Partial discharge diagnosis method, device, equipment and storage medium

Abstract

The application relates to a partial discharge diagnosis method and device, computer equipment, a computer readable storage medium and a computer program product. The method comprises the following steps: acquiring an acoustic-electric signal to be analyzed and task constraint information for providing constraint information, performing feature coding processing on the acoustic-electric signal to be analyzed to obtain an acoustic-electric coupling feature, performing feature transformation on the acoustic-electric coupling feature and aligning the acoustic-electric coupling feature to a preset state parameter domain to obtain a state analysis feature, then performing acoustic-electric joint diagnosis reasoning based on the task constraint information and the state analysis feature to obtain a partial discharge reasoning feature, and performing semantic generation based on the partial discharge reasoning feature to obtain a diagnosis result. The method can solve the acoustic-electric cross-modal feature space dislocation problem, significantly improve the precision and robustness of partial discharge diagnosis under complex working conditions, can synchronously output accurate partial discharge state description and fault labels, and significantly improves the accuracy and intelligent analysis depth of partial discharge diagnosis.

Description

Technical Field

[0001] This application relates to the field of power equipment management, and in particular to a partial discharge diagnosis method, apparatus, computer equipment, computer-readable storage medium, and computer program product. Background Technology

[0002] Partial discharge is an early manifestation of insulation defects in power equipment. Accurate diagnosis of the partial discharge state of equipment is of great significance for timely detection of potential insulation problems and ensuring the safe and stable operation of power equipment. Acoustic-electrical combined diagnostics, due to its integration of the complementary characteristics of acoustic and electrical signals, has become an important technical direction in the field of partial discharge diagnosis. However, existing acoustic-electrical combined partial discharge diagnostic technologies still have many shortcomings in practical applications and are difficult to meet the needs of accurate diagnosis under complex operating conditions.

[0003] Existing technologies, when processing combined acoustic and electrical signals, can only achieve simple splicing of underlying physical features or classification operations based on domain adaptation. They are unable to understand the correlation between different modal features when describing the same partial discharge defect, which leads to problems such as feature space misalignment and vague description of specific fault phenomena in partial discharge diagnostic models. Summary of the Invention

[0004] Based on this, it is necessary to provide an improved acoustic-electrical combined partial discharge diagnosis method, device, computer equipment, computer-readable storage medium, and computer program product to address the above-mentioned technical problems. By introducing a pre-trained state parameter domain and task constraint information guidance, heterogeneous acoustic and electrical features can be uniformly aligned to a preset state parameter semantic space, solving the problems of modal feature space misalignment and ambiguous description in traditional methods.

[0005] In a first aspect, this application provides a method for diagnosing partial discharge, comprising:

[0006] Acquire the acoustic and electrical signals to be analyzed and the task constraint information;

[0007] The acoustic-electric signal to be analyzed is subjected to feature encoding processing to obtain acoustic-electric coupling features;

[0008] The acoustic-electric coupling features are aligned to a preset state parameter domain to obtain state analysis features;

[0009] Based on the task constraint information and the state parsing features, acoustic-electric joint diagnostic reasoning is performed to obtain partial discharge reasoning features;

[0010] Semantic generation is performed based on the partial discharge inference features to obtain the diagnostic results of the acoustic-electric signal to be analyzed.

[0011] In one embodiment, aligning the acoustic-electric coupling features to a preset state parameter domain to obtain state analysis features includes:

[0012] The acoustic-electric coupling features are subjected to time-frequency band-level feature processing, and the acoustic-electric coupling features are transformed to the time-frequency state parameter domain to obtain time-frequency embedding features;

[0013] The time-frequency embedding features are subjected to time-frequency correlation extraction to obtain the state parsing features;

[0014] The step of performing acoustic-electric joint diagnostic reasoning based on the task constraint information and the state analysis features to obtain partial discharge reasoning features includes:

[0015] Based on the task constraint information, semantic association is performed with the state parsing features to obtain task-oriented partial inference features.

[0016] In one embodiment, the partial discharge diagnosis method is executed based on an acoustic-electric diagnostic network, which is composed of a pre-feature network and a post-inference network cascaded together. The pre-feature network is used to extract features and spatially align acoustic-electric signals, and output state analysis features. The post-inference network is used to perform inference based on task constraint information and state analysis features, and output partial discharge inference features.

[0017] The training steps for the pre-feature network and the post-inference network include:

[0018] Acquire the acoustic-electric coupling sample set from the basic operating condition library, the acoustic-electric coupling sample set from the target operating condition library, and the diagnostic description semantic set;

[0019] Based on the acoustic-electric coupling sample set of the basic working condition library and the acoustic-electric coupling sample set of the target working condition library, the first stage of training is carried out to align the acoustic-electric coupling features and state semantic features. While keeping the network topology weights of the post-inference network unchanged, the network topology weights of the pre-feature network are updated until the first model convergence condition is met.

[0020] Based on the acoustic-electric coupling sample set of the target working condition library and the diagnostic description semantic set, the network that meets the convergence condition of the first model is subjected to the second stage of training for partial discharge feature extraction of acoustic and electrical signals and recognition of acoustic and electrical signals.

[0021] In the second stage of training, the network topology weights of the preceding feature network are kept unchanged, and the network topology weights of the following inference network are updated until the second model convergence condition is met.

[0022] In one embodiment, the acoustic-electric coupling sample set of the basic operating condition library includes a first sample acoustic-electric signal, a first diagnostic guidance constraint, and a first state parameter reference value corresponding to the first sample acoustic-electric signal;

[0023] The target operating condition library's acoustic-electric coupling sample set includes a second sample acoustic-electric signal, a second diagnostic guidance constraint, and a second state parameter reference value corresponding to the second sample acoustic-electric signal.

[0024] The first phase of training includes:

[0025] Feature encoding and feature transformation are performed on the first sample acoustic-electric signal and the second sample acoustic-electric signal, respectively, to obtain the first state analytical feature and the second state analytical feature.

[0026] First sample training data is constructed based on the first state analysis features and the first diagnostic guidance constraints, and second sample training data is constructed based on the second state analysis features and the second diagnostic guidance constraints. Then, acoustic-electric joint diagnostic reasoning is performed based on the first sample training data and the second sample training data to obtain the first sample partial discharge state description and the second sample partial discharge state description, respectively.

[0027] The first diagnostic error is determined based on the parameter offset between the first sample partial discharge state description and the first state parameter reference value, and the parameter offset between the second sample partial discharge state description and the second state parameter reference value. The network topology weights of the preceding feature network are then updated based on the first diagnostic error.

[0028] In one embodiment, the acoustic-electric coupling sample set of the target operating condition library further includes the fault mode identification code of the second sample acoustic-electric signal, and the diagnostic description semantic set includes the sample diagnostic operation paradigm and the actual state parameter benchmark value.

[0029] The second phase of training includes:

[0030] The second sample acoustic-electric signal is subjected to feature encoding and feature transformation to obtain the third state analytical features;

[0031] Based on the third state analytical features and the sample diagnostic operation paradigm, a joint acoustic-electrical diagnostic reasoning is performed to obtain the third sample reasoning features, the third sample partial discharge state description, and the actual partial discharge state description.

[0032] Based on the third sample reasoning features, the label determination is performed to obtain the sample acoustic and electrical signal category determination result;

[0033] The second diagnostic error is determined based on the third sample partial discharge state description, the second state parameter reference value, the actual partial discharge state description, the actual state parameter reference value, the sample acoustic signal category determination result, and the fault mode identification code. The network topology weights of the post-inference network are then updated based on the second diagnostic error.

[0034] In one embodiment, both the first diagnostic guidance constraint and the second diagnostic guidance constraint include a discharge event calibration paradigm, a feature parameter extraction paradigm, and a fault mode determination paradigm.

[0035] The discharge event calibration paradigm is used for directional constraint to locate acoustic and electrical signal discharge events, and the state parameter reference value corresponding to the discharge event calibration paradigm has a spatial position reference value.

[0036] The feature parameter extraction paradigm is used for directional constraint to extract global feature parameters of acoustic and electrical signals, and the state parameter reference value corresponding to the feature parameter extraction paradigm has a feature parameter amplitude space.

[0037] The fault mode determination paradigm is used for fault mode determination of partial discharge mode of acoustic and electrical signals under directional constraints.

[0038] Secondly, this application also provides a partial discharge diagnostic device, comprising:

[0039] The information acquisition module is used to acquire the acoustic and electrical signals to be analyzed and the task constraint information.

[0040] The feature encoding module is used to perform feature encoding processing on the acoustic-electric signal to be analyzed to obtain acoustic-electric coupling features;

[0041] The feature alignment module is used to align the acoustic-electric coupling features to a preset state parameter domain to obtain state analysis features.

[0042] The diagnostic reasoning module is used to perform joint acoustic and electronic diagnostic reasoning based on the task constraint information and the state parsing features to obtain partial discharge reasoning features.

[0043] The result output module is used to generate semantics based on the partial discharge inference features to obtain the diagnostic results of the acoustic-electric signal to be analyzed.

[0044] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0045] Acquire the acoustic and electrical signals to be analyzed and the task constraint information;

[0046] The acoustic-electric signal to be analyzed is subjected to feature encoding processing to obtain acoustic-electric coupling features;

[0047] The acoustic-electric coupling features are aligned to a preset state parameter domain to obtain state analysis features;

[0048] Based on the task constraint information and the state parsing features, acoustic-electric joint diagnostic reasoning is performed to obtain partial discharge reasoning features;

[0049] Semantic generation is performed based on the partial discharge inference features to obtain the diagnostic results of the acoustic-electric signal to be analyzed.

[0050] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:

[0051] Acquire the acoustic and electrical signals to be analyzed and the task constraint information;

[0052] The acoustic-electric signal to be analyzed is subjected to feature encoding processing to obtain acoustic-electric coupling features;

[0053] The acoustic-electric coupling features are aligned to a preset state parameter domain to obtain state analysis features;

[0054] Based on the task constraint information and the state parsing features, acoustic-electric joint diagnostic reasoning is performed to obtain partial discharge reasoning features;

[0055] Semantic generation is performed based on the partial discharge inference features to obtain the diagnostic results of the acoustic-electric signal to be analyzed.

[0056] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, performs the following steps:

[0057] Acquire the acoustic and electrical signals to be analyzed and the task constraint information;

[0058] The acoustic-electric signal to be analyzed is subjected to feature encoding processing to obtain acoustic-electric coupling features;

[0059] The acoustic-electric coupling features are aligned to a preset state parameter domain to obtain state analysis features;

[0060] Based on the task constraint information and the state parsing features, acoustic-electric joint diagnostic reasoning is performed to obtain partial discharge reasoning features;

[0061] Semantic generation is performed based on the partial discharge inference features to obtain the diagnostic results of the acoustic-electric signal to be analyzed.

[0062] The aforementioned partial discharge diagnostic methods, devices, computer equipment, computer-readable storage media, and computer program products, by introducing task constraint information as constraints, uniformly align heterogeneous acoustic and electrical features to a preset state parameter semantic space, solving the problem of modal feature space misalignment in traditional methods, and realizing deep decoupling and fusion of acoustic and electrical signals at the physical state level.

[0063] Secondly, by using task text information to guide feature extraction, the traditional black-box classification model has been changed, enabling the diagnostic results to not only include simple category labels, but also generate corresponding status descriptions, thereby improving the guiding value of diagnostic conclusions in actual operation and maintenance.

[0064] Furthermore, since the alignment process is based on pre-trained physical state parameters, the model can filter out single-mode noise interference that is unrelated to the partial discharge state, and can still maintain a high diagnostic accuracy in field environments with low signal-to-noise ratios. Attached Figure Description

[0065] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0066] Figure 1 This is a diagram illustrating the application environment of a partial discharge diagnosis method in one embodiment;

[0067] Figure 2 This is a flowchart illustrating a partial discharge diagnosis method in one embodiment;

[0068] Figure 3 This is a flowchart of the two-stage training of a partial discharge diagnostic model in one embodiment;

[0069] Figure 4 This is a structural block diagram of a partial discharge diagnostic device in one embodiment;

[0070] Figure 5 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0071] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0072] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the embodiments, or any combination of multiple embodiments.

[0073] The partial discharge diagnosis method provided in this embodiment can be applied to, for example... Figure 1 The system environment shown is an intelligent system for power equipment condition monitoring and fault diagnosis. It is adaptable to on-site operation scenarios of power facilities such as substations, converter stations, and power line hubs. It can perform real-time or offline partial discharge diagnosis on various high-voltage power equipment such as gas-insulated switchboards (GIS), transformers, high-voltage switchgear, and cable terminals. It achieves integrated operation of on-site acquisition of acoustic and electrical signals, data transmission, and cloud-based intelligent diagnosis.

[0074] Terminal device 110 is deployed at the power equipment field monitoring point and serves as the signal acquisition and front-end transmission node for the diagnostic system. It mainly includes acoustic and electrical signal acquisition sensors (ultrasonic sensors, high-frequency current sensors, ultra-high frequency sensors, etc.), edge acquisition terminals, industrial-grade embedded controllers, and field communication modules. Among them, the acoustic and electrical signal acquisition sensors are directly attached to the measured part of the power equipment to complete the synchronous real-time acquisition of partial discharge acoustic signals and electrical signals. The edge acquisition terminals perform basic preprocessing such as power frequency filtering, wavelet denoising, and timestamp alignment on the raw signals. The industrial-grade embedded controller is responsible for format conversion and data encapsulation of the preprocessed acoustic and electrical signals. The field communication module (industrial Ethernet, 5G / 4G wireless communication, LoRa narrowband IoT, etc.) uploads the encapsulated acoustic and electrical signals to be analyzed and the field-configured task constraint information (such as fault location, parameter extraction, mode determination, etc.) to the server 120 according to the industrial communication protocol. At the same time, it can receive the diagnostic results issued by the server 120 and display them on the field display terminal. Terminal device 110 has industrial-grade characteristics such as strong electromagnetic interference resistance, wide temperature range operation, and dust and water resistance. It is suitable for complex physical and electromagnetic environments of substations, both outdoors and indoors. It supports synchronous acquisition of multi-channel signals, and the acquisition parameters can be flexibly configured according to on-site diagnostic needs.

[0075] The server-side 120 is deployed in the power company's back-end monitoring center or cloud data center. It is the core intelligent diagnostic node of the diagnostic system and mainly includes a data receiving server, a diagnostic computing server, a database server, and an application management server. These servers work together to receive, store, infer diagnostic results, and distribute results. The data receiving server receives the acoustic and electrical signals to be analyzed and the task constraint information uploaded by the terminal device 110 through a dedicated communication link, and completes data verification and classification storage. The database server pre-stores training sample data such as the basic operating condition database, the target operating condition database, and the diagnostic description semantic set, as well as basic parameters of power equipment and historical diagnostic records, providing data support for diagnostic reasoning. The diagnostic computing server is the core unit, with a built-in trained acoustic and electrical joint partial discharge diagnostic model. It sequentially performs acoustic and electrical signal feature encoding, feature transformation, joint diagnostic reasoning, and label determination according to this method to generate partial discharge status descriptions and acoustic and electrical signal labels. The application management server performs structured processing of the diagnostic results and can push them to multiple terminals such as the power operation and maintenance platform, mobile operation and maintenance terminals, and field display devices. It also supports historical query, statistical analysis, and fault warning of diagnostic results. The server 120 features high computing power, high reliability, and high scalability. It adopts a multi-core processor and a distributed computing architecture, which can simultaneously process the acoustic and electrical signal diagnostic tasks uploaded by multiple field terminal devices 110. It supports long-term storage and rapid retrieval of massive monitoring data, and is suitable for the partial discharge diagnostic needs of multiple devices and a wide range in the power system.

[0076] After the field terminal device 110 completes the acquisition and preprocessing of acoustic and electrical signals, it uploads the acoustic and electrical signals to be analyzed and the task constraint information to the backend server 120 in real time or at regular intervals through an encrypted communication link. After receiving the data, the server 120 automatically calls the diagnostic model to execute the partial discharge diagnosis process, completes feature processing, diagnostic reasoning and label determination within a preset time, and generates a complete partial discharge diagnosis result. The diagnosis result is sent from the server 120 to the field terminal device 110 and the power operation and maintenance related platform. At the same time, the server 120 can send instructions to the terminal device 110 to adjust the signal acquisition parameters and change the diagnostic task according to the field diagnosis needs, so as to realize the remote control of the diagnosis process.

[0077] In one exemplary embodiment, such as Figure 2 As shown, a partial discharge diagnosis method is provided, which can be applied to... Figure 1 Taking server-side 120 as an example, the explanation includes the following steps 210 to 250. Wherein:

[0078] Step 210: Obtain the acoustic-electric signal to be analyzed and the task constraint information.

[0079] The acoustic-electric signals to be analyzed are multimodal sensing data reflecting the physical phenomenon of partial discharge occurring inside or on the surface of the insulation medium of power equipment. In actual industrial environments, transient electrical signals and acoustic emission signals caused by insulation defects can be simultaneously acquired by pre-deploying high-frequency current transformers, ultra-high-frequency antennas, and ultrasonic sensors on the monitored power equipment, such as gas-insulated switchgear, high-voltage switchgear, or power transformers. During the acquisition of the acoustic-electric signals to be analyzed, a pre-processing data cleaning procedure can be configured. This involves performing strict time-series synchronization matching and adaptive filtering and noise reduction on the raw acoustic-electric data acquired by the sensors based on timestamps. This removes background white noise and periodic narrowband interference, adapting to the complex electromagnetic interference and mechanical vibration environment of substations, and ensuring that the final acoustic-electric signals input to the subsequent analysis stage have a high signal-to-noise ratio and high fidelity.

[0080] Task constraint information contains guiding data that instructs the system to perform specific-level analysis objectives and provides prerequisite logical constraints. In complex smart grid operation and maintenance scenarios, the partial discharge diagnosis needs faced by maintenance personnel are often multi-dimensional and dynamically changing. For example, under certain operating conditions, the focus may be on accurately identifying the insulation defect types of partial discharge, such as corona discharge, floating discharge, or surface discharge, while under other operating conditions, the focus may be on assessing the severity of discharge energy or calibrating the physical spatial location of the discharge event. Therefore, the process of obtaining task constraint information can be achieved by parsing instructions configured and issued by maintenance personnel through a human-machine interface terminal, or by automatically generating the current operating condition inspection strategy by connecting to the substation's higher-level equipment asset management system. This task constraint information embeds the aforementioned targeted diagnostic constraint information, aiming to inform the subsequent acoustic-electric joint diagnostic model which specific state parameter in the feature space should be focused on, thereby effectively preventing the model from getting lost in massive redundant features and providing solid target guidance for achieving a high-level state semantic description with logical reasoning.

[0081] For example, in a partial discharge diagnosis scenario for GIS equipment in a 110kV substation, the server first collects raw data through a sensor system deployed on the GIS equipment. At key monitoring points on the A-phase busbar of the GIS equipment, ultrasonic sensors and high-frequency current transformers (HFCTs) are installed. The ultrasonic sensors collect the mechanical vibration acoustic signals generated by partial discharge at a sampling frequency of 500kHz; the signals are time-domain waveform data, containing continuous vibration amplitudes over a 10-second duration. The HFCTs collect the pulse current signals generated by partial discharge with a bandwidth of 1MHz to 20MHz; the signals are discrete pulse sequences, containing pulse amplitude, phase, and time interval information. The sensors transmit real-time data to the server via an industrial Ethernet network. The server preprocesses the signals, including power frequency filtering and wavelet denoising to eliminate environmental interference, pulse shaping to remove glitches from the electrical signals, and aligning the timestamps of the acoustic and electrical signals to microsecond-level accuracy based on second pulse signals. The final result is the acoustic and electrical signal to be analyzed, containing time-synchronized acoustic waveform data and electrical pulse data.

[0082] Meanwhile, the server receives task constraint information configured by maintenance personnel through the human-machine interface of the substation monitoring system. This information clearly defines the diagnostic objectives and constraints, such as performing fault mode determination to identify metal tip discharge, floating potential discharge, or free particle discharge; outputting partial discharge charge quantity, phase distribution characteristics, and fault location coordinates; and presenting the results in the form of natural language descriptions and fault code labels. The server parses the task information into structured instructions to guide the diagnostic logic of subsequent models.

[0083] Step 220: Perform feature encoding processing on the acoustic-electric signal to be analyzed to obtain acoustic-electric coupling features.

[0084] Specifically, since the acoustic and electrical signals to be analyzed usually contain acoustic modal data and electrical modal data with significant differences in sampling rate and physical dimensions, the specific process of performing feature encoding processing on the acoustic and electrical signals to be analyzed preferably relies on a pre-configured feature encoding network, which can avoid feature space conflicts and dimensionality curse caused by direct splicing of heterogeneous data.

[0085] For example, for acoustic signals, the time-domain waveform is first converted into a time-frequency graph using a short-time Fourier transform to capture the frequency distribution (e.g., the characteristic frequency band of 100-150kHz) and energy changes of the partial discharge sound wave. Then, a multi-layer convolutional neural network is used to perform deep encoding on the time-frequency graph to extract local features such as the amplitude fluctuation pattern and the main frequency bandwidth of the sound wave. For electrical signals, a recurrent neural network is used to perform time-series modeling of the pulse sequence, learning the amplitude distribution (e.g., maximum pulse amplitude of 150mV), phase distribution (e.g., 42% of the phase interval is between 0 and 90°), and statistical characteristics of the pulse interval (e.g., average interval of 2ms). Subsequently, an attention mechanism is used to dynamically weight and fuse the acoustic and electrical features, where the weights are adaptively adjusted according to the signal-to-noise ratio, such as a weight of 0.4 for acoustic signals and 0.6 for electrical signals, to obtain "acoustic-electric coupling features" containing the spatiotemporal correlation information of acoustic and electrical signals, such as the time difference between acoustic and electrical pulses (the signal arrival time difference caused by the propagation delay of the sound wave) and the amplitude ratio (the linear correlation of the energy of acoustic and electrical signals), etc.

[0086] After multi-layer nonlinear spatial mapping and feature aggregation calculation, the acoustic-electric coupling features output by the network are no longer a simple splicing combination of the underlying physical representation data, but a high-dimensional abstract feature representation that deeply integrates multimodal implicit physical associations. This provides high-quality and non-redundant underlying data support for subsequent cross-space feature alignment and high-level semantic state reasoning.

[0087] Step 230: Align the acoustic-electric coupling features to the preset state parameter domain to obtain the state analysis features.

[0088] Among them, the state parameter domain is a high-dimensional potential space with specific physical guidance and semantic logic. Each feature manifold dimension within it implicitly corresponds to professional state parameters that can characterize the degree of insulation degradation of power equipment, such as discharge phase distribution, discharge amplitude variation rate, and discharge energy decay trend.

[0089] Specifically, by using a specific multi-scale pooling layer or time-frequency transform network structure, the acoustic-electric coupling features originally based on temporal and spatial dimensions are reconstructed and projected onto the time-frequency state parameter domain, thereby generating time-frequency embedded features. Deep time-frequency correlation extraction is then performed on these embedded features, aligning them to the state parameter domain. Environmental redundancy fluctuations and apparent interference in the original acoustic-electric signal are removed, resulting in structured and semantically pure state analysis features.

[0090] For example, firstly, the acoustic-electric coupling characteristics are processed at the time-frequency band level through a fully connected network, mapping the main frequency band characteristics of the acoustic signal to the "center frequency" (e.g., 125kHz) and "frequency band energy ratio" (e.g., 68%), and mapping the pulse characteristics of the electrical signal to the "maximum pulse amplitude" (150mV) and "amplitude standard deviation" (20mV); at the same time, the time domain characteristics are statistically quantized to obtain the "signal duration" (85% of effective pulses within 10s) and "pulse density" (average 120 pulses per second).

[0091] Subsequently, the correlation between the above time-frequency parameters is learned through graph attention network, such as the positive correlation between "center frequency" and "maximum pulse amplitude" (the higher the main frequency band energy, the larger the pulse amplitude) and the negative correlation between "amplitude standard deviation" and "phase distribution dispersion" (the higher the amplitude stability, the more concentrated the phase distribution).

[0092] Finally, the output state analysis features correspond to the key state parameters required for the diagnostic task, such as the partial discharge charge, phase centroid angle, and fault location coordinates—all directly interpretable physical quantities. This overcomes the representational misalignment and logical gap between the underlying physical feature space and the final partial discharge state semantics caused by directly utilizing the underlying multimodal physical features for black-box classification in existing deep learning diagnostic models.

[0093] Step 240: Perform acoustic-electric joint diagnostic reasoning based on task constraint information and state analysis features to obtain partial discharge reasoning features.

[0094] Among them, the partial discharge inference feature is a high-dimensional feature vector containing task-oriented semantics, with each dimension corresponding to the semantic concepts required for diagnostic inference. This feature simultaneously carries semantic information for generating a partial discharge state description and classification information for determining fault labels, realizing the interpretable feature expression required to convert objective physical quantities into diagnostic conclusions.

[0095] Specifically, after receiving highly structured state parsing features, the system uses the previously acquired task constraint information as a priori condition query vector. By calculating the correlation weights between the condition query vector and each dimension of the state parsing features, the system can adaptively and specifically aggregate discharge parameter dimensions that are highly relevant to the current diagnostic task. After multi-level deep logical reasoning, the system outputs partial discharge reasoning features stripped of irrelevant task interference.

[0096] For example, a pre-tuned diagnostic reasoning network (based on a Transformer architecture) is used to perform deep reasoning on state parsing features in conjunction with task constraint information. First, the state parsing features are semantically associated with constraint parameters in the diagnostic task (such as "charge amount" and "phase distribution"). The encoder learns the mapping rules between state parameters and fault modes. For example, when the charge amount is 650pC, the phase distribution is bipolar symmetrical (the proportions of the 0~90° and 180~270° intervals are 42% and 45%, respectively), and the fault location coordinates are 30cm from the left flange, the network identifies that this feature combination corresponds to the typical mode of "floating potential discharge".

[0097] Step 250: Semantic generation is performed based on partial discharge inference features to obtain the diagnostic results of the acoustic-electric signal to be analyzed.

[0098] The diagnostic results include a description of the partial discharge status and labels for acoustic and electrical signals.

[0099] A partial discharge (PD) state description can be a complete logical expression that can clearly illustrate the phase distribution characteristics, active evolution trend, and corresponding insulation degradation risk level of partial discharge. The PD inference features are input as high-dimensional contextual semantic latent variables to the decoding end. The decoding end will generate structured diagnostic conclusions node by node according to the normative context of the power grid equipment condition assessment guidelines, and output the PD state description.

[0100] For example, the decoder converts the inference results into a partial discharge state description in natural language, which includes quantitative parameters and qualitative conclusions. For instance, "Partial discharge signal was detected in the A-phase busbar of the GIS equipment, with a partial discharge charge of 650pC, a bipolar symmetrical phase distribution, a fault location 30cm from the left flange, and high discharge stability." The output partial discharge state description is no longer a single, isolated fault probability value, but a complete logical expression that can explain in detail the active evolution trend of the phase distribution characteristics of the partial discharge and the corresponding insulation degradation risk level. This provides maintenance personnel with extremely intuitive and in-depth decision support for formulating targeted power outage and troubleshooting strategies.

[0101] Acoustic-electric signal tags are structured data used to identify the partial discharge fault category to which the acoustic-electric signal to be analyzed belongs, including specific fault modes such as metal tip discharge, floating potential discharge, or free particle discharge, and their corresponding fault codes. Compared with the natural language form of partial discharge state description, acoustic-electric signal tags are presented in a standardized coding format, which facilitates automatic recording, retrieval, and statistical analysis by the system.

[0102] The partial discharge inference features output from the pre-processing steps already contain the aggregated and dimensionality-reduced physical attribute representations of the partial discharge fault. Therefore, the label determination process can be implemented using a multilayer perceptron and a normalized exponential function layer. The high-dimensional partial discharge inference features are mapped to a low-dimensional log-probability space corresponding to the preset fault category through fully connected operations. Subsequently, probability normalization calculations are performed to output the predicted probability distribution of each candidate fault mode. The category with the highest predicted probability value is then determined as the acoustic signal label for the acoustic signal to be analyzed.

[0103] For example, fault mode classification is performed based on the partial discharge inference features output by the diagnostic inference unit to obtain acoustic-electric signal labels. The inference features are then subjected to dimensionality reduction and nonlinear transformation using a multilayer perceptron to calculate the probability distributions of three typical fault types (metal tip discharge, floating potential discharge, and free particle discharge). For instance, the calculated probability is 85% for floating potential discharge, 12% for metal tip discharge, and 3% for free particle discharge. The category with the highest probability is selected as the acoustic-electric signal label, namely "floating potential discharge," and a corresponding fault code (e.g., "FM-02") is output.

[0104] In the above-mentioned partial discharge diagnosis method, by acquiring and introducing task constraint information as a constraint condition, the shortcomings of traditional partial discharge diagnosis models in feature extraction and classification that lack specific task orientation are effectively overcome. This enables the entire model's diagnostic reasoning process to be strictly guided by the targeted guidance of task constraint information, thereby significantly improving the accuracy and task relevance of the model in multimodal partial discharge diagnosis when dealing with variable and complex power grid conditions.

[0105] Secondly, this embodiment innovatively performs feature transformation on the acoustic-electric coupling features obtained after feature encoding processing, aligning them across spatial depth to the preset state parameter domain. This process effectively eliminates the spatial representation misalignment between the underlying multimodal physical acoustic-electric signals and the upper-level partial discharge state semantics, greatly enhancing the ability of state analysis features to characterize the essence of physical discharge phenomena and the semantic correlation, enabling the model to accurately capture the deep implicit correlations within the acoustic-electric signals.

[0106] Finally, this embodiment performs joint acoustic and electrical diagnostic reasoning based on task constraint information and state analysis features, which completely breaks through the black box limitation of most existing technologies that can only output a single isolated fault probability or classification label. While achieving accurate acoustic and electrical signal label determination, the generated diagnostic results include a partial discharge state description with high logical reasoning and interpretability, thereby providing power grid maintenance personnel with a more intuitive and detailed semantic reference for partial discharge state, significantly improving the analytical depth and human-computer interaction efficiency of intelligent operation and maintenance of power equipment.

[0107] Taking the partial discharge diagnosis of the A-phase busbar of the GIS equipment in a 110kV substation as an example, the server first executes step 210 to obtain the acoustic and electrical signals to be analyzed and the task constraint information.

[0108] Then, step 220 is executed to perform feature encoding processing on the acoustic-electric signal to be analyzed, obtaining the acoustic-electric coupling characteristics. The preprocessed acoustic-electric signal to be analyzed (including 10 seconds of time-synchronized acoustic waveform data and 1200 electrical pulse data) is converted into a time-frequency diagram through short-time Fourier transform, capturing the energy concentration characteristics of the partial discharge sound wave in the 100~150kHz frequency band (main frequency bandwidth 50kHz, frequency band energy ratio 68%), and extracting the amplitude fluctuation law of the sound wave (maximum amplitude 3.2V, amplitude standard deviation 0.5V); for the electrical signal, the pulse sequence is time-series modeled through a recurrent neural network to obtain the pulse amplitude distribution (maximum pulse amplitude 150mV, average amplitude 80mV), phase distribution (0~90° phase interval ratio 42%, 180~270° ratio 45%), and pulse interval statistical characteristics (average interval 2ms, interval standard deviation 0.5ms). The acoustic and electrical features are dynamically weighted and fused using an attention mechanism (with acoustic signal weight of 0.4 and electrical signal weight of 0.6 based on the on-site signal-to-noise ratio), resulting in acoustic-electric coupling features containing spatiotemporal correlation information, such as the time difference (20 μs) between acoustic and electrical pulses due to differences in propagation paths, and the linear correlation coefficient (0.72) of acoustic and electrical signal energy, among other key fusion features.

[0109] In one embodiment, step 230 is performed to transform the acoustic-electric coupling features, aligning the acoustic-electric coupling features to a preset state parameter domain to obtain state analysis features, including steps 232 to 234:

[0110] Step 232: Perform time-frequency band-level feature processing on the acoustic-electric coupling features, transform the acoustic-electric coupling features to the time-frequency state parameter domain, and obtain the time-frequency embedding features.

[0111] Specifically, the core time-frequency feature dimension is extracted from the acoustic-electric coupling characteristics; the extracted time-frequency features are mapped one by one to the corresponding dimension of the preset state parameter domain to complete the time-frequency segment-level feature transformation; after mapping, a time-frequency embedding feature that perfectly matches the time-frequency state parameter domain is obtained.

[0112] For example, the main frequency band features in the acoustic-electric coupling features in step 220 are mapped to time-frequency embedding features, such as the main frequency band center frequency of the acoustic signal being 125kHz and the frequency band energy accounting for 68%, and the main frequency band center frequency of the electrical signal being 6.5MHz and the frequency band energy accounting for 72%; the amplitude features are mapped to the maximum pulse amplitude (acoustic 3.2V, electrical 150mV) and the amplitude standard deviation (acoustic 0.5V, electrical 20mV).

[0113] Step 234: Extract time-frequency correlation from the time-frequency embedded features to obtain state analysis features.

[0114] Specifically, based on the physical laws of transformer partial discharge faults, the correlation of each dimension of the time-frequency embedding feature is verified. If the verification finds that a certain dimension of the feature does not conform to the physical correlation (such as the matching degree between the main frequency band and the charge quantity is <80%), the feature of that dimension is corrected based on the correlation coefficient. In this embodiment, the matching degree of each dimension is ≥98%, so no correction is required. After correlation extraction and verification, the final value of the time-frequency embedding feature is locked, and a state analysis feature that is aligned with the depth of the preset state parameter domain is generated.

[0115] For example, a graph attention network is used to learn the correlations between time-frequency parameters, such as the positive correlation between the center frequency of the main frequency band and the maximum pulse amplitude (correlation coefficient 0.85), and the negative correlation between the amplitude standard deviation and the phase distribution dispersion (correlation coefficient -0.6). The final output state analysis features are: the first dimension is the partial discharge charge of 650pC, the second dimension is the phase distribution centroid angle of 45°, the third dimension is the fault location coordinates (30cm from the left flange), and the fourth dimension is the discharge stability index of 0.85 (the closer to 1, the more stable). It should be understood that each dimension corresponds to a physically interpretable state parameter.

[0116] In one embodiment, step 240 performs acoustic-electric joint diagnostic reasoning based on task constraint information and state parsing features to obtain partial discharge reasoning features, including steps 242 to 244:

[0117] Step 242: Based on the task constraint information, perform logical reasoning on the state parsing features to obtain the partial discharge reasoning features.

[0118] Specifically, using task constraint information as a directional constraint, the parameters of each dimension of the state parsing features are correlated, filtered and refined, redundant information unrelated to the task is eliminated, core diagnostic features are retained, and finally, partial discharge reasoning features are generated.

[0119] For example, based on the pre-tuned Transformer architecture, the state parsing features are first semantically associated with the constraint parameters (charge, phase distribution, and position) in the task. The encoder learns the feature combination "charge 650pC + bipolar phase distribution (0~90° / 180~270° percentage 42% / 45%) + position 30cm", which highly matches the "floating potential discharge" mode, generating partial discharge inference features containing key information about the fault mode (such as the degree of discharge energy concentration, phase symmetry, etc.).

[0120] Step 244: Semantic generation is performed on the partial discharge reasoning features to form a partial discharge state description.

[0121] Specifically, based on the partial discharge inference characteristics and combined with the output constraints of task constraint information, the semantic generation module (which has built-in commonly used expression standards for power operation and maintenance) transforms the structured inference characteristics into natural, accurate, and semantic descriptions that meet the needs of operation and maintenance.

[0122] For example, based on the output format required by the task, the partial discharge inference characteristics are converted into a natural language partial discharge state description: "Partial discharge signal was detected in the A phase busbar of the GIS equipment. The partial discharge charge is 650pC, the phase distribution is bipolar symmetrical (42% for 0~90° and 45% for 180~270°), the fault location is 30cm from the left flange, and the discharge stability is high." Through the directional constraints of the task constraint information, it is ensured that the logical reasoning focuses on the operation and maintenance needs and avoids invalid reasoning.

[0123] Finally, the server integrates the partial discharge status description and acoustic signal tags output by the model into a diagnostic report, which is then pushed to the maintenance terminal through the substation monitoring system. The report includes information such as "Partial discharge status of GIS equipment A-phase busbar: floating potential discharge (FM-02), charge 650pC, fault location 30cm from the left flange, power outage maintenance recommended within 24 hours." Based on this report, maintenance personnel develop targeted maintenance strategies to achieve early warning and precise handling of partial discharge faults.

[0124] In one embodiment, such as Figure 3 The diagram shows a flowchart of the two-stage collaborative optimization training of the partial discharge diagnostic model. In the first stage, the front-end network is forced to learn the general low-level physical feature alignment mapping across operating conditions by freezing the back-end network. In the second stage, the front-end network is frozen to focus on training the back-end network's specific diagnostic task reasoning and semantic generation capabilities.

[0125] The partial discharge diagnosis method is based on an acoustic-electric diagnostic network, which consists of a pre-feature network and a post-inference network cascaded together. The pre-feature network is used to extract features and spatially align acoustic-electric signals, and output state analysis features. The post-inference network is used to perform inference based on task constraint information and state analysis features, and output partial discharge inference features.

[0126] The training steps for the pre-feature network and the post-inference network include steps 310 to 340:

[0127] Step 310: Obtain the acoustic-electric coupling sample set of the basic operating condition library, the acoustic-electric coupling sample set of the target operating condition library, and the diagnostic description semantic set.

[0128] The basic operating condition library consists of 5,000 sets of acoustic and electrical signals simulated in the laboratory. The acoustic and electrical coupling sample set of the basic operating condition library includes acoustic and electrical signals of typical faults (such as the acoustic signal of metal tip discharge with 55% energy in the 100~120kHz frequency band and the electrical signal with 60% energy in the 8~10MHz frequency band), diagnostic guidance constraints (such as "discharge event calibration" and "feature parameter extraction"), and state parameter benchmark values ​​for different fault types (such as charge of 500pC and phase distribution centroid angle of 30°).

[0129] The target operating condition database consists of 3,000 sets of measured samples synchronized with the substation field server. The acoustic-electric coupling sample set of the target operating condition database includes acoustic and electrical signals during the operation of GIS equipment (such as floating potential discharge signals with background noise), field diagnostic guidance constraints (such as "fault mode determination"), status parameter benchmark values ​​(such as charge 650pC, fault location 30cm) and fault mode identification codes (such as "FM-02" corresponding to floating potential discharge).

[0130] The diagnostic description semantic set consists of 2,000 text descriptions extracted from expert diagnostic reports, such as "metal tip discharge: charge 500±50pC, phase distribution is unipolar (0~90° accounts for 80%)", with each description associated with the actual state parameter benchmark value.

[0131] Step 320: Based on the acoustic-electric coupling sample set of the basic working condition library and the acoustic-electric coupling sample set of the target working condition library, perform the first stage training of aligning acoustic-electric coupling features and state semantic features. While keeping the network topology weights of the post-inference network unchanged, update the network topology weights of the pre-feature network until the first model convergence condition is met.

[0132] Specifically, the acoustic-electric signals from the basic operating condition database (such as metal tip discharge samples) and the target operating condition database (such as on-site floating potential discharge samples) are feature-encoded to obtain sample acoustic-electric coupling features. The network parameters of the diagnostic inference unit are frozen (the network topology weights are pre-adjusted based on general diagnostic knowledge). Only the acoustic-electric signal feature extraction unit and the feature space alignment unit are optimized, and their weights are updated. By minimizing the parameter offset between the state analysis features and the state parameter reference value (e.g., the deviation between the calculated charge value and the 500pC reference value ≤ 5%, and the fault location error ≤ 2cm), the acoustic-electric coupling features are accurately mapped to the state parameter domain, enabling the extracted features to be accurately mapped to the physical parameter domain. When the mean parameter offset of the basic operating condition database samples is ≤ 3% and the mean parameter offset of the target operating condition database samples is ≤ 5%, the first model convergence condition is met, and training stops.

[0133] Step 330: Based on the acoustic-electric coupling sample set and diagnostic description semantic set of the target working condition library, perform the second stage training of partial discharge feature extraction and acoustic-electric signal recognition for the network that meets the convergence condition of the first model.

[0134] The diagnostic description semantic set includes fault mode labels and expert-annotated status description text.

[0135] Specifically, the acoustic and electrical signals from the target operating condition database are input into the acoustic and electrical signal feature extraction unit (weight freezing) to obtain state analysis features; at the same time, the text description of the diagnostic description semantic set (such as "floating potential discharge: bipolar phase distribution") is input into the diagnostic reasoning unit.

[0136] Step 340: In the second stage of training, keep the network topology weights of the preceding feature network unchanged, and update the network topology weights of the following inference network until the convergence condition of the second model is met.

[0137] For example, the weights of the acoustic signal feature extraction unit and the feature space alignment unit are frozen, and only the weights of the diagnostic reasoning unit and the label recognition unit are updated; the diagnostic reasoning logic and label classification ability are optimized by jointly minimizing the "semantic deviation between the partial discharge state description and the diagnostic description semantic set" (e.g., the BLEU score of the generated description and the expert description is ≥0.9) and the "label recognition accuracy" (e.g., the fault mode identification code determination accuracy is ≥95%); when the semantic deviation is ≤3% and the label recognition accuracy is ≥95%, the second model convergence condition is met, and training is stopped.

[0138] The original diagnostic model, which meets the convergence condition of the second model after two stages of training, is determined as the final usable "acoustic-electric joint diagnostic model" and deployed to the substation real-time diagnostic system for online monitoring of partial discharge status of GIS equipment. Through the above process, the server realizes the migration of the model from "laboratory basic feature learning" to "adaptation to complex field conditions," ensuring that the model has both high diagnostic accuracy and strong robustness in practical applications.

[0139] In some embodiments, to improve the sensitivity of the feature encoding network to processing complex multimodal signals in the initial stage, the feature encoding network performing feature encoding processing can also undergo a specific pre-lower-level network tuning process before being formally put into the aforementioned two-stage co-optimization training, including steps 410 to 430, wherein:

[0140] Step 410: Obtain the sample acoustic-electric state association group.

[0141] This association group contains pairs of massive historical sample acoustic and electrical signals and their corresponding underlying basic state semantic tags. For example, the metal tip discharge sample contains acoustic and electrical signals, state parameter reference values, and tags. The acoustic signal has an energy ratio of 55% in the 100~120kHz frequency band and an electrical signal has an energy ratio of 60% in the 8~10MHz frequency band. The corresponding state parameter reference values ​​are "charge 500pC, phase distribution centroid angle 30°" and the corresponding tag is "metal tip discharge".

[0142] Step 420: Configure the state semantic feature network branches that are independent yet mutually constrained, and the acoustic-electric signal feature network branches based on multi-source dynamic weighting.

[0143] Among them, the state semantic feature sub-model extracts semantic features from the state parameter benchmark value, such as "high charge + unipolar phase distribution" corresponding to the semantics of metal tip discharge; the multi-source dynamically weighted acoustic-electric signal feature sub-model dynamically adjusts the acoustic-electric feature weights according to the sample signal-to-noise ratio, such as setting the weight to 0.7 when the electrical signal signal-to-noise ratio is high.

[0144] Step 430: The state semantic feature sub-model guides the mapping of acoustic and electrical signal features to state parameters, and the source dynamic weighting sub-model optimizes the weights to match the fused features with the labels, thereby completing the accurate mapping of acoustic and electrical signals to state parameters and the collaborative optimization of labels, and obtaining the trained acoustic and electrical signal feature extraction unit.

[0145] For example, mapping the acoustic-electric energy ratio to a charge of 500 pC with a deviation of ≤5% results in a metal tip discharge sample labeling accuracy of ≥95%. This effectively enables the model to consciously remove mechanical vibration noise or periodic electromagnetic interference unrelated to the partial discharge state at the lowest physical signal encoding stage. This allows the final output acoustic-electric coupling features to naturally evolve towards a higher semantic space, thereby greatly reducing the computational pressure on the subsequent feature alignment network and significantly shortening the overall model's convergence period.

[0146] In one embodiment, the acoustic-electric coupling sample set of the basic operating condition library includes a first sample acoustic-electric signal, a first diagnostic guidance constraint, and a first state parameter reference value corresponding to the first sample acoustic-electric signal.

[0147] The first sample acoustic-electric signal is a pure signal under ideal conditions, including acoustic signals collected by ultrasonic sensors (55% energy in the 100~150kHz frequency band, maximum amplitude 2.8V, no background noise) and electrical signals collected by HFCT (60% energy in the 8~10MHz frequency band, maximum pulse amplitude 120mV, phase concentrated in 0~90°). The first diagnostic guidance constraint is the "feature parameter extraction paradigm", which guides the diagnostic reasoning unit to extract global feature parameters (such as charge and phase distribution). The first state parameter benchmark value is the actual measurement value based on this paradigm, including a charge of 500pC (±20pC), a phase distribution centroid angle of 30° (80% of the phase interval in the 0~90° range), and a main frequency band center frequency of 125kHz (acoustic) / 8MHz (electric).

[0148] The acoustic-electric coupling sample set of the target working condition library includes the second sample acoustic-electric signal, the second diagnostic guidance constraint, and the second state parameter reference value corresponding to the second sample acoustic-electric signal.

[0149] The second sample acoustic-electric signal is an actual signal containing background noise, including acoustic signal superimposed with substation equipment vibration noise (signal-to-noise ratio 15dB), with 68% energy in the 100~150kHz frequency band (noise causing fluctuations of ±5%), and a maximum amplitude of 3.2V (including random noise spikes); the electrical signal is affected by cable interference, with pulse interval fluctuations of ±1ms, and phase distribution in the bipolar ranges of 0~90° and 180~270° (42% and 45% respectively); the second diagnostic guidance constraint is the "fault mode determination paradigm", which guides the diagnostic reasoning unit to associate fault type characteristics; the second state parameter benchmark value is the field measured value, including charge of 650pC (±30pC), fault location of 30cm (distance from the left flange, calibrated by a laser rangefinder), and discharge stability index of 0.85 (calculated based on amplitude standard deviation).

[0150] The first phase of training includes steps 510 to 550, in which:

[0151] Step 510: Extract the acoustic-electric coupling features of the sample.

[0152] The first sample acoustic-electric signal (metal tip discharge) and the second sample acoustic-electric signal (floating potential discharge) are respectively input into the acoustic-electric signal feature extraction unit of the original model. For the first sample acoustic-electric signal, the feature extraction unit captures the main frequency band of the acoustic signal from 100 to 150 kHz through short-time Fourier transform, extracts the amplitude fluctuation pattern through a convolutional network (maximum amplitude 2.8V, standard deviation 0.3V); learns the pulse sequence of the electrical signal through a recurrent network to obtain the phase distribution characteristics (80% from 0 to 90°), and then fuses them through an attention mechanism (acoustic-electric weight 0.5:0.5, weight balance under ideal signal), outputting the "first sample acoustic-electric coupling characteristics", which includes the main frequency band energy ratio (55% / 60%), phase concentration (0.92), and amplitude statistics (2.8V / 120mV). For the second sample acoustic-electric signal, the feature extraction unit extracts the energy proportion of the 100~150kHz frequency band of the acoustic signal (fluctuation ±2% after denoising) and the bipolar phase distribution features of the electrical signal after wavelet denoising to suppress noise. The weights are dynamically adjusted during fusion (the electrical signal has a high signal-to-noise ratio, so the weight is 0.6), and the "acoustic-electric coupling features of the second sample" are output, which include the proportion of the main frequency band after noise suppression (68% / 72%), the bipolar phase parameters (42% / 45%), and the amplitude standard deviation (0.5V / 20mV).

[0153] Step 520: Generate state parsing features, including performing feature encoding and feature transformation on the first sample acoustic-electric coupling features and the second sample acoustic-electric coupling features respectively, to obtain the first state parsing features and the second state parsing features.

[0154] The acoustic-electric coupling features of the first and second samples are input into the feature space alignment unit. For the coupling features of the first sample, the alignment unit first performs time-frequency band-level processing, including mapping the proportion of the main acoustic and electrical frequency bands to time-frequency embedded features (acoustic center frequency 125kHz, energy proportion 55%; electrical center frequency 8MHz, energy proportion 60%), and then learns the positive correlation between "main frequency band proportion - amplitude" (correlation coefficient 0.88) through a graph attention network, and outputs "first state analytical features", including charge 490pC, phase centroid angle 32°, and main frequency band center frequency 125kHz / 8MHz (deviations from the first state parameter reference values ​​are 2%, 6.7%, and 0%, respectively).

[0155] For the coupling characteristics of the second sample, the coupling characteristics after noise interference are aligned by the unit. The time-frequency embedding characteristics are the acoustic center frequency of 125kHz (accounting for 68%) and the electrical center frequency of 6.5MHz (accounting for 72%). The "second state analytical characteristics" are obtained through correlation analysis, including charge of 620pC, fault location of 33cm, and stability index of 0.82 (the deviations from the second state parameter reference values ​​are 4.6%, 10%, and 3.5%, respectively).

[0156] Step 530: Generate sample partial discharge state descriptions, including constructing first sample training data based on first state analytical features and first diagnostic guidance constraints, constructing second sample training data based on second state analytical features and second diagnostic guidance constraints, and performing acoustic-electric joint diagnostic reasoning based on the first sample training data and the second sample training data to obtain the first sample partial discharge state description and the second sample partial discharge state description, respectively.

[0157] The first state analysis features and the first diagnostic guidance constraints (“feature parameter extraction paradigm”) are input into the diagnostic inference unit (weights frozen, pre-adjusted based on general diagnostic knowledge). The inference unit outputs a description of the first sample partial discharge state, including “metal tip discharge signal, charge 490pC, phase distribution centroid angle 32°, acoustic main frequency band 125kHz (energy percentage 55%), electrical main frequency band 8MHz (energy percentage 60%)”. The second state analysis features and the second diagnostic guidance constraints (“fault mode determination paradigm”) are input into the diagnostic inference unit, and a description of the second sample partial discharge state is output, including “suspected floating potential discharge signal, charge 620pC, fault location 33cm, stability index 0.82”.

[0158] Step 540: Determine the first diagnostic error based on the parameter offset between the first sample partial discharge state description and the first state parameter reference value, and the parameter offset between the second sample partial discharge state description and the second state parameter reference value.

[0159] The "parameter offset" between the partial discharge state description and the state parameter baseline value of the sample is calculated. For the first sample, the charge offset = |490-500| / 500 = 2%, the phase centroid angle offset = |32-30| / 30 = 6.7%, and the average is 4.35%. For the second sample, the charge offset = |620-650| / 650 = 4.6%, the fault location offset = |33-30| / 30 = 10%, and the stability index offset = |0.82-0.85| / 0.85 = 3.5%, and the average is 6.03%. Combining the offsets of the two types of samples, the first diagnostic error is calculated as (4.35% × base sample weight 0.6 + 6.03% × target sample weight 0.4) = 5.02%.

[0160] Step 550: Update the network topology weights of the preceding feature network based on the first diagnostic error.

[0161] The network topology weights of the diagnostic inference unit are frozen (preserving the pre-tuned feature-semantic association logic). Based on the first diagnostic error (5.02%), the weights of the acoustic-electric signal feature extraction unit and the feature space alignment unit are updated. This includes adjusting the network parameters of the "acoustic-electric pulse time difference-position mapping" in the feature space alignment unit to reduce the position calculation error, given a 10% fault location offset in the second sample; and optimizing the encoding weights of the electrical signal phase distribution in the feature extraction unit, given a 6.7% phase offset in the first sample. After 200 iterations of training, the parameter offsets are calculated again, including a mean of 2.8% for the first sample (charge deviation 1.5%, phase deviation 3.2%) and a mean of 4.5% for the second sample (charge deviation 3%, position deviation 5%, stability index deviation 2.5%). The first diagnostic error is reduced to (2.8% × 0.6 + 4.5% × 0.4) = 3.48%, satisfying the first model convergence condition (error ≤ 5%), and training is stopped.

[0162] After the first collaborative optimization training, the acoustic-electric signal feature extraction unit can stably extract coupled features from ideal / actual acoustic-electric signals, and the feature space alignment unit can accurately map the coupled features to the state parameter domain (the average parameter offset is ≤4.5%), laying the feature foundation for the subsequent second-stage diagnosis-recognition training.

[0163] In one embodiment, the acoustic-electric coupling sample set of the target operating condition library also includes a fault mode identification code for the second sample acoustic-electric signal, and the diagnostic description semantic set includes the sample diagnostic operation paradigm and the actual state parameter benchmark value.

[0164] The second phase of training includes steps 610 to 650, in which:

[0165] Step 610: Perform feature encoding and feature transformation on the second sample acoustic-electric signal to obtain the third state analytical features.

[0166] Specifically, the second sample acoustic-electric signal is used as the input of the acoustic-electric signal feature extraction unit that meets the convergence condition of the first model to perform feature encoding processing, thereby obtaining the third sample acoustic-electric coupling feature; the third sample acoustic-electric coupling feature is used as the input of the feature space alignment unit that meets the convergence condition of the first model to perform feature transformation, thereby obtaining the third state analytical feature.

[0167] Step 620: Perform acoustic-electric joint diagnostic reasoning based on the third state analytical features and the sample diagnostic operation paradigm to obtain the third sample reasoning features, the third sample partial discharge state description, and the actual partial discharge state description.

[0168] Specifically, constrained by the sample diagnostic operation paradigm, the third-state analytical features (sample features optimized in the first stage, such as [544pC, 12.6cm, 0.88, 150kHz / 7.5MHz]) are subjected to dimensional logical reasoning to obtain the third-sample reasoning features; semantic generation is then performed on these reasoning features to obtain the third-sample partial discharge state description. Both are outputs of model training, while the actual partial discharge state description is the expert-annotated ground truth value inherent in the sample, used to compare calculation errors and optimize the model.

[0169] Step 630: Based on the inference features of the third sample, the label is determined to obtain the category determination result of the sample acoustic and electrical signals.

[0170] Specifically, based on the inference features of the third sample, the label determination is completed through classification matching and fault mode recognition, and the category determination result of the sample acoustic and electrical signals is output.

[0171] Step 640: Determine the second diagnostic error based on the third sample partial discharge state description, the second state parameter reference value, the actual partial discharge state description, the actual state parameter reference value, the sample acoustic signal category determination result, and the fault mode identification code; update the network topology weights of the post-inference network according to the second diagnostic error.

[0172] In short, the model output is compared with various benchmark values ​​and label ground truths to calculate the semantic error of the state description, the numerical error of the state parameters, and the classification error of the category labels. These errors are then weighted and fused to obtain the second diagnostic error. This error is then used as a supervision signal to backpropagate and update the network topology weights of the post-inference network, completing the second stage of training optimization. By constraining the sample diagnostic operation paradigm, the model inference is made to fit actual diagnostic needs (such as fault determination / parameter extraction), avoiding aimless generalization and improving the practicality of the diagnostic results. By comparing the model output (i.e., the third sample partial discharge state description / category determination) with the expert-annotated ground truths (i.e., the actual state description / fault identification code), the error is accurately calculated and the post-inference network is optimized, significantly improving the semantic accuracy of the state description and the matching degree of the label determination. Training is conducted based on the accurate features optimized in the first stage (i.e., the third state parsing features), avoiding feature bias from interfering with the optimization of the inference layer, ensuring that the post-inference network focuses on improving semantic generation and label determination capabilities, and improving the overall model diagnostic accuracy.

[0173] In one embodiment, both the first diagnostic guidance constraint and the second diagnostic guidance constraint include a discharge event calibration paradigm, a feature parameter extraction paradigm, and a fault mode determination paradigm.

[0174] First, the discharge event calibration paradigm is used for directional constraints to locate acoustic and electrical signal discharge events, and the state parameter reference value corresponding to the discharge event calibration paradigm has a spatial position reference value.

[0175] Specifically, the discharge event calibration paradigm orientation-guided diagnostic reasoning unit accurately locates the physical location of partial discharge in the acoustic and electrical signals, and its corresponding state parameter reference values ​​include clear spatial location reference values ​​(such as distance from the device reference point, three-dimensional coordinates, etc.).

[0176] The server needs to locate the partial discharge source within the A-phase busbar of the GIS equipment. At this time, the first diagnostic guidance constraint (basic operating condition library) or the second diagnostic guidance constraint (target operating condition library) in the task constraint information is configured as "discharge event calibration paradigm", and the specific instruction is "locate the partial discharge source based on the time difference of acoustic and electrical signals, output the axial distance (cm) from the left flange of the busbar, and the positioning error is ≤2cm".

[0177] The execution process involves installing three ultrasonic sensors (20cm apart, numbered S1, S2, and S3, 10cm, 30cm, and 50cm from the left flange, respectively) equidistantly along the axial direction on the outer wall of the GIS equipment to synchronously collect partial discharge acoustic signals. An HFCT sensor collects electrical signals as a time reference (the propagation speed of the electrical signal is close to the speed of light, considered as the time of partial discharge occurrence t0). The server receives the acoustic and electrical signals, including the electrical signal pulse trigger time t0 = 10:00:00.000000s; the times when S1, S2, and S3 collect the acoustic signals are t1 = 10:00:00.000020s, t2 = 10:00:00.000010s, and t3 = 10:00:00.000020s, respectively (the speed of sound in SF6 gas is approximately 340m / s, requiring approximately 588μs to propagate 20cm; the actual time difference varies depending on the location).

[0178] Subsequently, under the constraints of the "discharge event calibration paradigm," the diagnostic inference unit invokes the acoustic-electric time difference localization algorithm. Specifically, based on the electrical signal t0, the occurrence time of partial discharge is determined, and the arrival time differences of the signals from each acoustic sensor are calculated: Δt1 = t1 - t0 = 20 μs, Δt2 = t2 - t0 = 10 μs, and Δt3 = t3 - t0 = 20 μs. Combining the sensor spacing and sound velocity, a spatial localization equation is established. For example, if the distance from the partial discharge source to the left flange is x, and the distances from sensors S1 (10 cm), S2 (30 cm), and S3 (50 cm) to x are |x - 10|, |x - 30|, and |x - 50|, then the sound propagation time difference Δt2 - Δt1 = (|x - 30| - |x - 10|) / 340 m / s = -10 μs, solving for x yields x = 30 cm.

[0179] The state parameter reference value, including the first / second state parameter reference value corresponding to this paradigm, is the "spatial position reference value," which is the actual position of the partial discharge source measured on-site by a laser rangefinder, for example, 30±1cm (axial distance). The server compares the positioning result (30cm) with the reference value, and the deviation is 1cm≤2cm, which meets the positioning accuracy requirements.

[0180] Secondly, the feature parameter extraction paradigm is used for directional constraint to extract global feature parameters of acoustic and electrical signals, and the state parameter reference value corresponding to the feature parameter extraction paradigm has a feature parameter amplitude space.

[0181] Specifically, the feature parameter extraction paradigm-guided diagnostic reasoning unit extracts global feature parameters (such as charge, phase distribution, and main frequency band energy) that reflect the nature of partial discharge from the acoustic and electrical signals. The corresponding state parameter reference values ​​include the feature parameter amplitude space (i.e., the reasonable fluctuation range of the parameters).

[0182] The server needs to extract global characteristic parameters of the partial discharge signal from the GIS equipment for subsequent fault analysis. At this time, the diagnostic guidance constraint is configured as "characteristic parameter extraction paradigm", and the specific instructions are "extract partial discharge charge (pC), phase distribution ratio (%), and acoustic-electric main frequency band energy ratio (%), and the parameters must be within the measured amplitude range".

[0183] The execution process includes signal input, acquiring the partial discharge signal of the B-phase circuit breaker of the GIS equipment as the acoustic signal to be analyzed. The acoustic signal has concentrated energy in the 100~200kHz frequency band, and the electrical signal pulse phase is distributed in the 0~180° range.

[0184] Subsequently, under the constraints of the "feature parameter extraction paradigm," the diagnostic inference unit performs multi-dimensional feature extraction, including:

[0185] The charge quantity, including the charge quantity calibration curve based on the electrical signal pulse amplitude and charge quantity (laboratory pre-calibration, pulse amplitude of 150mV corresponds to a charge quantity of 650pC), is extracted as 645pC;

[0186] The phase distribution percentage includes the percentage of the electrical signal pulse in four phase intervals: 0~90°, 90~180°, 180~270°, and 270~360°, which are 42%, 38%, 10%, and 10%, respectively, with the main focusing interval 0~180° accounting for 80%.

[0187] The energy percentage of the main frequency band, including the Fourier transform of the acoustic signal, shows that the energy in the 120~160kHz band accounts for 72% of the total energy; and the energy in the 6~8MHz band accounts for 68% of the total energy when the frequency spectrum of the electrical signal is analyzed.

[0188] The state parameter reference value, at this time, is the "characteristic parameter amplitude space" corresponding to this paradigm, that is, the parameter fluctuation range measured on site, including charge 650±30pC (amplitude space 620~680pC), 0~180° phase ratio 75%~85% (amplitude space), acoustic main frequency band energy ratio 70%±5%, and electrical main frequency band energy ratio 65%±5%. The characteristic parameters extracted by the server (645pC, 80%, 72%, 68%) all fall within the amplitude space, meeting the extraction requirements.

[0189] Third, the fault mode determination paradigm is used for fault mode determination of partial discharge mode of acoustic and electrical signals under directional constraints.

[0190] Specifically, the fault mode determination paradigm-guided diagnostic reasoning unit matches the extracted feature parameters with the feature library of typical fault modes to achieve fault type identification. The corresponding state parameter benchmark values ​​include the feature parameter range of different fault modes (such as the unipolar phase distribution of metal tip discharge and the bipolar phase distribution of floating potential discharge).

[0191] The server needs to determine the fault mode (metal tip, floating potential, or free particle discharge) corresponding to the partial discharge signal of the GIS equipment. At this time, the diagnostic guidance constraint is configured as "fault mode determination paradigm", and the specific instruction is "based on feature parameter matching of typical fault mode feature library, output fault mode type and code".

[0192] The execution process includes input feature parameters, which are the features already obtained based on the feature parameter extraction paradigm, including charge 645pC (620~680pC), phase distribution 0~90° accounting for 42%, 180~270° accounting for 45% (bipolar), acoustic main frequency band 120~160kHz (energy accounting for 72%), and electrical main frequency band 6~8MHz (energy accounting for 68%).

[0193] Subsequently, under the constraints of the "fault mode determination paradigm," the diagnostic inference unit calls upon the fault mode feature library (pre-trained typical fault feature templates), including:

[0194] Metal tip discharge, including unipolar phase distribution (0~90° ≥70%), charge 300~500pC, and main acoustic frequency band 80~120kHz;

[0195] Floating potential discharge includes a bipolar phase distribution (0~90° and 180~270° each account for ≥40%), a charge of 500~700pC, and a main acoustic frequency band of 100~160kHz.

[0196] Free particle discharge includes phase distribution dispersion (each interval accounts for 20%~30%), charge amount of 200~400pC, and acoustic main frequency band of 50~100kHz.

[0197] The server matches the extracted feature parameters with templates in the library. The bipolar phase distribution (42% / 45%), charge amount of 645pC (500~700pC), and acoustic main frequency band of 120~160kHz all conform to the characteristics of "floating potential discharge".

[0198] The corresponding state parameter benchmark value for this paradigm is the "fault mode characteristic range," which is the characteristic parameter range of floating potential discharge, namely, bipolar phase ratio ≥40%, charge 500~700pC, and main acoustic frequency band 100~160kHz. If the server determines that the result matches this range, it outputs the fault mode type "floating potential discharge" and the code "FM-02."

[0199] In summary, the server achieves spatial localization of partial discharge sources (spatial location reference value 30±1cm) through the "discharge event calibration paradigm," extracts global feature parameters (charge quantity, phase ratio, etc. within the amplitude variation space) through the "feature parameter extraction paradigm," and completes fault type identification (matching typical fault feature ranges) through the "fault mode determination paradigm." These three diagnostic operation paradigms respectively constrain the logic of the diagnostic reasoning unit from three dimensions: spatial localization, feature quantization, and pattern recognition. Combined with the corresponding state parameter reference values ​​(spatial location, amplitude variation space, and fault feature range), this achieves multi-dimensional and precise constraints on partial discharge diagnosis of GIS equipment, providing clear guiding logic for subsequent diagnostic reasoning and tag identification.

[0200] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.

[0201] Based on the same inventive concept, this application also provides a partial discharge diagnostic device for implementing the partial discharge diagnostic method described above. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations in one or more embodiments of the partial discharge diagnostic device provided below can be found in the limitations of the partial discharge diagnostic method described above, and will not be repeated here.

[0202] In one exemplary embodiment, such as Figure 4 As shown, a partial discharge diagnostic device 400 is provided, including: an information acquisition module 401, a feature encoding module 403, a feature alignment module 405, a diagnostic reasoning module 407, and a result output module 409, wherein:

[0203] Information acquisition module 401 is used to acquire the acoustic and electrical signals to be analyzed and task constraint information;

[0204] Feature encoding module 403 is used to perform feature encoding processing on the acoustic-electric signal to be analyzed to obtain acoustic-electric coupling features;

[0205] The feature alignment module 405 is used to perform feature transformation on the acoustic-electric coupling features, align the acoustic-electric coupling features to a preset state parameter domain, and obtain state analysis features.

[0206] The diagnostic reasoning module 407 is used to perform joint acoustic and electronic diagnostic reasoning based on task constraint information and state analysis features to obtain partial discharge reasoning features.

[0207] The result output module 409 performs semantic generation based on the partial discharge inference features to obtain the diagnostic results of the acoustic-electric signal to be analyzed.

[0208] In one embodiment, the feature alignment module 405 performs feature transformation on the acoustic-electric coupling features, aligning the acoustic-electric coupling features to a preset state parameter domain to obtain state parsing features, including:

[0209] The acoustic-electric coupling features are processed at the time-frequency band level and transformed into the time-frequency state parameter domain to obtain the time-frequency embedding features.

[0210] Time-frequency correlation extraction is performed on the time-frequency embedded features to obtain state analysis features;

[0211] In the diagnostic reasoning module 407, joint acoustic and electrical diagnostic reasoning is performed based on task constraint information and state analysis features to obtain partial discharge reasoning features, including:

[0212] Logical reasoning is performed on the state parsing features based on task constraint information to obtain partial discharge reasoning features;

[0213] Semantic generation is performed on the partial release inference features to form a partial release state description.

[0214] In one embodiment, the partial discharge diagnosis method is executed based on an acoustic-electric diagnostic network, which consists of a pre-feature network and a post-inference network cascaded together. The pre-feature network is used to extract features and spatially align the acoustic-electric signals, outputting state analysis features; the post-inference network is used to perform inference based on task constraint information and state analysis features, outputting partial discharge inference features; the training steps for the pre-feature network and the post-inference network include:

[0215] Acquire the acoustic-electric coupling sample set from the basic operating condition library, the acoustic-electric coupling sample set from the target operating condition library, and the diagnostic description semantic set;

[0216] Based on the acoustic-electric coupling sample set of the basic working condition library and the acoustic-electric coupling sample set of the target working condition library, the first stage of training is carried out to align the acoustic-electric coupling features and state semantic features. While keeping the network topology weights of the post-inference network unchanged, the network topology weights of the pre-feature network are updated until the first model convergence condition is met.

[0217] Based on the acoustic-electric coupling sample set and diagnostic description semantic set of the target working condition library, the second stage training of the network that meets the convergence condition of the first model is carried out for partial discharge feature extraction of acoustic and electric signals and recognition of acoustic and electric signals.

[0218] In the second stage of training, the network topology weights of the front feature network are kept unchanged, while the network topology weights of the back inference network are updated until the convergence condition of the second model is met.

[0219] In one embodiment, the acoustic-electric coupling sample set of the basic operating condition library includes a first sample acoustic-electric signal, a first diagnostic guidance constraint, and a first state parameter reference value corresponding to the first sample acoustic-electric signal.

[0220] The acoustic-electric coupling sample set of the target working condition library includes the second sample acoustic-electric signal, the second diagnostic guidance constraint, and the second state parameter reference value corresponding to the second sample acoustic-electric signal.

[0221] The first phase of training includes:

[0222] Feature encoding and feature transformation are performed on the first sample acoustic-electric signal and the second sample acoustic-electric signal, respectively, to obtain the first state analytical features and the second state analytical features.

[0223] First sample training data is constructed based on first state analytical features and first diagnostic guidance constraints, and second sample training data is constructed based on second state analytical features and second diagnostic guidance constraints. Then, acoustic-electric joint diagnostic reasoning is performed based on the first sample training data and the second sample training data to obtain the first sample partial discharge state description and the second sample partial discharge state description, respectively.

[0224] The first diagnostic error is determined based on the parameter offset between the first sample partial discharge state description and the first state parameter reference value, and the parameter offset between the second sample partial discharge state description and the second state parameter reference value. The network topology weights of the preceding feature network are then updated based on the first diagnostic error.

[0225] In one embodiment, the acoustic-electric coupling sample set of the target operating condition library also includes the fault mode identification code of the second sample acoustic-electric signal, and the diagnostic description semantic set includes the sample diagnostic operation paradigm and the actual state parameter benchmark value.

[0226] The second phase of training includes:

[0227] The second sample acoustic-electric signal is subjected to feature encoding and feature transformation to obtain the third state analytical features;

[0228] Based on the third-state analytical features and the sample diagnostic operation paradigm, the acoustic-electric joint diagnostic reasoning is performed to obtain the third-sample reasoning features, the third-sample partial discharge state description, and the actual partial discharge state description.

[0229] The label determination result of the sample acoustic and electrical signal category is obtained by determining the label based on the inference features of the third sample.

[0230] The second diagnostic error is determined based on the third sample partial discharge state description, the second state parameter reference value, the actual partial discharge state description, the actual state parameter reference value, the sample acoustic and electrical signal category determination result, and the fault mode identification code. The network topology weights of the post-inference network are then updated based on the second diagnostic error.

[0231] In one embodiment, both the first diagnostic guidance constraint and the second diagnostic guidance constraint include a discharge event calibration paradigm, a feature parameter extraction paradigm, and a fault mode determination paradigm.

[0232] The discharge event calibration paradigm is used for directional constraint to locate acoustic and electrical signal discharge events, and the state parameter reference value corresponding to the discharge event calibration paradigm has a spatial position reference value.

[0233] The feature parameter extraction paradigm is used for global feature parameter extraction of acoustic and electrical signals under directional constraints, and the state parameter reference value corresponding to the feature parameter extraction paradigm has a feature parameter amplitude space.

[0234] The fault mode determination paradigm is used for fault mode determination of partial discharge mode of acoustic and electrical signals under directional constraints.

[0235] Each module in the aforementioned partial discharge diagnostic device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0236] In one exemplary embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 5 As shown, the computer device includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The input / output interface is used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, Near Field Communication (NFC), or other technologies. When executed by the processor, the computer program implements a partial discharge diagnostic method.

[0237] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0238] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0239] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above method embodiments.

[0240] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0241] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0242] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0243] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0244] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for diagnosing partial discharge, characterized in that, The method includes: Acquire the acoustic and electrical signals to be analyzed and the task constraint information; The acoustic-electric signal to be analyzed is subjected to feature encoding processing to obtain acoustic-electric coupling features; The acoustic-electric coupling features are aligned to a preset state parameter domain to obtain state analysis features; Based on the task constraint information and the state parsing features, acoustic-electric joint diagnostic reasoning is performed to obtain partial discharge reasoning features; Semantic generation is performed based on the partial discharge inference features to obtain the diagnostic results of the acoustic-electric signal to be analyzed.

2. The method according to claim 1, characterized in that, Aligning the acoustic-electric coupling features to a preset state parameter domain to obtain state analysis features includes: The acoustic-electric coupling features are subjected to time-frequency band-level feature processing, and the acoustic-electric coupling features are transformed to the time-frequency state parameter domain to obtain time-frequency embedding features; The time-frequency embedding features are subjected to time-frequency correlation extraction to obtain the state parsing features; The step of performing acoustic-electric joint diagnostic reasoning based on the task constraint information and the state analysis features to obtain partial discharge reasoning features includes: Based on the task constraint information, semantic association is performed with the state parsing features to obtain task-oriented partial inference features.

3. The method according to claim 1 or 2, characterized in that, The partial discharge diagnosis method is based on an acoustic-electric diagnostic network, which is composed of a pre-feature network and a post-inference network cascaded together. The pre-feature network is used to extract features and spatially align acoustic-electric signals, and output state analysis features. The post-inference network is used to perform inference based on task constraint information and state analysis features, and output partial discharge inference features. The training steps for the pre-feature network and the post-inference network include: Acquire the acoustic-electric coupling sample set from the basic operating condition library, the acoustic-electric coupling sample set from the target operating condition library, and the diagnostic description semantic set; Based on the acoustic-electric coupling sample set of the basic working condition library and the acoustic-electric coupling sample set of the target working condition library, the first stage of training is carried out to align the acoustic-electric coupling features and state semantic features. While keeping the network topology weights of the post-inference network unchanged, the network topology weights of the pre-feature network are updated until the first model convergence condition is met. Based on the acoustic-electric coupling sample set of the target working condition library and the diagnostic description semantic set, the network that meets the convergence condition of the first model is subjected to the second stage of training for partial discharge feature extraction of acoustic and electrical signals and recognition of acoustic and electrical signals. In the second stage of training, the network topology weights of the preceding feature network are kept unchanged, and the network topology weights of the following inference network are updated until the second model convergence condition is met.

4. The method according to claim 3, characterized in that, The acoustic-electric coupling sample set of the basic working condition library includes a first sample acoustic-electric signal, a first diagnostic guidance constraint, and a first state parameter reference value corresponding to the first sample acoustic-electric signal. The target operating condition library's acoustic-electric coupling sample set includes a second sample acoustic-electric signal, a second diagnostic guidance constraint, and a second state parameter reference value corresponding to the second sample acoustic-electric signal. The first phase of training includes: Feature encoding and feature transformation are performed on the first sample acoustic-electric signal and the second sample acoustic-electric signal, respectively, to obtain the first state analytical feature and the second state analytical feature. First sample training data is constructed based on the first state analysis features and the first diagnostic guidance constraints, and second sample training data is constructed based on the second state analysis features and the second diagnostic guidance constraints. Then, acoustic-electric joint diagnostic reasoning is performed based on the first sample training data and the second sample training data to obtain the first sample partial discharge state description and the second sample partial discharge state description, respectively. The first diagnostic error is determined based on the parameter offset between the first sample partial discharge state description and the first state parameter reference value, and the parameter offset between the second sample partial discharge state description and the second state parameter reference value. The network topology weights of the preceding feature network are then updated based on the first diagnostic error.

5. The method according to claim 4, characterized in that, The acoustic-electric coupling sample set of the target working condition library also includes the fault mode identification code of the second sample acoustic-electric signal, and the diagnostic description semantic set includes the sample diagnostic operation paradigm and the actual state parameter benchmark value. The second phase of training includes: The second sample acoustic-electric signal is subjected to feature encoding and feature transformation to obtain the third state analytical features; Based on the third state analytical features and the sample diagnostic operation paradigm, a joint acoustic-electrical diagnostic reasoning is performed to obtain the third sample reasoning features, the third sample partial discharge state description, and the actual partial discharge state description. Based on the third sample reasoning features, the label determination is performed to obtain the sample acoustic and electrical signal category determination result; The second diagnostic error is determined based on the third sample partial discharge state description, the second state parameter reference value, the actual partial discharge state description, the actual state parameter reference value, the sample acoustic signal category determination result, and the fault mode identification code. The network topology weights of the post-inference network are then updated based on the second diagnostic error.

6. The method according to claim 4, characterized in that, The first diagnostic guidance constraint and the second diagnostic guidance constraint include a discharge event calibration paradigm, a feature parameter extraction paradigm, and a fault mode determination paradigm. The discharge event calibration paradigm is used for directional constraint to locate acoustic and electrical signal discharge events, and the state parameter reference value corresponding to the discharge event calibration paradigm has a spatial position reference value. The feature parameter extraction paradigm is used for directional constraint to extract global feature parameters of acoustic and electrical signals, and the state parameter reference value corresponding to the feature parameter extraction paradigm has a feature parameter amplitude space. The fault mode determination paradigm is used for fault mode determination of partial discharge mode of acoustic and electrical signals under directional constraints.

7. A partial discharge diagnostic device, characterized in that, The device includes: The information acquisition module is used to acquire the acoustic and electrical signals to be analyzed and the task constraint information. The feature encoding module is used to perform feature encoding processing on the acoustic-electric signal to be analyzed to obtain acoustic-electric coupling features; The feature alignment module is used to align the acoustic-electric coupling features to a preset state parameter domain to obtain state analysis features. The diagnostic reasoning module is used to perform joint acoustic and electronic diagnostic reasoning based on the task constraint information and the state parsing features to obtain partial discharge reasoning features. The result output module is used to generate semantics based on the partial discharge inference features to obtain the diagnostic results of the acoustic-electric signal to be analyzed.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images