Power transformer discharge fault diagnosis method based on neural network

Abstract

The invention relates to the technical field of power transformer discharge fault diagnosis in a power supply system, in particular to a power transformer discharge fault diagnosis method based on a neural network, and the method comprises the steps: collecting a partial discharge signal, characteristic gas data in oil and power supply system operation parameters in real time through a sensor, and unifying timestamps; statistical aggregation is performed on the partial discharge signals, and standardization processing is performed on the electrical quantity parameters and the discharge characteristic parameters to generate fusion characteristic vectors; constructing an expansion mapping relation library containing laboratory simulation data and actual operation data; training a diagnosis model by adopting a mixed structure of a convolutional neural network and a long-short-term memory network; and deploying the model in an edge computing device, and dynamically verifying and optimizing the performance of the model through online learning. According to the method, the problems of low multi-source data fusion efficiency and insufficient generalization ability of a neural network model to a complex discharge mode are solved, and the operation stability of a power supply system is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power transformer discharge fault diagnosis in a power supply system, and particularly to a power transformer discharge fault diagnosis method based on a neural network. Background Art

[0002] As a core device for voltage transformation and power transmission in a power supply system, the integrity of the internal insulation structure of a power transformer directly affects the continuity of power supply in a distribution network; a discharge fault is a common early abnormal state during the insulation deterioration process, usually caused by factors such as long-term operation aging of insulation materials, distortion of local electric field distribution, or accumulation of mechanical stress, and is manifested in forms such as corona discharge, surface discharge, or floating discharge. In the operation scenario of a power supply system, a discharge phenomenon will be accompanied by the generation of electrical signals (such as partial discharge pulses, high-frequency currents) and non-electrical signals (such as acoustic emission waves, infrared thermal radiation), and there is an internal correlation between these signal characteristics and the discharge type, location, and severity; by collecting multi-dimensional signals through a sensor array deployed on the transformer body and the power supply line, extracting discharge-related feature quantities (such as pulse amplitude, frequency distribution, time interval) through signal processing techniques such as filtering and denoising, and combining with real-time operation data of the power supply system (such as load current, voltage fluctuation, ambient temperature and humidity), and using a pattern recognition algorithm to analyze the mapping law between the feature quantities and the fault type, the fault type can be discriminated and the risk level can be evaluated at the initial stage of discharge development, providing a basis for the operation and maintenance decision-making of the power supply system and avoiding transformer outage and power supply interruption accidents caused by insulation breakdown.

[0003] In the operation scenario of the power supply system, the core reason for the low efficiency of multi-source data fusion such as partial discharge patterns, oil chromatography, and electrical quantities lies in the significant differences in data characteristics: partial discharge signals (such as current pulses collected by high-frequency current coil transformers and voltage fluctuations measured by capacitive voltage dividers) have high-frequency and instantaneous characteristics, while oil chromatography data (such as the content of gases like hydrogen and acetylene) needs to be obtained through periodic oil sample collection (once every 30 minutes for partial discharge tests), and the mismatch in time scales makes it difficult to align the data; at the same time, the physical dimensions of electrical quantity parameters (such as load current and voltage amplitude) and discharge characteristic parameters (such as discharge energy and pulse frequency) are different (the former reflects the operation state of the system, and the latter reflects the degree of insulation deterioration), and the lack of a unified feature standardization method makes it difficult to directly establish an effective correlation. The insufficient generalization ability of the neural network model for complex discharge patterns stems from the limitations of the training data: the existing mapping relationship library is generated based on laboratory simulation experiments (such as partial discharge patterns, spark discharge patterns, and arc discharge with fixed gaps under the needle-plate electrode model), while in the actual power supply system, transformers may face working conditions not covered in the experiments, such as dynamic electric field distortion (such as voltage fluctuations caused by load mutations) and mixed discharge patterns (such as partial discharge superimposed on floating discharge). For example, in the experiment, arc discharge is defined as "a continuous and stable through-discharge channel remains for 10 seconds", but in actual operation, intermittent arcs (the discharge channel exists intermittently) may occur due to vibration, and such data is not included in the training set, so the model cannot accurately identify it, resulting in diagnostic deviations for complex fault types and ultimately reducing the operation stability of the power supply system. Summary of the Invention

[0004] Aiming at the deficiencies of the prior art, the present invention provides a method for diagnosing discharge faults of power transformers based on neural networks, which solves the problems of low data fusion efficiency of partial discharge patterns, oil chromatography, electrical quantities, etc. during the operation of the power supply system, and the insufficient generalization ability of the neural network model for complex discharge patterns, resulting in a reduction in the operation stability of the power supply system.

[0005] To solve the above technical problems, the specific technical solutions of the present invention are as follows: The method for diagnosing discharge faults of power transformers based on neural networks provided by the present invention includes: Step 1, through a sensor group deployed on the power transformer body and the power supply line, collect partial discharge signals, characteristic gas data in oil, and power supply system operation parameters in real time, calculate discharge characteristic parameters based on the partial discharge signals, and mark time stamps. The discharge characteristic parameters include pulse amplitude, frequency distribution, and number of discharges per unit time; Step 2, statistically aggregate the partial discharge signals collected in Step 1 according to the sampling period of the characteristic gas data in oil, and perform standardization processing on the power supply system operation parameters collected in Step 1 and the discharge characteristic parameters calculated in Step 1 to generate a fused feature vector; Step 3: Collect laboratory simulation data, where the laboratory simulation data includes fusion feature vectors corresponding to partial discharge patterns, spark discharge patterns, mixed discharge and other patterns, and collect actual operation data through the temporary monitoring device of the actually operating transformer. The actual operation data includes fusion feature vectors corresponding to patterns such as intermittent arcs under scenarios of sudden load changes and drastic temperature and humidity changes. Based on the fusion feature vectors generated in Step 2, the laboratory simulation data, and the actual operation data, construct an extended mapping relationship library. The extended mapping relationship library includes the association data between the fusion feature vectors and the discharge patterns, and the discharge patterns include mixed discharge and intermittent arc; Step 4: Use the fusion feature vectors in the extended mapping relationship library constructed in Step 3 as the input, and adopt a hybrid neural network structure including a convolutional neural network branch and a long short-term memory network branch to train the diagnostic model. The diagnostic model outputs the probability distribution of the discharge severity; Step 5: Deploy the diagnostic model trained in Step 4 to the edge computing device, use the fusion feature vectors generated in real time in Step 2 for discharge fault diagnosis, and update the parameters of the diagnostic model trained in Step 4 through an online learning mechanism, and simultaneously dynamically verify the diagnostic performance of the diagnostic model.

[0006] Further, for the power transformer discharge fault diagnosis method based on a neural network of the present invention, the sensor group includes a high-frequency current coil transformer for collecting partial discharge signals, a resistor-capacitor voltage divider for measuring the discharge voltage fluctuation, an on-line gas sensor for monitoring the characteristic gas data in the oil, and a SCADA system for obtaining the operation parameters of the power supply system; Among them, the high-frequency current coil transformer is installed on the grounding wire of the transformer bushing end screen, and the partial discharge signal in Step 1 is obtained by collecting the partial discharge high-frequency current signal; The resistor-capacitor voltage divider is connected in parallel to the high-voltage winding outlet end, and the collection of the partial discharge signal is supplemented by measuring the voltage fluctuation signal during the discharge process; The on-line gas sensor is integrated into the transformer oil flow circuit, and the characteristic gas data in the oil in Step 1 is obtained by monitoring the volume fractions of hydrogen, acetylene, total hydrocarbons, carbon monoxide, and carbon dioxide; The SCADA system is accessed through the Modbus protocol, and the operation parameters of the power supply system in Step 1 are obtained by obtaining the load current, voltage amplitude, and ambient temperature and humidity; The sensor group and the SCADA system are time synchronized through GPS or the network time protocol to make the timestamps of the partial discharge signal, the characteristic gas data in the oil, and the operation parameters consistent.

[0007] Further, the power transformer discharge fault diagnosis method based on a neural network according to the present invention further includes: local discharge signal statistical aggregation. Taking the sampling period of the characteristic gas data in the oil as a window, sliding window statistics are performed on the high-frequency current and voltage signals collected by the high-frequency current coil transformer and the capacitive voltage divider, and the average pulse amplitude, frequency distribution variance, and number of discharges per unit time within the window are extracted, converting the high-frequency instantaneous local discharge signal in step 1 into a statistic matching the time scale of the characteristic gas data in the oil; For the electrical quantity parameters and discharge characteristic parameters collected in step 1, standardization processing is performed. The load current obtained from the SCADA system and the discharge energy calculated based on the voltage and current signals are processed using the Z-score method to eliminate the dimensional differences of current and energy; Combining the characteristic gas data in the oil in step 1 with the standardized discharge characteristic parameters in step 2, the fusion feature vector is formed by combining the volume fraction of the characteristic gas in the oil monitored by the on-line gas sensor and the standardized discharge energy, and is used as the structured input data for associating the discharge mode in the extended mapping relationship library in step 3.

[0008] Further, the power transformer discharge fault diagnosis method based on a neural network according to the present invention further includes: laboratory simulation data is generated through a needle-plate electrode model, a floating electrode model, and a vibration interference device, including mixed discharge data of the fusion feature vector and the corresponding partial discharge mode, spark discharge mode, arc discharge mode, and partial discharge superimposed floating discharge model; Actual operation data is collected through a temporary monitoring device for discharge data in scenarios of load mutation and drastic changes in ambient temperature and humidity. The discharge data includes the fusion feature vector generated in step 2 and the corresponding complex discharge modes such as mixed discharge and intermittent arc; Based on the association requirement between the fusion feature vector generated in step 2 and the discharge mode, the extended mapping relationship library is composed of 70% laboratory simulation data and 30% actual operation data. Each piece of data includes the fusion feature vector generated in step 2, the discharge mode, the discharge severity, and the working condition label, providing a sample set including the association between multi-scenario fusion features and discharge modes for the neural network model training in step 4.

[0009] Further, the power transformer discharge fault diagnosis method based on a neural network according to the present invention further includes: the convolutional neural network branch inputs the temporal features of the local discharge signal in the extended mapping relationship library in step 3, extracts local features through 1 convolutional layer, reduces the dimension through 1 max pooling layer, and captures the short-term mode of the local discharge signal in the fusion feature vector in step 2; The long - short - term memory network branch inputs the time - series data of the characteristic gases in oil in the extended mapping relationship library in step 3, captures the long - term change trend of the gas content through 1 LSTM layer, and extracts the time - series correlation features of the characteristic gases in oil in the fusion feature vector in step 2; Based on the multi - modal fusion requirements of step 4 model training, the outputs of the CNN branch and the LSTM branch of the hybrid neural network are fused through a fully - connected layer, and the probability distribution of the discharge severity is output through the Softmax activation function, providing a classification basis for multi - dimensional feature fusion in the training of the diagnosis model in step 4.

[0010] Furthermore, the method for diagnosing power transformer discharge faults based on a neural network according to the present invention further includes: deploying the diagnosis model trained in step 4 on a substation edge - computing device, receiving the fusion feature vector generated in real - time in step 2, and outputting the probability distribution of the discharge severity and the corresponding operating condition label; The online learning mechanism accumulates 100 groups of new fusion feature vectors generated in step 2 through the edge - computing device, and fine - tunes the weight parameters of the fully - connected layer and the output layer of the hybrid neural network in step 5 using the stochastic gradient descent algorithm.

[0011] Furthermore, the method for diagnosing power transformer discharge faults based on a neural network according to the present invention further includes: the dynamic verification process extracts 100 groups of fusion feature vectors corresponding to complex operating conditions from the newly generated fusion feature vectors in step 2 every month as a test set, and inputs them into the diagnosis model deployed on the edge - computing device in step 5 to evaluate the diagnostic accuracy of the discharge severity; If the accuracy rate fails to reach the preset threshold for two consecutive months, re - execute step 3 and step 4 to generate an updated diagnosis model and replace the old model.

[0012] Advantages of the present invention; Through multi-source data collaborative acquisition and feature standardization processing, the present invention solves the problem of low fusion efficiency caused by time-scale mismatch and dimensional differences in data such as partial discharge patterns, oil chromatography, and electrical quantities: Step 1 realizes data time alignment through sensor deployment and time synchronization, and Step 2 generates a structured fusion feature vector through statistical aggregation and Z-score standardization to provide effective input for subsequent analysis; by expanding the mapping relationship library and designing a multi-modal neural network structure, the generalization ability of the model for complex discharge patterns is improved: Step 3 integrates laboratory simulation data and actual operation data to cover multi-scenario discharge patterns, and Step 4 uses a hybrid network of CNN and LSTM to fuse short-term discharge patterns and long-term gas trends to enhance the recognition ability for complex patterns such as hybrid discharge and intermittent arcs; through an online learning and dynamic verification mechanism, the adaptability of the model is maintained: Step 5 the model deployed on the edge computing device fine-tunes the parameters through incremental learning, combines periodic test set evaluation and model update, continuously adapts to the dynamic operating environment, reduces the risk of insulation breakdown caused by discharge fault recognition deviation, and ultimately improves the operating stability of the power supply system. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] In order to more clearly illustrate the technical solutions of the present invention, the drawings required for use in the embodiments will be briefly introduced below. Obviously, for those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on the drawings.

[0014] Figure 1 It is a flowchart of a method for diagnosing discharge faults in a power transformer based on a neural network provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0015] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the specific embodiments and corresponding drawings of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention. The technical solutions provided in each embodiment of the present invention are described in detail below with reference to the drawings. To better understand the objectives of the present invention, the present invention will be further described in detail below.

[0016] Please refer to Figure 1 , a method for diagnosing discharge faults in a power transformer based on a neural network provided by the present invention, Step 1: Through the sensor group deployed on the power transformer body and the power supply line, locally-discharged signals, characteristic gas data in oil, and power supply system operation parameters are collected in real time. Discharge characteristic parameters are calculated based on the locally-discharged signals and time stamps are marked. The discharge characteristic parameters include pulse amplitude, frequency distribution, and number of discharges per unit time. Step 2: The locally-discharged signals collected in Step 1 are statistically aggregated according to the sampling period of the characteristic gas data in oil. The power supply system operation parameters collected in Step 1 and the discharge characteristic parameters calculated in Step 1 are standardized to generate a fused feature vector. Step 3: Laboratory simulation data is collected. The laboratory simulation data includes fused feature vectors corresponding to locally-discharged patterns, spark discharge patterns, mixed discharge, etc. Actual operation data is collected through the temporary monitoring device of the actually-operating transformer. The actual operation data includes fused feature vectors corresponding to patterns such as intermittent arcs in scenarios of sudden load changes and drastic temperature and humidity changes. An extended mapping relationship library is constructed based on the fused feature vector generated in Step 2, the laboratory simulation data, and the actual operation data. The extended mapping relationship library includes associated data between the fused feature vector and the discharge pattern. The discharge patterns include mixed discharge and intermittent arc. Step 4: Using the fused feature vector in the extended mapping relationship library constructed in Step 3 as the input, a diagnostic model is trained using a hybrid neural network structure including a convolutional neural network branch and a long short-term memory network branch. The diagnostic model outputs the probability distribution of the discharge severity. Step 5: The diagnostic model trained in Step 4 is deployed on the edge computing device. The fused feature vector generated in real time in Step 2 is used for discharge fault diagnosis, and the parameters of the diagnostic model trained in Step 4 are updated through an online learning mechanism. At the same time, the diagnostic performance of the diagnostic model is dynamically verified.

[0017] The power transformer discharge fault diagnosis method based on neural network provided by the present invention realizes the accurate diagnosis of the power transformer discharge fault through a closed-loop process of multi-source data collection, feature fusion, model training, and dynamic optimization. The specific implementation and logical relationship of each step are as follows: First, in step 1, multi-source data is collected in real time by a sensor group deployed on the power transformer body and the power supply line. Specifically, the collection of partial discharge signals relies on high-frequency current coil transformers and capacitive voltage dividers: the high-frequency current coil transformer is installed on the grounding wire of the transformer bushing end screen to capture the high-frequency current signals generated by partial discharges (frequency range 100 kHz - 10 MHz); the capacitive voltage divider is connected in parallel to the outgoing end of the high-voltage winding to synchronously measure the voltage fluctuation signals during the discharge process. The collection of characteristic gas data in oil is completed by an on-line gas sensor integrated in the transformer oil flow circuit, which continuously monitors the volume fractions of 5 characteristic gases, namely hydrogen, acetylene, total hydrocarbons, carbon monoxide, and carbon dioxide, through photoacoustic spectroscopy technology. The operating parameters of the power supply system (such as load current, voltage amplitude, ambient temperature and humidity) are obtained by accessing the SCADA system through the Modbus protocol. The sensor group and the SCADA system are time-synchronized through GPS or the Network Time Protocol (NTP) to unify the timestamps of partial discharge signals (sampling frequency 1 MHz), characteristic gas data in oil (sampling period 5 minutes), and operating parameters (sampling period 1 minute), providing a synchronization reference for subsequent data alignment. This step covers discharge signals (electrical quantities), characteristic gases in oil (non-electrical quantities), and system operating states (environmental parameters) through the sensor group, and is the basis for subsequent feature processing.

[0018] In step 2, statistical aggregation and standardization processing are performed on the multi-source data collected in step 1 to generate a fused feature vector. Regarding the high-frequency instantaneous characteristics of partial discharge signals (such as high-frequency current and voltage pulses), with the 5-minute sampling period of the characteristic gas data in oil as the window, sliding window statistics are performed on the high-frequency current and voltage signals to extract the average pulse amplitude within the window (reflecting the discharge intensity), the variance of the frequency distribution (reflecting the discharge density), and the number of discharges per unit time (reflecting the discharge frequency), converting the high-frequency instantaneous signals into low-frequency statistics consistent with the time scale of the characteristic gas data in oil. For electrical quantity parameters (such as the load current obtained from the SCADA system) and discharge characteristic parameters (such as the discharge energy calculated by integrating voltage and current signals), the Z-score standardization method is used to eliminate the dimensional differences between current (A) and energy (J), making the two types of parameters comparable. Finally, the volume fractions of 5 characteristic gases in oil and the standardized discharge energy are combined to form a 6-dimensional fused feature vector (dimension definition: [H2, C2H2, total hydrocarbons, CO, CO2, discharge energy]), serving as the structured input data for subsequent mapping library construction. This step solves the key obstacles in multi-source data fusion through time scale alignment and dimension unification, providing effective input for model training.

[0019] Step 3 constructs an extended mapping relationship library based on the fused feature vectors generated in Step 2, covering multi-scenario discharge patterns. Laboratory simulation data is generated through a needle-plate electrode model (generating partial discharge patterns, spark discharge patterns, and arc discharge data), a floating electrode model (simulating floating discharges caused by winding looseness), and a vibration interference device (applying mechanical vibrations of 0.5g - 2g), including hybrid discharge data of a partial discharge superposed with a floating discharge model; actual operation data is collected through a temporary monitoring device for discharge data in scenarios of load mutations (such as voltage fluctuations of ±15%) and drastic changes in ambient temperature and humidity (temperature from 25°C to 85°C, humidity from 30% to 90%), and the content includes fused feature vectors and corresponding complex patterns such as hybrid discharges and intermittent arcs. The extended mapping relationship library is composed of a mixture of 70% laboratory data and 30% actual operation data. Each piece of data contains a fused feature vector, a discharge pattern (partial discharge pattern, spark discharge pattern, arc discharge pattern, hybrid discharge, intermittent arc), a discharge severity level (mild / moderate / severe), and a working condition label (such as "load mutation", "vibration interference"), providing a multi-scenario sample set for the model training in Step 4. This step solves the problem of incomplete coverage of working conditions in the traditional mapping library by supplementing actual operation data.

[0020] Step 4 uses the fused feature vectors in the extended mapping relationship library in Step 3 as inputs and trains a diagnostic model using a hybrid neural network structure. The convolutional neural network (CNN) branch inputs the temporal features of the partial discharge signal (average pulse amplitude, variance of frequency distribution, number of discharges), extracts local features (such as local patterns of pulse waveforms) through a 1-layer 3×3 convolutional layer, reduces the dimension through a 1-layer 2×2 max-pooling layer, and captures the short-term patterns of the discharge signal (such as changes in pulse sequence density); the long short-term memory network (LSTM) branch inputs the time series data of the characteristic gases in the oil (volume fractions of 5 gases), and captures the long-term change trends of gas content (such as the cumulative rate of acetylene content over time) through an LSTM layer with 16 hidden layer nodes. After the outputs of the two branches are fused through a fully connected layer (with 32 nodes), the probability distribution of the discharge severity level (mild: partial discharge; moderate: spark discharge; severe: arc discharge / hybrid discharge) is output through the Softmax activation function. During the training process, the cross-entropy loss function is used in combination with the gradient descent method to optimize the weight parameters, and a Dropout layer (dropout rate of 20%) is introduced to suppress overfitting. The early stopping mechanism is triggered when the error rate of the validation set does not decrease for 5 consecutive iterations. This step solves the problem of insufficient recognition of complex discharge patterns by traditional models through the fusion of short-term discharge patterns and long-term gas trends using a multi-modal network structure.

[0021] Step 5 deploys the diagnostic model trained in Step 4 to the substation edge computing device to achieve real-time diagnosis and dynamic optimization. The edge computing device receives the fused feature vectors generated in real time in Step 2 and outputs the probability distribution of the discharge severity and the corresponding operating condition labels (such as "load mutation scenario"). To maintain the adaptability of the model to the dynamic environment, every time the edge computing device accumulates 100 groups of new fused feature vectors (including complex operating condition data such as mixed discharges and intermittent arcs), the weight parameters of the fully connected layer and the output layer are fine-tuned using the stochastic gradient descent algorithm (learning rate 0.001). Every month, 100 groups of complex operating condition data are extracted from the newly generated fused feature vectors as the test set to evaluate the diagnostic accuracy of the model (goal: mild / moderate / severe recognition accuracy ≥ 90%); if the threshold is not reached for two consecutive months, Steps 3 (expand the mapping relationship library and supplement new operating condition data) and Step 4 (adjust the network structure, such as increasing the number of LSTM hidden layer nodes) are re-executed to generate a new version of the model and replace the old model. This step forms a closed loop of "data acquisition - feature processing - model training - online optimization" through online deployment, incremental learning, and dynamic verification.

[0022] Specifically, for the power transformer discharge fault diagnosis method based on a neural network according to the present invention, the sensor group includes four types, respectively corresponding to the partial discharge signals, dissolved gas analysis data in oil, and power supply system operation parameters collected in Step 1: a high-frequency current coil transformer for collecting partial discharge signals, a resistor-capacitor voltage divider for measuring the discharge voltage fluctuation, an online gas sensor for monitoring the dissolved gas analysis data in oil, and a SCADA system for obtaining the power supply system operation parameters; Among them, the high-frequency current coil transformer is installed on the grounding wire of the transformer bushing end screen to obtain the partial discharge signals in Step 1 by collecting the high-frequency current signals of partial discharges; the resistor-capacitor voltage divider is connected in parallel to the high-voltage winding outlet terminal to supplement the collection of partial discharge signals by measuring the voltage fluctuation signals during the discharge process; the online gas sensor is integrated into the transformer oil flow circuit to obtain the dissolved gas analysis data in oil in Step 1 by monitoring the volume fractions of hydrogen, acetylene, total hydrocarbons, carbon monoxide, and carbon dioxide; the SCADA system is accessed through the Modbus protocol to obtain the power supply system operation parameters in Step 1 by acquiring the load current, voltage amplitude, and ambient temperature and humidity. The sensor group and the SCADA system are time-synchronized through GPS or the network time protocol to make the timestamps of the partial discharge signals, dissolved gas analysis data in oil, and operation parameters consistent, providing a synchronization reference for "stamping timestamps on the data" in Step 1.

[0023] In step 1 of the present invention, in the process of deploying the sensor group and collecting data, by specifically selecting the sensor type, clarifying the installation location, and adopting a unified time synchronization mechanism, multi-source collaborative collection of partial discharge signals, characteristic gas data in oil, and power supply system operation parameters is achieved, providing basic support for subsequent data processing.

[0024] The high-frequency current coil transformer, as the core acquisition device for partial discharge signals, is installed on the grounding wire of the end screen of the power transformer bushing. This position is the main path for the partial discharge current to discharge to the ground, and the high-frequency current (frequency range 100 kHz - 10 MHz) generated by the discharge will propagate along this path. Therefore, deploying the high-frequency current coil transformer here can effectively capture the current pulse signal during the discharge process and directly monitor the partial discharge phenomenon.

[0025] The capacitive voltage divider is connected in parallel to the outlet of the high-voltage winding and is used to supplement the acquisition of voltage fluctuation signals during partial discharge. The outlet of the high-voltage winding is the area where the internal electric field of the transformer is concentrated. Partial discharge will cause transient voltage changes at this position. By connecting the capacitive voltage divider in parallel, the amplitude and frequency characteristics of the voltage fluctuation can be measured in real time, forming a complement with the current signal collected by the high-frequency current coil transformer, and jointly constituting the electrical characteristic data set of partial discharge.

[0026] The on-line gas sensor is integrated into the transformer oil flow circuit (such as the oil flow pipeline connecting the oil conservator and the main body), and continuously monitors the volume fractions of 5 characteristic gases, namely hydrogen (H2), acetylene (C2H2), total hydrocarbons (CH4 + C2H4 + C2H2, etc.), carbon monoxide (CO), and carbon dioxide (CO2), through photoacoustic spectroscopy detection technology. The oil flow circuit is the main channel for transformer oil circulation. Deploying the sensor here can obtain the dynamic changes of the dissolved gases in the oil in real time, reflect the generation and diffusion laws of gases during the deterioration process of insulating materials, and provide a reliable source for the collection of characteristic gas data in oil.

[0027] The SCADA system (Supervisory Control And Data Acquisition system) accesses the power supply system through the Modbus protocol and is used to obtain operation parameters such as load current, voltage amplitude, and ambient temperature and humidity. The Modbus protocol is a widely used serial communication protocol in the industrial field, supporting multi-device access and real-time data transmission. By accessing the SCADA system through this protocol, the real-time operation status data of the power supply system can be stably obtained, including current / voltage parameters reflecting the load situation and ambient temperature and humidity parameters affecting the insulation performance, providing a basis for subsequent analysis of the correlation between discharge faults and system operation status.

[0028] To align data in the time dimension, the sensor group and the SCADA system perform clock synchronization through GPS (Global Positioning System) or NTP (Network Time Protocol). GPS provides a high-precision time reference through satellite signals (error ≤ 1 μs) and is applicable to substation scenarios without network coverage; NTP synchronizes the clocks of various devices through the network (error ≤ 10 ms) and is applicable to substation scenarios with a deployed local area network. Through unified timestamp annotation, partial discharge signals (sampling frequency 1 MHz), dissolved gas in oil characteristic gas data (sampling period 5 minutes), and operating parameters (sampling period 1 minute) can be aligned in the time dimension during subsequent processing, avoiding feature matching deviations caused by time asynchronization, and providing time consistency guarantee for the statistical aggregation and feature fusion in Step 2.

[0029] Specifically, for the partial discharge signals collected in Step 1 in the power transformer discharge fault diagnosis method based on a neural network of the present invention, the statistical aggregation of the partial discharge signals uses the sampling period of the dissolved gas in oil characteristic gas data as the window, and performs sliding window statistics on the high-frequency current and voltage signals collected by the high-frequency current coil transformer and the capacitive voltage divider, extracting the average pulse amplitude, frequency distribution variance, and number of discharges per unit time within the window, and converting the high-frequency instantaneous partial discharge signals in Step 1 into statistics that match the time scale of the dissolved gas in oil characteristic gas data; For the electrical quantity parameters and discharge characteristic parameters collected in Step 1, the standardization process uses the Z-score method for the load current (belonging to the electrical quantity parameters in Step 1) obtained by the SCADA system and the discharge energy calculated based on the voltage and current signals (belonging to the discharge characteristic parameters in Step 1), eliminating the dimensional differences between current (A) and energy (J), and making the two types of parameters comparable; Combining the dissolved gas in oil characteristic gas data in Step 1 with the standardized discharge characteristic parameters in Step 2, the fusion feature vector is formed by combining the volume fractions of 5 dissolved gases in oil monitored by the on-line gas sensor (belonging to the dissolved gas in oil characteristic gas data in Step 1) and the standardized discharge energy, and serves as the structured input data for associating discharge patterns in the extended mapping relationship library in Step 3.

[0030] In the process of statistical aggregation, standardization processing, and fusion feature vector generation of multi-source data in Step 2 of the present invention, through time scale alignment, dimension unification, and multi-dimensional feature integration, it solves the fusion difficulties caused by the characteristic differences of partial discharge signals, dissolved gas in oil characteristic gas data, and electrical quantity parameters, and provides structured input data for the subsequent mapping library construction and model training.

[0031] Regarding the partial discharge signals collected in Step 1, there is a problem of time-scale mismatch between their high-frequency transient characteristics (such as the 1MHz sampled current signal collected by a high-frequency current coil transformer and the voltage fluctuation signal collected by a capacitive voltage divider) and the low-frequency periodicity of the characteristic gas data in oil (such as the 5-minute sampling period of an on-line gas sensor). Therefore, in this step, a 5-minute sampling period of the characteristic gas data in oil is used as a sliding window to perform statistical aggregation on the high-frequency current and voltage signals: within each 5-minute window, all high-frequency signal points within the window are traversed, and the average value of the pulse amplitude (reflecting the overall level of the discharge intensity during this time period), the variance of the frequency distribution (reflecting the density of the discharge pulses on the time axis, the larger the variance, the more concentrated the discharge), and the number of discharges per unit time (i.e., the number of pulses within the window divided by 5 minutes, reflecting the frequency of the discharge activity) are calculated. Through this operation, the high-frequency transient partial discharge signals are converted into low-frequency statistics (such as a set of statistical values every 5 minutes) that are consistent with the time scale of the characteristic gas data in oil, providing a basis for time alignment for subsequent multi-source data fusion.

[0032] Regarding the electrical quantity parameters and discharge characteristic parameters collected in Step 1, the two cannot be directly correlated and analyzed due to different physical dimensions (the unit of load current is ampere, and the unit of discharge energy is joule). In this step, the Z-score normalization method is used to process the two types of parameters: first, calculate the mean and standard deviation of the load current (the real-time acquisition value from the SCADA system), and convert each current value into a dimensionless value of "(current value - mean) / standard deviation"; at the same time, calculate the mean and standard deviation of the discharge energy (reflecting the total energy released during the discharge process) obtained by integrating the voltage and current signals, and convert each energy value into the same dimensionless form. Through this normalization operation, the load current and the discharge energy are mapped to the same dimension space (mean is 0, standard deviation is 1), making the two types of parameters comparable and providing a guarantee for numerical consistency for subsequent feature fusion.

[0033] After completing the statistical aggregation and standardization processes, this step combines the in-oil characteristic gas data in Step 1 with the standardized discharge characteristic parameters to generate a fused feature vector. Specifically, the volume fractions of five in-oil characteristic gases (hydrogen H2, acetylene C2H2, total hydrocarbons, carbon monoxide CO, and carbon dioxide CO2) monitored by on-line gas sensors, which are non-electrical characteristics directly reflecting the degree of insulation deterioration, are combined with the standardized discharge energy (an electrical characteristic reflecting the discharge intensity) in a fixed order to form a 6-dimensional fused feature vector (dimension definition: [H2 volume fraction, C2H2 volume fraction, total hydrocarbon volume fraction, CO volume fraction, CO2 volume fraction, standardized discharge energy]). This fused feature vector serves as the structured input data for the extended mapping relationship library in Step 3, integrating the physical meanings (gas content, discharge intensity) and numerical characteristics (time alignment, consistent dimension) of multi-source data into a single feature space, providing effective input for establishing the correlation between the fused features and discharge patterns in the subsequent steps.

[0034] Specifically, the method for diagnosing discharge faults in power transformers based on neural networks described in the present invention To cover the multi-scenario data required for constructing the mapping library in Step 3, the laboratory simulation data is generated through a needle-plate electrode model (generating partial discharge patterns, spark discharge patterns, and arc discharge data), a floating electrode model (simulating floating discharges caused by winding looseness), and a vibration interference device (applying mechanical vibrations), including the fused feature vector generated in Step 2 and the corresponding mixed discharge data of partial discharge patterns, spark discharge patterns, arc discharge patterns, and a model of partial discharge superimposed on floating discharges; To supplement the actual operating condition data not covered in the laboratory, the actual operating data is collected through a temporary monitoring device for discharge data under load mutation (voltage fluctuation scenarios) and drastic changes in ambient temperature and humidity. The discharge data includes the fused feature vector generated in Step 2 and the corresponding complex discharge patterns such as mixed discharges and intermittent arcs; Based on the correlation requirements between the fused feature vector generated in Step 2 and the discharge patterns, the extended mapping relationship library is composed of a mixture of 70% laboratory simulation data and 30% actual operating data. Each piece of data includes the fused feature vector generated in Step 2, the discharge pattern (partial discharge pattern, spark discharge pattern, arc discharge pattern, mixed discharge, intermittent arc), the discharge severity (mild / moderate / severe), and the operating condition label (load mutation, vibration interference), directly providing a sample set including the correlation between multi-scenario fused features and discharge patterns for the neural network model training in Step 4.

[0035] In the construction process of the extended mapping relationship library in step 3 of the present invention, through the collaborative acquisition of laboratory simulation data and actual operation data, covering multi-scenario discharge modes, it solves the problem of insufficient model generalization ability caused by incomplete coverage of working conditions in the traditional mapping library, and provides multi-dimensional sample support for the neural network model training in step 4.

[0036] The generation of laboratory simulation data aims to cover basic and typical hybrid discharge modes, and is realized through the collaboration of three types of experimental devices: The needle-plate electrode model uses a pure tungsten needle electrode and a brass plate electrode to form a fixed gap (such as a gap distance of 2 mm), simulating partial discharge (with low discharge energy and stable pulse frequency), spark discharge (with a short-circuited discharge channel and medium energy), and arc discharge (with a continuous through-discharge channel and high energy) caused by electric field concentration inside the insulating material; The floating electrode model adds a movable floating electrode on the basis of the needle-plate electrode (simulating a floating conductor caused by the loosening of the winding spacer block), and by adjusting the distance between the floating electrode and the main electrode (such as 1 - 3 mm), generates hybrid discharge data of the partial discharge superposed floating discharge model (characterized by an increase in pulse amplitude fluctuation and irregular frequency); The vibration interference device applies mechanical vibration of 0.5g - 2g (g is the acceleration of gravity) to the electrode system through an electromagnetic vibrator, simulating the intermittent phenomenon of the discharge channel caused by mechanical stress during the operation of the transformer, and generating intermittent arc data (the discharge channel exists for 2 - 8 seconds, and the energy fluctuates periodically). While generating the discharge signal, the above experimental devices synchronously collect the fusion feature vectors generated in step 2 (such as the gas volume fraction and standardized discharge energy within a 5-minute window), and label the corresponding discharge modes (partial discharge mode, spark discharge, etc.), forming a laboratory simulation data set.

[0037] The acquisition of actual operation data aims to supplement complex working conditions not covered in the laboratory, and is realized by deploying temporary monitoring devices on the operating transformer. The acquisition scenarios include: load mutation scenarios (such as voltage fluctuations of ±15% caused by the parallel / separation operation of the transformer), at this time, the internal electric field distribution of the transformer changes dynamically, and it is easy to trigger hybrid discharge (partial discharge superposed floating discharge); scenarios of drastic changes in ambient temperature and humidity (such as in a high-temperature and high-humidity environment in summer, the temperature rises from 25°C to 85°C, and the humidity rises from 30% to 90%), at this time, the performance degradation of the insulating material accelerates, and it is easy to trigger intermittent arc (the discharge channel intermittently exists due to thermal expansion / contraction). The temporary monitoring device synchronously collects the fusion feature vectors generated in step 2 (such as the rapid increase in the volume fractions of hydrogen and acetylene and the periodic fluctuation of the standardized discharge energy in a high-temperature and high-humidity environment), and identifies the corresponding complex discharge modes (hybrid discharge, intermittent arc) through manual annotation or an expert system, forming an actual operation data set.

[0038] The construction of the extended mapping relationship library is based on the association requirements between the fused feature vectors generated in step 2 and the discharge patterns. Laboratory simulation data (accounting for 70%) and actual operation data (accounting for 30%) are mixed and integrated. The ratio of 7:3 is selected to balance the controllability of laboratory data (clear discharge patterns and stable features) and the authenticity of actual data (covering dynamic working conditions and complex features): 70% of laboratory data enables sufficient learning of the basic discharge patterns, and 30% of actual data supplements the feature diversity under complex working conditions. Each piece of data contains four parts: the fused feature vectors generated in step 2 (6-dimensional, such as [H2 = 50 ppm, C2H2 = 10 ppm, total hydrocarbons = 80 ppm, CO = 200 ppm, CO2 = 1500 ppm, normalized discharge energy = 1.2]), discharge patterns (partial discharge pattern, spark discharge pattern, arc discharge pattern, mixed discharge, intermittent arc), discharge severity (mild: partial discharge; moderate: spark discharge; severe: arc discharge / mixed discharge), and working condition labels (such as "load mutation", "vibration interference", "high temperature and high humidity"). Among them, the working condition labels are used to label the specific scenarios where the data is generated (such as the scenario corresponding to a voltage fluctuation of ±15% for "load mutation"), helping the model learn the discharge rules under different working conditions.

[0039] Specifically, for the power transformer discharge fault diagnosis method based on neural network described in the present invention, as the core input processing module in step 4 of model training, the convolutional neural network (CNN) branch inputs the temporal features of the partial discharge signals in the extended mapping relationship library in step 3 (average pulse amplitude, variance of frequency distribution, number of discharges, which are components of the fused feature vectors in step 2), extracts local features (such as local patterns of pulse waveforms) through 1 convolutional layer, reduces the dimension through 1 max-pooling layer (retaining key features), and captures the short-term patterns of the partial discharge signals in the fused feature vectors in step 2 (such as changes in pulse sequence density); As another core input processing module in step 4 of model training, the long short-term memory network (LSTM) branch inputs the time series data of the characteristic gases in oil in the extended mapping relationship library in step 3 (volume fractions of hydrogen, acetylene, total hydrocarbons, carbon monoxide, and carbon dioxide, which are components of the fused feature vectors in step 2), captures the long-term change trends of gas contents (such as the cumulative rate of acetylene content over time) through 1 LSTM layer, and extracts the temporal correlation features of the characteristic gases in oil in the fused feature vectors in step 2; Based on the multi-modal fusion requirements of step 4 model training, the outputs of the CNN branch and the LSTM branch of the hybrid neural network are fused through a fully connected layer (integrating short-term discharge patterns and long-term gas trends), and the probability distribution of discharge severity is output through the Softmax activation function (mild: partial discharge; moderate: spark discharge; severe: arc discharge / hybrid discharge), providing a classification basis for multi-dimensional feature fusion in the diagnostic model training of step 4.

[0040] In the construction and training process of the hybrid neural network in step 4 of the present invention, through the multi-modal branch design of the convolutional neural network (CNN) and the long short-term memory network (LSTM), the short-term patterns of partial discharge signals and the long-term trends of characteristic gases in oil are extracted respectively. After feature fusion, the probability distribution of discharge severity is output, providing a multi-dimensional classification basis for diagnostic model training.

[0041] The convolutional neural network (CNN) branch, as a short-term pattern extraction module for partial discharge signals, takes the time series features of partial discharge signals in the extended mapping relationship library of step 3 as its input, specifically including the average pulse amplitude (reflecting the average level of discharge intensity), the variance of frequency distribution (reflecting the concentration degree of discharge pulses on the time axis), and the number of discharges per unit time (reflecting the frequency of discharge activities) in the fusion feature vector of step 2. This branch uses 1 convolutional layer (such as a 3×3 convolutional kernel) to extract local features from the input time series features: the convolutional kernel slides on the input sequence, and captures the local correlation of features at adjacent time points through weighted summation operations (such as the increasing or decreasing pattern of pulse amplitudes in consecutive multiple time windows), and outputs a feature map containing the short-term pattern of partial discharge. Subsequently, 1 max pooling layer (such as a 2×2 pooling kernel) downsamples the feature map, retains the key features by selecting the maximum value in the local area (such as the distribution position of high-amplitude pulses), and at the same time reduces the feature dimension and computational complexity. Through this process, the CNN branch can effectively capture the short-term patterns of partial discharge signals in the fusion feature vector of step 2 (such as the sudden change of the pulse sequence from sparse to dense, and the change of the amplitude from stable to fluctuating), providing the model with feature information reflecting the instantaneous state of discharge.

[0042] The long short-term memory network (LSTM) branch, as the long-term trend extraction module for characteristic gases in oil, takes as input the time series data of characteristic gases in oil from the extended mapping relationship library in step 3, specifically including the volume fractions of hydrogen (H2), acetylene (C2H2), total hydrocarbons, carbon monoxide (CO), and carbon dioxide (CO2) in the fused feature vector in step 2. The LSTM layer performs sequence modeling on the time series data through internal forget gates, input gates, and output gate mechanisms: the forget gate determines whether to retain or discard historical information (such as the acetylene content in the previous 5 time windows), the input gate updates the new information at the current time step (such as the acetylene increment in the current window), and the output gate integrates historical and current information to generate a hidden state. Through this process, the LSTM branch can capture the long-term change trends of characteristic gases in oil (such as the continuous increase in acetylene content over time, reflecting the continuous thermal decomposition of insulating materials; the periodic fluctuations in carbon dioxide content, reflecting the impact of environmental temperature and humidity changes on gas diffusion), providing temporal correlation features for the model that reflect the process of insulation deterioration.

[0043] Based on the requirement of multi-modal feature fusion, the outputs of the CNN branch and the LSTM branch are integrated through a fully connected layer and then the probability distribution of the discharge severity is output through the Softmax activation function. Specifically, the short-term pattern features (such as sudden changes in pulse density) output by the CNN branch and the long-term trend features (such as continuous increase in acetylene) output by the LSTM branch are input into the fully connected layer (such as 32 neurons), and cross-modal fusion of the features is achieved through linear transformation and non-linear activation (such as the ReLU function) to extract the correlation relationship between the two types of features (such as sudden changes in pulse density accompanied by a rapid increase in acetylene, which may correspond to arc discharge). Finally, the Softmax activation function maps the fused features to the interval [0,1] and outputs the probability distribution of the 3 types of discharge severity (mild: partial discharge, probability 0.8; moderate: spark discharge, probability 0.15; severe: arc discharge / mixed discharge, probability 0.05), corresponding to the early, middle, and late stages of insulation deterioration respectively.

[0044] Specifically, the neural network-based power transformer discharge fault diagnosis method described in the present invention, as the specific implementation of "deploying the trained diagnosis model on the edge computing device" in step 5, the diagnosis model trained in step 4 is deployed on the substation edge computing device, receives the fused feature vector generated in real time in step 2 (which is the real-time output of "generating the fused feature vector for subsequent analysis" in step 2), and outputs the probability distribution of the discharge severity and the corresponding operating condition label (such as "load mutation scenario"); As a specific implementation of "updating model parameters through the online learning mechanism" in step 5, the online learning mechanism accumulates 100 groups of new fusion feature vectors generated in step 2 (including complex working condition data such as mixed discharge and intermittent arc, which are new data for "generating fusion feature vectors for subsequent analysis" in step 2) through edge computing devices, and uses the stochastic gradient descent algorithm to fine-tune the weight parameters of the fully connected layer and the output layer of the hybrid neural network in step 5 to maintain the adaptability of the diagnostic model in step 5 to the dynamic operating environment.

[0045] In step 5 of the present invention, the deployment and online learning mechanism of the diagnostic model realize the online application and dynamic optimization of the diagnostic model through the real-time data processing ability and incremental learning algorithm of edge computing devices, and solve the problem of the decline in the generalization ability of traditional models caused by changes in the operating environment.

[0046] The deployment of the diagnostic model aims to achieve real-time fault diagnosis, and deploys the hybrid neural network model trained in step 4 on the edge computing device of the substation (such as the edge computing unit of the integrated automation system). This device is usually deployed in the control cabinet in the substation and has low-latency data processing ability (response time ≤ 100 ms), and supports real-time communication with sensors and SCADA systems (through Modbus or MQTT protocols). The deployment process includes model parameter loading (transferring the trained weight parameters from the training server to the edge computing device), interface configuration (defining the input format of the fusion feature vector and the output protocol of the diagnostic result), and operating environment initialization (installing a deep learning framework such as TensorFlow Lite to optimize the model inference speed). After the deployment is completed, the edge computing device receives the fusion feature vectors generated in step 2 in real time (transmitted through industrial Ethernet, with the frequency consistent with the feature generation period in step 2, that is, one group every 5 minutes), calls the model inference interface to output the probability distribution of the discharge severity (such as "partial discharge probability 85%, spark discharge probability 10%, arc discharge / mixed discharge probability 5%") and the corresponding working condition labels (such as "load mutation scenario", "vibration interference scenario"), and displays the diagnostic results through the human-machine interface (HMI) for the reference of operation and maintenance personnel.

[0047] The online learning mechanism aims to keep the model adaptable to the dynamic operating environment and realizes the incremental update of model parameters through the local computing resources of edge computing devices. Specifically, the edge computing device is built with a data cache module. Every time a new set of fused feature vectors generated in step 2 (including complex working condition data such as mixed discharge and intermittent arc) is received, validity verification is carried out (such as excluding missing data or outliers), and the valid data is stored in the cache. When 100 sets of valid data are accumulated in the cache, the online learning process is triggered: the random gradient descent (SGD) algorithm is used to fine-tune the weight parameters of the fully connected layer and the output layer of the hybrid neural network (the learning rate is set to 0.001 to avoid overfitting). The reason for choosing to fine-tune the fully connected layer and the output layer is that these two layers are directly responsible for the fusion of multi-modal features and classification decisions, and are more sensitive to the adaptive adjustment of new working conditions; while the underlying feature extraction layers (convolutional layer, LSTM layer) of the CNN and LSTM branches do not need to be frequently adjusted because they have learned the basic discharge mode features. Through this fine-tuning operation, the model can gradually learn the feature patterns under new working conditions (such as the co-growth law of the volume fractions of hydrogen and acetylene in a high-temperature and high-humidity environment), and improve its adaptability to the dynamic operating environment.

[0048] Specifically, the method for diagnosing power transformer discharge faults based on a neural network according to the present invention is the specific implementation of "dynamically verifying the diagnostic performance of the model" in step 5. In the dynamic verification process, 100 sets of fused feature vectors corresponding to complex working conditions (mixed discharge, intermittent arc) are extracted from the fused feature vectors newly generated in step 2 every month (which belong to the real-time data of "generating fused feature vectors for subsequent analysis" in step 2) as the test set, and input into the diagnostic model deployed on the edge computing device in step 5 to evaluate the diagnostic accuracy of the discharge severity. Based on the goal of "keeping the model adaptable to the dynamic operating environment" in step 5, if the accuracy rate fails to reach the preset threshold for two consecutive months, steps 3 (expanding the mapping relationship library and supplementing the associated data between the fused feature vectors of new working conditions and discharge modes to provide more comprehensive training samples for step 4) and step 4 (adjusting the structure of the hybrid neural network, such as increasing the number of LSTM hidden layer nodes to improve the model's feature extraction ability for new working conditions) are re-executed, and an updated diagnostic model is generated and replaces the old model.

[0049] The process of dynamically verifying the diagnostic performance of the model in step 5 of the present invention enables the diagnostic model to continuously adapt to the dynamic operating environment of the power transformer through periodic test set evaluation and model iterative optimization, and solves the problem of the decay of diagnostic performance caused by changes in working conditions in traditional models.

[0050] The implementation of the dynamic verification process focuses on evaluating the generalization ability of the model under complex working conditions, and is specifically achieved through the following steps: Every month, from the newly generated fusion feature vectors in Step 2 (i.e., the 6D feature data collected and processed in real time, including the volume fractions of hydrogen and acetylene and the standardized discharge energy, etc.), complex working condition data is screened according to the working condition labels (such as samples labeled as "mixed discharge" or "intermittent arc"), and 100 groups are randomly sampled as the test set. The scale of choosing 100 groups is based on statistical significance considerations (a sample size of ≥30 can meet the requirements of the central limit theorem). After the test set is generated, it is input into the diagnostic model deployed on the substation edge computing device in Step 5, and the model inference interface is called to output the probability distribution of the discharge severity (such as "arc discharge / mixed discharge probability 90%"), and it is compared with the true severity manually labeled (determined by experts based on discharge characteristics and historical cases), and the diagnostic accuracy rate (number of correctly classified samples / total number of samples) is calculated. This accuracy rate reflects the model's recognition ability for recent complex working conditions and is the core indicator for evaluating the model's adaptability.

[0051] When the diagnostic accuracy rates for two consecutive months do not reach the preset threshold (such as 90% required by industry standards or the benchmark value determined based on historical training data), the model update process is triggered. The update process first re-executes Step 3 to expand the mapping relationship library: By deploying temporary monitoring devices on the operating transformers, discharge data under new working conditions is collected (such as the newly added "high-frequency vibration + sudden load increase" composite scenario), and the corresponding fusion feature vectors are synchronously generated (such as the abnormal fluctuation of the hydrogen volume fraction and the high-frequency pulse of the standardized discharge energy in the vibration scenario), and new discharge modes (such as "vibration-induced mixed discharge") and severities (such as "severe") are labeled, and the new data is supplemented to the expanded mapping relationship library according to a ratio of 3:7 (actual operation data: laboratory simulation data).

[0052] Subsequently, Step 4 is re-executed to adjust the hybrid neural network structure and retrain the model. The specific way of structure adjustment is determined according to the characteristics of the new working conditions: If the new data shows that the long-term trend of the characteristic gases in the oil is more complex (such as the acetylene content shows multi-stage growth), the number of hidden layer nodes in the LSTM layer is increased (such as from 16 to 24) to enhance the ability to capture long-term sequence dependencies; if the short-term pattern of the partial discharge signal is more discrete (such as the mutation frequency of the pulse density increases), the number of convolutional kernels in the CNN layer is increased (such as from 16 to 32) to enhance the ability to extract local features. The adjusted network structure is retrained using the expanded mapping relationship library (training set: validation set = 8:2), and the same optimization strategy as in Step 4 is adopted (cross-entropy loss function, gradient descent method, Dropout to suppress overfitting) until the error rate of the validation set stabilizes.

[0053] Finally, deploy the newly trained model to the edge computing device to replace the old model. Before replacement, verify the diagnostic accuracy of the new model for the historical test set and the new operating condition test set through offline testing (both need to be ≥ the preset threshold).

[0054] The present invention solves the problem of low data fusion efficiency of partial discharge patterns, oil chromatography, electrical quantities, etc. through multi-source data collaborative acquisition and feature standardization processing. Step 1: Deploy high-frequency current coil transformers, capacitive voltage dividers, on-line gas sensors, and SCADA systems to collect partial discharge signals (high-frequency current, voltage fluctuations), characteristic gas data in oil (volume fractions of hydrogen, acetylene, etc.), and power supply system operation parameters (load current, ambient temperature and humidity) respectively, and unify the timestamps through the GPS or NTP protocol to achieve time alignment of multi-source data. Step 2: Take the sampling period of the characteristic gas data in oil as the window, perform sliding window statistics on the partial discharge signals (extract statistical quantities such as the average pulse amplitude and frequency distribution variance), and convert the high-frequency instantaneous signals into low-frequency statistical quantities; at the same time, use the Z-score method to standardize the electrical quantity parameters (load current) and discharge characteristic parameters (discharge energy) to eliminate the dimension difference, and generate a 6-dimensional fusion feature vector containing gas volume fractions and standardized discharge energy, providing a structured input for multi-source data fusion.

[0055] Explanation of technical feature terms of the present invention Partial discharge signal: Electrical signals such as high-frequency current and voltage fluctuations generated when the internal insulating material of a power transformer undergoes local breakdown due to factors such as aging, electric field distortion, or mechanical stress. Such signals are early manifestations of insulation deterioration, directly reflecting the instantaneous characteristics of discharge (such as intensity, frequency), and are the core electrical basis for diagnosing faults.

[0056] Characteristic gas data in oil: Gas volume fraction data generated by the deterioration of insulating materials (such as oil-paper) during the discharge or thermal decomposition process of transformer oil, mainly including hydrogen (H2), acetylene (C2H2), total hydrocarbons (sum of hydrocarbon gases such as methane and ethylene), carbon monoxide (CO), and carbon dioxide (CO2). The generation of these gases is directly related to the degree of insulation deterioration and is a non-electrical characteristic for judging the type and severity of discharge.

[0057] Power supply system operation parameters: External environment and load data reflecting the operating state of the transformer, including load current (reflecting the magnitude of the electrical load), voltage amplitude (reflecting the voltage stability of the system), ambient temperature and humidity (affecting the performance of insulating materials), etc. These parameters are used to analyze the correlation between discharge faults and the system operating state and assist in locating the fault scenario.

[0058] Timestamp: A unified time identifier (such as "2025-05-21 10:00:00") marked for multi-source data (partial discharge signals, characteristic gas data in oil, operating parameters). Synchronize the clocks of each sensor through GPS or Network Time Protocol to align different types of data in the time dimension, facilitating the subsequent analysis of multi-dimensional information within the same time period.

[0059] Statistical aggregation: For high-frequency and instantaneous partial discharge signals (such as high-frequency current and voltage pulses), using the sampling period of the characteristic gas data in oil (such as 5 minutes) as a window, the process of extracting statistical quantities such as the average pulse amplitude within the window (reflecting the overall level of discharge intensity), the variance of frequency distribution (reflecting the density of discharges), and the number of discharges per unit time (reflecting the frequency of discharge activities). This operation converts high-frequency signals into low-frequency statistics that match the time scale of the gas data in oil, solving the problem of mismatched time scales of multi-source data.

[0060] Standardization processing: The process of eliminating the dimensional differences of electrical quantity parameters with different physical dimensions (such as load current, unit: ampere) and discharge characteristic parameters (such as discharge energy, unit: joule) using methods such as Z-score. For example, convert the load current and discharge energy into dimensionless values with a mean of 0 and a standard deviation of 1 respectively, making the two types of data comparable and facilitating subsequent feature fusion.

[0061] Fusion feature vector: A structured data formed by combining the characteristic gas data in oil (such as the volume fractions of hydrogen and acetylene) and the standardized discharge characteristic parameters (such as standardized discharge energy) in a fixed order (such as [H2, C2H2, total hydrocarbons, CO, CO2, standardized discharge energy]). This vector integrates the physical meanings (gas content, discharge intensity) and numerical characteristics (time alignment, unified dimension) of multi-source data, serving as the input of the neural network model to establish the correlation between features and discharge patterns.

[0062] Extended mapping relationship library: A sample set composed of a mixture of laboratory simulation data and actual operation data. Laboratory simulation data is obtained through devices such as the needle-plate electrode model (generating partial discharge patterns and spark discharge data) and the floating electrode model (generating mixed discharge data); actual operation data is collected through temporary monitoring devices for complex discharge data (such as intermittent arcs) in scenarios such as load mutations and drastic changes in temperature and humidity. Each piece of data in the library contains a fusion feature vector, a discharge pattern (partial discharge pattern, mixed discharge, etc.), the severity of the discharge (mild / moderate / severe), and a working condition label (such as "load mutation"), providing multi-scenario sample support for model training.

[0063] Convolutional Neural Network (CNN) Branch: A neural network module used to extract short-term patterns of partial discharge signals. The temporal features of the input partial discharge signals (such as the average pulse amplitude and the variance of frequency distribution) are inputted. Through the convolutional layer, local features (such as the change in the density of the pulse waveform) are captured. After reducing the dimension through the pooling layer, the characteristic information reflecting the instantaneous discharge state (such as the sudden change of the pulse sequence from sparse to dense) is outputted.

[0064] Long Short-Term Memory Network (LSTM) Branch: A neural network module used to capture the long-term change trend of characteristic gases in oil. The time series data of the characteristic gases in oil (such as the volume fraction of acetylene) are inputted. Through the internal gating mechanism (forget gate, input gate, output gate), the time series is processed to extract the temporal correlation features of the gas content (such as the cumulative rate of acetylene over time), reflecting the process of insulation deterioration.

[0065] Hybrid Neural Network Structure: A network model composed of the CNN branch and the LSTM branch. The CNN extracts short-term patterns of partial discharge signals (such as the change in pulse density), and the LSTM captures the long-term trend of gases in oil (such as the continuous growth of acetylene). After the outputs of the two branches are fused through the fully connected layer, the probability distribution of the discharge severity (such as "mild: local discharge probability 85%") is outputted through the Softmax activation function, realizing the collaborative learning of multi-modal features.

[0066] Edge Computing Device: A low-latency data processing device deployed in the substation (such as the edge computing unit of the integrated automation system), used to receive the fused feature vectors in real time and run the diagnostic model. Its advantage lies in reducing data transmission latency, supporting the real-time output of the discharge severity and operating condition labels (such as "load mutation scenario"), and meeting the fast response requirements of the power system.

[0067] Online Learning Mechanism: The process in which the edge computing device fine-tunes the weight parameters of the fully connected layer and the output layer of the hybrid neural network by accumulating newly acquired fused feature vectors (such as mixed discharge and intermittent arc data) using the stochastic gradient descent algorithm. This mechanism enables the model to gradually learn the features of new operating conditions (such as the collaborative growth law of hydrogen and acetylene in high-temperature and high-humidity environments), maintaining its adaptability to the dynamic operating environment.

[0068] Dynamic Verification Process: The process of extracting complex operating condition data (such as mixed discharge and intermittent arc) from the newly generated fused feature vectors as the test set every month and inputting them into the diagnostic model to evaluate the accuracy. If the preset threshold (such as 90%) is not reached for two consecutive months, then the extended mapping relationship library is rebuilt (supplementing new operating condition data) and the network structure is adjusted (such as increasing the number of LSTM hidden layer nodes), generating a new version of the model to replace the old model, so that the model can continuously adapt to environmental changes. The specific implementation of the present invention focuses on the multi-source data fusion for power transformer discharge fault diagnosis, the improvement of model generalization ability, and the dynamic adaptability optimization. Combining with the actual operation scenario of the power supply system, the technical solution described in the claims is achieved through the following technical means: In the multi-source data collaborative acquisition link, sensor groups are deployed on the power transformer body and the power supply line. Among them, the high-frequency current coil current transformer is installed on the grounding wire of the transformer bushing end screen. This position is the main path for the partial discharge current to discharge to the ground, and can effectively capture the high-frequency current signal generated by partial discharge (frequency range 100 kHz - 10 MHz); the capacitive voltage divider is connected in parallel to the outlet end of the high-voltage winding. This area is an electric field concentration area, and partial discharge will cause voltage transient changes. The capacitive voltage divider can synchronously measure the voltage fluctuation signal during the discharge process, forming a complement with the high-frequency current signal. The characteristic gas data in the oil is collected by the on-line gas sensor integrated in the transformer oil flow loop (such as the oil flow pipeline connecting the oil conservator and the body), and the volume fractions of 5 characteristic gases, namely hydrogen (H2), acetylene (C2H2), total hydrocarbons (the sum of hydrocarbon gases such as methane and ethylene), carbon monoxide (CO), and carbon dioxide (CO2), are continuously monitored using photoacoustic spectroscopy technology to reflect the dynamic changes of gases during the deterioration process of insulating materials. The operation parameters of the power supply system (load current, voltage amplitude, ambient temperature and humidity) are obtained by accessing the SCADA system through the Modbus protocol. The Modbus protocol supports multi-device real-time communication and can stably obtain the system operation status data. All sensors and the SCADA system are synchronized in time through GPS (in a network-free scenario, time error ≤ 1 μs) or NTP (in a networked scenario, time error ≤ 10 ms), so that the partial discharge signal (sampling frequency 1 MHz), the characteristic gas data in the oil (sampling period 5 minutes), and the operation parameters (sampling period 1 minute) are marked with a unified timestamp, providing a time reference for subsequent data alignment.

[0069] In the multi-source data feature fusion stage, for the high-frequency transient characteristics of partial discharge signals (such as high-frequency current and voltage pulses), with the 5-minute sampling period of the characteristic gas data in oil as the sliding window, statistical aggregation is performed on the high-frequency current and voltage signals: within each 5-minute window, calculate the average pulse amplitude (reflecting the overall level of discharge intensity during this time period), the variance of the frequency distribution (reflecting the density of discharge pulses on the time axis), and the number of discharges per unit time (the number of pulses in the window divided by 5 minutes, reflecting the frequency of discharge activities), and convert the high-frequency transient signals into low-frequency statistics with the same time scale as the characteristic gas data in oil. For the electrical quantity parameters (load current obtained from the SCADA system, unit: ampere) and discharge characteristic parameters (discharge energy calculated based on the integration of voltage and current signals, unit: joule), the Z-score normalization method is used to eliminate the dimension difference: calculate the mean and standard deviation of the load current and discharge energy respectively, and convert the original value into a dimensionless form of "(original value - mean) / standard deviation", making the two types of parameters comparable. Finally, combine the volume fractions of 5 characteristic gases in oil with the normalized discharge energy to form a 6-dimensional fusion feature vector (dimension definition: [H2 volume fraction, C2H2 volume fraction, total hydrocarbon volume fraction, CO volume fraction, CO2 volume fraction, normalized discharge energy]), which is used as the structured input data for the subsequent mapping library construction.

[0070] In the construction process of the extended mapping relationship library, laboratory simulation data is generated through a needle-plate electrode model, a floating electrode model, and a vibration interference device: The needle-plate electrode model consists of a pure tungsten needle electrode and a brass plate electrode with a fixed gap (such as 2 mm), simulating partial discharge (low energy, stable pulse), spark discharge (medium energy, short breakdown), and arc discharge (high energy, continuous penetration); The floating electrode model adds a movable floating electrode on the basis of the needle-plate electrode (simulating the loosening of winding pads), and adjusts the distance between the floating electrode and the main electrode (such as 1 - 3 mm) to generate mixed discharge data of the partial discharge superposed floating discharge model (large pulse amplitude fluctuation, irregular frequency); The vibration interference device applies mechanical vibration of 0.5g - 2g (g is the acceleration of gravity) through an electromagnetic exciter to simulate the intermittent phenomenon of the discharge channel caused by mechanical stress, generating intermittent arc data (the discharge channel exists for 2 - 8 seconds, and the energy fluctuates periodically). The actual operation data is collected by a temporary monitoring device for the discharge data in scenarios of load mutation (such as voltage fluctuation of ±15% caused by transformer paralleling / disconnection operation) and drastic changes in ambient temperature and humidity (temperature from 25°C to 85°C, humidity from 30% to 90%), and the fusion feature vectors and corresponding complex patterns such as mixed discharge and intermittent arc are synchronously recorded. The extended mapping relationship library is composed of 70% laboratory data and 30% actual data. Each piece of data contains a fusion feature vector, a discharge mode (partial discharge mode, spark discharge mode, arc discharge mode, mixed discharge, intermittent arc), a discharge severity level (mild: partial discharge; moderate: spark discharge; severe: arc discharge / mixed discharge), and a working condition label (such as "load mutation", "vibration interference", "high temperature and high humidity"), providing a multi-scenario sample set for model training.

[0071] In the training process of the hybrid neural network model, a hybrid structure of convolutional neural network (CNN) and long short-term memory network (LSTM) is adopted. The CNN branch inputs the temporal features of partial discharge signals (average pulse amplitude, variance of frequency distribution, number of discharges per unit time), extracts local features (such as the increasing or decreasing pattern of pulse amplitude within a continuous time window) through a 1-layer 3×3 convolutional layer, reduces the dimension through a 1-layer 2×2 max pooling layer (retaining key features such as the distribution position of high-amplitude pulses), and captures the short-term patterns of partial discharge signals (such as the sudden change of pulse sequence from sparse to dense). The LSTM branch inputs the time series data of characteristic gases in oil (volume fractions of H2, C2H2, total hydrocarbons, CO, CO2), and through an LSTM layer with 16 hidden layer nodes, uses the forgetting gate (retaining historical gas content information), input gate (updating the current gas increment), and output gate (integrating historical and current information) mechanisms to capture the long-term change trend of gas content (such as the continuous increase of acetylene content over time, reflecting the continuous thermal decomposition of insulating materials). After the outputs of the two branches are fused through a fully connected layer (with 32 nodes), the probability distribution of discharge severity (mild, moderate, severe) is output through the Softmax activation function. During the training process, the cross-entropy loss function is used in combination with the gradient descent method to optimize the weight parameters, and a Dropout layer (dropout rate 20%) is introduced to suppress overfitting. The early stopping mechanism is triggered when the error rate of the validation set does not decrease for 5 consecutive iterations.

[0072] In the model deployment and dynamic optimization phase, the trained hybrid neural network model is deployed on the edge computing device of the substation (such as the edge computing unit of the integrated automation system). This device supports real-time communication with sensors and the SCADA system through the Modbus / MQTT protocol (response time ≤ 100 ms), receives the real-time generated fused feature vectors (one set every 5 minutes), and calls the model inference interface to output the probability distribution of the discharge severity (such as "partial discharge probability 85%, spark discharge probability 10%, arc discharge / mixed discharge probability 5%") and the corresponding operating condition label (such as "load mutation scenario"). To maintain the adaptability of the model to the dynamic environment, the edge computing device is built with a data cache module. Every time a new set of fused feature vectors (including complex operating condition data such as mixed discharge and intermittent arc) is received, it performs validity verification (excluding missing data or outliers). When the cache accumulates 100 sets of valid data, the weight parameters of the fully connected layer and the output layer are fine-tuned using the stochastic gradient descent algorithm (learning rate 0.001) (the underlying convolutional layer and LSTM layer do not need to be frequently adjusted as they have learned the basic features). Every month, from the newly generated fused feature vectors, complex operating condition data (mixed discharge, intermittent arc) is screened according to the operating condition label, and 100 sets are randomly selected as the test set to input the model to evaluate the diagnostic accuracy rate (target ≥ 90%); if the threshold is not reached for two consecutive months, the mapping library construction is re-executed (supplementing new operating condition data such as "high-frequency vibration + sudden load increase") and model training is performed (such as increasing the number of LSTM hidden layer nodes to 24 to improve the ability to capture long-term trends), generating a new version of the model and replacing the old model.

[0073] Through the above implementation methods, the present invention realizes the time alignment and dimension unification of multi-source data, and solves the problem of low data fusion efficiency of partial discharge patterns, oil chromatography, electrical quantities, etc.; by expanding the mapping relationship library to cover multi-scenario discharge patterns, combining the hybrid neural network to fuse short-term discharge patterns and long-term gas trends, the generalization ability of the model to complex discharge patterns is improved; through edge deployment, online learning and dynamic verification, a closed-loop optimization mechanism is formed, enabling the model to continuously adapt to the dynamic operating environment of power transformers, and ultimately improving the operating stability of the power supply system.

[0074] Aiming at the problem of insufficient generalization ability of neural network models for complex discharge patterns, the present invention improves the model learning ability by expanding the mapping relationship library and designing a multi-modal network structure. In step 3, an extended mapping relationship library is constructed, integrating laboratory simulation data (partial discharge patterns generated by the needle-plate electrode model, spark discharge data, and hybrid discharge data generated by the floating electrode model) and actual operation data (intermittent arc data under scenarios of load mutation and drastic changes in ambient temperature and humidity), covering 5 types of patterns such as partial discharge patterns and hybrid discharges, and labeling the discharge severity and operating condition labels to provide a multi-scenario sample set for model training. In step 4, a hybrid neural network structure is adopted: the CNN branch extracts the short-term patterns of partial discharge signals (such as changes in pulse density), and the LSTM branch captures the long-term trends of characteristic gases in oil (such as the cumulative rate of acetylene content). The outputs of the two branches are fused through a fully connected layer and then the severity probability distribution is output through Softmax to achieve the collaborative learning of multi-modal features and enhance the model's recognition ability for complex patterns.

[0075] To ensure the continuous adaptability of the model to the dynamic operating environment, the present invention forms a closed-loop optimization mechanism through online learning and dynamic verification. In step 5, the trained model is deployed on the edge computing device of the substation to receive the fused feature vectors in real time and output the diagnostic results; every time 100 new fused feature vectors (including complex operating condition data) are accumulated, the weight parameters of the fully connected layer and the output layer are fine-tuned using the stochastic gradient descent algorithm to achieve incremental update of the model parameters. Every month, 100 sets of complex operating condition data are extracted as a test set to evaluate the diagnostic accuracy. If the threshold is not reached for two consecutive months, steps 3 (expanding the mapping relationship library) and 4 (adjusting the network structure) are re-executed to generate a new version of the model and replace the old model. This mechanism enables the model to continuously learn the features of new operating conditions, avoid the attenuation of diagnostic performance caused by environmental changes, and ultimately improve the operating stability of the power supply system.

Claims

1. A method for diagnosing discharge faults of power transformers based on neural networks, characterized in that, Including: Step 1: Through the sensor group deployed on the power transformer body and the power supply line, locally-discharged signals, characteristic gas data in oil, and power supply system operation parameters are collected in real time. Discharge characteristic parameters are calculated based on the locally-discharged signals and time stamps are marked. The discharge characteristic parameters include pulse amplitude, frequency distribution, and number of discharges per unit time; Step 2: The locally-discharged signals collected in Step 1 are statistically aggregated according to the sampling period of the characteristic gas data in oil. The power supply system operation parameters collected in Step 1 and the discharge characteristic parameters calculated in Step 1 are standardized to generate a fused feature vector; Step 3: Laboratory simulation data is collected. The laboratory simulation data includes fused feature vectors corresponding to partial discharge modes, spark discharge modes, and mixed discharge modes. Actual operation data is collected through the temporary monitoring device of the actual operation transformer. The actual operation data includes fused feature vectors corresponding to the intermittent arc mode under scenarios of sudden load change and drastic temperature and humidity change. An extended mapping relationship library is constructed based on the fused feature vector generated in Step 2, the laboratory simulation data, and the actual operation data. The extended mapping relationship library includes associated data between the fused feature vector and the discharge mode. The discharge modes include mixed discharge and intermittent arc; Step 4: Using the fused feature vector in the extended mapping relationship library constructed in Step 3 as the input, a diagnostic model is trained using a hybrid neural network structure including a convolutional neural network branch and a long short-term memory network branch. The diagnostic model outputs the probability distribution of the discharge severity; Step 5: The diagnostic model trained in Step 4 is deployed on the edge computing device. The fused feature vector generated in real time in Step 2 is used for discharge fault diagnosis, and the parameters of the diagnostic model trained in Step 4 are updated through an online learning mechanism.

2. The power transformer discharge fault diagnosis method based on a neural network according to claim 1, wherein, The sensor group includes a high-frequency current coil transformer for collecting locally-discharged signals, a resistor-capacitor voltage divider for measuring the discharge voltage fluctuation, an on-line gas sensor for monitoring the characteristic gas data in oil, and a SCADA system for obtaining the power supply system operation parameters; Among them, the high-frequency current coil transformer is installed on the grounding wire of the transformer bushing end screen, and the locally-discharged signal in Step 1 is obtained by collecting the locally-discharged high-frequency current signal; The resistor-capacitor voltage divider is connected in parallel to the high-voltage winding outlet end, and the collection of the locally-discharged signal is supplemented by measuring the voltage fluctuation signal during the discharge process; The on-line gas sensor is integrated into the transformer oil flow loop, and the characteristic gas data in oil in Step 1 is obtained by monitoring the volume fractions of hydrogen, acetylene, total hydrocarbons, carbon monoxide, and carbon dioxide; The SCADA system is accessed through the Modbus protocol, and the power supply system operation parameters in Step 1 are obtained by acquiring the load current, voltage amplitude, and ambient temperature and humidity; The sensor group and the SCADA system are time-synchronized through GPS or the network time protocol to make the time stamps of the locally-discharged signals, the characteristic gas data in oil, and the operation parameters consistent.

3. The power transformer discharge fault diagnosis method based on a neural network according to claim 2, wherein Also including: The statistical aggregation of partial discharge signals uses the sampling period of the characteristic gas data in oil as a window, and performs sliding window statistics on the high-frequency current and voltage signals collected by the high-frequency current coil transformer and the capacitive voltage divider, extracts the average pulse amplitude, frequency distribution variance, and number of discharges per unit time within the window, and converts the high-frequency instantaneous partial discharge signals in Step 1 into statistics that match the time scale of the characteristic gas data in oil; For the electrical quantity parameters and discharge characteristic parameters collected in Step 1, standardization processing uses the Z-score method for the load current obtained from the SCADA system and the discharge energy calculated based on voltage and current signals to eliminate the dimensional differences of current and energy; Combine the characteristic gas data in oil in Step 1 with the standardized discharge characteristic parameters in Step 2. The fusion feature vector is formed by combining the volume fraction of the characteristic gas in oil monitored by the on-line gas sensor and the standardized discharge energy, and serves as the structured input data for associating discharge patterns in the extended mapping relationship library in Step 3.

4. The method for diagnosing discharge faults of a power transformer based on a neural network according to claim 3, characterized in that, It also includes: Laboratory simulation data is generated through a needle-plate electrode model, a floating electrode model, and a vibration interference device. The laboratory simulation data includes fusion feature vectors and corresponding partial discharge patterns, spark discharge patterns, arc discharge patterns, and mixed discharge data of a partial discharge superimposed floating discharge model; Actual operation data is collected through a temporary monitoring device for discharge data under scenarios of load mutation and drastic changes in ambient temperature and humidity. The discharge data includes the fusion feature vectors generated in Step 2 and corresponding mixed discharges and intermittent arc discharge patterns; Based on the association requirements between the fusion feature vectors and discharge patterns generated in Step 2, the extended mapping relationship library is composed of 70% laboratory simulation data and 30% actual operation data. Each piece of data includes the fusion feature vectors generated in Step 2, discharge patterns, discharge severity, and operating condition labels, providing a sample set including the association between multi-scenario fusion features and discharge patterns for the neural network model training in Step 4.

5. The power transformer discharge fault diagnosis method based on a neural network according to claim 4, wherein It also includes: The convolutional neural network branch inputs the temporal features of the partial discharge signals in the extended mapping relationship library in Step 3, extracts local features through 1 convolutional layer, reduces the dimension through 1 max-pooling layer, and captures the short-term patterns of the partial discharge signals in the fusion feature vectors in Step 2; The long short-term memory network branch inputs the time series data of the characteristic gas in oil in the extended mapping relationship library in Step 3, captures the long-term change trend of the gas content through 1 LSTM layer, and extracts the temporal correlation features of the characteristic gas in oil in the fusion feature vectors in Step 2; Based on the multi-modal fusion requirements of the model training in Step 4, the outputs of the CNN branch and the LSTM branch of the hybrid neural network are fused through a fully connected layer, and the probability distribution of discharge severity is output through the Softmax activation function, providing a classification basis for multi-dimensional feature fusion for the diagnostic model training in Step 4.

6. The method for diagnosing discharge faults of a power transformer based on a neural network according to claim 5, wherein It also includes: The diagnostic model trained in Step 4 is deployed on the substation edge computing device, receives the fusion feature vectors generated in real time in Step 2, and outputs the probability distribution of discharge severity and the corresponding operating condition labels; The online learning mechanism accumulates 100 groups of new fusion feature vectors generated in step 2 through edge computing devices, and uses the stochastic gradient descent algorithm to fine-tune the weight parameters of the fully connected layer and the output layer of the hybrid neural network in step 5.

7. The method for diagnosing discharge faults of a power transformer based on a neural network according to claim 6, wherein, It also includes: The dynamic verification process extracts 100 groups of fusion feature vectors corresponding to complex working conditions from the newly generated fusion feature vectors in step 2 every month as the test set, and inputs them into the diagnostic model deployed on the edge computing device in step 5 to evaluate the diagnostic accuracy of the discharge severity; If the accuracy rate fails to reach the preset threshold for two consecutive months, re-execute step 3 and step 4 to generate an updated diagnostic model and replace the old model.

Images