, the following steps are included:

[0243] The mechanism of parallel execution of multiple optimal symbols on the target UHF data; for the first The optimal symbol mechanism is used to verify the comprehensive performance of the set. Equipment compatibility Input data quality and target data coverage The product is normalized to obtain the fusion weight. :

[0244] ;

[0245] Based on the fusion weight Calculate the fusion score for each partial discharge type k. :

[0246] ;

[0247] In the formula, For the first The optimal symbol mechanism determines the diagnostic probability or normalized matching degree of type k;

[0248] Preset weight threshold, select fusion weight The high-weighted selection mechanism, which prioritizes symbols above a preset weight threshold, is used in the fusion scoring. The maximum value is selected as the highest fusion score. When the partial discharge type determinations of multiple high-weight preferred symbol mechanisms are consistent, and the partial discharge type is the same as the partial discharge type corresponding to the highest fusion score, and the highest fusion score reaches the alarm threshold, the alarm conclusion of the partial discharge type is output. Otherwise, the status of pending review is output, and the conflict mechanism and evidence are displayed. The evidence includes mechanism identifier, input channel and time window, key feature value, judgment threshold corresponding to key feature, mechanism trigger threshold and alarm threshold, original mechanism matching degree, calibration status, diagnostic probability output when the probability calibration condition is met, and complete judgment path.

[0249] The online diagnostic results are generated as follows: the type corresponding to the highest fusion score is taken as the partial discharge type; the partial discharge status is determined by the highest fusion score and the alarm threshold; based on the risk score... The severity of partial discharge is determined by the interval; suspected discharge locations or regions are determined based on multi-channel arrival time difference, installation location, inter-channel energy attenuation ratio, and propagation consistency; the mechanisms and characteristics that contribute to the fusion score reaching a preset contribution threshold are respectively identified as trigger symbol mechanisms and key features, and the node determination path is output; it should be noted that the suspected discharge location can be determined by arrival time difference, amplitude attenuation, cross-correlation peak value, propagation path model, sensor topology constraints, or multi-source evidence. For devices with uncertain propagation speed or path, segments, cavities, intervals, or confidence regions can be output instead of forcing the output of a single coordinate. For single-sensor scenarios, the spatial positioning node can be omitted, and only the partial discharge type, severity, and trend are output.

[0250] The online diagnostic results also include the discharge development trend: based on the number of pulses, window energy, characteristic frequency band energy, phase concentration degree, and the original mechanism matching degree of the preferred symbol mechanism in the corresponding window output in the formally activated state for K consecutive time windows, linear regression or robust regression is performed with the window corresponding time as the independent variable, the regression slope is used as the growth rate, the difference between adjacent regression slopes divided by the corresponding time interval is used as the growth acceleration, and the growth rate, growth acceleration and the number of continuous windows in the same direction are used as trend features, where K is not less than 3;

[0251] Provided that data quality and sample coverage reach the corresponding thresholds, if at least two preferred symbol mechanisms in the officially enabled state support the same trend, or if a high-weight preferred symbol mechanism continuously supports the trend, then the discharge development trend is further determined based on the trend characteristics: when the confidence interval of the growth rate falls into the preset zero-value neighborhood and the number of continuous windows reaches the stability determination threshold, the output is stable; if the stability determination condition is not met, and at least L windows in K windows change in the same direction, and the confidence interval of the corresponding growth rate does not contain zero, the output is rising, rapidly rising, or falling based on the sign of the growth rate and the comparison results of the growth rate, growth acceleration, and preset rapid rise threshold; if the corresponding determination condition is not met, the output is uncertain, where 1≤L≤K. Candidate symbol mechanisms, verification mechanisms, and grayscale operation mechanisms do not participate in the determination of the formal discharge development trend;

[0252] Using a preferred symbol mechanism that includes time-series nodes, trend characteristics are extrapolated according to the current growth rate and growth acceleration within a preset prediction time H after the current moment. The extrapolated mechanism fusion score, normalized growth rate, growth acceleration and duration are weighted and mapped to 0 to 100 to obtain the trend risk score. [t, t+H] is used as the prediction time window, and the moment when the trend risk score first reaches the preset risk threshold is taken as the expected trigger time. The uncertainty range of the expected trigger time is given based on the regression residual or resampling results.

[0253] The method described in this application can use only a CPU, or any combination of a CPU with a GPU, NPU, FPGA, DSP, or other acceleration units. The specific timing can be determined based on operator support, data size, expected latency, GPU or RAM usage, power consumption, and task priority. The station-side can perform low-latency screening and alarms, while the central station can perform complex backtesting, mechanism distillation, and model training; both ends maintain consistency through mechanism versions and verification values. Diagnostic results can be output through a comprehensive online monitoring platform, station-side terminals, mobile inspection terminals, or interface services.

[0254] Example 1: Online identification of partial discharge in GIS.

[0255] UHF monitoring data from 24 GIS devices in 6 substations were selected, including 12,600 normal samples and 2,850 partial discharge samples confirmed on-site. The discharge types included floating potential, metal particles, sharp points, and insulator surface discharge. The tags were obtained from live-line testing, power outage tests, disassembly inspections, and expert review.

[0256] Data was divided into training, validation, and test sets in approximately a 6:2:2 ratio based on device identification, with data from the same device not crossing dataset boundaries. An ultra-high frequency sensor with an effective bandwidth of 300–1500 MHz was used, with a sampling frequency of 5 GS / s, a single sampling window of 2 μs, and a trigger threshold equal to the sum of the mean background noise and six times its standard deviation. DC removal, bandpass filtering, narrowband interference suppression, baseline correction, and saturation identification were performed on the waveforms, preserving the calibration amplitude sequence in mV and generating a normalized amplitude sequence for mechanistic calculations.

[0257] Figure 2 The process of extracting window-level amplitude and moving average amplitude from UHF waveforms is shown, which can then be further used to extract frequency band energy ratio, pulse rise time, number of effective pulses, PRPD phase convergence, multi-channel correlation coefficient and amplitude statistical characteristics.

[0258] Based on the changes in partial discharge characteristics over time, the continuous data was divided into three intervals: before the slope, in the middle of the slope, and after the slope. Figures 3 to 5 The pulse amplitude distribution for the three intervals is shown separately. The horizontal axis represents the pulse amplitude in mV; the vertical axis represents the sample distribution proportion in %; the green bars represent the sample proportion for each amplitude interval, and the red curve represents the smooth distribution trend. From Figure 3 It can be seen that the amplitude in the slope-front region is mainly distributed in the range of −10mV to 10mV, with a relatively dispersed distribution and no obvious peak, which can be used as a background baseline. Figure 4 It can be seen that the amplitude range in the middle section of the slope extends to approximately -120mV to 110mV, forming a peak near 0mV with long tails on both sides, indicating an increase in high-amplitude abnormal pulses and enhanced partial discharge activity. From Figure 5It can be seen that the amplitude in the post-slope region is mainly concentrated in the range of 0mV to 15mV and is distributed to the right, indicating that the partial discharge has entered a new relatively stable state that is different from the baseline before the slope.

[0259] In this embodiment, the mean amplitude, variance, skewness, kurtosis, quantile, tail proportion, and continuous window rate of change of the three intervals are input into the target mechanism library along with energy, pulse number, and phase concentration.

[0260] 320 candidate symbol mechanisms were generated through rule templates, historical mechanism combinations, and parameter search. After syntax and type validation, historical backtesting, merging of equivalent mechanisms, complexity constraints, and cross-device stability screening, 28 preferred symbol mechanisms were obtained. Node type, child node index, node parameters, and effective node mask were encoded into input tensors, and parallel execution was performed through operator merging, common subtree reuse, and precondition pruning.

[0261] In an environment with an Intel Xeon Gold 6330 processor, NVIDIA A10 GPU, and 64GB of memory, the test set recall rate was 94.3%, and the false positive rate was 3.1%. The average latency for a single batch of 64 samples was 18.6ms, and the P95 latency was 27.4ms, representing a speedup of 7.8 times compared to serial interpretation execution. The system outputs partial discharge status, type, severity, fusion risk score, key features, and decision path; when the fusion score is insufficient, there is a mechanism conflict, or the data quality is low, the system outputs "Pending Review".

[0262] Example 2: Analysis of partial discharge trends and suspected locations in transformers and parallel reactors.

[0263] A 500kV oil-immersed shunt reactor was selected, and four UHF sensors were installed on the high-voltage side, middle, low-voltage side, and bottom of the enclosure. The effective frequency band of the sensors is 300–1200MHz, the sampling frequency is 5GS / s, the single sampling window is 2μs, and the synchronization error between channels is no greater than 0.4ns.

[0264] The multi-channel waveforms are time-aligned, DC removed, bandpass filtered, narrowband interference suppressed, pulse detected, and baseline corrected. Data batches are formed with a 10s statistical window and a 5s sliding step size. The window energy, the energy ratio of the 400-800MHz frequency band, the number of effective pulses, the PRPD phase convergence, the time difference of arrival, the channel correlation coefficient, and the amplitude distribution characteristics are calculated.

[0265] This embodiment adopts Figures 3 to 5 The method of segmenting the slope into front, middle, and back sections is shown. Figure 3 A dynamic baseline is established using the mean, variance, quantiles, and tail proportions of the amplitude in the slope-front interval shown; when a continuous window is presented... Figure 4When the amplitude range expands, the long tail strengthens, and the number of high-amplitude pulses increases, it is determined that the partial discharge has entered the mid-slope growth stage; when the characteristic rate of change decreases, but the amplitude center, skewness, and proportion of high-amplitude samples still differ from the pre-slope baseline, and exhibit... Figure 5 When the positive amplitude clustering characteristic shown is displayed, it is determined that the slope has entered the stable stage.

[0266] The system will Figures 3 to 5 The amplitude distribution changes reflected, along with frequency band energy, pulse quantity, phase clustering, and multi-channel correlation, are used to input the trend diagnosis mechanism. When the normalized feature slope is greater than 0.035 for six consecutive windows, the slope confidence interval does not contain zero, and the mechanism matching degree is consistently higher than 0.75, an "upward" trend is output; when the mechanism fusion score exceeds 0.85 and remains at 4 consecutive windows, a level 3 trend risk alarm is output.

[0267] The positioning mechanism is based on the arrival time difference, installation location, amplitude attenuation, cross-correlation peak value, and propagation consistency of the four sensors. The output power supply is located near the high-voltage lead shielding structure, with a suspected area radius of 0.45m. Power-off inspection revealed a metal tip of approximately 3.2mm and traces of partial discharge at this location. The actual defect is located within the suspected area, with a positioning error of 0.31m.

[0268] After 72 hours of continuous monitoring following the maintenance, the target frequency band energy percentage decreased by 68.4% compared to before the maintenance, the number of effective pulses decreased by 81.7%, and the mechanism matching degree decreased from 0.89 to 0.23. The system output a "decreasing" trend and returned to normal monitoring status. These results demonstrate that the present invention can combine... Figures 3 to 5 The amplitude distribution changes shown enable partial discharge trend judgment, risk warning, and suspected location analysis.

[0269] A second embodiment of the present invention also provides a system for online diagnosis of partial discharge in electrical equipment, comprising:

[0270] The initialization unit is configured to acquire the ultra-high frequency waveform of at least one sensor in the high-voltage electrical equipment and form an initial dataset. The initial dataset is then preprocessed to obtain data batches for parallel processing.

[0271] The feature unit is configured to calculate the partial discharge features of each waveform window in the data batch, the partial discharge features including time-domain features, frequency-domain features, time-frequency features, phase features, pulse statistics features, and multi-channel spatial features;

[0272] The matching unit is configured to organize the partial discharge features and the pre-constructed partial discharge-specific symbol system into an input tensor, and generate a tensor computation graph based on the dependency relationship, node type and topological level of the symbol tree. The initial matching result is obtained through batch processing, merging of similar operators, reuse of common subtrees, precondition pruning and heterogeneous parallelism.

[0273] The distillation unit is configured to perform probability calibration on the initial matching result to obtain a reliable feedback label for the initial matching result, and to perform historical backtesting and symbol distillation on the initial matching result based on historical UHF data and the reliable feedback label to obtain the preferred symbol mechanism and the candidate symbol mechanism.

[0274] The output unit is configured to construct a formal online diagnostic environment based on the preferred symbol mechanism in a formally enabled state, and output the online diagnostic results.

[0275] A third embodiment of the present invention also provides an electronic device, the electronic device comprising:

[0276] At least one processor; and,

[0277] The memory is communicatively connected to the at least one processor; wherein,

[0278] The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the online diagnostic method for partial discharge of electrical equipment according to any of the foregoing embodiments.

[0279] The fourth embodiment of the present invention also provides a non-transitory computer-readable storage medium storing computer instructions for causing the computer to perform the online diagnostic method for partial discharge of electrical equipment as described in any of the foregoing embodiments.

[0280] The fifth embodiment of the present invention also provides a computer program product, which includes a computing program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions that, when executed by a computer, cause the computer to perform the online diagnostic method for partial discharge of electrical equipment according to any of the foregoing embodiments.

[0281] Figure 6 The illustration shows a method or device 1000 implementing an embodiment of the present invention. In some embodiments, more or fewer devices may be included than illustrated. In some embodiments, it may be implemented using a single or multiple devices. In some embodiments, it may be implemented using cloud-based or distributed devices.

[0282] like Figure 6As shown, device 1000 includes a processor 1001 for performing various appropriate operations and processes based on programs and / or data stored in read-only memory (ROM) 1002 or programs and / or data loaded from storage portion 1008 into random access memory (RAM) 1003. Processor 1001 may be a multi-core processor or may contain multiple processors. In some embodiments, processor 1001 may include a general-purpose main processor and one or more special coprocessors, such as a central processing unit (CPU), graphics processing unit (GPU), neural network processor (NPU), digital signal processor (DSP), etc. Various programs and data required for the operation of device 1000 are also stored in RAM 1003. Processor 1001, ROM 1002, and RAM 1003 are interconnected via bus 1004. Input / output (I / O) interface 1005 is also connected to bus 1004.

[0283] The processor and memory described above are used together to execute a program stored in the memory. When the program is executed by a computer, it can implement the methods, steps, or functions described in the above embodiments.

[0284] The following components are connected to I / O interface 1005: an input section 1006 including a keyboard, mouse, touchscreen, etc.; an output section 1007 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 1008 including a hard disk, etc.; and a communication section 1009 including a network interface card such as a LAN card, modem, etc. The communication section 1009 performs communication processing via a network such as the Internet. A drive 1010 is also connected to I / O interface 1005 as needed. A removable medium 1011, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on drive 1010 as needed so that computer programs read from it can be installed into storage section 1008 as needed. Figure 6 The diagram only shows a portion of the components and does not imply that the device 1000 only includes... Figure 6 The components shown.

[0285] The systems, devices, modules, or units described in the above embodiments can be implemented by a computer or its associated components. The computer may be, for example, a mobile terminal, smartphone, personal computer, laptop computer, in-vehicle human-machine interface device, personal digital assistant, media player, navigation device, game console, tablet computer, wearable device, smart TV, Internet of Things system, smart home, industrial computer, server, or a combination thereof.

[0286] Although not shown, in this embodiment of the invention, a computer-readable storage medium is provided having a computer program / instructions stored thereon, which, when executed by a processor, implements the method for online diagnosis of partial discharge of electrical equipment as described in the embodiment.

[0287] Storage media in embodiments of the present invention include articles that are permanent and non-permanent, removable and non-removable, capable of storing information by any method or technology. Examples of storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0288] Although not shown, embodiments of the present invention also provide a computer program product, including: a computer program / instructions that, when executed by a processor, implement the method for online diagnosis of partial discharge of electrical equipment as described in the embodiments.

[0289] The methods, programs, systems, apparatuses, etc., in embodiments of the present invention can be executed or implemented in one or more networked computers, or practiced in a distributed computing environment. In the embodiments of this specification, in these distributed computing environments, tasks can be performed by remote processing devices connected via a communication network.

[0290] Those skilled in the art will understand that the embodiments described in this specification can be provided as methods, systems, or computer program products. Therefore, those skilled in the art will realize that the functional modules / units or controllers and related method steps described in the above embodiments can be implemented in software, hardware, or a combination of both.

[0291] Unless explicitly stated otherwise, the actions or steps of the methods and procedures described in the embodiments of the present invention do not necessarily have to be performed in a specific order and can still achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0292] This document describes several embodiments of the present invention; however, for the sake of brevity, the descriptions of the embodiments are not exhaustive, and identical or similar features or parts between the embodiments may be omitted. In this document, "one embodiment," "some embodiments," "example," "specific example," or "some examples" refers to embodiments applicable to at least one, but not all, of the present invention. The above terms do not necessarily refer to the same embodiments or examples. Without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described herein, as well as the features of the different embodiments or examples.

[0293] The exemplary systems and methods of the present invention have been specifically shown and described with reference to the above embodiments, which are merely examples of the best mode for implementing the systems and methods. Those skilled in the art will understand that various changes can be made to the embodiments of the systems and methods described herein without departing from the spirit and scope of the invention as defined in the appended claims when implementing the systems and / or methods.

[0294] Furthermore, the method, apparatus, electronic device, and storage medium for online diagnosis of partial discharge in electrical equipment according to the present invention can also be implemented in the following ways:

[0295] (1) A method for online diagnosis of partial discharge in electrical equipment, characterized by comprising the following steps:

[0296] Step S1: Collect the UHF waveform of at least one sensor in the high-voltage electrical equipment and form an initial dataset. Preprocess the initial dataset to obtain data batches for parallel processing.

[0297] Step S2: Calculate the partial discharge characteristics of each waveform window in the data batch. The partial discharge characteristics include time domain characteristics, frequency domain characteristics, time-frequency characteristics, phase characteristics, pulse statistical characteristics, and multi-channel spatial characteristics.

[0298] Step S3: Organize the partial discharge features and the pre-constructed partial discharge-specific symbol system into an input tensor, and generate a tensor computation graph according to the dependency relationship, node type and topology level of the symbol tree. Through batch processing, merging of similar operators, reuse of common subtrees, precondition pruning and heterogeneous parallelism, the initial matching result is obtained.

[0299] Step S4: Perform probability calibration on the initial matching result to obtain a reliable feedback label for the initial matching result. Based on historical high frequency data and the reliable feedback label, perform historical backtesting and symbol distillation on the initial matching result to obtain the preferred symbol mechanism and candidate symbol mechanism.

[0300] Step S5: Construct a formal online diagnostic environment based on the preferred symbol mechanism that is in a formally enabled state, and output the online diagnostic results.

[0301] (2) According to the method described in (1), the initial dataset in step S1 is characterized in that... include:

[0302] ;

[0303] In the formula, It is a waveform sampling sequence; For the time of data collection; The sampling frequency; For sensor or channel identification; It is the power frequency phase; For equipment identification;

[0304] During the preprocessing, the UHF waveform is subjected to preprocessing including DC component removal, bandpass filtering, power frequency and narrowband interference suppression, background noise estimation, pulse triggering, baseline correction, amplitude normalization and / or abnormal saturation identification, and the data batch is obtained.

[0305] For multi-channel data, time alignment is performed based on a hardware clock, synchronization pulse, or correction signal, and waveform windows are divided according to a fixed length, pulse trigger, or sliding window method to obtain the data batch.

[0306] (3) According to the method described in (1), the step S2 of obtaining the partial discharge characteristics includes the following steps:

[0307] Assuming the waveform window is , for the shape window After performing a short-time Fourier transform or wavelet transform, a coefficient matrix that varies with time and frequency is obtained, and the amplitude, energy and energy ratio of each time-frequency unit are used as the time-frequency characteristics of the waveform window.

[0308] The time-domain characteristics of the waveform window include the root mean square value. and window energy , wherein the root mean square value include:

[0309] ;

[0310] The window energy include:

[0311] ;

[0312] The frequency domain characteristics of the waveform window are represented by the spectral proportion. This indicates the percentage of the spectrum within the specified frequency band [f1, f2]. include:

[0313] ;

[0314] In the formula, The spectrum of the ultra-high frequency waveform. To prevent positive numbers from being divided by zero;

[0315] The multi-channel spatial characteristics of the waveform window are represented by the arrival time difference of the multi-channel pulses. This indicates the arrival time difference of the multi-channel pulses. include:

[0316] ;

[0317] In the formula, for Arrival time of channel pulse; for Arrival time of channel pulse;

[0318] The pulse statistical features include pulse count, peak value, rise time, pulse width, pulse interval, and power frequency phase clustering, as well as the statistical values ​​of the pulse statistical features within a continuous time window;

[0319] Based on channel and channels The window energy is used to obtain the inter-channel energy attenuation ratio. :

[0320] ;

[0321] Calculate the channel within the preset time delay search range. and channels The normalized cross-correlation of the waveforms is used to obtain the cross-correlation peak value, which is then used as the correlation coefficient. The time delay corresponding to the peak cross-correlation value is used as the arrival time difference. ;

[0322] Based on the arrival time difference Align the multi-channel waveforms and calculate the correlation coefficient of any channel pair. and normalized residuals And calculate propagation consistency ,include:

[0323] ;

[0324] Among them, normalized residuals For channel The waveform and the amplitude-scaled channel The L2 norm of the difference between aligned waveforms divided by the L2 norm of channel i waveform and sum; The value range is [0, 1];

[0325] The power frequency phase, amplitude, and occurrence period of the pulse are mapped to a PRPD / PRPS spectrum, and the phase distribution and power frequency phase clustering are extracted from the PRPD / PRPS spectrum as the phase features.

[0326] (4) According to the method described in (1), the partial discharge special symbol system in step S3 includes waveform nodes, sensor nodes, time window nodes, time domain operation nodes, frequency domain operation nodes, phase nodes, statistical nodes, spatial correlation nodes, condition nodes, logic nodes, timing nodes and result nodes, and each node is predefined with node name, number of parameters, input type, output type, allowed sub-nodes, numerical range and execution function;

[0327] It also includes the initial symbol mechanism, which refers to the partial discharge diagnostic expression composed of registered nodes, parameters, dependencies and execution semantics. The partial discharge diagnostic expression includes at least feature calculation, condition judgment and result output. Each intermediate node of the partial discharge diagnostic expression stores the node value, threshold, input window and execution state.

[0328] The steps involved in generating the tensor computation graph are as follows:

[0329] Step A1: Generate candidate symbol mechanisms based on the initial symbol mechanism, including:

[0330] By selecting or searching for each configurable parameter from the allowed range using a preset parameterized rule template, one or more candidate symbol mechanisms are generated. The preset parameterized rule template is a symbolic expression skeleton that represents the verified partial discharge diagnostic rules as having a fixed node type and logical structure, and is configurable with threshold, frequency band, time window and applicable device range.

[0331] Alternatively, select a subtree with the same input and output types from historical effective symbol mechanisms, perform subtree cross-combination at type-compatible node positions, and mutate at least one of the threshold, frequency band, time window, comparison operator, or logical connection relationship to generate candidate symbol mechanisms;

[0332] Alternatively, a trained symbol mechanism generation model can encode the device type, target partial discharge type, available channels, feature vocabulary, upper limit of expression tree depth, and evaluation target as task conditions. Expressions are generated symbol by symbol using preorder traversal, postorder traversal, or parenthesized function forms. Syntax masks and type constraints are used during generation until a result node is generated and all parameters to be supplemented are closed, yielding candidate symbol mechanisms. During symbol generation, nodes matching the output type are selected from the registered symbol vocabulary based on the data type required by the input slot to be filled in the current parent node. Nodes with more parameters than the remaining number of generateable symbols, nodes whose minimum depth to complete exceeds the remaining tree depth, and nodes that violate node safety rules or create circular dependencies are removed. The remaining nodes are then used as the allowed set for the next symbol, and other nodes are masked as unselectable.

[0333] Step A2: Verify the syntax, type, and security of the candidate symbol mechanisms; normalize the expressions of the candidate symbol mechanisms that pass the verification; review the executability of the normalized expressions; and generate mechanism hash values ​​based on the normalized expressions that pass the review.

[0334] The expression normalization includes:

[0335] Calculate and replace the subtree containing only constants with the result;

[0336] Sort the operands of the commutative operation according to the hash values ​​of their child nodes;

[0337] Replace two consecutive NOT statements with the original node;

[0338] For branches where the condition is always true or always false, only the reachable branches are retained;

[0339] In addition, map synonym node names to standard names in the registry;

[0340] The verification is used to identify always true, always false, and / or ineffective branches in the candidate symbol mechanism, and to process them as objects for expression normalization. The verification process includes:

[0341] Check whether the symbols of the candidate symbol mechanism belong to the registered symbol vocabulary;

[0342] Check whether the brackets and expression tree structure of the candidate symbol mechanism are complete;

[0343] Check whether the number of node parameters for the candidate symbol mechanism is correct;

[0344] Check whether the node input type and output type of the candidate symbol mechanism match;

[0345] Check whether the candidate symbol mechanism has division by zero, empty window and / or invalid frequency band;

[0346] Check whether the candidate symbol mechanism has circular references and / or excessive expression tree depth;

[0347] Check whether the candidate symbol mechanism invokes unregistered and / or disallowed operators;

[0348] Step A3: Organize the B data batches, C sensor channels, W time windows, and F partial discharge features into the input tensor. :

[0349] ;

[0350] Where X[b, c, w, f] represents the f-th normalized feature of the b-th batch, c-th channel, and w-th window, and a valid value mask with the same shape as X is set to identify missing or abnormal data;

[0351] Step A4: Using the mechanism hash value as the retrieval key, group the candidate symbol mechanisms that have passed the verification. Retain one duplicate candidate symbol mechanism with the same mechanism hash value and consistent normalized expression. Organize the retained M candidate symbol mechanism descriptions into a mechanism parameter tensor. :

[0352] ;

[0353] Each candidate symbol mechanism contains at most N nodes and nodes in D dimensions; each mechanism parameter tensor It should include at least node type encoding, child node index, threshold or frequency band parameter, time window parameter, and valid node mask;

[0354] Step A5: Based on the input tensor and the mechanism parameter tensor The initial matching result is obtained, wherein the input tensor Provide the feature values ​​and mechanistic parameter tensors of the sample to be diagnosed. Provides node structures and parameters for performing calculations and determinations on the aforementioned feature values;

[0355] The node types and dependency indices in the mechanism parameter tensor P are read according to the topological hierarchy of the symbolic expression tree, and nodes of the same type or with the same execution semantics are compiled into batch tensor operators; the batch tensor operators are derived from the input tensor. Alternatively, the intermediate results from the previous topology layer can be used as input, and the node values ​​can be calculated by combining the node parameters in the mechanism parameter tensor P. The nodes are then connected according to the dependency relationship to form a directed acyclic tensor computation graph, and finally, the initial matching results corresponding to B batches of data and M candidate symbol mechanisms are output.

[0356] (5) According to the method described in (4), the tensor computation graph is executed on the parallel computing unit in step S3 to obtain the initial matching result, and the parallel computing unit executes:

[0357] When executing the tensor computation graph, the dimension-corresponding data batches, monitored devices, sensor channels, waveform windows, partial discharge characteristics and / or candidate symbol mechanisms are processed in batches and heterogeneous parallel processing.

[0358] By implementing precondition pruning, nodes whose computational cost is lower than a preset cost threshold and whose Boolean output is a necessary condition for the execution of subsequent branches are executed first. The preconditions include pulse count, amplitude, data quality, and / or device applicability conditions. For batches, channels, or window samples that do not meet the preconditions, as well as subsequent subtree branches of the corresponding candidate symbol mechanism, Boolean masks are set, and frequency domain, time-frequency, and / or multi-channel correlation calculations are skipped.

[0359] For nodes of the same type, operator merging is performed through merging of similar operators and reuse of common subtrees; intermediate results are reused for nodes of the same common subtree.

[0360] In addition, the initial matching result, intermediate node value, execution delay, computational resource consumption and / or abnormal status of each candidate symbol mechanism are recorded.

[0361] (6) According to the method described in (1), the characteristic is that, in step S4, the first The degree of matching between the candidate symbol mechanism described in the previous section and the current partial discharge data is used as the original mechanism matching degree. And based on the original mechanism matching degree Get the trigger determination result ;

[0362] The original mechanism matching degree include:

[0363] ;

[0364] The trigger determination result include:

[0365] ;

[0366] in, Indicates the first The degree of matching between the candidate symbol mechanism described in the article and the current partial discharge data; This represents the minimum matching degree required for the candidate symbol mechanism to produce a positive determination, and serves as the trigger threshold.

[0367] When similar equipment and similar discharge tasks have no less than a preset number of labeled and calibrated samples, and both positive and negative samples reach a preset proportion, and pass the sample distribution consistency check, a logistic regression calibration, or order-preserving regression, or temperature scaling method is used to fit the probability mapping function. , obtained the Diagnostic probability of candidate symbol mechanism :

[0368] ;

[0369] Diagnostic probability is calculated using resampling methods or calibrated residual quantiles. When the confidence interval does not meet the calibration conditions, the output mechanism matching degree is... Risk Score and the "Uncalibrated" label;

[0370] Based on mechanism matching degree Diagnostic probability Data quality coefficient Severity coefficient Calculate risk score ,include:

[0371] ;

[0372] in, These are the non-negative weights corresponding to mechanism matching degree, diagnostic probability, data quality, and severity, respectively. This serves as a probability calibration status indicator; it indicates when the diagnostic probability has been calibrated. ,otherwise ; The risk score The value range is 0 to 100;

[0373] When the ratio between the preferred symbol mechanism and the candidate symbol mechanism does not meet the preferred threshold, some parameters of the candidate symbol mechanism are adjusted, historical backtesting and symbol distillation are performed again, and the preferred symbol mechanism and the candidate symbol mechanism are updated. The target mechanism library is then updated based on the updated preferred symbol mechanism.

[0374] When adjusting some parameters of the candidate symbol mechanism, including: feature threshold, feature frequency band, time window length, node parameters, logical connection relationship, mechanism fusion weight, applicable device range and / or enabled status.

[0375] (7) According to the method described in (1), the preferred symbol mechanism in the formally enabled state is compiled and loaded into the formal online diagnostic environment; the candidate symbol mechanism, the verification mechanism and the gray-scale operation mechanism are executed in a bypass environment isolated from the formal alarm, and only evaluation records with verification identifiers are generated.

[0376] The target mechanism library includes: mechanism identifier, symbolic expression tree, tensor computation graph, applicable device type, applicable sensor type, target partial discharge type, feature threshold and time window, data version, historical backtesting index, expression complexity, mechanism creation source, mechanism version number and / or mechanism activation status;

[0377] The target mechanism library has at least candidate mechanisms, mechanisms under verification, gray-scale operation mechanisms, and states of official activation, deactivation, and rollback. When a gray-scale operation mechanism meets the conditions and is reviewed, it is switched to the official activation state.

[0378] When performing online partial discharge diagnosis based on the aforementioned formal online diagnostic environment, the following steps are included:

[0379] The mechanism of parallel execution of multiple optimal symbols on the target UHF data; for the first The optimal symbol mechanism is used to verify the comprehensive performance of the set. Equipment compatibility Input data quality and target data coverage The product is normalized to obtain the fusion weight. :

[0380] ;

[0381] Based on the fusion weight Calculate the fusion score for each partial discharge type k. :

[0382] ;

[0383] In the formula, For the first The optimal symbol mechanism determines the diagnostic probability or normalized matching degree of type k;

[0384] Preset weight threshold, select fusion weight The high-weighted selection mechanism, which prioritizes symbols above a preset weight threshold, is used in the fusion scoring. The maximum value is selected as the highest fusion score. When the partial discharge type determinations of multiple high-weight preferred symbol mechanisms are consistent, and the partial discharge type is the same as the partial discharge type corresponding to the highest fusion score, and the highest fusion score reaches the alarm threshold, the alarm conclusion of the partial discharge type is output. Otherwise, the status of pending review is output, and the conflict mechanism and evidence are displayed. The evidence includes mechanism identifier, input channel and time window, key feature value, judgment threshold corresponding to key feature, mechanism trigger threshold and alarm threshold, original mechanism matching degree, calibration status, diagnostic probability output when the probability calibration condition is met, and complete judgment path.

[0385] The online diagnostic results are generated as follows: the type corresponding to the highest fusion score is taken as the partial discharge type; the partial discharge status is determined by the highest fusion score and the alarm threshold; based on the risk score... The severity of partial discharge is determined by the interval; the suspected discharge location or region is determined based on the arrival time difference of multiple channels, installation location, energy attenuation ratio between channels, and propagation consistency; the mechanism and characteristics that contribute to the fusion score to the preset contribution threshold are respectively determined as the trigger symbol mechanism and key characteristics, and the node judgment path is output.

[0386] The online diagnostic results also include the discharge development trend: based on the number of pulses, window energy, characteristic frequency band energy and phase concentration of K consecutive time windows, and the matching degree of the original mechanism output by the preferred symbol mechanism in the corresponding window when it is in the officially activated state, linear regression or robust regression is performed with the window corresponding time as the independent variable, the regression slope is used as the growth rate, the difference between adjacent regression slopes divided by the corresponding time interval is used as the growth acceleration, and the growth rate, growth acceleration and the number of continuous windows in the same direction are used as trend features, where K is not less than 3;

[0387] Provided that the data quality and sample coverage reach the corresponding thresholds, if at least two preferred symbol mechanisms in the officially enabled state support the same trend, or if a high-weight preferred symbol mechanism continuously supports the trend, then the discharge development trend is determined according to the characteristics of the trend: when the confidence interval of the growth rate falls into the preset zero value neighborhood and the number of continuous windows reaches the stability determination threshold, a stable discharge development trend is output; when the stability determination condition is not met, and at least L windows in K windows change in the same direction and the confidence interval of the corresponding growth rate does not contain zero, the discharge development trend of rising, rapidly rising, or falling is output according to the positive or negative sign of the growth rate and the comparison result of the growth rate, growth acceleration, and preset rapid rise threshold; when the corresponding determination condition is not met, an uncertain discharge development trend is output, where 1≤L≤K;

[0388] Using a preferred symbol mechanism that includes time-series nodes, trend characteristics are extrapolated according to the current growth rate and growth acceleration within a preset prediction time H after the current moment. The extrapolated mechanism fusion score, normalized growth rate, growth acceleration and duration are weighted and mapped to 0 to 100 to obtain the trend risk score. [t, t+H] is used as the prediction time window, and the moment when the trend risk score first reaches the preset risk threshold is taken as the expected trigger time. The uncertainty range of the expected trigger time is given based on the regression residual or resampling results.

[0389] (8) A system for online diagnosis of partial discharge in electrical equipment, characterized in that, based on any one of (1)-(7), it comprises:

[0390] The initialization unit is configured to acquire the ultra-high frequency waveform of at least one sensor in the high-voltage electrical equipment and form an initial dataset. The initial dataset is then preprocessed to obtain data batches for parallel processing.

[0391] The feature unit is configured to calculate the partial discharge features of each waveform window in the data batch, the partial discharge features including time-domain features, frequency-domain features, time-frequency features, phase features, pulse statistics features, and multi-channel spatial features;

[0392] The matching unit is configured to organize the partial discharge features and the pre-constructed partial discharge-specific symbol system into an input tensor, and generate a tensor computation graph based on the dependency relationship, node type and topological level of the symbol tree. The initial matching result is obtained through batch processing, merging of similar operators, reuse of common subtrees, precondition pruning and heterogeneous parallelism.

[0393] The distillation unit is configured to perform probability calibration on the initial matching result to obtain a reliable feedback label for the initial matching result, and to perform historical backtesting and symbol distillation on the initial matching result based on historical UHF data and the reliable feedback label to obtain the preferred symbol mechanism and the candidate symbol mechanism.

[0394] The output unit is configured to construct a formal online diagnostic environment based on the preferred symbol mechanism in a formally enabled state, and output the online diagnostic results.

[0395] (9) The system according to (8) is characterized in that the initialization unit is further configured as the initial dataset in step S1. include:

[0396] ;

[0397] In the formula, It is a waveform sampling sequence; For the time of data collection; The sampling frequency; For sensor or channel identification; It is the power frequency phase; For equipment identification;

[0398] During the preprocessing, the UHF waveform is subjected to preprocessing including DC component removal, bandpass filtering, power frequency and narrowband interference suppression, background noise estimation, pulse triggering, baseline correction, amplitude normalization and / or abnormal saturation identification, and the data batch is obtained.

[0399] For multi-channel data, time alignment is performed based on a hardware clock, synchronization pulse, or correction signal, and waveform windows are divided according to a fixed length, pulse trigger, or sliding window method to obtain the data batch.

[0400] (10) The system according to (8) is characterized in that, when the feature unit is further configured to obtain the partial discharge feature in step S2, the following steps are included:

[0401] Assuming the waveform window is , for the shape window After performing a short-time Fourier transform or wavelet transform, a coefficient matrix that varies with time and frequency is obtained, and the amplitude, energy and energy ratio of each time-frequency unit are used as the time-frequency characteristics of the waveform window.

[0402] The time-domain characteristics of the waveform window include the root mean square value. and window energy , wherein the root mean square value include:

[0403] ;

[0404] The window energy include:

[0405] ;

[0406] The frequency domain characteristics of the waveform window are represented by the spectral proportion. This indicates the percentage of the spectrum within the specified frequency band [f1, f2]. include:

[0407] ;

[0408] In the formula, The spectrum of the ultra-high frequency waveform. To prevent positive numbers from being divided by zero;

[0409] The multi-channel spatial characteristics of the waveform window are represented by the arrival time difference of the multi-channel pulses. This indicates the arrival time difference of the multi-channel pulses. include:

[0410] ;

[0411] In the formula, for Arrival time of channel pulse; for Arrival time of channel pulse;

[0412] The pulse statistical features include pulse count, peak value, rise time, pulse width, pulse interval, and power frequency phase clustering, as well as the statistical values ​​of the pulse statistical features within a continuous time window;

[0413] Based on channel and channels The window energy is used to obtain the inter-channel energy attenuation ratio. :

[0414] ;

[0415] Calculate the channel within the preset time delay search range. and channels The normalized cross-correlation of the waveforms is used to obtain the cross-correlation peak value, which is then used as the correlation coefficient. The time delay corresponding to the peak cross-correlation value is used as the arrival time difference. ;

[0416] Based on the arrival time difference Align the multi-channel waveforms and calculate the correlation coefficient of any channel pair. and normalized residuals And calculate propagation consistency ,include:

[0417] ;

[0418] Among them, normalized residuals For channel The waveform and the amplitude-scaled channel The L2 norm of the difference between aligned waveforms divided by the L2 norm of channel i waveform and sum; The value range is [0, 1];

[0419] The power frequency phase, amplitude, and occurrence period of the pulse are mapped to a PRPD / PRPS spectrum, and the phase distribution and power frequency phase clustering are extracted from the PRPD / PRPS spectrum as the phase features.

[0420] (11) According to the system described in (8), the matching unit is further configured such that the partial discharge special symbol system in step S3 includes waveform nodes, sensor nodes, time window nodes, time domain operation nodes, frequency domain operation nodes, phase nodes, statistical nodes, spatial correlation nodes, condition nodes, logic nodes, timing nodes and result nodes, and each node is predefined with node name, number of parameters, input type, output type, allowed sub-nodes, numerical range and execution function;

[0421] It also includes the initial symbol mechanism, which refers to the partial discharge diagnostic expression composed of registered nodes, parameters, dependencies and execution semantics. The partial discharge diagnostic expression includes at least feature calculation, condition judgment and result output. Each intermediate node of the partial discharge diagnostic expression stores the node value, threshold, input window and execution state.

[0422] The steps involved in generating the tensor computation graph are as follows:

[0423] Step A1: Generate candidate symbol mechanisms based on the initial symbol mechanism, including:

[0424] By selecting or searching for each configurable parameter from the allowed range using a preset parameterized rule template, one or more candidate symbol mechanisms are generated. The preset parameterized rule template is a symbolic expression skeleton that represents the verified partial discharge diagnostic rules as having a fixed node type and logical structure, and is configurable with threshold, frequency band, time window and applicable device range.

[0425] Alternatively, select a subtree with the same input and output types from historical effective symbol mechanisms, perform subtree cross-combination at type-compatible node positions, and mutate at least one of the threshold, frequency band, time window, comparison operator, or logical connection relationship to generate candidate symbol mechanisms;

[0426] Alternatively, a trained symbol mechanism generation model can encode the device type, target partial discharge type, available channels, feature vocabulary, upper limit of expression tree depth, and evaluation target as task conditions. Expressions are generated symbol by symbol using preorder traversal, postorder traversal, or parenthesized function forms. Syntax masks and type constraints are used during generation until a result node is generated and all parameters to be supplemented are closed, yielding candidate symbol mechanisms. During symbol generation, nodes matching the output type are selected from the registered symbol vocabulary based on the data type required by the input slot to be filled in the current parent node. Nodes with more parameters than the remaining number of generateable symbols, nodes whose minimum depth to complete exceeds the remaining tree depth, and nodes that violate node safety rules or create circular dependencies are removed. The remaining nodes are then used as the allowed set for the next symbol, and other nodes are masked as unselectable.

[0427] Step A2: Verify the syntax, type, and security of the candidate symbol mechanisms; normalize the expressions of the candidate symbol mechanisms that pass the verification; review the executability of the normalized expressions; and generate mechanism hash values ​​based on the normalized expressions that pass the review.

[0428] The expression normalization includes:

[0429] Calculate and replace the subtree containing only constants with the result;

[0430] Sort the operands of the commutative operation according to the hash values ​​of their child nodes;

[0431] Replace two consecutive NOT statements with the original node;

[0432] For branches where the condition is always true or always false, only the reachable branches are retained;

[0433] In addition, map synonym node names to standard names in the registry;

[0434] The verification is used to identify always true, always false, and / or ineffective branches in the candidate symbol mechanism, and to process them as objects for expression normalization. The verification process includes:

[0435] Check whether the symbols of the candidate symbol mechanism belong to the registered symbol vocabulary;

[0436] Check whether the brackets and expression tree structure of the candidate symbol mechanism are complete;

[0437] Check whether the number of node parameters for the candidate symbol mechanism is correct;

[0438] Check whether the node input type and output type of the candidate symbol mechanism match;

[0439] Check whether the candidate symbol mechanism has division by zero, empty window and / or invalid frequency band;

[0440] Check whether the candidate symbol mechanism has circular references and / or excessive expression tree depth;

[0441] Check whether the candidate symbol mechanism invokes unregistered and / or disallowed operators;

[0442] Step A3: Organize the B data batches, C sensor channels, W time windows, and F partial discharge features into the input tensor. :

[0443] ;

[0444] Where X[b, c, w, f] represents the f-th normalized feature of the b-th batch, c-th channel, and w-th window, and a valid value mask with the same shape as X is set to identify missing or abnormal data;

[0445] Step A4: Using the mechanism hash value as the retrieval key, group the candidate symbol mechanisms that have passed the verification. Retain one duplicate candidate symbol mechanism with the same mechanism hash value and consistent normalized expression. Organize the retained M candidate symbol mechanism descriptions into a mechanism parameter tensor. :

[0446] ;

[0447] Each candidate symbol mechanism contains at most N nodes and nodes in D dimensions; each mechanism parameter tensor It should include at least node type encoding, child node index, threshold or frequency band parameter, time window parameter, and valid node mask;

[0448] Step A5: Based on the input tensor and the mechanism parameter tensor The initial matching result is obtained, wherein the input tensor Provide the feature values ​​and mechanistic parameter tensors of the sample to be diagnosed. Provides node structures and parameters for performing calculations and determinations on the aforementioned feature values;

[0449] The node types and dependency indices in the mechanism parameter tensor P are read according to the topological hierarchy of the symbolic expression tree, and nodes of the same type or with the same execution semantics are compiled into batch tensor operators; the batch tensor operators are derived from the input tensor. Alternatively, the intermediate results from the previous topology layer can be used as input, and the node values ​​can be calculated by combining the node parameters in the mechanism parameter tensor P. The nodes are then connected according to the dependency relationship to form a directed acyclic tensor computation graph, and finally, the initial matching results corresponding to B batches of data and M candidate symbol mechanisms are output.

[0450] (12) According to the system described in (11), the matching unit is further configured to execute the tensor computation graph on the parallel computing unit in step S3 to obtain the initial matching result, wherein the parallel computing unit executes:

[0451] When executing the tensor computation graph, the dimension-corresponding data batches, monitored devices, sensor channels, waveform windows, partial discharge characteristics and / or candidate symbol mechanisms are processed in batches and heterogeneous parallel processing.

[0452] By implementing precondition pruning, nodes whose computational cost is lower than a preset cost threshold and whose Boolean output is a necessary condition for the execution of subsequent branches are executed first. The preconditions include pulse count, amplitude, data quality, and / or device applicability conditions. For batches, channels, or window samples that do not meet the preconditions, as well as subsequent subtree branches of the corresponding candidate symbol mechanism, Boolean masks are set, and frequency domain, time-frequency, and / or multi-channel correlation calculations are skipped.

[0453] For nodes of the same type, operator merging is performed through merging of similar operators and reuse of common subtrees; intermediate results are reused for nodes of the same common subtree.

[0454] In addition, the initial matching result, intermediate node value, execution delay, computational resource consumption and / or abnormal status of each candidate symbol mechanism are recorded.

[0455] (13) According to the system of (8), the distillation unit is further configured to, in step S4, distill the first The degree of matching between the candidate symbol mechanism described in the previous section and the current partial discharge data is used as the original mechanism matching degree. And based on the original mechanism matching degree Get the trigger determination result ;

[0456] The original mechanism matching degree include:

[0457] ;

[0458] The trigger determination result include:

[0459] ;

[0460] in, Indicates the first The degree of matching between the candidate symbol mechanism described in the article and the current partial discharge data; This represents the minimum matching degree required for the candidate symbol mechanism to produce a positive determination, and serves as the trigger threshold.

[0461] When similar equipment and similar discharge tasks have no less than a preset number of labeled and calibrated samples, and both positive and negative samples reach a preset proportion, and pass the sample distribution consistency check, a logistic regression calibration, or order-preserving regression, or temperature scaling method is used to fit the probability mapping function. , obtained the Diagnostic probability of candidate symbol mechanism :

[0462] ;

[0463] Diagnostic probability is calculated using resampling methods or calibrated residual quantiles. When the confidence interval does not meet the calibration conditions, the output mechanism matching degree is... Risk Score and the "Uncalibrated" label;

[0464] Based on mechanism matching degree Diagnostic probability Data quality coefficient Severity coefficient Calculate risk score ,include:

[0465] ;

[0466] in, These are the non-negative weights corresponding to mechanism matching degree, diagnostic probability, data quality, and severity, respectively. This serves as a probability calibration status indicator; it indicates when the diagnostic probability has been calibrated. ,otherwise ; The risk score The value range is 0 to 100;

[0467] When the ratio between the preferred symbol mechanism and the candidate symbol mechanism does not meet the preferred threshold, some parameters of the candidate symbol mechanism are adjusted, historical backtesting and symbol distillation are performed again, and the preferred symbol mechanism and the candidate symbol mechanism are updated. The target mechanism library is then updated based on the updated preferred symbol mechanism.

[0468] When adjusting some parameters of the candidate symbol mechanism, including: feature threshold, feature frequency band, time window length, node parameters, logical connection relationship, mechanism fusion weight, applicable device range and / or enabled status.

[0469] (14) According to the system described in (8), the output unit is further configured to compile and load the preferred symbol mechanism in the formally enabled state into the formal online diagnostic environment in step S5; the candidate symbol mechanism, the verification mechanism and the gray-scale operation mechanism are executed in a bypass environment isolated from the formal alarm, and only an evaluation record with a verification identifier is generated.

[0470] The target mechanism library includes: mechanism identifier, symbolic expression tree, tensor computation graph, applicable device type, applicable sensor type, target partial discharge type, feature threshold and time window, data version, historical backtesting index, expression complexity, mechanism creation source, mechanism version number and / or mechanism activation status;

[0471] The target mechanism library has at least candidate mechanisms, mechanisms under verification, gray-scale operation mechanisms, and states of official activation, deactivation, and rollback. When a gray-scale operation mechanism meets the conditions and is reviewed, it is switched to the official activation state.

[0472] When performing online partial discharge diagnosis based on the aforementioned formal online diagnostic environment, the following steps are included:

[0473] The mechanism of parallel execution of multiple optimal symbols on the target UHF data; for the first The optimal symbol mechanism is used to verify the comprehensive performance of the set. Equipment compatibility Input data quality and target data coverage The product is normalized to obtain the fusion weight. :

[0474] ;

[0475] Based on the fusion weight Calculate the fusion score for each partial discharge type k. :

[0476] ;

[0477] In the formula, For the first The optimal symbol mechanism determines the diagnostic probability or normalized matching degree of type k;

[0478] Preset weight threshold, select fusion weight The high-weighted selection mechanism, which prioritizes symbols above a preset weight threshold, is used in the fusion scoring. The maximum value is selected as the highest fusion score. When the partial discharge type determinations of multiple high-weight preferred symbol mechanisms are consistent, and the partial discharge type is the same as the partial discharge type corresponding to the highest fusion score, and the highest fusion score reaches the alarm threshold, the alarm conclusion of the partial discharge type is output. Otherwise, the status of pending review is output, and the conflict mechanism and evidence are displayed. The evidence includes mechanism identifier, input channel and time window, key feature value, judgment threshold corresponding to key feature, mechanism trigger threshold and alarm threshold, original mechanism matching degree, calibration status, diagnostic probability output when the probability calibration condition is met, and complete judgment path.

[0479] The online diagnostic results are generated as follows: the type corresponding to the highest fusion score is taken as the partial discharge type; the partial discharge status is determined by the highest fusion score and the alarm threshold; based on the risk score... The severity of partial discharge is determined by the interval; the suspected discharge location or region is determined based on the arrival time difference of multiple channels, installation location, energy attenuation ratio between channels, and propagation consistency; the mechanism and characteristics that contribute to the fusion score to the preset contribution threshold are respectively determined as the trigger symbol mechanism and key characteristics, and the node judgment path is output.

[0480] The online diagnostic results also include the discharge development trend: based on the number of pulses, window energy, characteristic frequency band energy and phase concentration of K consecutive time windows, and the matching degree of the original mechanism output by the preferred symbol mechanism in the corresponding window when it is in the officially activated state, linear regression or robust regression is performed with the window corresponding time as the independent variable, the regression slope is used as the growth rate, the difference between adjacent regression slopes divided by the corresponding time interval is used as the growth acceleration, and the growth rate, growth acceleration and the number of continuous windows in the same direction are used as trend features, where K is not less than 3;

[0481] Provided that the data quality and sample coverage reach the corresponding thresholds, if at least two preferred symbol mechanisms in the officially enabled state support the same trend, or if a high-weight preferred symbol mechanism continuously supports the trend, then the discharge development trend is determined according to the characteristics of the trend: when the confidence interval of the growth rate falls into the preset zero value neighborhood and the number of continuous windows reaches the stability determination threshold, a stable discharge development trend is output; when the stability determination condition is not met, and at least L windows in K windows change in the same direction and the confidence interval of the corresponding growth rate does not contain zero, the discharge development trend of rising, rapidly rising, or falling is output according to the positive or negative sign of the growth rate and the comparison result of the growth rate, growth acceleration, and preset rapid rise threshold; when the corresponding determination condition is not met, an uncertain discharge development trend is output, where 1≤L≤K;

[0482] Using a preferred symbol mechanism that includes time-series nodes, trend characteristics are extrapolated according to the current growth rate and growth acceleration within a preset prediction time H after the current moment. The extrapolated mechanism fusion score, normalized growth rate, growth acceleration and duration are weighted and mapped to 0 to 100 to obtain the trend risk score. [t, t+H] is used as the prediction time window, and the moment when the trend risk score first reaches the preset risk threshold is taken as the expected trigger time. The uncertainty range of the expected trigger time is given based on the regression residual or resampling results.

[0483] (15) An electronic device, characterized in that the electronic device comprises:

[0484] At least one processor; and,

[0485] The memory is communicatively connected to the at least one processor; wherein,

[0486] The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method for online diagnosis of partial discharge of electrical equipment as described in any one of (1) to (7).

[0487] (16) A non-transitory computer-readable storage medium, characterized in that the non-transitory computer-readable storage medium stores computer instructions for causing the computer to perform the method for online diagnosis of partial discharge of electrical equipment as described in any one of (1) to (7).

Claims

1. A method for online diagnosis of partial discharge in electrical equipment, characterized in that, Includes the following steps: Step S1: Collect the UHF waveform of at least one sensor in the high-voltage electrical equipment and form an initial dataset. Preprocess the initial dataset to obtain data batches for parallel processing. Step S2: Calculate the partial discharge characteristics of each waveform window in the data batch. The partial discharge characteristics include time domain characteristics, frequency domain characteristics, time-frequency characteristics, phase characteristics, pulse statistical characteristics, and multi-channel spatial characteristics. Step S3: Organize the partial discharge features and the pre-constructed partial discharge-specific symbol system into an input tensor, and generate a tensor computation graph according to the dependency relationship, node type and topology level of the symbol tree. Through batch processing, merging of similar operators, reuse of common subtrees, precondition pruning and heterogeneous parallelism, the initial matching result is obtained. Step S4: Perform probability calibration on the initial matching result to obtain a reliable feedback label for the initial matching result. Based on historical high frequency data and the reliable feedback label, perform historical backtesting and symbol distillation on the initial matching result to obtain the preferred symbol mechanism and candidate symbol mechanism. Step S5: Construct a formal online diagnostic environment based on the preferred symbol mechanism that is in a formally enabled state, and output the online diagnostic results.

2. The method for online diagnosis of partial discharge in electrical equipment according to claim 1, characterized in that, The initial dataset in step S1 include: ； In the formula, It is a waveform sampling sequence; For the time of data collection; The sampling frequency; For sensor or channel identification; It is the power frequency phase; For equipment identification; During the preprocessing, the UHF waveform is subjected to preprocessing including DC component removal, bandpass filtering, power frequency and narrowband interference suppression, background noise estimation, pulse triggering, baseline correction, amplitude normalization and / or abnormal saturation identification, and the data batch is obtained. For multi-channel data, time alignment is performed based on a hardware clock, synchronization pulse, or correction signal, and waveform windows are divided according to a fixed length, pulse trigger, or sliding window method to obtain the data batch.

3. The method for online diagnosis of partial discharge in electrical equipment according to claim 1, characterized in that, Obtaining the partial discharge characteristics in step S2 includes the following steps: Assuming the waveform window is , for the shape window After performing a short-time Fourier transform or wavelet transform, a coefficient matrix that varies with time and frequency is obtained, and the amplitude, energy and energy ratio of each time-frequency unit are used as the time-frequency characteristics of the waveform window. The time-domain characteristics of the waveform window include the root mean square value. and window energy , wherein the root mean square value include: ； The window energy include: ； The frequency domain characteristics of the waveform window are represented by the spectral proportion. This indicates the percentage of the spectrum within the specified frequency band [f1, f2]. include: ； In the formula, For the spectrum of ultra-high frequency waveforms, To prevent positive numbers from being divided by zero; The multi-channel spatial characteristics of the waveform window are represented by the arrival time difference of the multi-channel pulses. This indicates the arrival time difference of the multi-channel pulses. include: ； In the formula, for Arrival time of channel pulse; for Arrival time of channel pulse; The pulse statistical features include pulse count, peak value, rise time, pulse width, pulse interval, and power frequency phase clustering, as well as the statistical values ​​of the pulse statistical features within a continuous time window; Based on channel and channels The window energy is used to obtain the inter-channel energy attenuation ratio. : ； Calculate the channel within the preset time delay search range. and channels The normalized cross-correlation of the waveforms yields the cross-correlation peak value, which is then used as the correlation coefficient. The time delay corresponding to the peak cross-correlation value is used as the arrival time difference. ; Based on the arrival time difference Align the multi-channel waveforms and calculate the correlation coefficient of any channel pair. and normalized residuals And calculate propagation consistency ,include: ； Among them, normalized residuals For channel The waveform and the amplitude-scaled channel The L2 norm of the difference between aligned waveforms divided by the L2 norm of channel i waveform and sum; The value range is [0, 1]; The power frequency phase, amplitude, and occurrence period of the pulse are mapped to a PRPD / PRPS spectrum, and the phase distribution and power frequency phase clustering are extracted from the PRPD / PRPS spectrum as the phase features.

4. The method for online diagnosis of partial discharge in electrical equipment according to claim 1, characterized in that, The partial discharge-specific symbol system in step S3 includes waveform nodes, sensor nodes, time window nodes, time domain operation nodes, frequency domain operation nodes, phase nodes, statistical nodes, spatial correlation nodes, condition nodes, logic nodes, timing nodes, and result nodes. Each type of node is predefined with node name, number of parameters, input type, output type, allowed sub-nodes, numerical range, and execution function. It also includes the initial symbol mechanism, which refers to the partial discharge diagnostic expression composed of registered nodes, parameters, dependencies and execution semantics. The partial discharge diagnostic expression includes at least feature calculation, condition judgment and result output. Each intermediate node of the partial discharge diagnostic expression stores the node value, threshold, input window and execution state. The steps involved in generating the tensor computation graph are as follows: Step A1: Generate candidate symbol mechanisms based on the initial symbol mechanism, including: By selecting or searching for each configurable parameter from the allowed range using a preset parameterized rule template, one or more candidate symbol mechanisms are generated. The preset parameterized rule template is a symbolic expression skeleton that represents the verified partial discharge diagnostic rules as having a fixed node type and logical structure, and is configurable with threshold, frequency band, time window and applicable device range. Alternatively, select a subtree with the same input and output types from historical effective symbol mechanisms, perform subtree cross-combination at type-compatible node positions, and mutate at least one of the threshold, frequency band, time window, comparison operator, or logical connection relationship to generate candidate symbol mechanisms; Alternatively, a trained symbol mechanism generation model can encode the device type, target partial discharge type, available channels, feature vocabulary, upper limit of expression tree depth, and evaluation target as task conditions. Expressions are generated symbol by symbol using preorder traversal, postorder traversal, or parenthesized function forms. Syntax masks and type constraints are used during generation until a result node is generated and all parameters to be supplemented are closed, yielding candidate symbol mechanisms. During symbol generation, nodes matching the output type are selected from the registered symbol vocabulary based on the data type required by the input slot to be filled in the current parent node. Nodes with more parameters than the remaining number of generateable symbols, nodes whose minimum depth to complete exceeds the remaining tree depth, and nodes that violate node safety rules or create circular dependencies are removed. The remaining nodes are then used as the allowed set for the next symbol, and other nodes are masked as unselectable. Step A2: Verify the syntax, type, and security of the candidate symbol mechanisms; normalize the expressions of the candidate symbol mechanisms that pass the verification; review the executability of the normalized expressions; and generate mechanism hash values ​​based on the normalized expressions that pass the review. The expression normalization includes: Calculate and replace the subtree containing only constants with the result; Sort the operands of the commutative operation according to the hash values ​​of their child nodes; Replace two consecutive logical NOTs with the original node; For branches where the condition is always true or always false, only the reachable branches are retained; In addition, map synonym node names to standard names in the registry; The verification is used to identify always true, always false, and / or ineffective branches in the candidate symbol mechanism, and to process them as objects for expression normalization. The verification process includes: Check whether the symbols of the candidate symbol mechanism belong to the registered symbol vocabulary; Check whether the brackets and expression tree structure of the candidate symbol mechanism are complete; Check whether the number of node parameters for the candidate symbol mechanism is correct; Check whether the node input type and output type of the candidate symbol mechanism match; Check whether the candidate symbol mechanism has division by zero, empty window and / or invalid frequency band; Check whether the candidate symbol mechanism has circular references and / or excessive expression tree depth; Check whether the candidate symbol mechanism invokes unregistered and / or disallowed operators; Step A3: Organize the B data batches, C sensor channels, W time windows, and F partial discharge features into the input tensor. : ； Where X[b, c, w, f] represents the f-th normalized feature of the b-th batch, c-th channel, and w-th window, and a valid value mask with the same shape as X is set to identify missing or abnormal data; Step A4: Using the mechanism hash value as the retrieval key, group the candidate symbol mechanisms that have passed the verification. Retain one duplicate candidate symbol mechanism with the same mechanism hash value and consistent normalized expression. Organize the retained M candidate symbol mechanism descriptions into a mechanism parameter tensor. : ； Each candidate symbol mechanism contains at most N nodes and nodes in D dimensions; each mechanism parameter tensor It should include at least node type encoding, child node index, threshold or frequency band parameter, time window parameter, and valid node mask; Step A5: Based on the input tensor and the mechanism parameter tensor The initial matching result is obtained, wherein the input tensor Provide the feature values ​​and mechanistic parameter tensors of the sample to be diagnosed. Provides node structures and parameters for performing calculations and determinations on the aforementioned feature values; The node types and dependency indices in the mechanism parameter tensor P are read according to the topological hierarchy of the symbolic expression tree, and nodes of the same type or with the same execution semantics are compiled into batch tensor operators; the batch tensor operators are derived from the input tensor. Alternatively, the intermediate results from the previous topology layer can be used as input, and the node values ​​can be calculated by combining the node parameters in the mechanism parameter tensor P. The nodes are then connected according to the dependency relationship to form a directed acyclic tensor computation graph, and finally, the initial matching results corresponding to B batches of data and M candidate symbol mechanisms are output.

5. The method for online diagnosis of partial discharge in electrical equipment according to claim 4, characterized in that, In step S3, when the tensor computation graph is executed on the parallel computing unit to obtain the initial matching result, the parallel computing unit executes: When executing the tensor computation graph, the dimension-corresponding data batches, monitored devices, sensor channels, waveform windows, partial discharge characteristics and / or candidate symbol mechanisms are processed in batches and heterogeneous parallel processing. By implementing precondition pruning, nodes whose computational cost is lower than a preset cost threshold and whose Boolean output is a necessary condition for the execution of subsequent branches are executed first. The preconditions include pulse count, amplitude, data quality, and / or device applicability conditions. For batches, channels, or window samples that do not meet the preconditions, as well as subsequent subtree branches of the corresponding candidate symbol mechanism, Boolean masks are set, and frequency domain, time-frequency, and / or multi-channel correlation calculations are skipped. For nodes of the same type, operator merging is performed through merging of similar operators and reuse of common subtrees; intermediate results are reused for nodes of the same common subtree. In addition, the initial matching result, intermediate node value, execution delay, computational resource consumption and / or abnormal status of each candidate symbol mechanism are recorded.

6. The method for online diagnosis of partial discharge in electrical equipment according to claim 1, characterized in that, In step S4, the first The degree of matching between the candidate symbol mechanism described in the previous section and the current partial discharge data is used as the original mechanism matching degree. And based on the original mechanism matching degree Get the trigger determination result ; The original mechanism matching degree include: ； The trigger determination result include: ； in, Indicates the first The degree of matching between the candidate symbol mechanism described in the article and the current partial discharge data; This represents the minimum matching degree required for the candidate symbol mechanism to produce a positive determination, and serves as the trigger threshold. When similar equipment and similar discharge tasks have no less than a preset number of labeled and calibrated samples, and both positive and negative samples reach a preset proportion, and pass the sample distribution consistency check, a logistic regression calibration, or order-preserving regression, or temperature scaling method is used to fit the probability mapping function. , obtained the Diagnostic probability of candidate symbol mechanism : ； Diagnostic probability is calculated using resampling methods or calibrated residual quantiles. When the confidence interval does not meet the calibration conditions, the output mechanism matching degree is... Risk Score and the "Uncalibrated" label; Based on mechanism matching degree Diagnostic probability Data quality coefficient Severity coefficient Calculate risk score ,include: ； in, These are the non-negative weights corresponding to mechanism matching degree, diagnostic probability, data quality, and severity, respectively. This serves as a probability calibration status indicator; it indicates when the diagnostic probability has been calibrated. ,otherwise ; The risk score The value range is 0 to 100; When the ratio between the preferred symbol mechanism and the candidate symbol mechanism does not meet the preferred threshold, some parameters of the candidate symbol mechanism are adjusted, historical backtesting and symbol distillation are performed again, and the preferred symbol mechanism and the candidate symbol mechanism are updated. The target mechanism library is then updated based on the updated preferred symbol mechanism. When adjusting some parameters of the candidate symbol mechanism, including: feature threshold, feature frequency band, time window length, node parameters, logical connection relationship, mechanism fusion weight, applicable device range and / or enabled status.

7. The method for online diagnosis of partial discharge in electrical equipment according to claim 1, characterized in that, In step S5, the preferred symbol mechanism that is in the officially enabled state is compiled and loaded into the official online diagnostic environment; the candidate symbol mechanism, the verification mechanism and the gray-scale operation mechanism are executed in a bypass environment isolated from the official alarm, and only evaluation records with verification identifiers are generated. The target mechanism library includes: mechanism identifier, symbolic expression tree, tensor computation graph, applicable device type, applicable sensor type, target partial discharge type, feature threshold and time window, data version, historical backtesting index, expression complexity, mechanism creation source, mechanism version number and / or mechanism activation status; The target mechanism library has at least candidate mechanisms, mechanisms under verification, gray-scale operation mechanisms, and states of official activation, deactivation, and rollback. When a gray-scale operation mechanism meets the conditions and is reviewed, it is switched to the official activation state. When performing online partial discharge diagnosis based on the aforementioned formal online diagnostic environment, the following steps are included: The mechanism of parallel execution of multiple optimal symbols on the target UHF data; for the first The optimal symbol mechanism is used to verify the comprehensive performance of the set. Equipment compatibility Input data quality and target data coverage The product is normalized to obtain the fusion weight. : ； Based on the fusion weight Calculate the fusion score for each partial discharge type k. : ； In the formula, For the first The optimal symbol mechanism determines the diagnostic probability or normalized matching degree of type k; Preset weight threshold, select fusion weight The high-weighted selection mechanism, which prioritizes symbols above a preset weight threshold, is used in the fusion scoring. The maximum value is selected as the highest fusion score. When the partial discharge type determinations of multiple high-weight preferred symbol mechanisms are consistent, and the partial discharge type is the same as the partial discharge type corresponding to the highest fusion score, and the highest fusion score reaches the alarm threshold, the alarm conclusion of the partial discharge type is output. Otherwise, the status of pending review is output, and the conflict mechanism and evidence are displayed. The evidence includes mechanism identifier, input channel and time window, key feature value, judgment threshold corresponding to key feature, mechanism trigger threshold and alarm threshold, original mechanism matching degree, calibration status, diagnostic probability output when the probability calibration condition is met, and complete judgment path. The online diagnostic results are generated as follows: the type corresponding to the highest fusion score is taken as the partial discharge type; the partial discharge status is determined by the highest fusion score and the alarm threshold; based on the risk score... The severity of partial discharge is determined by the interval; the suspected discharge location or region is determined based on the arrival time difference of multiple channels, installation location, energy attenuation ratio between channels, and propagation consistency; the mechanism and characteristics that contribute to the fusion score to the preset contribution threshold are respectively determined as the trigger symbol mechanism and key characteristics, and the node judgment path is output. The online diagnostic results also include the discharge development trend: based on the number of pulses, window energy, characteristic frequency band energy and phase concentration of K consecutive time windows, and the matching degree of the original mechanism output by the preferred symbol mechanism in the corresponding window when it is in the officially activated state, linear regression or robust regression is performed with the window corresponding time as the independent variable, the regression slope is used as the growth rate, the difference between adjacent regression slopes divided by the corresponding time interval is used as the growth acceleration, and the growth rate, growth acceleration and the number of continuous windows in the same direction are used as trend features, where K is not less than 3; Provided that the data quality and sample coverage reach the corresponding thresholds, if at least two preferred symbol mechanisms in the officially enabled state support the same trend, or if a high-weight preferred symbol mechanism continuously supports the trend, then the discharge development trend is determined according to the characteristics of the trend: when the confidence interval of the growth rate falls into the preset zero value neighborhood and the number of continuous windows reaches the stability determination threshold, a stable discharge development trend is output; when the stability determination condition is not met, and at least L windows in K windows change in the same direction and the confidence interval of the corresponding growth rate does not contain zero, the discharge development trend of rising, rapidly rising, or falling is output according to the positive or negative sign of the growth rate and the comparison result of the growth rate, growth acceleration, and preset rapid rise threshold; when the corresponding determination condition is not met, an uncertain discharge development trend is output, where 1≤L≤K; Using a preferred symbol mechanism that includes time-series nodes, trend characteristics are extrapolated according to the current growth rate and growth acceleration within a preset prediction time H after the current moment. The extrapolated mechanism fusion score, normalized growth rate, growth acceleration and duration are weighted and mapped to 0 to 100 to obtain the trend risk score. [t, t+H] is used as the prediction time window, and the moment when the trend risk score first reaches the preset risk threshold is taken as the expected trigger time. The uncertainty range of the expected trigger time is given based on the regression residual or resampling results.

8. A system for online diagnosis of partial discharge in electrical equipment, characterized in that, The method for online diagnosis of partial discharge in electrical equipment according to any one of claims 1-7 includes: The initialization unit is configured to acquire the ultra-high frequency waveform of at least one sensor in the high-voltage electrical equipment and form an initial dataset. The initial dataset is then preprocessed to obtain data batches for parallel processing. The feature unit is configured to calculate the partial discharge features of each waveform window in the data batch, the partial discharge features including time-domain features, frequency-domain features, time-frequency features, phase features, pulse statistics features, and multi-channel spatial features; The matching unit is configured to organize the partial discharge features and the pre-constructed partial discharge-specific symbol system into an input tensor, and generate a tensor computation graph based on the dependency relationship, node type and topological level of the symbol tree. The initial matching result is obtained through batch processing, merging of similar operators, reuse of common subtrees, precondition pruning and heterogeneous parallelism. The distillation unit is configured to perform probability calibration on the initial matching result to obtain a reliable feedback label for the initial matching result, and to perform historical backtesting and symbol distillation on the initial matching result based on historical UHF data and the reliable feedback label to obtain the preferred symbol mechanism and the candidate symbol mechanism. The output unit is configured to construct a formal online diagnostic environment based on the preferred symbol mechanism in a formally enabled state, and output the online diagnostic results.

9. An electronic device, characterized in that, The electronic device includes: At least one processor; and, The memory is communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method for online diagnosis of partial discharge of electrical equipment as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores computer instructions for causing the computer to perform the method for online diagnosis of partial discharge of electrical equipment as described in any one of claims 1 to 7.

Images