A partial discharge pattern recognition method and device based on an attention mechanism and a convolutional neural network

Abstract

The application relates to the technical field of power equipment monitoring, and discloses a partial discharge mode recognition method and device based on an attention mechanism and a convolutional neural network. The application aims to solve the technical problems of the existing partial discharge diagnosis method, such as untimely early warning, single-model feature recognition dependence, and insufficient root cause analysis capability. The application realizes accurate tracking from early warning to critical state by constructing a multi-modal feature library covering the whole life cycle of discharge and combining a dynamic frequency acquisition strategy. The core lies in that the discharge type is accurately recognized, the discharge source is located, and the physical parameter deviation of the equipment leading to the discharge is inversely deduced by utilizing multi-physical data cooperative positioning and inversion comprehensive analysis. By constructing a closed-loop diagnosis system from signal acquisition, intelligent analysis to root cause tracing, the accuracy and depth of the partial discharge diagnosis are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of power equipment monitoring technology, and more specifically, to a method and apparatus for partial discharge pattern recognition based on attention mechanisms and convolutional neural networks. Background Technology

[0002] Power equipment, especially high-voltage critical devices such as transformers and gas-insulated switchgear, is a core component ensuring the safe and stable operation of the power system. During long-term operation, the internal insulation structure of these devices accumulates damage due to the combined effects of electrical, thermal, mechanical stress, and environmental factors. Analyzing the partial discharge phenomena caused by this accumulated damage within the equipment can accurately identify the type of discharge, locate its source, and assess its severity. This plays an irreplaceable role in providing early warning of potential catastrophic failures, guiding precise predictive maintenance, and ensuring the safe and stable operation of the power system.

[0003] In recent years, with the rapid development of artificial intelligence technology, deep learning models, represented by convolutional neural networks and attention mechanisms, have provided powerful tools for processing complex signal and image data. Convolutional neural networks, with their excellent local feature extraction and spatial hierarchical modeling capabilities, have achieved great success in the field of image recognition, providing new possibilities for converting partial discharge signals into two-dimensional maps (such as phase-resolved partial discharge maps) and performing intelligent pattern recognition. Attention mechanisms, on the other hand, mimic the focusing ability of human vision. They enable models to automatically and selectively focus on the key parts of the input data most relevant to the current task when processing massive amounts of information, while ignoring secondary or irrelevant noise information. This holds promise for solving the problems of low signal-to-noise ratio and complex interference in partial discharge field monitoring.

[0004] Existing partial discharge diagnostic procedures often focus on analyzing obvious, already occurring discharge events, lacking effective tracking and correlation analysis of the complete evolution of a fault from its initial stage to its final form. Furthermore, some methods rely excessively on single data modes or fixed identification models, which challenges their accuracy and robustness when faced with complex field conditions, intertwined discharge phenomena, or atypical signal characteristics. More importantly, most diagnostic methods only provide a phenomenological classification result (such as "internal discharge" or "surface discharge"), failing to deeply trace the underlying physical causes of the discharge (such as the deterioration of local material parameters or changes in microstructure). This limits the diagnostic conclusions' ability to guide precise and efficient equipment maintenance. Summary of the Invention

[0005] In view of the aforementioned existing problems, the present invention is proposed.

[0006] Therefore, this invention provides a partial discharge pattern recognition method and apparatus based on attention mechanism and convolutional neural network, which solves the technical problems of untimely early warning, reliance on a single model for feature recognition, and insufficient root cause analysis capability in existing partial discharge diagnosis methods.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a method for partial discharge pattern recognition based on attention mechanisms and convolutional neural networks, which includes the following steps: S1: Environment Construction: Deploy a physical sensing array in the monitoring environment and construct a three-dimensional spatial model of the monitoring environment; establish an index association between the spatial model and the confidential database of core parameters of the equipment, and calibrate the spatial coordinates of each sensor in the physical sensing array in the spatial model; establish a discharge feature database. S2: Data Acquisition: The physical sensing array acquires low-frequency data, combines it with the discharge feature library to identify early warning features in the low-frequency acquired data, and then increases the acquisition frequency of the physical sensing array to medium frequency; combines it with the discharge feature library to identify critical features in the medium-frequency acquired data, and then increases the acquisition frequency of the physical sensing array from medium frequency to high frequency; after capturing discharge event data, the acquisition frequency of the physical sensing array is reduced to medium frequency to acquire post-discharge effect data, and then it is restored to low-frequency acquisition mode; S3: Data Analysis: Identify the discharge type by combining the discharge feature library and the event data of the entire discharge process; locate the spatial region where the discharge source is located by combining the synchronous data of the physical sensing array and the spatial model; initiate a targeted data request to the confidential library to obtain the core parameters of the device corresponding to the spatial region; S4: Root Cause Analysis: Correlate the core parameters with the discharge feature library and the complete discharge process data collected by the physical sensing array to reverse-engineer the deviation of physical parameters that caused the discharge; generate a discharge event analysis report.

[0008] As a preferred embodiment of the partial discharge pattern recognition method based on attention mechanism and convolutional neural network described in this invention, the physical sensing array in step S1 environment construction includes: a gas component analyzer, an ultraviolet imager, an ozone sensor, a high-frequency vibration sensor, and a broadband radio frequency antenna for monitoring the early warning features and the critical features; a high-frequency electromagnetic probe for capturing electromagnetic signals during the discharge process; and an ultrasonic stress wave sensor for capturing acoustic signals.

[0009] As a preferred embodiment of the partial discharge pattern recognition method based on attention mechanism and convolutional neural network described in this invention, wherein: the discharge feature library in step S1 environment construction is a structure divided according to the time processing flow, specifically including: A warning feature set storing weak signal characteristics that characterize the initial deviation of the insulation state; A critical feature set containing critical thresholds for various monitoring parameters that indicate an impending discharge; A set of discharge event features that defines experimentally calibrated, standardized waveform characteristics associated with a specific discharge type; A set of aftereffect characteristics that clarifies the verification benchmark corresponding to various aftereffect data when the physical parameter deviation is reverse-engineered.

[0010] As a preferred embodiment of the partial discharge pattern recognition method based on attention mechanism and convolutional neural network described in this invention, the early warning feature set is classified and stored according to the estimated discharge type, specifically including: A safety threshold for hydrogen concentration, specific to internal discharge and used for comparison with actual data collected by the gas composition analyzer; The high-frequency vibration sensor identifies abnormal spectral components that differ significantly from the stored vibration data inherent to the device's operation, indicating potential discharge of floating potentials. The specific frequency band spectrum rise of the radio frequency background noise detected by the broadband radio frequency antenna reflects that the corona discharge is in the early stage; The composite characteristics of the early stage of surface discharge include the appearance of stable ultraviolet spots in the image captured by the ultraviolet imager, and the ozone sensor detecting that the local ozone concentration on the device surface exceeds the safety threshold.

[0011] As a preferred embodiment of the partial discharge pattern recognition method based on attention mechanism and convolutional neural network described in this invention, the method includes: after identifying warning features, tracking the evolution trend of the identified warning features, and in subsequent mid-frequency monitoring, including other subsequently identified warning features in the tracking range; the critical feature set defines the triggering conditions for increasing the monitoring frequency from mid-frequency to high-frequency, and the triggering condition is satisfied when any feature in the tracking range first meets its corresponding critical criterion; the critical feature set specifically includes: The threshold values ​​for the growth rate of hydrogen concentration, the threshold values ​​for the energy accumulation rate of the abnormal spectral components, the threshold values ​​for the growth slope of the signal amplitude of the radio frequency spectrum rise, the threshold values ​​for the signal intensity change rate and spatial coverage expansion rate of ozone concentration, and the threshold values ​​for the signal intensity change rate and spatial coverage expansion rate of the ultraviolet spot.

[0012] As a preferred embodiment of the partial discharge pattern recognition method based on attention mechanism and convolutional neural network described in this invention, during the event data acquisition of the entire discharge process, after the first critical feature is identified, all early warning features and critical features will continue to be identified in parallel. Furthermore, the identification of the discharge type in step S3 data analysis specifically includes: combining the critical feature set in the discharge feature library, pairing the critical features successfully identified in the discharge process event data, and then defining the set of all possible discharge types as a candidate discharge type set; then, processing the electromagnetic wave data collected by the high-frequency electromagnetic probe and the stress wave data collected by the ultrasonic stress wave sensor in the discharge event data to form a multimodal discharge feature set; by comparing the similarity between the multimodal discharge feature set and the standardized features in the discharge feature library that belong to the candidate discharge type, and combining the verification results based on the post-discharge effect data, determining the final discharge type in the candidate discharge type set.

[0013] As a preferred embodiment of the partial discharge pattern recognition method based on attention mechanism and convolutional neural network described in this invention, the identification of the spatial region where the discharge occurs in step S3 data analysis specifically includes: taking the earliest signal start time of the earliest recorded discharge data from the high-frequency electromagnetic probe as the zero time of the discharge event; calculating the absolute propagation time of the sound wave generated by the discharge to each ultrasonic stress wave sensor based on the zero time of the discharge event and in combination with the synchronous data collected by the ultrasonic stress wave sensor; and determining the spatial region where the discharge source is located through time difference positioning analysis.

[0014] As a preferred embodiment of the partial discharge pattern recognition method based on attention mechanism and convolutional neural network described in this invention, the process of deducing the physical parameter deviation of the discharge in step S4 specifically includes: analyzing the pulse amplitude and rise time of the electromagnetic wave signal in the discharge event data to obtain the energy level and discharge medium state of the current discharge event; using the energy level and the medium state as initial conditions, simulating the sound wave propagation process in a local physical model composed of the obtained core parameters of the device to generate a theoretical acoustic waveform under the initial conditions; comparing the theoretical acoustic waveform with the actual acoustic waveform in the discharge event data to identify the differences between the two in arrival time, amplitude attenuation, and waveform dispersion; iteratively adjusting the physical parameters until the deviation between the simulated acoustic waveform and the actual acoustic waveform is less than a preset value, and determining the finally adjusted physical parameters as the physical parameter deviation causing the discharge.

[0015] As a preferred embodiment of the partial discharge pattern recognition method based on attention mechanism and convolutional neural network described in this invention, the discharge event analysis report is an interactive four-dimensional simulation scene, presented as a three-dimensional view that can be observed and zoomed from multiple perspectives; the interactive four-dimensional simulation scene has a timeline that can be dragged arbitrarily, used to review the discharge process and to conduct a detailed examination of the spatiotemporal profile of the discharge process; the interactive four-dimensional simulation scene also has a discharge analysis panel, used to centrally present the final type of the discharge event identified, the physical parameter deviation analysis as the root cause of the discharge, and the risk level assessment and operation and maintenance handling suggestions generated accordingly.

[0016] The present invention also provides a method and apparatus for partial discharge pattern recognition based on attention mechanism and convolutional neural network, used to perform the above method, specifically including: A processor is used to execute program instructions and is responsible for the core logic of calculation, analysis, and decision-making; A memory used to store computer program instructions and temporary data; A data storage unit persistently stores a three-dimensional spatial model, a discharge characteristic library, and a confidential library of core device parameters; A communication interface is used for data communication with the physical sensing array.

[0017] The beneficial effects of this invention are as follows: By constructing a multimodal feature library covering the entire life cycle of discharge and combining it with a dynamic frequency acquisition strategy, accurate tracking from early warning to critical state is achieved. Its core lies in accurately identifying the discharge type and locating the discharge source through multi-physical data collaborative positioning and inversion comprehensive analysis, and deducing the deviation of the equipment's physical parameters causing the discharge. By constructing a closed-loop diagnostic system from signal acquisition and intelligent analysis to root cause tracing, the accuracy and depth of partial discharge diagnosis are significantly improved. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of a partial discharge pattern recognition method based on attention mechanism and convolutional neural network.

[0020] Figure 2 This is a flowchart of the dynamic data acquisition process during discharge.

[0021] Figure 3This is a flowchart for comprehensive discharge diagnosis and source tracing.

[0022] Figure 4 This is a schematic diagram of the functional modules of a partial discharge pattern recognition system based on attention mechanism and convolutional neural network.

[0023] Figure 5 This is a schematic diagram of a partial discharge pattern recognition device based on attention mechanism and convolutional neural network. Detailed Implementation

[0024] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0025] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0026] Secondly, the term "one embodiment" or "example" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the invention. The appearance of an embodiment in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that mutually excludes other embodiments.

[0027] Example 1 Reference Figures 1-3 This is one embodiment of the present invention, which provides a partial discharge pattern recognition method based on attention mechanism and convolutional neural network, including the following steps: S1. Environment Construction: Deploy a physical sensing array in the monitoring environment and construct a three-dimensional spatial model of the monitoring environment; establish an index association between the spatial model and the confidential database of core equipment parameters; calibrate the spatial coordinates of each sensor in the physical sensing array in the spatial model; and establish a discharge characteristic database. Physical sensing arrays include gas component analyzers, ultraviolet imagers, ozone sensors, high-frequency vibration sensors, and broadband radio frequency antennas used to monitor early warning and critical characteristics; high-frequency electromagnetic probes used to capture electromagnetic signals during discharge; and ultrasonic stress wave sensors used to capture acoustic signals. The discharge characteristic library is a structure divided according to the time processing flow, specifically including: A warning feature set storing weak signal characteristics that characterize the initial deviation of the insulation state; A critical feature set containing critical thresholds for various monitoring parameters that indicate an impending discharge; A set of discharge event features that defines experimentally calibrated, standardized waveform characteristics associated with a specific discharge type; A set of aftereffect characteristics that clarifies the verification benchmark corresponding to various aftereffect data when the physical parameter deviation is reversed; The early warning feature set is categorized and stored according to the estimated discharge type, specifically including: A unique safety threshold for hydrogen concentration, specific to internal discharge, used for comparison with actual data collected by a gas composition analyzer; Abnormal spectral components identified by high-frequency vibration sensors that differ significantly from the stored vibration data inherent to device operation and indicate floating potential discharge; The specific frequency band spectrum rise of the radio frequency background noise detected by the broadband radio frequency antenna reflects that the corona discharge is in the early stage; The composite characteristics of the early stage of surface discharge include: a stable ultraviolet spot appears in the image captured by the ultraviolet imager, and the ozone sensor detects that the local ozone concentration on the surface of the device exceeds the safety threshold. Critical feature sets include: hydrogen concentration growth rate threshold, abnormal spectral component energy accumulation rate threshold, radio frequency spectrum rise signal amplitude growth slope threshold, ozone concentration signal intensity change rate threshold and spatial coverage area expansion rate threshold, and ultraviolet spot signal intensity change rate threshold and spatial coverage area expansion rate threshold.

[0028] The process involves first installing sensors at key locations within the monitored indoor high-voltage equipment environment. Multiple high-frequency electromagnetic probes and ultrasonic stress wave sensors are closely mounted on the high-voltage insulation components, bushing roots, and near cavity joints to maximize the capture efficiency of potential discharge signals. An online dissolved gas composition analyzer is integrated into the equipment's insulation medium circulation system. Multiple ozone sensors are then distributed in a grid pattern within the indoor environment. A high-frequency vibration sensor is installed on the equipment's mechanical operating or fastening structures. A broadband RF antenna and a solar-blind ultraviolet imager are positioned in advantageous locations that cover most of the equipment surface. During sensor installation, a three-dimensional spatial positioning tool is used to accurately measure and record the sensors' three-dimensional coordinates in a unified coordinate system anchored to corners of the monitored environment.

[0029] Then, a 3D laser scanner is used to comprehensively scan the entire monitoring environment, generating high-precision 3D point cloud data. In the modeling software, the main body representing the power equipment is separated from the background such as walls and the ground. Subsequently, different functional areas of the equipment are semantically labeled in the 3D point cloud view and assigned unique numbers. This number is then linked to a unique index link with a separate, strictly access-controlled confidential database of the equipment's core parameters. The confidential database pre-stores detailed internal geometric models of the equipment and the physical parameters of various component materials (such as epoxy resin, insulating oil, SF6 gas, and aluminum alloy) under different operating conditions. Finally, the previously recorded 3D coordinates of all sensors are precisely loaded and calibrated to their corresponding positions in this 3D spatial model, forming visualized sensor nodes.

[0030] The root cause of internal discharge is the presence of air gaps or bubbles within the solid or liquid insulating medium, formed due to manufacturing defects or accumulated material damage. Because the dielectric strength of the gas is much lower than that of the surrounding medium, these air gaps will experience electric field strengths far exceeding normal levels under the influence of an electric field. Even in the very early stages before the air gaps have fully formed stable breakdown, the localized high electric field is sufficient to cause the chemical bonds in the insulating oil or solid insulating material to break, producing trace amounts of hydrogen gas. Therefore, monitoring the slow, continuous increase in hydrogen concentration is an effective means of capturing the earliest nascent state of internal discharge. When the discharge activity inside the air gap intensifies, it accelerates the rate of damage to the surrounding insulating material. At this point, the hydrogen production rate will significantly increase, indicating that insulation degradation has entered an accelerated phase and discharge is imminent. Because the discharge is encased within the insulating medium, the resulting electromagnetic pulse signal is usually quite steep and appears as regularly shaped discharge clusters symmetrically distributed across the positive and negative half-cycles of the power frequency voltage on the phase-resolved partial discharge spectrum. During high-energy internal discharge, a large amount of characteristic gas (such as acetylene C2H2) is generated instantaneously. Therefore, detecting a sharp pulse of gas concentration, distinct from a slowly increasing one, after a discharge event using a gas composition analyzer is strong evidence to verify high-energy internal discharge.

[0031] Contamination of insulator surfaces due to moisture, dust accumulation, etc., can lead to excessively large tangential electric fields, eventually forming sliding or crawling surface discharges at the gas-solid interface. Even extremely weak surface micro-discharges or corona discharges can ionize the surrounding air to produce ozone (O3), which emits light in the ultraviolet band. When surface contamination worsens or humidity increases, discharge activity becomes more intense and the affected area expands, leading to a rapid increase in ozone concentration or a rapid spread of ultraviolet spots accompanied by increased brightness. Its waveform characteristics are usually strongly correlated with the phase of the power frequency voltage, often appearing as asymmetrical, comet-shaped or rabbit-ear-shaped discharge clusters concentrated only in a certain polarity or phase range on phase-resolved partial discharge maps. The surface discharge channel generates Joule heating, leaving a temporary hot spot on the insulator surface. By capturing the cooling process of this hot spot after the discharge ends using an infrared thermal imager, the energy level and severity of the surface discharge can be verified in reverse.

[0032] Near high-voltage conductors with extremely small radii of curvature (such as wire burrs or metal tips), electric field lines are highly concentrated, causing the local electric field strength to exceed the breakdown threshold of air, resulting in a self-sustaining corona discharge on the conductor surface. Corona discharge does not form a complete breakdown channel but continuously generates electromagnetic interference and chemical byproducts. The plasma oscillations during corona discharge produce a relatively stable and continuous broadband radio frequency radiation. The earliest and most sensitive electrical indicator of a stable corona source is the presence of persistent bulges or rises in background noise at a certain frequency band (e.g., tens to hundreds of megahertz) that cannot be attributed to external interference, detected by a broadband radio frequency antenna in the absence of obvious pulse signals. Increased equipment voltage, increased air humidity, or decreased air pressure all lead to a decrease in the corona initiation voltage or an increase in intensity, which directly reflects an increase in radio frequency radiation power. Therefore, a rapid increase in the magnitude of the spectral rise within a short period indicates that the corona is developing into a more severe state, and may even transform into a more destructive spark discharge. Its typical electrical pulse signal manifests as numerous, dense, low-amplitude pulse clusters with extremely high repetition rates. These pulses usually only appear in a single polarity half-cycle of the power frequency voltage. On a phase-resolved partial discharge spectrum, they form a nearly vertical stripe distribution close to the longitudinal axis. Corona discharge is a relatively steady-state discharge, so its aftereffects are not obvious; only its accompanying effects can be observed. If a radio frequency antenna detects corona characteristics, and simultaneously an ultraviolet imager continuously captures an ultraviolet spot at a fixed location, and the ozone sensor reading near that location is high, this high spatial overlap of multiple physical phenomena is decisive evidence confirming the presence of corona discharge at that location.

[0033] Inside the equipment, a suspended body is repeatedly charged and discharged in an alternating electric field due to capacitive coupling. When the electric field strength between it and surrounding components at different potentials is sufficiently high, spark discharge, also known as suspended potential discharge, occurs. The formation of suspended bodies often stems from loosening or fracture of mechanical structures. Before electrical discharge occurs, this mechanical defect alters the overall vibration characteristics of the equipment. The presence of new, unexplained resonance peaks in the vibration spectrum detected by high-frequency vibration sensors is the earliest clue to tracing the mechanical causes of suspended potential discharge. As the loosening of a component worsens, its displacement and impact energy under normal power frequency vibration increase. Monitoring a rapid accumulation or increase in the energy (or square of the amplitude) of this abnormal vibration frequency component within a short period indicates that the mechanical defect is rapidly deteriorating, and the suspended body is about to reach a critical state capable of generating stable electrical discharge. Because the suspended body is charged with opposite charges during the positive and negative half-cycles, it discharges to both conductors (possibly the high-voltage end or the ground end), resulting in two similarly shaped discharge clusters symmetrical about the voltage zero point on the spectrum. The position and orientation of the suspended body will constantly change slightly with mechanical vibration. Its discharge gap is not fixed. After a dense discharge activity ends, random pulses (acoustic or electrical signals) with extremely low energy and completely irregular time intervals may continue to appear. This high degree of randomness is strong evidence that it is different from other types of discharge, and can be used to verify that the event was caused by a suspended body with a non-fixed and randomly positioned position.

[0034] Organizing the discharge feature library according to a structure that aligns with the chronological order of the diagnostic process—including early warning feature sets, critical feature sets, discharge event feature sets, and after-effect feature sets—offers significant technical advantages. This structure enables the system to achieve highly intelligent and efficient data retrieval during diagnostic tasks. In the low-frequency monitoring phase of routine inspections, the system only needs to load and compare the computationally minimal early warning feature sets. After identifying an early warning and transitioning to mid-frequency monitoring, the focus shifts to loading and comparing critical feature sets to assess the rate of risk evolution. Only after confirming an impending discharge and switching to high-frequency capture does the system invoke the most computationally complex discharge event and after-effect feature sets for in-depth type identification and root cause analysis. This on-demand, phased loading mechanism not only significantly reduces the system's computational resource consumption during daily monitoring but, more importantly, ensures complete synchronization between the entire diagnostic process and the physical development of the discharge phenomenon. This guarantees the use of the right tools and the analysis of the right data at the right time, thereby greatly improving the real-time performance, accuracy, and logical rigor of the diagnosis.

[0035] S2. Data Acquisition: The physical sensing array acquires low-frequency data, combines it with the discharge feature library to identify early warning features in the low-frequency data, and then increases the acquisition frequency of the physical sensing array to medium frequency; combining it with the discharge feature library to identify critical features in the medium-frequency data, and then increases the acquisition frequency of the physical sensing array from medium frequency to high frequency; after capturing discharge event data, the acquisition frequency of the physical sensing array is reduced to medium frequency to acquire post-discharge effect data, and then it is restored to low-frequency acquisition mode. After identifying the warning features, the evolution trend of the identified warning features is tracked, and other warning features identified later are also included in the tracking range during the subsequent mid-frequency monitoring process; the critical feature set defines the triggering conditions for upgrading the monitoring frequency from mid-frequency to high-frequency. The triggering condition is met when any feature in the tracking range first meets its corresponding critical criterion. During the data acquisition of the entire discharge process, after the first critical feature is identified, all early warning features and critical features will continue to be identified in parallel.

[0036] The core of this mechanism lies in its intelligent and dynamic matching of resource allocation (such as computing power, storage bandwidth, and sensor power consumption) with the physical stages of fault development. When most devices are in a stable period, routine inspections are performed with minimal resource consumption. Once any weak early warning signs are identified, the monitoring level is automatically escalated, and all potential anomalies are tracked in parallel to ensure no possibility of fault evolution is overlooked. Only when a discharge risk is deemed imminent will all resources be mobilized for high-precision transient event capture. This scheduling strategy greatly optimizes operational efficiency and constructs a seamless data chain throughout the entire discharge process, providing comprehensive data support covering the entire discharge lifecycle for subsequent high-reliability diagnostic analysis.

[0037] S3. Data Analysis: Identify the discharge type by combining the discharge feature library and the event data of the entire discharge process; locate the spatial region where the power source is located by combining the synchronous data of the physical sensing array and the spatial model; and send a targeted data request to the confidential library to obtain the core parameters of the equipment corresponding to the spatial region. The discharge type is identified, and the critical feature set in the discharge feature library is used to pair the critical features successfully identified in the event data of the entire discharge process. Then, the set of all possible discharge types is defined as the candidate discharge type set. Then, the electromagnetic wave data collected by the high-frequency electromagnetic probe and the stress wave data collected by the ultrasonic stress wave sensor in the discharge event data are processed separately to form a multimodal discharge feature set. By comparing the similarity between the multimodal discharge feature set and the standardized features belonging to the candidate discharge types in the discharge feature library, and combining the verification results based on the post-discharge effect data, the final discharge type is determined in the candidate discharge type set. The spatial region where the discharge occurs is identified. The signal start time of the earliest recorded discharge data from the high-frequency electromagnetic probe is taken as the zero time of the discharge event. Based on the zero time of the discharge event and combined with the synchronous data collected by the ultrasonic stress wave sensor, the absolute propagation time of the sound wave generated by the discharge to each ultrasonic stress wave sensor is calculated. The spatial region where the discharge source is located is determined through time difference positioning analysis.

[0038] The system analyzes the event data of the entire discharge process, associates and matches all successfully identified critical features with the critical feature set in the discharge feature library, thereby quickly defining a set of candidate discharge types that includes all discharge types that meet the conditions, and then retrieves the feature data of the corresponding discharge type from the discharge feature library.

[0039] First, feature vector sets of electromagnetic wave data and acoustic wave data are extracted using Mel frequency cepstral coefficients, and then similarity matching is performed between them and the feature vectors in the feature data.

[0040] By performing a Mel-frequency cepstral coefficient extraction process on the discharge data of each sensor, a feature vector that accurately describes its spectral characteristics can be obtained. This process specifically includes the following steps: a. Pre-emphasis: Applying a high-pass filter to the original digital signal to boost the energy of the high-frequency components and compensate for the natural attenuation of the high-frequency energy.

[0041] b. Framing and windowing: The continuous data signal is divided into a series of short frames, and a window function is applied to each frame to reduce spectral leakage caused by subsequent Fourier transform.

[0042] c. Fast Fourier Transform: Perform a Fast Fourier Transform on each frame of the windowed signal to transform it from the time domain to the frequency domain, and obtain the spectrum of the signal for that frame.

[0043] d. Mel filtering: The energy of the spectrogram is filtered through a set of triangular filters (i.e., the Mel filter bank). This filter bank is characterized by a dense distribution in the low-frequency region and a sparse distribution in the high-frequency region. The output of this step is the logarithmic energy spectrum of each frame.

[0044] e. Discrete Cosine Transform: Perform a discrete cosine transform on the logarithmic energy spectrum of each frame. Since adjacent energy values ​​in the energy spectrum are correlated, the discrete cosine transform can effectively discorrelate these energy values ​​and concentrate their energy on a few coefficients.

[0045] f. Coefficient extraction: Usually, the first 12-13 coefficients of the discrete cosine transform result are taken as the feature vector of the Mel frequency cepstral coefficients of the frame data.

[0046] The comparison of feature vectors mainly uses cosine similarity.

[0047] in, This represents the feature vector extracted from the actual acquired signal; This is a standardized feature vector for a candidate discharge type.

[0048] By comparing the cosine of the angle between the actual vector and the candidate vector in spatial coordinates, the standard feature vector with the highest similarity to the actual vector set can be selected, and a discharge type can be determined in terms of feature vector.

[0049] Phase-resolved partial discharge maps were plotted using electromagnetic wave data and then input into a pre-trained image recognition model based on convolutional neural networks and attention mechanisms. The models were then analyzed for texture, shape, and bright spot distribution patterns, and compared with standard map patterns of preliminarily identified candidate types for in-depth analysis and recognition.

[0050] After a double comparison, the most likely discharge type can be obtained. Then, the verification benchmark corresponding to the initially confirmed discharge type, as defined in the aftereffect feature set, is fitted and verified with the discharge aftereffect data actually collected in this event. When the critical features, feature value similarity, image pattern recognition, and aftereffect verification all point to the same result, the discharge type can be determined as the final identification result for this event.

[0051] The spatial region where the discharge source is located is finally calculated by solving a set of nonlinear equations about the propagation path of sound waves in the medium.

[0052] in,( , , ( ) represents the three-dimensional spatial coordinates of the unknown discharge source to be solved; , , )(in =1,2,..., ) represents the first element that has been precisely calibrated in the three-dimensional spatial model. The three-dimensional spatial coordinates of an ultrasonic stress wave transmitter; It is the speed of sound in air, taken as 340. ; (in =1,2,..., ) is the first The absolute propagation time of an ultrasonic stress wave transmitter.

[0053] Ideally, all equations should be simultaneously satisfied by a single discharge source coordinate. However, actual measurements involve numerous uncertainties, leading to a system of equations without a definite solution. Therefore, a grouping method is adopted, solving for every three sets of equations to obtain a possible discharge source coordinate. After solving for all possible discharge source coordinates, the region where these coordinates are located is the potential discharge source spatial region.

[0054] S4. Root Cause Analysis: Correlate the core parameters with the discharge feature library and the complete discharge process data collected by the physical sensing array to reverse-engineer the deviation of the physical parameters that caused the discharge; generate a discharge event analysis report. The deviation of physical parameters of the discharge is deduced. The pulse amplitude and rise time of the electromagnetic wave signal in the discharge event data are analyzed to obtain the energy level and state of the discharge medium in this discharge event. Using the energy level and state of the medium as initial conditions, the sound wave propagation process is simulated in a local physical model composed of the core parameters of the acquired equipment to generate theoretical acoustic waveforms under the initial conditions. The theoretical acoustic waveforms are compared with the actual acoustic waveforms in the discharge event data to identify the differences between the two in terms of arrival time, amplitude attenuation, and waveform dispersion. The physical parameters are iteratively adjusted until the deviation between the simulated acoustic waveform and the actual acoustic waveform is less than a preset value. The finally adjusted physical parameters are determined as the deviation of physical parameters causing the discharge. The discharge event analysis report is an interactive four-dimensional simulation scene, presented as a three-dimensional view that can be observed and zoomed from multiple perspectives. The interactive four-dimensional simulation scene has a draggable timeline for reviewing the discharge process and for detailed examination of the spatiotemporal profile of the discharge process. The interactive four-dimensional simulation scene also has a discharge analysis panel, which centrally presents the final type of the discharge event identified, the physical parameter deviation analysis as the root cause of the discharge, and the risk level assessment and operation and maintenance handling suggestions generated accordingly.

[0055] Specifically, by analyzing high-frequency electromagnetic wave signals to obtain pulse amplitude and rise time, the energy level (e.g., 500 pC) and discharge medium state (e.g., gas discharge under standard atmospheric pressure) of this discharge event are quickly determined by referring to the discharge feature library, and this is used as the initial conditions for the simulation process.

[0056] After receiving the core parameters of the device corresponding to the power supply area, the system first simulates the sound wave propagation process using initial conditions to generate a theoretical acoustic waveform under healthy conditions. By comparing it with the actual acoustic waveform, the differences between the two in multiple dimensions such as arrival time, amplitude attenuation, and waveform dispersion are quantified.

[0057] Parameters are adjusted based on the aftereffect feature set of the discharge feature library. Priority is given to adjusting the physical parameters most relevant to the observed discrepancies (e.g., adjusting the moisture content of the insulating paperboard if waveform dispersion is large), until a coarse-tuned optimal value is found that minimizes the error between the simulated and actual waveforms in the discrepancies. Then, by fine-tuning all parameters one by one, the simulation parameters are continuously converged to the actual data, ultimately finding a set of optimal multi-parameter solutions that maximizes the overall similarity between the simulated and actual waveforms across all dimensions (arrival time, amplitude attenuation, waveform dispersion, etc.). Even if the final result is not a perfect match, it represents the theoretically closest physical state obtainable under the current model. This finalized set of parameters is considered the fundamental physical parameter deviation causing the discharge.

[0058] in, Represents the moment The amplitude of the acoustic waveform generated by the digital twin model simulation; Represents the same moment The amplitude of the acoustic waveform actually acquired by the sensor; calculation results It is the root mean square error of the two sets of waves. The smaller the value, the higher the similarity between the two waveforms.

[0059] The core of the discharge event analysis report is an interactive four-dimensional simulation scenario. Within a simplified three-dimensional view, it recapsulates the entire discharge process through interactive nodes and a draggable timeline. The scenario also includes a discharge analysis panel, which centrally presents the final type of the event identified, as well as the physical parameter deviation analysis derived from the previous steps that serves as the root cause. Furthermore, this panel provides a comprehensive risk level assessment and operational / maintenance recommendations. The risk assessment considers factors such as: the energy level of the discharge (higher energy, greater risk), the inherent danger of the identified discharge type (e.g., internal discharge is generally more dangerous than corona discharge), the degree of deviation in derived physical parameters (e.g., the risk is significantly different between exceeding the standard moisture content by 10% and exceeding it by 50%), and the failure development speed recorded during the evolution trend tracking phase (faster development, higher risk). Based on these comprehensive factors, the system provides a final, highly valuable operational recommendation.

[0060] in, It is a normalized discharge energy level; It is the inherent hazard factor of the discharge type (e.g., 1.0 for internal discharge, 0.7 for surface discharge, and 0.2 for corona discharge). This refers to the relative severity of the deviation in physical parameters (e.g., (measured moisture content - standard moisture content) / standard moisture content). The growth rate, which represents the critical characteristic, has also been normalized. Representing their respective weighting coefficients (e.g.); the calculated The comprehensive risk score representing this discharge will determine the discharge risk level based on the score results.

[0061] Example 2 Reference Figure 4 These are two embodiments of the present invention. This embodiment provides a partial discharge pattern recognition system based on an attention mechanism and a convolutional neural network. The system is typically deployed on a high-performance computing server to execute the method steps in Embodiment 1.

[0062] The system's software architecture consists of a set of highly collaborative functional modules, specifically including: Time synchronization module: Provides a unified time reference for the sensors in the physical sensing array, and adds high-precision timestamps to the collected multi-source data to generate a multi-source data stream that can be processed synchronously; Discharge phase identification module: Combines discharge feature library to analyze multi-source data streams, identify early warning features, critical features, and discharge process data; adjusts the acquisition frequency of physical sensing array according to the identification results; records all data from the appearance of the first critical feature to the end of the post-discharge effect acquisition to form full discharge process event data; Candidate type screening module: Combining the discharge feature library, it analyzes all successfully identified critical features in the event data of the entire discharge process, and establishes a candidate discharge type set based on the identification results; Data feature extraction module: Extracts feature values ​​of electromagnetic wave signals from high-frequency electromagnetic probes at each location to establish a discharge electromagnetic feature set; extracts feature values ​​of stress wave signals from ultrasonic stress wave sensors at each location to establish a discharge stress feature set. Discharge identification module: Receives a set of candidate discharge types, then obtains the electromagnetic and stress features of the candidate discharge types from the discharge feature library, compares them with the discharge electromagnetic feature set and the discharge stress feature set, and determines the discharge type to be verified with the highest matching degree; Post-effect verification module: Combining the post-discharge effect data in the event data of the entire discharge process, performing post-effect feature matching on the discharge type to be verified, and finally determining the discharge type; Collaborative positioning module: Receives electromagnetic wave signals and stress wave signals from the discharge process, which are synchronously processed by the time synchronization module, calculates the spatial region where the discharge source is located, and initiates a targeted data request to the equipment core parameter confidential database based on the spatial region to obtain the equipment core parameters of the spatial region; The physical inversion module receives the determined discharge type, spatial region, and core equipment parameters, and iteratively corrects and inversely deduces the deviations in the physical parameters that caused the discharge. Report generation module: Receives the determined discharge type, spatial region, and physical parameter deviations, performs comprehensive analysis, and generates a discharge event analysis report.

[0063] Upon system startup, the time synchronization module is activated first. This module uses a high-precision internal clock source and a precise time protocol (such as IEEE 1588) to distribute a unified time reference to each sensor in the physical sensing array, instructing all sensors to append a high-precision timestamp to each data point they collect. In this way, all subsequent data entering the system carries a unified time coordinate, forming a multi-source data stream that can be precisely synchronized.

[0064] The discharge phase identification module begins receiving and processing these multi-source data streams in a polling manner. In the system's low-frequency inspection mode, this module continuously calls the early warning feature set in the discharge feature library to compare the data. Once an early warning feature is identified, it immediately sends an up-frequency command to the physical sensing array, putting it into mid-frequency monitoring mode. In mid-frequency mode, the module then calls the critical feature set for comparison. When any tracked feature meets the critical criteria, the module again sends an up-frequency command to the high-frequency event capture mode. After capturing the discharge event and completing the mid-frequency aftereffect acquisition, the module packages all the data recorded during this period into a complete "discharge process event data" file and distributes it to the subsequent analysis modules.

[0065] The candidate type filtering module parses all successfully identified critical feature records in the data and associates them with the discharge feature library. It extracts the possible discharge types corresponding to each critical feature and finally summarizes them into a set of candidate discharge types containing all possibilities, which is then sent to the discharge identification module.

[0066] The data feature extraction module processes the core discharge event portion of the entire process data. It extracts PRPD spectra and waveform feature vectors from the electromagnetic wave signals collected by all high-frequency electromagnetic probes to form a discharge electromagnetic feature set; and extracts waveform feature vectors from the stress wave signals collected by all ultrasonic stress wave sensors to form a discharge stress feature set.

[0067] After receiving the candidate discharge type set, discharge electromagnetic feature set, and discharge stress feature set, the discharge identification module begins preliminary identification. It sequentially retrieves the standardized electromagnetic and stress features corresponding to each discharge type in the candidate set from the discharge feature database. Then, it calculates the matching degree between the actual feature set and the standardized feature set using a similarity algorithm (such as cosine similarity). Finally, the module selects the discharge type with the highest overall matching degree as a "discharge type to be verified" and passes it to the post-effect verification module.

[0068] Upon receiving the discharge type to be verified, the aftereffect verification module retrieves the corresponding aftereffect verification benchmark (e.g., a specific temperature cooling curve model) from the discharge feature library and fits it with the actual aftereffect data in the full-process data. Only when the fit exceeds a preset confidence threshold will the module finally confirm the discharge type and output it as the final identification result.

[0069] In parallel, the collaborative positioning module first identifies the earliest signal start point from the data of the high-frequency electromagnetic probe as the global discharge event zero moment. Then, based on this zero moment, it calculates the absolute propagation time of the sound wave reaching each ultrasonic stress wave sensor and solves a multi-point positioning equation system to determine the spatial region of the discharge source in the three-dimensional spatial model. After obtaining the positioning results, the module immediately initiates a directional data request to the equipment's core parameter confidential database to obtain detailed equipment core parameters for that spatial region.

[0070] After receiving the final determined discharge type, discharge source space region, and acquired core equipment parameters, the physical inversion module initiates the root cause tracing process. It uses a two-stage iterative correction algorithm based on a digital twin model to reverse-engineer a set of physical parameters that best reproduce the actual signal and indicate any deviations.

[0071] The report generation module gathers the final discharge type, discharge source spatial region, and deduced physical parameter deviations from all the aforementioned analysis modules. After conducting a comprehensive risk assessment, it constructs and outputs a final discharge event analysis report that includes an interactive four-dimensional simulation scenario and a discharge analysis panel. This completes a full diagnostic process.

[0072] Example 3 Reference Figure 5 These are three embodiments of the present invention. This embodiment provides a partial discharge pattern recognition device based on an attention mechanism and a convolutional neural network. This device is used to execute the method described in Embodiment 1 and serves as a carrier of the system described in Embodiment 2.

[0073] Specifically, it includes: A processor is used to execute program instructions and is responsible for the core logic of calculation, analysis, and decision-making; A memory used to store computer program instructions and temporary data; A data storage unit persistently stores a three-dimensional spatial model, a discharge characteristic library, and a confidential library of core device parameters; A communication interface is used for data communication with the physical sensing array.

[0074] The processor executes computer program instructions stored in memory to implement all the core computational, analytical, and decision-making logic of the method of this invention. Specifically, the processor is responsible for executing the logic of analyzing multi-source data streams to identify early warning characteristics and critical characteristics, the algorithm of processing discharge event data to identify discharge types and locate discharge sources, and the complex calculation of iterative correction based on a digital twin model to reverse-engineer physical parameter deviations.

[0075] The memory is configured to provide high-speed read and write space while the processor executes program instructions, for temporarily storing intermediate data processed by each module, including real-time acquisition data being compared, candidate discharge type sets, and simulation waveform data during iterative correction.

[0076] The data storage unit is used to persistently store the core knowledge and model library upon which this method depends for operation. Specifically, the data storage unit is divided into different storage areas to store the three-dimensional spatial model that serves as the benchmark for all spatial analyses, the discharge feature library that embeds complete diagnostic knowledge logic, and the confidential library of core device parameters used for in-depth root cause tracing.

[0077] The sensing interface serves as the channel for this device to interact with the external physical world. This interface is responsible for bidirectional data communication with the physical sensing array deployed in the monitoring environment. Specifically, it receives multi-source raw signals collected by the array in real time and appends high-precision timestamps, then transmits these signals to the processor for analysis. Simultaneously, it accurately sends the acquisition frequency adjustment commands generated by the processor based on diagnostic logic to the corresponding sensors in the array. Furthermore, it is responsible for presenting discharge event analysis reports to the interactive terminal (display).

[0078] Those skilled in the art will understand that the memory can be random access memory, while the data storage unit can be a non-volatile storage medium such as a hard disk, solid-state drive, or network-attached storage. Both can also be different logical partitions integrated on the same physical chip or circuit board.

[0079] In summary, by constructing a multimodal feature library covering the entire lifecycle of discharge and combining it with a dynamic frequency acquisition strategy, accurate tracking from early warning to critical states was achieved. The core of this approach lies in using multi-physical data collaborative positioning and inversion comprehensive analysis to accurately identify the discharge type, locate the discharge source, and inversely deduce the deviations in the physical parameters of the equipment causing the discharge. By constructing a closed-loop diagnostic system from signal acquisition and intelligent analysis to root cause tracing, the accuracy and depth of partial discharge diagnosis were significantly improved.

[0080] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A partial discharge pattern recognition method based on attention mechanism and convolutional neural network, characterized in that, Performed by a computer device, including the following steps: S1: Environment Construction: Deploy a physical sensing array in the monitoring environment and construct a three-dimensional spatial model of the monitoring environment; establish an index association between the spatial model and the confidential database of core parameters of the equipment, and calibrate the spatial coordinates of each sensor in the physical sensing array in the spatial model; establish a discharge feature database. S2: Data Acquisition: The physical sensing array acquires low-frequency data, combines it with the discharge feature library to identify early warning features in the low-frequency acquired data, and then increases the acquisition frequency of the physical sensing array to medium frequency; combines it with the discharge feature library to identify critical features in the medium-frequency acquired data, and then increases the acquisition frequency of the physical sensing array from medium frequency to high frequency; after capturing discharge event data, the acquisition frequency of the physical sensing array is reduced to medium frequency to acquire post-discharge effect data, and then it is restored to low-frequency acquisition mode; S3: Data Analysis: Identify the discharge type by combining the discharge feature library and the event data of the entire discharge process; locate the spatial region where the discharge source is located by combining the synchronous data of the physical sensing array and the spatial model; initiate a targeted data request to the confidential library to obtain the core parameters of the device corresponding to the spatial region; S4: Root Cause Analysis: Correlate the core parameters with the discharge feature library and the complete discharge process data collected by the physical sensing array to reverse-engineer the deviation of physical parameters that caused the discharge; generate a discharge event analysis report.

2. The partial discharge pattern recognition method based on attention mechanism and convolutional neural network according to claim 1, characterized in that, The physical sensing array in step S1 environment construction includes: a gas component analyzer, an ultraviolet imager, an ozone sensor, a high-frequency vibration sensor, and a broadband radio frequency antenna for monitoring the early warning features and the critical features; a high-frequency electromagnetic probe for capturing electromagnetic signals during the discharge process; and an ultrasonic stress wave sensor for capturing acoustic signals.

3. The partial discharge pattern recognition method based on attention mechanism and convolutional neural network according to claim 2, characterized in that, The discharge feature library mentioned in step S1, which involves environment construction, is a structure divided according to the time processing flow, specifically including: A warning feature set storing weak signal characteristics that characterize the initial deviation of the insulation state; A critical feature set containing critical thresholds for various monitoring parameters that indicate an impending discharge; A set of discharge event features that defines experimentally calibrated, standardized waveform characteristics associated with a specific discharge type; A set of aftereffect characteristics that clarifies the verification benchmark corresponding to various aftereffect data when the physical parameter deviation is reverse-engineered.

4. The partial discharge pattern recognition method based on attention mechanism and convolutional neural network according to claim 3, characterized in that, The early warning feature set is classified and stored according to the estimated discharge type, specifically including: A safety threshold for hydrogen concentration, specific to internal discharge and used for comparison with actual data collected by the gas composition analyzer; The high-frequency vibration sensor identifies abnormal spectral components that differ significantly from the stored vibration data inherent to the device's operation, indicating potential discharge of floating potentials. The specific frequency band spectrum rise of the radio frequency background noise detected by the broadband radio frequency antenna reflects that the corona discharge is in the early stage; The composite characteristics of the early stage of surface discharge include the appearance of stable ultraviolet spots in the image captured by the ultraviolet imager, and the ozone sensor detecting that the local ozone concentration on the device surface exceeds the safety threshold.

5. The partial discharge pattern recognition method based on attention mechanism and convolutional neural network according to claim 4, characterized in that, After identifying the warning features, the evolution trend of the identified warning features is tracked, and other subsequently identified warning features are also included in the tracking range during the intermediate frequency monitoring process. The critical feature set defines the triggering conditions for upgrading the monitoring frequency from intermediate frequency to high frequency. The triggering condition is met when any feature within the tracking range first satisfies its corresponding critical criterion. The critical feature set specifically includes: The threshold values ​​for the growth rate of hydrogen concentration, the threshold values ​​for the energy accumulation rate of the abnormal spectral components, the threshold values ​​for the growth slope of the signal amplitude of the radio frequency spectrum rise, the threshold values ​​for the signal intensity change rate and spatial coverage expansion rate of ozone concentration, and the threshold values ​​for the signal intensity change rate and spatial coverage expansion rate of the ultraviolet spot.

6. The partial discharge pattern recognition method based on attention mechanism and convolutional neural network according to claim 5, characterized in that, During the data acquisition of the entire discharge process, after the first critical feature is identified, all early warning features and critical features will continue to be identified in parallel. Furthermore, the identification of the discharge type in step S3 data analysis specifically includes: combining the critical feature set in the discharge feature library, pairing the critical features successfully identified in the discharge process event data, and then defining the set of all possible discharge types as a candidate discharge type set; then, processing the electromagnetic wave data collected by the high-frequency electromagnetic probe and the stress wave data collected by the ultrasonic stress wave sensor in the discharge event data to form a multimodal discharge feature set; by comparing the similarity between the multimodal discharge feature set and the standardized features in the discharge feature library that belong to the candidate discharge type, and combining the verification results based on the post-discharge effect data, determining the final discharge type in the candidate discharge type set.

7. The partial discharge pattern recognition method based on attention mechanism and convolutional neural network according to claim 6, characterized in that, The identification of the spatial region where the discharge occurs in step S3 data analysis specifically includes: taking the earliest signal start time of the earliest high-frequency electromagnetic probe that recorded the discharge data as the zero time of the discharge event; calculating the absolute propagation time of the sound wave generated by the discharge to each ultrasonic stress wave sensor based on the zero time of the discharge event and in combination with the synchronous data collected by the ultrasonic stress wave sensor; and determining the spatial region where the discharge source is located through time difference positioning analysis.

8. The partial discharge pattern recognition method based on attention mechanism and convolutional neural network according to claim 7, characterized in that, The process of deriving the physical parameter deviation of the discharge in step S4 specifically includes: analyzing the pulse amplitude and rise time of the electromagnetic wave signal in the discharge event data to obtain the energy level and discharge medium state of this discharge event; using the energy level and medium state as initial conditions, simulating the sound wave propagation process in a local physical model composed of the obtained core parameters of the equipment to generate a theoretical acoustic waveform under the initial conditions; comparing the theoretical acoustic waveform with the actual acoustic waveform in the discharge event data to identify the differences between the two in arrival time, amplitude attenuation, and waveform dispersion; and iteratively adjusting the physical parameters until the deviation between the simulated acoustic waveform and the actual acoustic waveform is less than a preset value, and finally determining the adjusted physical parameters as the physical parameter deviation causing the discharge.

9. The partial discharge pattern recognition method based on attention mechanism and convolutional neural network according to claim 8, characterized in that, The discharge event analysis report is an interactive four-dimensional simulation scene, presented as a three-dimensional view that can be observed and zoomed from multiple perspectives. The interactive four-dimensional simulation scene has a draggable timeline for reviewing the discharge process and for detailed examination of the spatiotemporal profile of the discharge process. The interactive four-dimensional simulation scene also has a discharge analysis panel, which centrally presents the final type of the discharge event identified, the physical parameter deviation analysis as the root cause of the discharge, and the risk level assessment and operation and maintenance suggestions generated accordingly.

10. A partial discharge pattern recognition device based on attention mechanism and convolutional neural network, characterized in that, The apparatus is used to perform the method according to any one of claims 1 to 9, specifically comprising: A processor is used to execute program instructions and is responsible for the core logic of calculation, analysis, and decision-making; A memory used to store computer program instructions and temporary data; A data storage unit persistently stores a three-dimensional spatial model, a discharge characteristic library, and a confidential library of core device parameters; A communication interface is used for data communication with the physical sensing array.

Images