Fault tracing method and device of transformer, computer equipment, readable storage medium and program product

Abstract

The invention relates to a fault tracing method and device of a transformer, computer equipment, a computer readable storage medium and a computer program product. The method comprises the following steps: acquiring physical monitoring data and chemical monitoring data of a transformer; when the physical monitoring data or the chemical monitoring data indicate that the transformer has a fault event, performing fault positioning by using the physical monitoring data and the chemical monitoring data to obtain a fault positioning result; fusing the physical monitoring data and the chemical monitoring data by using an attention mechanism to obtain cross-domain fusion features of the fault event; matching the cross-domain fusion features with a fault knowledge graph to obtain a candidate fault set corresponding to the fault event; obtaining the confidence of each candidate fault in the candidate fault set according to the cross-domain fusion features by using a trained fault recognition model; and obtaining a fault traceability result of the fault event according to the fault positioning result and the confidence coefficient of each candidate fault. By adopting the method, efficient and accurate traceability of the transformer fault can be realized.

Description

Technical Field

[0001] This application relates to the field of power safety technology, and in particular to a method, apparatus, computer equipment, computer-readable storage medium, and computer program product for tracing the fault source of a transformer. Background Technology

[0002] In transformer fault diagnosis, multiple parameters, such as dissolved gas analysis in oil (DGA) and partial discharge data, are typically analyzed independently, and the fault location process is often separated from the fault tracing analysis process. Consequently, when tracing transformer faults, related technologies struggle to comprehensively consider the correlations between different parameters and effectively combine fault characteristics with specific fault locations. This limits the accuracy and efficiency of fault tracing results, hindering efficient and accurate fault tracing of transformers. Summary of the Invention

[0003] Therefore, it is necessary to provide a method, apparatus, computer equipment, computer-readable storage medium, and computer program product for tracing the source of transformer faults in response to the above-mentioned technical problems.

[0004] Firstly, this application provides a method for tracing the source of transformer faults, including:

[0005] Obtain physical and chemical monitoring data of the transformer;

[0006] When the physical monitoring data or the chemical monitoring data indicates that a fault event has occurred in the transformer, the physical monitoring data and the chemical monitoring data are used to locate the fault in the transformer to obtain the fault location result of the fault event;

[0007] The physical monitoring data and the chemical monitoring data are fused using an attention mechanism to obtain the cross-domain fusion features of the fault event;

[0008] The cross-domain fusion features are matched with the fault knowledge graph to obtain the candidate fault set corresponding to the fault event;

[0009] Using the trained fault identification model, the confidence level of each candidate fault in the candidate fault set is obtained based on the cross-domain fusion features;

[0010] Based on the fault location results and the confidence levels of each candidate fault, the fault source tracing results of the fault event are obtained.

[0011] In one embodiment, when the physical monitoring data or the chemical monitoring data indicates that the transformer has experienced a fault event, the step of using the physical monitoring data and the chemical monitoring data to locate the fault in the transformer and obtain the fault location result of the fault event includes: when the physical monitoring data or the chemical monitoring data indicates that the transformer has experienced a fault event, obtaining the fault type corresponding to the fault event based on the physical monitoring data and the chemical monitoring data; and using the fault location method corresponding to the fault type, obtaining the fault location result corresponding to the fault event based on the physical monitoring data.

[0012] In one embodiment, the physical monitoring data includes waveform monitoring data; before using the physical monitoring data and the chemical monitoring data to locate the transformer fault when the physical monitoring data or the chemical monitoring data indicates a transformer fault, and obtaining the transformer fault location result, the method includes: when a pulse signal appears in the waveform monitoring data, calculating the signal source location of the pulse signal based on the location of each sensor corresponding to the waveform monitoring data, the reception time of each sensor for the pulse signal, and the default medium parameters of the transformer; updating the reception time of each sensor for the pulse signal based on the signal source location, the location of each sensor, and the real-time medium parameters of the transformer; and updating the reception time of each sensor for the pulse signal based on the real-time medium parameters of the transformer. The signal source position of the pulse signal is updated using parameters, the positions of each sensor, and the updated reception time. If the distance between the updated signal source position and the previously calculated signal source position is not less than a distance threshold, the step of updating the reception time of each sensor for the pulse signal based on the signal source position, the positions of each sensor, and the real-time dielectric parameters of the transformer is re-executed. If the distance between the updated signal source position and the previously calculated signal source position is less than a distance threshold, the waveform monitoring data collected by each sensor is time-aligned based on the updated reception time of each sensor to obtain aligned waveform monitoring data. The physical monitoring data is updated based on the aligned waveform monitoring data.

[0013] In one embodiment, when the physical monitoring data or the chemical monitoring data indicates that the transformer has experienced a fault event, the step of using the physical monitoring data and the chemical monitoring data to locate the transformer fault and obtain the fault location result of the fault event includes: when the physical monitoring data or the chemical monitoring data indicates that the transformer has experienced a fault event, obtaining the fault type corresponding to the fault event based on the physical monitoring data and the chemical monitoring data; when the fault type indicates that the fault event is associated with a pulse signal appearing in at least one of the waveform monitoring data, obtaining the signal source position of the pulse signal in each of the waveform monitoring data and the position of each of the sensors; obtaining the candidate fault space range of the transformer based on the signal source position; calculating the fault source probability of each candidate fault point within the candidate fault space range based on the real-time dielectric parameters of the transformer, the position of each of the sensors, and the signal source position of the pulse signal; and obtaining the fault location result of the fault event based on the candidate fault point with the highest fault source probability.

[0014] In one embodiment, obtaining the fault tracing result of the fault event based on the fault location result and the confidence level of each of the candidate faults includes: obtaining a high-confidence fault of the transformer based on the confidence level of each of the candidate faults; if the fault location result matches the high-confidence fault, obtaining the fault tracing result of the fault event based on the fault location result and the high-confidence fault; if the fault location result does not match the high-confidence fault, obtaining supplementary fault data of the transformer; and obtaining the fault tracing result of the fault event based on the fault location result, the supplementary fault data, and each of the candidate faults.

[0015] In one embodiment, the trained fault identification model is obtained through the following steps: acquiring a first fault identification model trained using industrial equipment fault samples; adjusting the model parameters of the first fault identification model using transformer fault samples to obtain the trained fault identification model.

[0016] Secondly, this application also provides a fault tracing device for a transformer, comprising:

[0017] The fault location module is used to locate the transformer fault using the physical monitoring data and the chemical monitoring data when the physical monitoring data or the chemical monitoring data indicates that a fault event has occurred in the transformer, and to obtain the fault location result of the fault event.

[0018] The data fusion module is used to fuse the physical monitoring data and the chemical monitoring data using an attention mechanism to obtain the cross-domain fusion features of the fault event;

[0019] The feature matching module is used to match the cross-domain fusion features with the fault knowledge graph to obtain the candidate fault set corresponding to the fault event;

[0020] The fault identification module is used to obtain the confidence level of each candidate fault in the candidate fault set based on the cross-domain fusion features using a trained fault identification model.

[0021] The result acquisition module is used to obtain the fault tracing result of the fault event based on the fault location result and the confidence level of each candidate fault.

[0022] 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:

[0023] Obtain physical and chemical monitoring data of the transformer;

[0024] When the physical monitoring data or the chemical monitoring data indicates that a fault event has occurred in the transformer, the physical monitoring data and the chemical monitoring data are used to locate the fault in the transformer to obtain the fault location result of the fault event;

[0025] The physical monitoring data and the chemical monitoring data are fused using an attention mechanism to obtain the cross-domain fusion features of the fault event;

[0026] The cross-domain fusion features are matched with the fault knowledge graph to obtain the candidate fault set corresponding to the fault event;

[0027] Using the trained fault identification model, the confidence level of each candidate fault in the candidate fault set is obtained based on the cross-domain fusion features;

[0028] Based on the fault location results and the confidence levels of each candidate fault, the fault source tracing results of the fault event are obtained.

[0029] 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:

[0030] Obtain physical and chemical monitoring data of the transformer;

[0031] When the physical monitoring data or the chemical monitoring data indicates that a fault event has occurred in the transformer, the physical monitoring data and the chemical monitoring data are used to locate the fault in the transformer to obtain the fault location result of the fault event;

[0032] The physical monitoring data and the chemical monitoring data are fused using an attention mechanism to obtain the cross-domain fusion features of the fault event;

[0033] The cross-domain fusion features are matched with the fault knowledge graph to obtain the candidate fault set corresponding to the fault event;

[0034] Using the trained fault identification model, the confidence level of each candidate fault in the candidate fault set is obtained based on the cross-domain fusion features;

[0035] Based on the fault location results and the confidence levels of each candidate fault, the fault source tracing results of the fault event are obtained.

[0036] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, performs the following steps:

[0037] Obtain physical and chemical monitoring data of the transformer;

[0038] When the physical monitoring data or the chemical monitoring data indicates that a fault event has occurred in the transformer, the physical monitoring data and the chemical monitoring data are used to locate the fault in the transformer to obtain the fault location result of the fault event;

[0039] The physical monitoring data and the chemical monitoring data are fused using an attention mechanism to obtain the cross-domain fusion features of the fault event;

[0040] The cross-domain fusion features are matched with the fault knowledge graph to obtain the candidate fault set corresponding to the fault event;

[0041] Using the trained fault identification model, the confidence level of each candidate fault in the candidate fault set is obtained based on the cross-domain fusion features;

[0042] Based on the fault location results and the confidence levels of each candidate fault, the fault source tracing results of the fault event are obtained.

[0043] The aforementioned transformer fault tracing method, apparatus, computer equipment, computer-readable storage medium, and computer program product first acquire physical and chemical monitoring data of the transformer. When either physical or chemical monitoring data indicates a transformer fault event, the fault is located using both physical and chemical monitoring data to obtain the fault location result. Subsequently, an attention mechanism is used to fuse the physical and chemical monitoring data to obtain cross-domain fusion features of the fault event. These cross-domain fusion features are then matched with a fault knowledge graph to obtain a candidate fault set corresponding to the fault event. A trained fault identification model is then used to obtain the confidence level of each candidate fault in the candidate fault set based on the cross-domain fusion features. Finally, based on the fault location result and the confidence levels of each candidate fault, the fault tracing result of the fault event is obtained. This scheme, by simultaneously considering both physical and chemical monitoring data of the transformer to monitor transformer fault occurrences and combining both types of monitoring data for fault location when a fault event is detected, can fully consider the potential physical and chemical effects of transformer faults, achieving accurate monitoring and location of fault events. Simultaneously, by fusing physical and chemical monitoring data using an attention mechanism and applying the resulting cross-domain fusion features to fault identification, deep fusion and collaborative reasoning of multi-source heterogeneous data can be achieved. Specifically, by first matching the cross-domain fusion features with a fault knowledge graph to obtain a candidate fault set, and then using a fault identification model to infer the confidence level of each candidate fault based on the cross-domain fusion features, accurate fault event identification can be achieved by combining expert experience and machine learning. Subsequently, by combining the fault location results and the confidence levels of each candidate fault, accurate fault tracing results can be obtained, enabling efficient and accurate tracing of fault events occurring in transformers. Attached Figure Description

[0044] 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.

[0045] Figure 1 This is a flowchart illustrating a transformer fault tracing method in one embodiment;

[0046] Figure 2 This is a schematic diagram of the process for updating physical monitoring data in one embodiment;

[0047] Figure 3 This is a flowchart illustrating the process of obtaining fault location results in one embodiment;

[0048] Figure 4 This is a flowchart illustrating the process of obtaining fault tracing results in one embodiment;

[0049] Figure 5 This is a schematic diagram of a dual-channel spatiotemporal fusion architecture in one embodiment;

[0050] Figure 6 This is a flowchart illustrating a transformer fault tracing method in another embodiment;

[0051] Figure 7 This is a schematic diagram of the processing flow of the source tracing analysis engine in one embodiment;

[0052] Figure 8 This is a structural block diagram of a transformer fault tracing device in one embodiment;

[0053] Figure 9 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0054] 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.

[0055] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various objects, but these objects are not limited by these terms. These terms are only used to distinguish the first object from the second object. 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.

[0056] In one embodiment, such as Figure 1 As shown, a method for tracing the source of transformer faults is provided. This embodiment illustrates the application of this method to a server. It is understood that this method can also be applied to a terminal, or to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0057] Step S101: Obtain physical monitoring data and chemical monitoring data of the transformer.

[0058] The physical monitoring data of the transformer can be acquired using various sensors installed on the transformer. These may include, but are not limited to, vibration signal data acquired using vibration sensors, ultra-high frequency signal data acquired using ultra-high frequency sensor arrays, and ultrasonic signal data acquired using ultrasonic sensor arrays. For example, the sensor network installed on the transformer can be equipped with a nanosecond-level synchronous clock. The clock system uses the IEEE 1588 precision time protocol to construct the topology, and utilizes a master-slave clock architecture to compress the time synchronization error between sensors to within ±100 nanoseconds. Furthermore, the sensor placement within the sensor network ensures that at least four sensors can receive direct wave signals regardless of the location of the partial discharge point on the winding.

[0059] The chemical monitoring data for transformers may include dissolved gas analysis in oil (DGA) data.

[0060] For example, a medium parameter monitoring unit can also be installed on the transformer, which can collect the transformer's oil temperature, oil density and insulation paper moisture content in real time.

[0061] Step S102: When physical monitoring data or chemical monitoring data indicates that a transformer fault event has occurred, the physical monitoring data and chemical monitoring data are used to locate the fault in the transformer and obtain the fault location result of the fault event.

[0062] Specifically, real-time detection of whether a fault event has occurred in the transformer can be performed based on physical monitoring data or chemical monitoring data. For example, the circumstances indicating a transformer fault event may include, but are not limited to, situations in physical monitoring data such as ultrasonic pulses exceeding a threshold, a surge in ultra-high frequency signal energy, or abnormal vibration signals, or situations in chemical monitoring data such as abnormal increases in the content of one or more gases, the presence or rapid increase of acetylene (C2H2), or abnormal proportions of specific gases.

[0063] In this step, when at least one of the physical monitoring data and chemical monitoring data indicates a transformer fault event, the physical and chemical monitoring data can be used to locate the fault in the transformer, thereby obtaining the fault location result. For example, this step can first determine the fault type corresponding to the fault event based on the physical and chemical monitoring data, and then use the fault location method corresponding to the fault type to obtain the fault location result based on the physical monitoring data. For instance, for a discharge fault, the location of the discharge event can be calculated using a location method based on the time difference of arrival of the acoustic-electric signals in the physical monitoring data. As another example, for a deformation fault, the location of the deformation event can be calculated based on the mathematical relationship between the change in the resonant frequency (Δf) of the vibration signal in the physical monitoring data and the spatial location.

[0064] Step S103: The attention mechanism is used to fuse physical monitoring data and chemical monitoring data to obtain cross-domain fusion features of the fault event.

[0065] Specifically, in this step, physical features can be obtained first from physical monitoring data, and chemical features can be obtained from chemical monitoring data. Then, attention mechanism is used to fuse the physical and chemical features to obtain cross-domain fusion features of the fault event.

[0066] For physical monitoring data containing multiple signal data points, time alignment processing can be performed first to align the various signal data points in time. Subsequently, time-domain, frequency-domain, and time-frequency-domain analyses can be performed on the aligned raw waveforms of various sensors on each sensor channel to extract physical features valuable for fault diagnosis and location. These features may include, but are not limited to, pulse peak amplitude, pulse rise time, pulse width, ultrasonic transit time, and vibration energy. Thus, a physical feature matrix can be constructed based on the physical monitoring data, with dimensions N×T×C. Here, N represents the number of sensors used to collect the physical monitoring data, T represents the time length of the physical monitoring data, and C represents the number of feature channels. For example, when different types of sensors correspond to different numbers of physical features, they can be projected onto a shared semantic space with a feature channel count of C, such as a 256-dimensional shared semantic space.

[0067] For example, in order to improve the accuracy of transformer fault diagnosis, the three monitoring quantities of oil temperature, load and high-frequency pulse current can be embedded in the physical feature matrix. For example, real-time sound velocity dynamic correction can be performed using oil temperature to improve positioning accuracy; partial discharge normalization can be used to eliminate misjudgments caused by load changes; and pulse counting features can be added to the physical feature matrix to realize the quantification of discharge intensity.

[0068] For chemical monitoring data, a multimodal feature extractor can be used to extract various features from DGA data, such as, but not limited to, C2H2 concentration, H2 concentration, CH4 concentration, CO / CO2 concentration ratio, etc., thereby forming a C'-dimensional chemical feature vector.

[0069] This involves mapping the physical feature matrix and chemical feature vectors to a semantic space of the same dimension, unifying their dimensions to achieve feature adaptation. Then, an attention mechanism can be used, treating the physical feature matrix as the "Query" and the chemical feature vectors as the "Key" and "Value," respectively, to perform attention calculations, thereby obtaining cross-domain fused features. For example, the calculation process of cross-domain fused features can be represented as follows:

[0070]

[0071] In the formula, Q represents the projection of the physical characteristic matrix, and K and V represent the projections of the chemical characteristic vectors.

[0072] Step S104: Match the cross-domain fusion features with the fault knowledge graph to obtain the candidate fault set corresponding to the fault event.

[0073] The fault knowledge graph can be constructed based on transformer fault diagnosis standards such as the IEC 60599 standard for dissolved gas analysis in transformer oil and the three-ratio method standard, as well as expert experience. In this step, a preliminary set of candidate faults ranked by probability can be obtained by matching cross-domain fusion features with the fault knowledge graph. For example, it can include: D2: high-energy discharge (arc discharge), D1: low-energy discharge (partial discharge), and T3: high-temperature overheating.

[0074] For example, the fault knowledge graph can be embedded to obtain the vector representation of each triple. The similarity between the vector representation of each triple and the cross-domain fusion feature is then calculated in the same vector space. Based on the similarity between the cross-domain fusion feature and each vector representation, a candidate fault set corresponding to the fault event is obtained. For example, the candidate faults in the candidate fault set can be either the top preset number of faults with the highest similarity, or faults with a similarity greater than a preset threshold.

[0075] Step S105: Using the trained fault identification model, the confidence level of each candidate fault in the candidate fault set is obtained based on the cross-domain fusion features.

[0076] The trained fault identification model can be a trained deep neural network model that infers the confidence level of various transformer faults based on input features and outputs the confidence level for each fault. In this step, cross-domain fusion features can be input into the model to obtain the confidence levels corresponding to various faults output by the model, and the confidence levels corresponding to each candidate fault in the candidate fault set can be selected from these. A higher confidence level for a candidate fault indicates that the fault event is more likely to correspond to that candidate fault. For example, the trained fault identification model used in this step can be a model fine-tuned by small-sample transfer learning using transformer fault samples, based on a fault identification model trained using a public dataset.

[0077] Step S106: Based on the fault location results and the confidence level of each candidate fault, obtain the fault source tracing results of the fault event.

[0078] Based on the fault location results and the confidence levels of each candidate fault obtained in the aforementioned process, the fault tracing result of the fault event can be comprehensively obtained. For example, high-confidence faults can be selected based on the confidence levels of each candidate fault, and then matched with the fault location results. When a high-confidence fault matches the fault location result, the fault tracing result of the fault event can be obtained based on the high-confidence fault and the fault location result. For example, when the fault location result indicates "fault coordinates are located at the tap changer," and the candidate fault with the highest confidence level is "high-energy discharge," the two match, and the fault tracing result of the fault event can be obtained based on both. When the high-confidence fault does not match the fault location result, other candidate faults can be considered to further request or wait for more data for a final judgment. For example, when the fault location result indicates "fault point is in the iron core," and the candidate fault with the highest confidence level is "arc discharge," the two do not match because oil sample arc discharge is uncommon in the iron core location. At this point, other faults in the candidate fault set can be reassessed, such as complex faults like "localized high temperature and discharge caused by multiple grounding points in the core," and more supplementary fault data (such as trend analysis) can be requested or awaited for a final judgment. Ultimately, the fault source tracing results of the fault event can be obtained based on the fault location results, supplementary fault data, and each candidate fault.

[0079] For example, the server can generate a final report of the fault event based on the fault tracing results. The report may include fault tracing analysis (including fault location and fault type), confidence levels of various candidate faults, and related data (such as physical monitoring data, chemical monitoring data, and supplementary fault data corresponding to the fault event).

[0080] The aforementioned transformer fault tracing method monitors transformer faults by simultaneously considering both physical and chemical monitoring data. When a fault event is detected, the method combines both sets of data for fault location, fully considering the potential physical and chemical impacts of transformer faults and achieving accurate monitoring and location of fault events. Furthermore, by utilizing an attention mechanism to fuse physical and chemical monitoring data and applying the resulting cross-domain fusion features to fault identification, deep fusion and collaborative reasoning of multi-source heterogeneous data are achieved. Specifically, by first matching the cross-domain fusion features with a fault knowledge graph to obtain a candidate fault set, and then using a fault identification model to infer the confidence level of each candidate fault based on the cross-domain fusion features, the method combines expert experience and machine learning to accurately identify fault events. Finally, by combining the fault location results and the confidence levels of each candidate fault, accurate fault tracing results can be obtained, achieving efficient and accurate tracing of transformer fault events.

[0081] In an exemplary embodiment, when physical monitoring data or chemical monitoring data indicates that a transformer has experienced a fault event, the transformer is located using the physical monitoring data and chemical monitoring data to obtain the fault location result of the fault event. This includes: when physical monitoring data or chemical monitoring data indicates that a transformer has experienced a fault event, obtaining the fault type corresponding to the fault event based on the physical monitoring data and chemical monitoring data; and using the fault location method corresponding to the fault type, obtaining the fault location result corresponding to the fault event based on the physical monitoring data.

[0082] Specifically, when locating a fault, one can first determine the type of fault event based on physical and chemical monitoring data, and then use the fault location method corresponding to that fault type to locate the fault.

[0083] For example, when a pulse is detected in the UHF or ultrasonic signal in the physical monitoring data, or when the chemical monitoring data indicates a C2H2 concentration >50ppm, a discharge-type fault in the transformer can be determined. The location of the discharge event can then be calculated using a location method based on the signal arrival time difference, utilizing the ultrasonic or UHF signal. As another example, when a 100 / 200Hz characteristic frequency shift is detected in the vibration signal in the physical monitoring data, a deformation-type fault in the transformer can be determined. The location of the deformation event can then be calculated based on the mathematical relationship between the change in the resonant frequency (Δf) of the vibration signal and its spatial location. Furthermore, when the chemical monitoring data indicates that the dissolved gas in the transformer oil is dominated by CH4 or C2H4, or when the index obtained after envelope integration of the vibration signal exceeds a set threshold, an overheating-type fault in the transformer can be determined. An overheating area scan of the transformer can then be performed to determine the location of the overheating event.

[0084] It is understandable that one or more faults can occur simultaneously in a transformer. When multiple types of faults occur, fault location methods for each fault type can be used to locate the fault event. For example, in a multi-fault scenario, fault probability density fields (i.e., the probability of each type of fault occurring at each point in the transformer space) corresponding to different fault types can be generated using various fault location methods. These probability density fields are then weighted and superimposed to form a preliminary comprehensive probability density field. Subsequently, the spatial gradient of this comprehensive probability density field can be calculated to obtain the gradient field, where the gradient field points in the direction of the fastest increase in probability, and zero gradient points correspond to probability peaks (i.e., candidate fault locations). Thus, by finding the zero-crossing points of the gradient field, multiple peak points in the comprehensive probability density field can be identified, each peak point potentially corresponding to a fault source. Then, by combining the fault probability density fields for each fault type, the fault type corresponding to each peak point can be determined.

[0085] When fault sources do not overlap, the locations of each fault source obtained from the localization can be directly output as the fault location result corresponding to the fault event. However, when an identified peak region actually contains contributions from more than one fault type, the distribution of the fault probability density fields of each fault type in that region can be reviewed retrospectively, and the fault source locations corresponding to each fault type can be further determined within that peak region. For example, if a peak region contains contributions from both discharge and overheating faults, and the discharge probability density gradient and overheating probability density gradient may point to slightly different sub-peak locations, the overlapping peak region can be divided into different "attraction domains" along the gradient descent direction, with each attraction domain corresponding to a most probable fault source type. For instance, when there are overlapping fault source regions, the fault source location of the highest priority fault type can be used as the primary output for that overlapping region according to a preset priority rule, while the fault source locations of lower priority fault types can be used as secondary or supplementary information output for that overlapping region. Alternatively, the overlapping region can be labeled as a "composite fault." For example, assuming the fault types are arranged in order of priority from high to low as discharge, overheating, and deformation, when a certain peak region (i.e., the overlapping region) is a mixture of the contributions of discharge faults and overheating faults, the sub-peak position corresponding to the discharge fault can be selected as the main output of the overlapping region, while the sub-peak position corresponding to the overheating fault can be output as secondary or additional information, or labeled as "composite fault".

[0086] In this embodiment, by first classifying fault events and then using corresponding fault location methods to locate the fault sources, different types of faults that may occur in the transformer can be fully considered, enabling accurate location and analysis of fault conditions. Furthermore, this embodiment also considers the simultaneous occurrence of multiple types of faults in the same area, separating overlapping fault source locations to achieve more precise fault event location.

[0087] In one exemplary embodiment, such as Figure 2 As shown, physical monitoring data includes waveform monitoring data; before obtaining the transformer fault location result by using physical and chemical monitoring data to locate the fault in the transformer when physical or chemical monitoring data indicates a transformer fault, the following may be included:

[0088] Step S201: When a pulse signal appears in the waveform monitoring data, calculate the signal source location of the pulse signal based on the position of each sensor corresponding to the waveform monitoring data, the receiving time of each sensor for the pulse signal, and the default dielectric parameters of the transformer.

[0089] Specifically, the physical monitoring data may include one or more waveform monitoring data, which may include, but are not limited to, vibration signal data, ultra-high frequency signal data, and ultrasonic signal data. For each waveform signal, a multi-sensor parallel acquisition method can be used for detection, thereby acquiring multiple waveform monitoring data corresponding to each type of waveform signal. In this embodiment, when a pulse signal appears in one or more types of waveform monitoring data, the pulse signal can be used to time-align multiple waveform monitoring data of the same type.

[0090] For example, taking ultrasound signal data as an example, when a pulse signal appears in the ultrasound signal data, the direct wave can be identified first. The specific judgment principle is as follows: when there is a pulse signal with a significantly larger amplitude in the ultrasound signal data, this pulse signal can be considered the direct wave; when there are multiple pulse signals with similar amplitudes in the ultrasound signal data, the pulse signal with the shorter arrival time can be considered the direct wave. The time of the first peak of the direct wave in each ultrasound signal data can be used as the reception time of the pulse signal for each ultrasound sensor. Subsequently, the reception time difference between different ultrasound sensors can be calculated, where the reception time difference between ultrasound sensor i and ultrasound sensor j can be expressed as: In the formula For the time difference of reception, Let be the time it takes for ultrasonic sensor i to receive the pulse signal. This represents the reception time of the pulse signal by ultrasonic sensor j. The location of each ultrasonic sensor can also be obtained; for example, for ultrasonic sensor i, its corresponding spatial coordinates can be acquired. .

[0091] Wherein, it is assumed that the spatial coordinates of the signal source location of the pulse signal are... Then the pulse signal originates from the signal source position. propagation to the observation point The time is ,but arrive distance It can be represented as: ,in For ultrasound signals in arrive The propagation speed in the medium between them. Based on this, assuming there are a total of 4 ultrasonic sensors, the following set of equations can be obtained:

[0092]

[0093] in, .

[0094] Based on the above set of equations, the default dielectric parameters of the transformer can be used to... The default propagation speed is set, and the location of the pulse signal source is determined using the least squares method based on this speed. .

[0095] Step S202: Update the pulse signal reception time of each sensor according to the signal source location, the location of each sensor, and the real-time dielectric parameters of the transformer.

[0096] In this step, the location of the signal source can be calculated previously. The system also monitors the location of each sensor and the real-time dielectric parameters of the transformer, updating the receiving time of each sensor for the pulse signal.

[0097] For example, a dielectric parameter monitoring unit can be installed on the transformer to collect parameters such as oil temperature, oil density, and moisture content of the insulating paper in real time, thereby obtaining the transformer's real-time dielectric parameters. Based on these real-time dielectric parameters, the transmission speed of various types of waveform signals within the transformer can be corrected. For example, for ultrasonic signals, an insulating oil sound velocity model can be established to describe the relationship between the propagation speed of ultrasound and the dielectric parameters of the insulating oil. This model can incorporate a temperature compensation formula. ( The propagation speed of ultrasound is corrected for (in Celsius). For example, for vibration signals, the steel plate of a transformer enclosure can be used as the object of vibration wave observation, and a steel plate vibration wave model can be established to describe the propagation characteristics of the vibration wave in the transformer. This model can incorporate a correction factor for the elastic modulus. ( (This refers to the trace water content), and this coefficient is used to correct the propagation speed of the vibration wave.

[0098] For example, taking the ultrasound signal data from the aforementioned steps as an example, it can be determined based on the location of the signal source. The position of each ultrasonic sensor And the real-time dielectric parameters of the transformer, to obtain the ultrasonic signal in With each speed of propagation between Subsequently, the initial propagation delay of each ultrasonic sensor can be calculated separately: And calculate the theoretical propagation delay of each ultrasonic sensor: Subsequently, time-shift compensation can be performed on each ultrasonic signal data using a finite impulse response filter, with the compensation amount being... Subsequently, coherent energy focusing analysis can be performed on each time-shifted ultrasonic signal data. When the peak value of the cross-correlation function of the ultrasonic sensor array exceeds the threshold, the reception time of each ultrasonic sensor for the pulse signal can be updated based on the final time shift of each ultrasonic signal data. The updated reception time was obtained. .

[0099] Step S203: Update the signal source position of the pulse signal according to the real-time dielectric parameters of the transformer, the position of each sensor, and the updated reception time.

[0100] Among them, the propagation speed is obtained based on the real-time dielectric parameters of the transformer. The position of each sensor and the updated reception time A new system of equations can be constructed again:

[0101]

[0102] in, .

[0103] Therefore, based on the above set of equations, the new signal source location of the pulse signal can be solved using the least squares method. This allows us to obtain the updated location of the signal source.

[0104] Specifically, after each update of the signal source position, the distance between that position and the previously calculated signal source position can be calculated, and this distance can be compared with a preset distance threshold. If the distance is not less than the distance threshold, step S204 can be executed; if the distance is less than the distance threshold, step S205 can be executed. For example, the internal space of the transformer can be discretized into a million-level voxel grid. When the distance between two calculated signal source positions is less than the grid resolution, it can be determined that the distance is less than the distance threshold.

[0105] Step S204: If the distance between the updated signal source position and the previously calculated signal source position is not less than the distance threshold, the step of updating the receiving time of each sensor for the pulse signal based on the signal source position, the position of each sensor, and the real-time medium parameters of the transformer is executed again.

[0106] If the distance between the updated signal source position and the previously calculated signal source position is not less than the distance threshold, steps S202 and S203 can be repeated to update the signal source position and the receiving time of each sensor for the pulse signal again, and to recalculate the distance between the signal source positions before and after the update, until the distance is less than the distance threshold, and then proceed to step S205.

[0107] Step S205: If the distance between the updated signal source position and the previously calculated signal source position is less than a distance threshold, the waveform monitoring data collected by each sensor is time-aligned according to the updated reception time of each sensor to obtain aligned waveform monitoring data.

[0108] In cases where the distance between the updated signal source location and the previously calculated signal source location is less than a distance threshold, the waveform monitoring data collected by each sensor can be time-aligned based on the last updated reception time of each sensor, thus aligning them on the time axis to obtain aligned waveform monitoring data.

[0109] Step S206: Update the physical monitoring data based on the aligned waveform monitoring data.

[0110] After obtaining the aligned waveform monitoring data, the physical monitoring data can be updated so that the updated physical monitoring data includes the aligned waveform monitoring data.

[0111] It is understood that although this embodiment uses ultrasonic signal data as an example for detailed explanation, other waveform monitoring data in physical monitoring data (such as vibration signal data, ultra-high frequency signal data, etc.) can be aligned in the manner described in this embodiment.

[0112] In this embodiment, by using a positioning method based on receiving time difference to perform time synchronization and preliminary positioning of pulse signals on waveform monitoring data collected by distributed sensors, the problem of signal spatiotemporal asynchrony caused by spatial location differences and propagation medium delay of distributed sensors can be solved. A precise mapping relationship between physical coordinates, signal propagation, and time axis is established through time difference compensation, and preliminary fault location is performed to provide a reference for subsequent fault location.

[0113] In one exemplary embodiment, such as Figure 3 As shown, when physical or chemical monitoring data indicate a transformer fault event, the fault location of the transformer is performed using both physical and chemical monitoring data to obtain the fault location result, which may include:

[0114] Step S301: When physical monitoring data or chemical monitoring data indicates that a transformer fault event has occurred, the fault type corresponding to the fault event is obtained based on the physical monitoring data and chemical monitoring data.

[0115] Specifically, when locating a fault, the fault type can be determined first based on physical and chemical monitoring data, and then the fault location method corresponding to that fault type can be used for fault location. The specific method for determining the fault type can be found in the descriptions in the preceding embodiments, and will not be repeated here.

[0116] Step S302: If the fault type indicates that the fault event is associated with a pulse signal in at least one waveform monitoring data, obtain the signal source location of the pulse signal in each waveform monitoring data and the location of each sensor.

[0117] Different types of fault events can cause pulse signals in one or more waveform monitoring data, or they may not cause pulse signals in the waveform monitoring data. For example, when the fault type of the fault event is a discharge fault, it can cause pulse signals in various waveform monitoring data such as vibration signal data, ultra-high frequency signal data, and ultrasonic signal data, so this type of fault event can be associated with the pulse signals of these waveform monitoring data; as another example, when the fault type of the fault event is a mechanical fault, it can be associated with the pulse signals of vibration signal data; as yet another example, when the fault type of the fault event is an overheating fault, it is unlikely to cause pulse signals in the waveform monitoring data, so it can be determined that it is not associated with the pulse signals of the waveform monitoring data.

[0118] Specifically, when the fault type of a fault event indicates that the event is associated with a pulse signal appearing in at least one waveform monitoring data, it can be determined that the pulse signal appearing in the associated waveform monitoring data is likely related to the event. This allows the acquisition of the signal source location of the pulse signal in each type of waveform monitoring data associated with the event, as well as the reception time of the pulse signal by each sensor. The signal source location is the location of the last updated pulse data obtained during the alignment process of that type of waveform monitoring data, and the reception time of the pulse signal by each sensor corresponding to that type of waveform monitoring data is the last updated reception time obtained during the alignment process. For example, for a discharge fault event, the signal source location of the pulse signal corresponding to vibration signal data, ultra-high frequency signal data, and ultrasonic signal data, as well as the reception time of the pulse signal by each sensor, can be acquired separately.

[0119] Step S303: Based on the location of the signal source, obtain the spatial range of candidate faults of the transformer.

[0120] In this step, based on the previously obtained signal source locations, a candidate fault space range can be determined within the transformer space, encompassing the locations of each signal source corresponding to the fault event. For example, a bounding box containing one or more signal source locations corresponding to the fault event can be constructed, and this bounding box can be expanded according to a preset distance parameter to obtain the candidate fault space range.

[0121] Step S304: Calculate the fault source probability of each candidate fault point within the candidate fault space range based on the real-time dielectric parameters of the transformer, the positions of each sensor, and the signal source position of the pulse signal.

[0122] Specifically, the candidate fault space can be divided into multiple grids, discretized into equally spaced voxels, and the center of each voxel can be used as a candidate fault point. Subsequently, the fault source probability of each candidate fault point can be calculated based on the real-time dielectric parameters of the transformer, the positions of each sensor, and the reception time of the pulse signal by each sensor. The fault source probability of a candidate fault point can be the probability that the point is the location of the fault source of the fault event; a higher fault source probability indicates that the candidate fault point is more likely to be the location of the fault source of the fault event. In this step, the resolution of the grid can be higher than the grid resolution used in the signal alignment process.

[0123] In this process, the transmission speed of various types of waveform signals within the transformer can be corrected based on the real-time dielectric parameters of the transformer. For example, for ultrasonic signals, an insulating oil sound velocity model can be established to describe the relationship between the propagation speed of ultrasonic waves and the dielectric parameters of the insulating oil. This model can incorporate a temperature compensation formula. ( The propagation speed of ultrasound is corrected for (in Celsius). For example, for vibration signals, the steel plate of a transformer enclosure can be used as the object of vibration wave observation, and a steel plate vibration wave model can be established to describe the propagation characteristics of the vibration wave in the transformer. This model can incorporate a correction factor for the elastic modulus. ( The coefficient (based on the trace water content) is used to correct the propagation velocity of the vibration wave. Therefore, in this step, the propagation velocity of various waveform monitoring signals associated with the fault event in the transformer can be obtained based on the transformer's real-time dielectric parameters. .

[0124] Among them, the propagation speed of various waveform monitoring signals based on fault event correlation in the transformer Sensor locations can be constructed. Candidate Fault Points and signal arrival time difference Mapping relationship between them:

[0125]

[0126] In the formula, Location of the fault source To the sensor location Euclidean distance, It is the signal arrival time difference between the signal occurrence time and the sensor's pulse signal reception time.

[0127] Following the same principle, the theoretical time difference of arrival of the pulse signal for each sensor can be calculated based on the signal source location of the pulse signal in each waveform monitoring data and the location of each sensor. .

[0128] Subsequently, the theoretical signal arrival time difference of each sensor for the pulse signal can be used as a basis. Signal arrival time difference Calculate the fault source probability for each candidate fault point:

[0129]

[0130] In the formula, Candidate Fault Points The probability of the fault source, This represents the number of sensors corresponding to all waveform monitoring data associated with the fault event. Let i be the weight of the i-th sensor. This represents the standard deviation. Each type of sensor can correspond to a different standard deviation. This can eliminate the differences in time scale between different types of sensors.

[0131] Among them, the weight of each sensor The system can dynamically adjust its weighting based on the sensor's real-time signal-to-noise ratio, sensor health status, and current fault assumptions. For example, when the diagnostic process focuses on discharge faults, the system can automatically increase the weighting of ultrasonic and ultra-high frequency sensors while decreasing the weighting of vibration sensors, which are less sensitive to discharge. If a sensor signal is significantly abnormal (e.g., excessive residual), its weight will be adjusted. It will be lowered or even set to zero to avoid erroneous data contaminating the fusion result.

[0132] Step S305: Based on the candidate fault point with the highest probability of fault source, obtain the fault location result of the fault event.

[0133] In this process, after obtaining the fault source probability of each candidate fault point, the candidate fault point with the highest fault source probability can be taken as the fault source location of the fault event, thereby obtaining the fault location result of the fault event. For example, the fault location result can be expressed as: .

[0134] In this embodiment, for fault events associated with pulse signals appearing in waveform monitoring data, the fault source location is further searched based on the signal source location of the pulse signal obtained in the previous spatiotemporal alignment. The fault source probability of candidate fault points is calculated by fusing data from all relevant sensors, and the candidate fault point with the highest probability is selected as the fault source location, which can achieve high-precision positioning of fault events.

[0135] In one exemplary embodiment, such as Figure 4 As shown, based on the fault location results and the confidence levels of each candidate fault, the fault source tracing results of the fault event can be obtained, which may include:

[0136] Step S401: Based on the confidence level of each candidate fault, obtain the high-confidence fault of the transformer.

[0137] This process involves filtering candidate faults based on their confidence levels to identify high-confidence faults in the transformer. For example, a preset confidence threshold can be used to filter faults, with those having a confidence level higher than the threshold classified as high-confidence faults. Alternatively, a preset number of candidate faults with the highest confidence levels can also be selected as high-confidence faults.

[0138] After identifying high-confidence faults, each high-confidence fault can be matched with the fault location results. If a match is found, proceed to step S402; otherwise, proceed to step S403. For example, if the location of a high-confidence fault conflicts with the fault location results, it can be determined that the two do not match; otherwise, it can be determined that they match.

[0139] Step S402: If the fault location result matches the high-confidence fault, obtain the fault tracing result of the fault event based on the fault location result and the high-confidence fault.

[0140] In cases where the fault location result matches a high-confidence fault, the fault tracing result of the fault event can be obtained based on the fault location result and the high-confidence fault. This fault tracing result may include the fault location result and the matching high-confidence fault, as well as the confidence level of the high-confidence fault.

[0141] Step S403: If the fault location result does not match the high-confidence fault, obtain supplementary fault data for the transformer.

[0142] In cases where the fault location results do not match the high-confidence fault, supplementary fault data for the transformer can be obtained to assist in fault tracing. For example, supplementary fault data can be physical and chemical monitoring data over a longer period, such as historical data from further back or newly acquired monitoring data. For example, supplementary fault data can also be one or more transformer indicator data, such as trend analysis results obtained after statistical analysis of physical and / or chemical monitoring data.

[0143] Step S404: Based on the fault location results, supplementary fault data, and each candidate fault, obtain the fault source tracing results of the fault event.

[0144] In cases where the fault location result does not match the high-confidence fault, the fault location result and supplementary fault data can be combined to re-evaluate each candidate fault to ultimately determine the most likely fault corresponding to the fault event. For example, when the fault location result indicates "the fault point is in the iron core," while the candidate fault with the highest confidence is "arc discharge," a mismatch occurs because arc discharge in oil samples is uncommon at the iron core location. In this case, other faults in the candidate fault set can be re-evaluated, such as considering composite faults like "multiple grounding points in the iron core causing localized high temperatures and discharge," and more supplementary fault data (such as trend analysis) can be requested or awaited for a final judgment. Ultimately, the fault source tracing result of the fault event can be obtained based on the fault location result, supplementary fault data, and each candidate fault. This fault source tracing result can include fault source analysis (including fault location and fault type), the confidence levels of various candidate faults, and related data (such as physical monitoring data, chemical monitoring data, and supplementary fault data corresponding to the fault event).

[0145] In this embodiment, by screening high-confidence faults in the transformer and matching high-confidence faults with fault locations, and further obtaining supplementary fault data and reconsidering each candidate fault for cases where the two do not match, more accurate fault tracing results can be obtained.

[0146] In an exemplary embodiment, the trained fault identification model is obtained through the following steps: obtaining a first fault identification model trained using industrial equipment fault samples; adjusting the model parameters of the first fault identification model using transformer fault samples to obtain the trained fault identification model.

[0147] Specifically, to overcome the bottleneck of scarce transformer fault samples, this embodiment employs a small-sample transfer learning method. Based on a pre-trained first fault identification model, its model parameters are adjusted to obtain a trained fault identification model. This method uses a "pre-training-fine-tuning" paradigm, which leverages cross-device fault feature transfer to address the problem of insufficient transformer fault samples.

[0148] For example, a fault identification model to be trained can be constructed first. This model can be a deep neural network model including a feature extractor and an output layer. In the pre-training stage, a large number of industrial equipment fault samples can be obtained using publicly available datasets. These industrial equipment fault samples are then used to train the fault identification model to be trained, resulting in a first fault identification model. For example, the industrial equipment fault samples can be bearing fault data samples, which may include bearing feature data and corresponding fault labels. Bearing fault data samples are chosen as the source domain data because they contain fault characteristics such as vibration and impact of rotating machinery, and have feature similarities (e.g., high-frequency impact pulses) to mechanical faults such as loose transformer windings and core vibration, exhibiting strong transferability.

[0149] For example, in the domain adaptation phase, the distribution difference between the source domain (bearing data) and the target domain (transformer data) can be measured by the Maximum Mean Discrepancy (MMD), as shown in the following formula:

[0150]

[0151] In the formula, X represents the source domain sample, and Y represents the target domain sample. This is the feature mapping function.

[0152] Specifically, by adjusting the model parameters of the feature extractor in the first fault identification model with the goal of minimizing the MMD (Mean Differential Damage), the feature extractor can be adapted to the characteristic patterns unique to transformers. Simultaneously, the top-level parameters of the first fault identification model can be adjusted using a small number of transformer fault samples during the fine-tuning stage. This allows the model to shift from "identifying bearing faults" to "identifying transformer faults," resulting in a trained fault identification model. The trained fault identification model can then utilize the feature extractor to further extract cross-domain fusion features of the transformer to obtain high-dimensional features. The output layer then uses these high-dimensional features to output the confidence levels for various transformer faults.

[0153] In one exemplary embodiment, a method for tracing the source of transformer faults is provided.

[0154] Specifically, the fault tracing method in this embodiment can utilize, for example... Figure 5The dual-channel spatiotemporal fusion architecture shown is implemented. This architecture includes a physical signal channel and a chemical signal channel. The physical signal channel integrates multi-source data (i.e., physical monitoring data) collected by various sensors such as vibration sensors, partial discharge sensor arrays, and ultrasonic sensor arrays. A spatiotemporal alignment module handles the time synchronization of the multi-source data and generates a spatiotemporal feature matrix, which is then input into the cross-domain feature fusion module. Simultaneously, the chemical signal channel performs online monitoring of DGA data (i.e., chemical monitoring data) and processes the DGA data using a multimodal feature extractor to generate chemical feature vectors, which are then input into the cross-domain feature fusion module. The cross-domain feature fusion module fuses the spatiotemporal feature matrix with the chemical feature vectors and inputs the resulting fused features into the source tracing analysis engine. Simultaneously, the spatiotemporal feature matrix and chemical feature vectors are input into the fault location engine, which performs fault location inference and inputs the resulting fault location result into the source tracing analysis engine. The source tracing analysis engine can infer the cause of the fault based on the cross-domain fusion features and the fault location result, forming an integrated "location-source tracing" diagnostic process.

[0155] Among them, such as Figure 6 As shown, the workflow of this architecture may include the following steps:

[0156] Step 1: Parallel signal processing.

[0157] Among them, the physical signal channel can collect various data such as vibration sensor, ultra-high frequency sensor array, and ultrasonic sensor array data, and align them using the spatiotemporal alignment module, and output the spatiotemporal feature matrix.

[0158] For example, the spatiotemporal alignment module can utilize spatiotemporal registration technology to solve the problem of signal spatiotemporal asynchrony caused by spatial location differences and propagation medium delays in distributed sensors. Through time difference compensation algorithms for acoustic and electrical signals and vibration waves, it can establish a precise mapping relationship between physical coordinates, signal propagation, and time axis, and simultaneously perform preliminary fault location to provide a reference for subsequent fault location. In the case of pulse signals in waveform monitoring data (such as vibration signal data, UHF signal data, and ultrasonic signal data), the spatiotemporal alignment module can calculate the signal source position of the pulse signal based on the position of each sensor corresponding to the waveform monitoring data, the receiving time of each sensor for the pulse signal, and the default dielectric parameters of the transformer; update the receiving time of each sensor for the pulse signal based on the signal source position, the position of each sensor, and the real-time dielectric parameters of the transformer; update the signal source position of the pulse signal based on the real-time dielectric parameters of the transformer, the position of each sensor, and the updated receiving time; if the distance between the updated signal source position and the previously calculated signal source position is not less than a distance threshold, the step of updating the receiving time of each sensor for the pulse signal based on the signal source position, the position of each sensor, and the real-time dielectric parameters of the transformer is repeated; if the distance between the updated signal source position and the previously calculated signal source position is less than a distance threshold, the waveform monitoring data collected by each sensor is time-aligned based on the updated receiving time of each sensor to obtain aligned waveform monitoring data; and update the physical monitoring data based on the aligned waveform monitoring data.

[0159] The spatiotemporal alignment module can construct dynamic propagation models, which can be pre-designed with multiphysics equations. For example, the sound velocity model of insulating oil uses the temperature compensation formula v=1400+3.6ΔT (T is the temperature rise in Celsius), and the steel plate vibration wave model introduces the elastic modulus correction coefficient k=5020-1.2T-0.05M (M is the trace moisture content). Therefore, during spatiotemporal alignment using spatiotemporal registration technology, the actual propagation speed of various waveforms in the medium can be obtained based on the real-time dielectric parameters of the transformer.

[0160] To improve the accuracy of transformer fault diagnosis, embedded processing is employed for three monitored parameters: oil temperature, load, and high-frequency pulse current. Real-time dynamic correction of sound velocity is performed using oil temperature to enhance positioning accuracy; partial discharge normalization eliminates misjudgments caused by load changes; and pulse counting features are added to the spatiotemporal feature matrix to quantify discharge intensity. For example, the spatiotemporal alignment module can be equipped with a residual monitoring unit, which calculates the time difference residuals |δt| of each sensor. When the residuals exceed twice the standard deviation for three consecutive sampling periods, a sensor health status flag is triggered. Faulty sensor data is automatically replaced with spatial interpolation results from neighboring sensors. Simultaneously, a temperature drift compensation module can be included to recalibrate the sound velocity model every 30 seconds and immediately trigger a model update when the oil temperature changes by more than 1°C.

[0161] Ultimately, time alignment and spatial mapping are achieved through the spatiotemporal registration signal matrix, providing a centimeter-level positioning foundation for subsequent time difference back projection.

[0162] Among them, the chemical signal channel can use a multimodal feature extractor to process DGA data and output a 16-dimensional chemical feature vector.

[0163] Step 2: Cross-domain feature fusion.

[0164] The cross-domain feature fusion module can map features from different structures to a semantic space of the same dimension, unifying the dimensions of the spatiotemporal feature matrix and the chemical feature vector to achieve feature adaptation. Then, cross-domain attention calculation is performed, using the spatiotemporal feature matrix as the "Query" and the chemical feature vector as the "Key" and "Value" to perform attention calculation, thereby obtaining cross-domain fused features.

[0165] Step 3: Analyze engine decisions.

[0166] Among them, the fault location engine and the source analysis engine can be used together to participate in the decision-making of the fault source tracing results of fault events.

[0167] For example, when physical or chemical monitoring data indicates a transformer fault event, the fault location engine can determine the fault type based on the physical and chemical monitoring data. Then, using the fault location method corresponding to the fault type, it obtains the fault location result based on the physical monitoring data. Specifically, when the fault type indicates a fault event associated with a pulse signal from at least one waveform monitoring data point, the fault location engine can call a time-difference back-projection algorithm for precise calculation. By comparing the theoretical and actual time differences, a probability heatmap is generated, and the fault coordinates are determined by finding the global peak of the energy field. For multi-fault scenarios, the algorithm introduces a hierarchical decision-making mechanism: for example, when a pulse signal is detected and the C2H2 concentration is >50ppm, a discharge location process is initiated using a 1cm high-resolution grid; if a 100 / 200Hz characteristic frequency offset exists, the winding deformation location module is activated, and the deformation position probability distribution is inferred from the resonant frequency shift Δf; after performing envelope integration on the broadband vibration signal, an overheated region scan is triggered. In case of spatial conflict, overlapping sources are separated based on the fault probability density gradient field, with the discharge point prioritized over the overheated region.

[0168] Among them, the source tracing analysis engine can infer the cause of a fault based on cross-domain fusion features and fault location results. For example, such as... Figure 7 As shown, the source tracing analysis engine can perform source tracing analysis through knowledge graphs and neural networks. It can map cross-domain fusion features to specific fault root cause categories (such as "multi-point grounding of the iron core," "inter-turn short circuit in the winding," "aging of insulation paper," etc.). The engine's output is not only a simple classification but also combines location information for comprehensive reasoning. For example, when the location engine determines that the fault point is in the tap changer area and the source tracing engine judges it to be a discharge fault, the system can confidently give the final diagnostic conclusion that "poor contact of the tap changer caused the discharge."

[0169] For example, the source tracing analysis engine can first match the cross-domain fusion features with the fault knowledge graph to obtain the candidate fault set corresponding to the fault event, then use the trained fault identification model to obtain the confidence of each candidate fault in the candidate fault set based on the cross-domain fusion features, and then obtain the fault source tracing result of the fault event based on the fault location result and the confidence of each candidate fault.

[0170] The fault identification model used in the source tracing analysis engine can be a model fine-tuned by small-sample transfer learning using transformer fault samples, based on a fault identification model trained on a public dataset. Specifically, a large number of industrial equipment fault samples (e.g., bearing fault samples) can be trained using a public dataset. These industrial equipment fault samples are then used to train the fault identification model to obtain a first fault identification model. The model parameters of the first fault identification model are then adjusted using transformer fault samples to obtain the trained fault identification model. The process of adjusting the model parameters of the first fault identification model can include a domain adaptation stage and a fine-tuning stage. In the domain adaptation stage, the maximum mean discrepancy (MMD) can be used to measure the distribution difference between the source domain (bearing data) and the target domain (transformer data). The model parameters of the feature extractor in the first fault identification model are adjusted with the goal of minimizing the MMD, so that the feature extractor adapts to the unique feature patterns of the transformer. During the fine-tuning phase, a small number of transformer fault samples can be used to adjust the top-level parameters of the first fault identification model, thereby enabling the model to shift from "identifying bearing faults" to "identifying transformer faults" and obtain a trained fault identification model.

[0171] The source tracing analysis engine can perform final arbitration by combining the confidence levels of each candidate fault with the fault location results from the fault location engine. For example, based on the confidence levels of each candidate fault, a high-confidence fault of the transformer can be obtained. If the fault location result matches the high-confidence fault, the source tracing result of the fault event is obtained based on the fault location result and the high-confidence fault. If the fault location result does not match the high-confidence fault, supplementary fault data of the transformer is obtained. Based on the fault location result, the supplementary fault data, and each candidate fault, the source tracing result of the fault event is obtained. For instance, when the fault location engine reports "fault coordinates are located at the tap changer," and the high-confidence fault is "high-energy discharge," it can be determined that the two match, thus leading to the final diagnostic conclusion that "discharge is caused by poor contact of the tap changer." As another example, when the fault location engine reports "fault point is in the core," and the high-confidence fault is "arc discharge," the engine will activate a conflict resolution mechanism. At this point, since arc discharge in the oil sample is uncommon at the core location, the source tracing analysis engine can lower the confidence level of "arc discharge" and reassess the knowledge graph, considering complex faults such as "localized high temperature and discharge caused by multiple grounding points in the core." Simultaneously, more data (such as trend analysis) can be requested or awaited for a final judgment. Ultimately, the source tracing analysis engine will generate a final report, which includes: fault source tracing analysis, confidence levels, and correlation data.

[0172] It should be understood that although the steps in the flowcharts of the embodiments described above 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 embodiments described above 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.

[0173] Based on the same inventive concept, this application also provides a transformer fault tracing device for implementing the above-described transformer fault tracing method. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations of one or more transformer fault tracing device embodiments provided below can be found in the limitations of the transformer fault tracing method described above, and will not be repeated here.

[0174] In one exemplary embodiment, such as Figure 8 As shown, a fault tracing device for a transformer is provided, comprising:

[0175] The data acquisition module 801 is used to acquire physical monitoring data and chemical monitoring data of the transformer;

[0176] The fault location module 802 is used to locate the fault in the transformer using the physical monitoring data and the chemical monitoring data when the physical monitoring data or the chemical monitoring data indicates that a fault event has occurred in the transformer, and to obtain the fault location result of the fault event.

[0177] The data fusion module 803 is used to fuse the physical monitoring data and the chemical monitoring data using an attention mechanism to obtain the cross-domain fusion features of the fault event;

[0178] The feature matching module 804 is used to match the cross-domain fusion features with the fault knowledge graph to obtain a candidate fault set corresponding to the fault event;

[0179] The fault identification module 805 is used to obtain the confidence level of each candidate fault in the candidate fault set based on the cross-domain fusion features using a trained fault identification model.

[0180] The result acquisition module 806 is used to obtain the fault tracing result of the fault event based on the fault location result and the confidence level of each candidate fault.

[0181] In an exemplary embodiment, the fault location module 802 is configured to: when the physical monitoring data or the chemical monitoring data indicates that a fault event has occurred in the transformer, obtain the fault type corresponding to the fault event based on the physical monitoring data and the chemical monitoring data; and use the fault location method corresponding to the fault type to obtain the fault location result corresponding to the fault event based on the physical monitoring data.

[0182] In an exemplary embodiment, the physical monitoring data includes waveform monitoring data; the device further includes: a pulse positioning module, configured to calculate the signal source position of the pulse signal based on the position of each sensor corresponding to the waveform monitoring data, the reception time of each sensor for the pulse signal, and the default dielectric parameters of the transformer when a pulse signal is detected in the waveform monitoring data; a time update module, configured to update the reception time of each sensor for the pulse signal based on the signal source position, the position of each sensor, and the real-time dielectric parameters of the transformer; and a positioning update module, configured to update the pulse signal based on the real-time dielectric parameters of the transformer, the position of each sensor, and the updated reception time. The system comprises: a signal source location module; an iterative execution module, configured to re-execute the step of updating the reception time of each sensor for the pulse signal based on the signal source location, the location of each sensor, and the real-time dielectric parameters of the transformer, provided that the distance between the updated signal source location and the previously calculated signal source location is not less than a distance threshold; a data alignment module, configured to perform time alignment processing on the waveform monitoring data collected by each sensor based on the updated reception time of each sensor, provided that the distance between the updated signal source location and the previously calculated signal source location is less than a distance threshold, to obtain aligned waveform monitoring data; and a data update module, configured to update the physical monitoring data based on the aligned waveform monitoring data.

[0183] In an exemplary embodiment, the fault location module 802 is configured to: when the physical monitoring data or the chemical monitoring data indicates that a fault event has occurred in the transformer, obtain the fault type corresponding to the fault event based on the physical monitoring data and the chemical monitoring data; when the fault type indicates that the fault event is associated with a pulse signal appearing in at least one of the waveform monitoring data, obtain the signal source location of the pulse signal in each of the waveform monitoring data and the location of each of the sensors; obtain the candidate fault space range of the transformer based on the signal source location; calculate the fault source probability of each candidate fault point within the candidate fault space range based on the real-time dielectric parameters of the transformer, the location of each of the sensors, and the signal source location of the pulse signal; and obtain the fault location result of the fault event based on the candidate fault point with the highest fault source probability.

[0184] In an exemplary embodiment, the result acquisition module 806 is configured to: obtain a high-confidence fault of the transformer based on the confidence level of each of the candidate faults; if the fault location result matches the high-confidence fault, obtain a fault tracing result of the fault event based on the fault location result and the high-confidence fault; if the fault location result does not match the high-confidence fault, obtain supplementary fault data of the transformer; and obtain a fault tracing result of the fault event based on the fault location result, the supplementary fault data, and each of the candidate faults.

[0185] In an exemplary embodiment, the trained fault identification model is obtained through the following steps: acquiring a first fault identification model trained using industrial equipment fault samples; adjusting the model parameters of the first fault identification model using transformer fault samples to obtain the trained fault identification model.

[0186] Each module in the aforementioned transformer fault tracing device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, so that the processor can call and execute the corresponding operations of each module.

[0187] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 9As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media. The database stores physical and chemical monitoring data of the transformer. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for tracing the fault source of a transformer.

[0188] Those skilled in the art will understand that Figure 9 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.

[0189] 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.

[0190] 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.

[0191] 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.

[0192] 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.

[0193] 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, and when executed, it 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.

[0194] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this application. The above embodiments only illustrate several implementation methods of this application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of this application. It should be noted that for those skilled in the art, several modifications and improvements can be made 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 of fault tracing of a transformer, characterized by, The method comprises: acquiring physical monitoring data and chemical monitoring data of a transformer; when the physical monitoring data or the chemical monitoring data indicates that a fault event occurs in the transformer, performing fault positioning on the transformer by using the physical monitoring data and the chemical monitoring data to obtain a fault positioning result of the fault event; fusing the physical monitoring data and the chemical monitoring data by using an attention mechanism to obtain cross-domain fusion features of the fault event; matching the cross-domain fusion features with a fault knowledge graph to obtain a candidate fault set corresponding to the fault event; obtaining a confidence of each candidate fault in the candidate fault set according to the cross-domain fusion features by using a trained fault identification model; obtaining a fault tracing result of the fault event according to the fault positioning result and the confidence of each candidate fault.

2. The method of claim 1, wherein, The method comprises: when the physical monitoring data or the chemical monitoring data indicates that a fault event occurs in the transformer, obtaining a fault type corresponding to the fault event according to the physical monitoring data and the chemical monitoring data; obtaining a fault positioning result corresponding to the fault event according to the physical monitoring data by using a fault positioning method corresponding to the fault type.

3. The method of claim 1, wherein, The physical monitoring data comprises waveform monitoring data; before the physical monitoring data or the chemical monitoring data indicates that a fault occurs in the transformer, performing fault positioning on the transformer by using the physical monitoring data and the chemical monitoring data to obtain a fault positioning result of the transformer, the method comprises: when a pulse signal appears in the waveform monitoring data, calculating a signal source position of the pulse signal according to a position of each sensor corresponding to the waveform monitoring data, a receiving time of the pulse signal by each sensor, and a default medium parameter of the transformer; updating the receiving time of the pulse signal by each sensor according to the signal source position, the position of each sensor, and a real-time medium parameter of the transformer; updating the signal source position of the pulse signal according to the real-time medium parameter of the transformer and the position of each sensor and the updated receiving time; when a distance between the updated signal source position and a previously calculated signal source position is not less than a distance threshold, re-executing the step of updating the receiving time of the pulse signal by each sensor according to the signal source position, the position of each sensor, and a real-time medium parameter of the transformer; when the distance between the updated signal source position and the previously calculated signal source position is less than the distance threshold, performing time alignment processing on the waveform monitoring data collected by each sensor according to the updated receiving time of each sensor to obtain aligned waveform monitoring data; updating the physical monitoring data according to the aligned waveform monitoring data.

4. The method of claim 3, wherein, when the physical monitoring data or the chemical monitoring data indicates that the transformer has a fault event, performing fault location on the transformer by using the physical monitoring data and the chemical monitoring data to obtain a fault location result of the fault event, including: when the physical monitoring data or the chemical monitoring data indicates that the transformer has a fault event, obtaining a fault type corresponding to the fault event according to the physical monitoring data and the chemical monitoring data; in a case where the fault type indicates that the fault event is associated with a pulse signal appearing in at least one of the waveform monitoring data, obtaining the signal source position of the pulse signal in each of the waveform monitoring data, and the position of each of the sensors; based on the signal source position, obtaining a candidate fault space range of the transformer; according to real-time medium parameters of the transformer, the position of each of the sensors, and the signal source position of the pulse signal, calculating a fault source probability of each candidate fault point in the candidate fault space range; obtaining a fault location result of the fault event according to the candidate fault point with the highest fault source probability.

5. The method of claim 1, wherein, the fault source probability, including: obtaining a high-confidence fault of the transformer according to the confidence of each candidate fault; in a case where the fault location result matches the high-confidence fault, obtaining a fault tracing result of the fault event according to the fault location result and the high-confidence fault; in a case where the fault location result does not match the high-confidence fault, obtaining fault supplementary data of the transformer; obtaining a fault tracing result of the fault event according to the fault location result, the fault supplementary data, and each candidate fault.

6. The method according to any one of claims 1 to 5, characterized in that, the trained fault identification model is trained by the following steps: obtaining a first fault identification model trained by using industrial equipment fault samples; adjusting model parameters of the first fault identification model by using transformer fault samples to obtain the trained fault identification model.

7. A fault tracing device for a transformer, characterized by The device comprises: a data acquisition module configured to acquire physical monitoring data and chemical monitoring data of a transformer; a fault location module configured to, when the physical monitoring data or the chemical monitoring data indicates that the transformer has a fault event, perform fault location on the transformer by using the physical monitoring data and the chemical monitoring data to obtain a fault location result of the fault event; a data fusion module configured to fuse the physical monitoring data and the chemical monitoring data by using an attention mechanism to obtain cross-domain fusion features of the fault event; a feature matching module configured to match the cross-domain fusion features with a fault knowledge graph to obtain a candidate fault set corresponding to the fault event; a fault identification module configured to obtain a confidence of each candidate fault in the candidate fault set according to the cross-domain fusion features by using a trained fault identification model. An outcome obtaining module is configured to obtain a fault tracing result of the fault event according to the fault location result and the confidence of each candidate fault.

8. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that, The computer program, when executed by the processor, implements the steps of the method of any one of claims 1 to 6.

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

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

Images