A method for diagnosing power transformer discharge faults based on neural network

Abstract

The present invention relates to the technical field of power transformer discharge fault diagnosis in power supply systems, and in particular to a neural network-based power transformer discharge fault diagnosis method. The method comprises collecting partial discharge signals, characteristic gas data in oil, and power supply system operating parameters in real time through sensors and unifying the timestamps; statistically aggregating the partial discharge signals, and standardizing the electrical quantity parameters and discharge characteristic parameters to generate a fused feature vector; constructing an extended mapping relationship library containing laboratory simulation data and actual operating data; training a diagnostic model using a hybrid structure of a convolutional neural network and a long-short-term memory network; and deploying the model on edge computing devices and optimizing model performance through online learning and dynamic verification. This method addresses the problems of low multi-source data fusion efficiency and the insufficient generalization ability of neural network models for complex discharge patterns, thereby improving the operational stability of the power supply system.

Description

Technical Field

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

[0002] Power transformers are core equipment for voltage conversion and power transmission in power supply systems. The integrity of their internal insulation structures directly impacts the continuity of power distribution networks. Discharge faults are a common early abnormality in the process of insulation degradation, typically caused by factors such as long-term aging of insulation materials, local electric field distortion, or accumulated mechanical stress. They manifest as corona discharge, creeping discharge, or suspended discharge. During power supply system operation, discharges are accompanied by electrical signals (such as partial discharge pulses and high-frequency currents) and non-electrical signals (such as acoustic emission waves and infrared thermal radiation). These signal characteristics are intrinsically correlated with the type, location, and severity of the discharge. Sensor arrays deployed on the transformer and power supply lines collect multi-dimensional signals. Signal processing techniques such as filtering and denoising are used to extract discharge-related features (such as pulse amplitude, frequency distribution, and time interval). Combined with real-time power supply system operating data (such as load current, voltage fluctuations, and ambient temperature and humidity), pattern recognition algorithms are used to analyze the mapping between these features and fault types. This allows for early identification of fault types and risk assessments at the early stages of discharge development, providing a basis for power supply system operation and maintenance decisions and preventing transformer outages and power outages caused by insulation breakdown.

[0003] In the power supply system operation scenario, the core reason for the low efficiency of multi-source data fusion such as partial discharge patterns, oil chromatograms, 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 resistor-capacitor dividers) have high-frequency and instantaneous characteristics, while oil chromatogram data (such as the content of gases such as hydrogen and acetylene) must be obtained through periodic oil sampling (partial discharge tests every 30 minutes). The mismatch in time scales makes data alignment difficult. 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 system operating status, and the latter reflects the degree of insulation degradation). The lack of a unified feature standardization method makes it difficult to directly establish an effective association. The lack of generalization ability of neural network models for complex discharge patterns stems from the limitations of training data: the existing mapping relationship library is generated based on laboratory simulation experiments (such as local discharge mode, spark discharge mode, and arc discharge at a fixed gap under the needle-plate electrode model), while the transformer in the actual power supply system may face dynamic electric field distortion (such as voltage fluctuations caused by sudden load changes), mixed discharge modes (such as local discharge superimposed on suspended discharge), and other working conditions not covered in the experiment. For example, arc discharge is defined in the experiment as "a continuous and stable penetrating discharge channel maintained for 10 seconds", but intermittent arcs caused by vibration may occur in actual operation (the discharge channel exists intermittently). Such data is not included in the training set, and the model cannot accurately identify it, resulting in diagnostic deviations for complex fault types, ultimately reducing the operational stability of the power supply system. Summary of the Invention

[0004] In view of the shortcomings of the existing technology, the present invention provides a neural network-based power transformer discharge fault diagnosis method to solve the problems of low efficiency in the fusion of local discharge patterns, oil chromatography, electrical quantities and other data during the operation of the power supply system, and the poor generalization ability of the neural network model for complex discharge patterns, which causes the power supply system to have reduced operational stability.

[0005] In order to solve the above technical problems, the specific technical solutions of the present invention are as follows:

[0006] The present invention provides a method for diagnosing a power transformer discharge fault based on a neural network, comprising:

[0007] Step 1: Using a sensor group deployed on the power transformer and power supply lines, partial discharge signals, characteristic gas data in oil, and power supply system operating parameters are collected in real time. Discharge characteristic parameters are calculated based on the partial discharge signals and timestamped. The discharge characteristic parameters include pulse amplitude, frequency distribution, and the number of discharges per unit time.

[0008] 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 standardize the power supply system operating parameters collected in step 1 and the discharge characteristic parameters calculated in step 1 to generate a fused feature vector;

[0009] Step 3: Collect laboratory simulation data, which includes fused feature vectors corresponding to partial discharge mode, spark discharge mode, mixed discharge mode, and other modes. Collect actual operation data through a temporary monitoring device of an actual operating transformer. The actual operation data includes fused feature vectors corresponding to modes such as intermittent arcs under scenarios of sudden load changes and drastic changes in temperature and humidity. Based on the fused feature vectors generated in step 2, the laboratory simulation data, and the actual operation data, an extended mapping relationship library is constructed. The extended mapping relationship library includes association data between the fused feature vectors and discharge modes, and the discharge modes include mixed discharge and intermittent arcs.

[0010] Step 4: using the fused feature vector in the extended mapping relationship library constructed in Step 3 as input, a hybrid neural network structure including a convolutional neural network branch and a long short-term memory network branch is used to train a diagnostic model, wherein the diagnostic model outputs a probability distribution of discharge severity;

[0011] In step 5, the diagnostic model trained in step 4 is deployed on the edge computing device, and the fused feature vector generated in real time in step 2 is used to diagnose the discharge fault. The parameters of the diagnostic model trained in step 4 are updated through the online learning mechanism, and the diagnostic performance of the diagnostic model is dynamically verified.

[0012] Furthermore, in the neural network-based power transformer discharge fault diagnosis method described in the present invention, the sensor group includes a high-frequency current coil mutual inductor for collecting partial discharge signals, a resistance-capacitance voltage divider for measuring discharge voltage fluctuations, an online gas sensor for monitoring characteristic gas data in oil, and a SCADA system for obtaining power supply system operating parameters;

[0013] The high-frequency current coil mutual inductor 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;

[0014] The RC voltage divider is connected in parallel to the high voltage winding output terminal, and the voltage fluctuation signal during the discharge process is measured to supplement the collection of partial discharge signals.

[0015] The online gas sensor is integrated into the transformer oil flow loop to acquire the characteristic gas data in the oil in step 1 by monitoring the volume fractions of hydrogen, acetylene, total hydrocarbons, carbon monoxide, and carbon dioxide.

[0016] The SCADA system is connected via the Modbus protocol and acquires the power supply system operating parameters in step 1 by obtaining the load current, voltage amplitude, and ambient temperature and humidity.

[0017] The sensor group and SCADA system are synchronized through GPS or network time protocol to ensure that the timestamps of partial discharge signals, characteristic gas data in oil and operating parameters are consistent.

[0018] Furthermore, the neural network-based power transformer discharge fault diagnosis method of the present invention further includes: performing sliding window statistics on the high-frequency current and voltage signals collected by the high-frequency current coil mutual inductor and the resistance-capacitance voltage divider using the sampling period of the characteristic gas data in the oil as a window for statistical aggregation of partial discharge signals, extracting the pulse amplitude average, frequency distribution variance, and number of discharges per unit time within the window, and converting the high-frequency instantaneous partial discharge signal in step 1 into a statistic that matches the time scale of the characteristic gas data in the oil;

[0019] For the electrical quantity parameters and discharge characteristic parameters collected in step 1, the load current obtained by the SCADA system and the discharge energy calculated based on the voltage and current signals are standardized using the Z-score method to eliminate the dimensional differences between current and energy;

[0020] The characteristic gas data in the oil in step 1 is combined 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 online gas sensor and the standardized discharge energy, which serves as the structured input data for the associated discharge mode in the expanded mapping relationship library in step 3.

[0021] Furthermore, the neural network-based power transformer discharge fault diagnosis method of the present invention further includes: laboratory simulation data generated by a needle-plate electrode model, a suspended electrode model, and a vibration interference device, including fusion feature vectors and corresponding partial discharge modes, spark discharge modes, arc discharge modes, and mixed discharge data of a partial discharge superimposed suspended discharge model;

[0022] Actual operation data is collected through temporary monitoring devices under load mutation 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 modes such as mixed discharge and intermittent arc.

[0023] Based on the association requirements between the fused feature vector generated in step 2 and the discharge mode, the extended mapping relationship library is composed of a mixture of 70% laboratory simulation data and 30% actual operation data. Each piece of data includes the fused feature vector generated in step 2, the discharge mode, the discharge severity and the operating 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.

[0024] Furthermore, the neural network-based power transformer discharge fault diagnosis method of the present invention further includes: a convolutional neural network branch inputs the temporal features of the partial discharge signal in the expanded mapping relationship library in step 3, extracts local features through a convolutional layer, reduces the dimension through a maximum pooling layer, and captures the short-term pattern of the partial discharge signal in the fused feature vector in step 2;

[0025] The long short-term memory network branch inputs the time series data of characteristic gases in oil from the expanded mapping relationship library in step 3, captures the long-term variation trend of gas content through a single LSTM layer, and extracts the temporal correlation features of characteristic gases in oil from the fused feature vector in step 2.

[0026] Based on the multimodal 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 the 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.

[0027] Furthermore, the neural network-based power transformer discharge fault diagnosis method of the present invention further includes: deploying the diagnostic model trained in step 4 on a substation edge computing device, receiving the fused feature vector generated in real time in step 2, and outputting a discharge severity probability distribution and a corresponding operating condition label;

[0028] The online learning mechanism accumulates 100 sets of new fused 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 output layer of the hybrid neural network in step 5.

[0029] Furthermore, the neural network-based power transformer discharge fault diagnosis method of the present invention further includes: a dynamic verification process extracting 100 sets of fused feature vectors corresponding to complex working conditions from the newly generated fused feature vectors in step 2 each month as a test set, and inputting the test set into the diagnostic model deployed on the edge computing device in step 5 to evaluate the discharge severity diagnosis accuracy;

[0030] If the accuracy rate does not reach the preset threshold for two consecutive months, re-execute steps 3 and 4 to generate an updated diagnostic model and replace the old model.

[0031] Beneficial effects of the present invention:

[0032] The present invention solves the problem of low fusion efficiency of data such as partial discharge patterns, oil chromatography, and electrical quantities due to time scale mismatch and dimensional difference through collaborative collection and feature standardization of multi-source data: Step 1 realizes data time alignment through sensor deployment and time synchronization; Step 2 generates structured fusion feature vectors through statistical aggregation and Z-score standardization to provide effective input for subsequent analysis; By expanding the mapping relationship library and multimodal neural network structure design, the model's generalization ability for complex discharge patterns is improved: Step 3 integrates laboratory simulation data and actual operation data to cover multi-scenario discharge patterns; Step 4 uses CNN and LSTM hybrid networks to fuse short-term discharge patterns and long-term gas trends to enhance the recognition ability of complex patterns such as mixed discharge and intermittent arcs; Maintain model adaptability through online learning and dynamic verification mechanism: Step 5 The model deployed on the edge computing device fine-tunes 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 identification deviation, and ultimately improves the operation stability of the power supply system. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the technical solution of the present invention, the following is a brief introduction to the drawings required for use in the embodiments. Obviously, for ordinary technicians in this field, other drawings can be obtained based on the drawings without paying any creative labor.

[0034] Figure 1 The present invention provides a flowchart of a method for diagnosing discharge faults of power transformers based on a neural network. DETAILED DESCRIPTION

[0035] In order to make the purpose, 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 of the present invention and the corresponding drawings. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. The technical solutions provided by each embodiment of the present invention are described in detail below in conjunction with the drawings. In order to better understand the purpose of the present invention, the present invention is further described in detail below.

[0036] See also Figure 1 The present invention provides a method for diagnosing power transformer discharge faults based on a neural network.

[0037] Step 1: Using a sensor group deployed on the power transformer and power supply lines, partial discharge signals, characteristic gas data in oil, and power supply system operating parameters are collected in real time. Discharge characteristic parameters are calculated based on the partial discharge signals and timestamped. The discharge characteristic parameters include pulse amplitude, frequency distribution, and the number of discharges per unit time.

[0038] 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 standardize the power supply system operating parameters collected in step 1 and the discharge characteristic parameters calculated in step 1 to generate a fused feature vector;

[0039] Step 3: Collect laboratory simulation data, which includes fused feature vectors corresponding to partial discharge mode, spark discharge mode, mixed discharge mode, and other modes. Collect actual operation data through a temporary monitoring device of an actual operating transformer. The actual operation data includes fused feature vectors corresponding to modes such as intermittent arcs under scenarios of sudden load changes and drastic changes in temperature and humidity. Based on the fused feature vectors generated in step 2, the laboratory simulation data, and the actual operation data, an extended mapping relationship library is constructed. The extended mapping relationship library includes association data between the fused feature vectors and discharge modes, and the discharge modes include mixed discharge and intermittent arcs.

[0040] Step 4: using the fused feature vector in the extended mapping relationship library constructed in Step 3 as input, a hybrid neural network structure including a convolutional neural network branch and a long short-term memory network branch is used to train a diagnostic model, wherein the diagnostic model outputs a probability distribution of discharge severity;

[0041] In step 5, the diagnostic model trained in step 4 is deployed on the edge computing device, and the fused feature vector generated in real time in step 2 is used to diagnose the discharge fault. The parameters of the diagnostic model trained in step 4 are updated through the online learning mechanism, and the diagnostic performance of the diagnostic model is dynamically verified.

[0042] The neural network-based power transformer discharge fault diagnosis method provided by this invention achieves accurate diagnosis of power transformer discharge faults through a closed-loop process of multi-source data acquisition, feature fusion, model training, and dynamic optimization. The specific implementation and logical relationship of each step are as follows:

[0043] First, in step 1, sensors deployed on the power transformer and power lines collect multi-source data in real time. Specifically, partial discharge signals are collected using a high-frequency current transformer and a resistor-capacitor voltage divider. The high-frequency current transformer, installed on the transformer bushing's end shield grounding wire, captures the high-frequency current signal (frequency range 100kHz-10MHz) generated by partial discharge. The resistor-capacitor voltage divider, connected in parallel to the high-voltage winding output, simultaneously measures the voltage fluctuation signal during the discharge process. Data on characteristic gases in the oil are collected using an online gas sensor integrated into the transformer's oil flow loop. This sensor continuously monitors the volume fractions of five characteristic gases: hydrogen, acetylene, total hydrocarbons, carbon monoxide, and carbon dioxide using photoacoustic spectroscopy. Power supply system operating parameters (such as load current, voltage amplitude, and ambient temperature and humidity) are accessed through the SCADA system via the Modbus protocol. The sensor group and the SCADA system synchronize their time using GPS or Network Time Protocol (NTP), aligning 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). This provides a synchronization benchmark for subsequent data alignment. This step, which covers the discharge signal (electrical quantity), gas in oil (non-electrical quantity), and system operating status (environmental parameters) through the sensor group, lays the foundation for subsequent feature processing.

[0044] Step 2 statistically aggregates and normalizes the multi-source data collected in step 1 to generate a fused feature vector. Targeting the high-frequency transient characteristics of partial discharge signals (such as high-frequency current and voltage pulses), a sliding window statistical analysis is performed on the high-frequency current and voltage signals, using the 5-minute sampling period of the oil-based characteristic gas data as a window. The average pulse amplitude (reflecting discharge intensity), the frequency distribution variance (reflecting discharge density), and the number of discharges per unit time (reflecting discharge frequency) within the window are extracted. This high-frequency transient signal is converted into low-frequency statistics consistent with the time scale of the oil-based characteristic gas data. For electrical parameters (such as load current acquired from the SCADA system) and discharge characteristic parameters (such as discharge energy calculated by integrating voltage and current signals), the Z-score normalization method is used to eliminate the dimensional differences between current (A) and energy (J), making the two parameters comparable. Finally, the volume fractions of the five oil-based characteristic gases are combined with the normalized discharge energy to form a 6-dimensional fused feature vector (dimensionality definition: [H2, C2H2, total hydrocarbons, CO, CO2, discharge energy]), which serves as the structured input data for the subsequent mapping library construction. This step solves the key obstacle of multi-source data fusion by aligning time scales and unifying dimensions, providing effective input for model training.

[0045] Step 3 builds an extended mapping relationship library based on the fused feature vectors generated in Step 2, covering discharge patterns in multiple scenarios. Laboratory simulation data is generated using a needle-plate electrode model (generating partial discharge patterns, spark discharge patterns, and arc discharge data), a suspended electrode model (simulating suspended discharge caused by loose windings), and a vibration interference device (applying 0.5g-2g mechanical vibration). This includes mixed discharge data from partial discharge and suspended discharge models. Actual operation data is collected using temporary monitoring devices under scenarios involving sudden load changes (e.g., voltage fluctuations of ±15%) and dramatic changes in ambient temperature and humidity (temperature from 25°C to 85°C, humidity from 30% to 90%). This data includes fused feature vectors and corresponding complex patterns such as mixed discharges and intermittent arcs. The expanded mapping library consists of a mixture of 70% laboratory data and 30% actual operating data. Each data entry includes a fused feature vector, discharge mode (partial discharge, spark discharge, arc discharge, mixed discharge, intermittent arc), discharge severity (mild / moderate / severe), and operating condition labels (such as "load mutation" and "vibration interference"). This provides a multi-scenario sample set for model training in step 4. This step, by supplementing actual operating data, addresses the incomplete operating condition coverage of traditional mapping libraries.

[0046] In step 4, the fused feature vector from the expanded mapping library in step 3 is used as input to train the diagnostic model using a hybrid neural network architecture. The convolutional neural network (CNN) branch receives the temporal characteristics of the partial discharge signal (pulse amplitude mean, frequency distribution variance, and number of discharges). A 3×3 convolutional layer extracts local features (such as the local pattern of the pulse waveform). A 2×2 max pooling layer reduces the dimensionality to capture short-term patterns in the discharge signal (such as changes in the density of the pulse sequence). The long short-term memory (LSTM) branch receives time series data of characteristic gases in the oil (volume fractions of five gases) and a 16-node LSTM layer captures long-term trends in gas content (such as the accumulation rate of acetylene content over time). The outputs of the two branches are fused through a fully connected layer (32 nodes) and a softmax activation function is used to output a probability distribution of discharge severity (mild: partial discharge; moderate: spark discharge; severe: arc discharge / mixed discharge). During training, the cross-entropy loss function combined with gradient descent was used to optimize weight parameters. A dropout layer (with a 20% dropout rate) was introduced to mitigate overfitting. Early stopping was triggered when the validation set error rate did not decrease after five consecutive iterations. This step, through a multimodal network structure, integrated short-term discharge patterns with long-term gas trends, addressing the inability of traditional models to recognize complex discharge patterns.

[0047] In step 5, the diagnostic model trained in step 4 is deployed on the substation edge computing device for real-time diagnosis and dynamic optimization. The edge computing device receives the fused feature vector generated in real time in step 2 and outputs a probability distribution of discharge severity and the corresponding operating condition label (e.g., "load sudden change scenario"). To maintain the model's adaptability to dynamic environments, the edge computing device fine-tunes the weight parameters of the fully connected layer and output layer using a stochastic gradient descent algorithm (learning rate 0.001) every 100 sets of newly fused feature vectors (including complex operating condition data such as mixed discharges and intermittent arcs). Monthly, 100 sets of complex operating condition data are extracted from the newly generated fused feature vectors as a test set to evaluate the model's diagnostic accuracy (target: ≥90% accuracy for identifying mild, moderate, and severe discharges). If the threshold is not met for two consecutive months, steps 3 (expanding the mapping library and adding new operating condition data) and 4 (adjusting the network structure, such as increasing the number of nodes in the LSTM hidden layer) are repeated to generate a new version of the model and replace the old one. This step forms a closed loop of "data collection-feature processing-model training-online optimization" through online deployment, incremental learning and dynamic verification.

[0048] Specifically, in the neural network-based power transformer discharge fault diagnosis method of the present invention, the sensor group includes four types, corresponding to the partial discharge signal, characteristic gas data in oil, and power supply system operating parameters collected in step 1: a high-frequency current coil mutual inductor for collecting partial discharge signals, a resistance-capacitance voltage divider for measuring discharge voltage fluctuations, an online gas sensor for monitoring characteristic gas data in oil, and a SCADA system for obtaining power supply system operating parameters;

[0049] Among them, the high-frequency current coil mutual inductor 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 high-frequency current signal of the partial discharge; the resistance-capacitance voltage divider is connected in parallel to the high-voltage winding output terminal, and the voltage fluctuation signal during the discharge process is measured to supplement the collection of the partial discharge signal; the online gas sensor is integrated into the transformer oil flow loop, 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 connected through the Modbus protocol, and the power supply system operating parameters in step 1 are obtained by obtaining the load current, voltage amplitude, and ambient temperature and humidity;

[0050] The sensor group and SCADA system are synchronized through GPS or network time protocol to make the timestamps of partial discharge signals, characteristic gas data in oil and operating parameters consistent, providing a synchronization benchmark for "marking timestamps for data" in step 1.

[0051] The deployment and data acquisition process of the sensor group in step 1 of the present invention realizes multi-source collaborative acquisition of partial discharge signals, characteristic gas data in oil, and power supply system operating parameters by targeted selection of sensor types, clear installation locations, and adoption of a unified time synchronization mechanism, providing basic support for subsequent data processing.

[0052] A high-frequency current coil transformer, the core device for collecting partial discharge signals, is installed at the grounding wire at the end shield of the power transformer's bushing. This location is the primary path for partial discharge current to discharge to the ground. The high-frequency current (100kHz-10MHz) generated by the discharge propagates along this path. Therefore, deploying a high-frequency current coil transformer at this location effectively captures the current pulse signals during the discharge process, enabling direct monitoring of partial discharge.

[0053] A RC divider connected in parallel to the high-voltage winding outlet is used to supplement the voltage fluctuation signal collected during partial discharge. The high-voltage winding outlet is the area of ​​concentrated electric field within the transformer. Partial discharge can cause transient voltage changes at this location. The parallel RC divider allows for real-time measurement of the amplitude and frequency characteristics of the voltage fluctuations. This complements the current signal collected by the high-frequency current coil transformer, forming a dataset of electrical characteristics of partial discharge.

[0054] Online gas sensors are integrated into the transformer's oil flow loop (e.g., the oil pipe connecting the oil pillow to the transformer body). Using photoacoustic spectroscopy, they continuously monitor the volume fractions of five characteristic gases: hydrogen (H2), acetylene (C2H2), total hydrocarbons (CH4+C2H4+C2H2, etc.), carbon monoxide (CO), and carbon dioxide (CO2). The oil flow loop is the primary channel for transformer oil circulation. Deploying sensors here allows real-time monitoring of the dynamic changes in dissolved gases in the oil, reflecting the generation and diffusion patterns of gases during insulation material degradation and providing a reliable source for collecting characteristic gas data in the oil.

[0055] The SCADA system (Supervisory Control and Data Acquisition System) is connected to the power supply system via the Modbus protocol to obtain operating parameters such as load current, voltage amplitude, and ambient temperature and humidity. Modbus is a serial communication protocol widely used in industry, supporting multi-device access and real-time data transmission. Connecting to the SCADA system through this protocol allows for stable acquisition of real-time operating status data for the power supply system, including current and voltage parameters reflecting load conditions and ambient temperature and humidity parameters that affect insulation performance. This provides a basis for subsequent analysis of the correlation between discharge faults and system operating status.

[0056] To align data temporally, the sensor group and SCADA system synchronize their clocks using GPS (Global Positioning System) or NTP (Network Time Protocol). GPS uses satellite signals to provide a high-precision time reference (error ≤ 1μs) and is suitable for substations without network coverage. NTP synchronizes device clocks over the network (error ≤ 10ms) and is suitable for substations with local area networks. By using unified timestamps, partial discharge signals (sampling frequency 1MHz), characteristic gas data in oil (sampling period 5 minutes), and operating parameters (sampling period 1 minute) can be aligned temporally during subsequent processing, avoiding feature matching errors caused by time asynchrony and ensuring temporal consistency for statistical aggregation and feature fusion in step 2.

[0057] Specifically, the neural network-based power transformer discharge fault diagnosis method of the present invention, for the partial discharge signal collected in step 1, the partial discharge signal statistical aggregation uses the sampling period of the characteristic gas data in the oil as a window, performs sliding window statistics on the high-frequency current and voltage signals collected by the high-frequency current coil mutual inductor and the resistance-capacitance voltage divider, extracts the pulse amplitude average, frequency distribution variance, and number of discharges per unit time within the window, and converts the high-frequency instantaneous partial discharge signal in step 1 into a statistic that matches the time scale of the characteristic gas data in the oil;

[0058] For the electrical quantity parameters and discharge characteristic parameters collected in step 1, the standardization process uses the Z-score method on the load current obtained by the SCADA system (an electrical quantity parameter in step 1) and the discharge energy calculated based on the voltage and current signals (a discharge characteristic parameter in step 1), eliminating the dimensional difference between current (A) and energy (J), making the two parameters comparable;

[0059] The characteristic gas data in oil in step 1 is combined with the standardized discharge characteristic parameters in step 2. The fused feature vector is formed by combining the volume fractions of the five characteristic gases in oil monitored by the online gas sensor (belonging to the characteristic gas data in oil in step 1) with the standardized discharge energy, and serves as the structured input data for the associated discharge mode in the expanded mapping relationship library in step 3.

[0060] The statistical aggregation, standardization, and fusion feature vector generation process of multi-source data in step 2 of the present invention solves the difficulty in fusing partial discharge signals, characteristic gas data in oil, and electrical parameters due to their differences in characteristics through time scale alignment, dimensional unification, and multi-dimensional feature integration, thereby providing structured input data for subsequent mapping library construction and model training.

[0061] The high-frequency transient characteristics of the partial discharge signals collected in step 1 (such as the 1MHz current sampling signal from the high-frequency current coil transformer and the voltage fluctuation signal from the resistor-capacitor divider) exhibit a time scale mismatch with the low-frequency periodicity of the oil-based characteristic gas data (such as the 5-minute sampling period of the online gas sensor). Therefore, this step uses the 5-minute sampling period of the oil-based characteristic gas data as a sliding window to statistically aggregate the high-frequency current and voltage signals. Within each 5-minute window, all high-frequency signal points within the window are traversed to calculate the average pulse amplitude (reflecting the overall level of discharge intensity during that period), the variance of the frequency distribution (reflecting the density of discharge pulses along the time axis; a larger variance indicates a more concentrated 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 discharge activity). This operation converts the high-frequency transient partial discharge signals into low-frequency statistics consistent with the time scale of the oil-based characteristic gas data (e.g., a set of statistical values ​​every 5 minutes), providing a foundation for time alignment for subsequent multi-source data fusion.

[0062] The electrical quantity parameters and discharge characteristic parameters collected in step 1 cannot be directly correlated for analysis due to their different physical dimensions (load current in amperes, discharge energy in joules). This step uses the Z-score normalization method to process these two types of parameters: first, the mean and standard deviation of the load current (real-time data collected from the SCADA system) are calculated, converting each current value into a dimensionless value of "(current value - mean) / standard deviation". Simultaneously, 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 are calculated, converting each energy value into the same dimensionless form. Through this normalization operation, the load current and discharge energy are mapped to the same dimensional space (mean = 0, standard deviation = 1), making the two types of parameters comparable and ensuring numerical consistency for subsequent feature fusion.

[0063] After statistical aggregation and normalization, this step combines the oil characteristic gas data from step 1 with the standardized discharge characteristic parameters to generate a fused feature vector. Specifically, the volume fractions of five oil characteristic gases (hydrogen H2, acetylene C2H2, total hydrocarbons, carbon monoxide CO, and carbon dioxide CO2) monitored by online gas sensors, serving as non-electrical features directly reflecting the degree of insulation degradation, are combined with the standardized discharge energy (an electrical feature reflecting discharge intensity) in a fixed order to form a six-dimensional fused feature vector (dimensionality definition: [H2 volume fraction, C2H2 volume fraction, total hydrocarbon volume fraction, CO volume fraction, CO2 volume fraction, normalized discharge energy]). This fused feature vector serves as structured input data for expanding the mapping relationship library in step 3. It integrates the physical meaning (gas content, discharge intensity) and numerical characteristics (time alignment and dimensionality uniformity) of multi-source data into a single feature space, providing effective input for subsequently establishing a correlation between the fused features and discharge patterns.

[0064] Specifically, the neural network-based power transformer discharge fault diagnosis method of the present invention is:

[0065] To cover the multi-scenario data required for building the mapping library in step 3, the laboratory simulation data is generated using a needle-plate electrode model (generating partial discharge patterns, spark discharge patterns, and arc discharge data), a suspended electrode model (simulating suspended discharge caused by loose windings), and a vibration interference device (imposing mechanical vibration). This includes the fused feature vector generated in step 2 and the corresponding partial discharge patterns, spark discharge patterns, arc discharge patterns, and mixed discharge data from the partial discharge superimposed suspended discharge model.

[0066] To supplement the actual operating data not covered by the laboratory, the actual operating data is collected through temporary monitoring devices under load mutation (voltage fluctuation scenario) 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 modes such as mixed discharge and intermittent arc.

[0067] Based on the association requirements between the fused feature vector generated in step 2 and the discharge mode, the extended mapping relationship library is composed of a mixture of 70% laboratory simulation data and 30% actual operation data. Each data includes the fused feature vector generated in step 2, the discharge mode (partial discharge mode, spark discharge mode, arc discharge mode, 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 association between multi-scenario fusion features and discharge modes for the neural network model training in step 4.

[0068] The construction process of the expanded mapping relationship library in step 3 of the present invention covers multiple discharge modes through the coordinated collection of laboratory simulation data and actual operation data, solves the problem of insufficient model generalization ability caused by incomplete working condition coverage of the traditional mapping library, and provides multi-dimensional sample support for the neural network model training in step 4.

[0069] The generation of laboratory simulation data aims to cover basic and typical mixed discharge modes, and is achieved through the collaborative efforts 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 (e.g., a gap distance of 2 mm) to simulate local discharge (low discharge energy, stable pulse frequency), spark discharge (brief breakdown of the discharge channel, medium energy), and arc discharge (continuous penetrating discharge channel, high energy) caused by electric field concentration within the insulating material; the floating electrode model adds a movable floating electrode to the needle-plate electrode (simulating a floating conductor caused by loose winding spacers). By adjusting the distance between the floating electrode and the main electrode (e.g., 1-3 mm), mixed discharge data of the local discharge superimposed on the floating discharge model is generated (manifested by increased pulse amplitude fluctuations and irregular frequency); the vibration interference device applies mechanical vibrations of 0.5g-2g (g is the acceleration of gravity) to the electrode system through an electromagnetic exciter, simulating the discontinuity of the discharge channel caused by mechanical stress during transformer operation, and generating intermittent arc data (the discharge channel lasts 2-8 seconds, and the energy fluctuates periodically). While generating the discharge signal, the experimental device simultaneously collects the fused feature vectors generated in step 2 (such as the gas volume fraction within a 5-minute window and the normalized discharge energy) and annotates the corresponding discharge modes (partial discharge mode, spark discharge, etc.) to form a laboratory simulation data set.

[0070] Actual operating data collection aims to cover complex operating conditions not covered in the laboratory. This is achieved by deploying temporary monitoring devices on operating transformers. These scenarios include: sudden load changes (e.g., voltage fluctuations of ±15% caused by transformer paralleling / de-coupling operations), where the internal electric field distribution of the transformer dynamically changes, potentially leading to mixed discharges (local discharges superimposed on suspended discharges); and drastic changes in ambient temperature and humidity (e.g., in high-temperature, high-humidity summer environments, where the temperature rises from 25°C to 85°C and the humidity rises from 30% to 90%), where insulation material degradation accelerates and intermittent arcing (discharge channels remain intermittent due to thermal expansion / contraction). The temporary monitoring devices simultaneously collect the fused feature vectors generated in step 2 (e.g., the rapid increase in hydrogen and acetylene volume fractions in high-temperature, high-humidity environments, and the periodic fluctuations in normalized discharge energy). The corresponding complex discharge patterns (mixed discharges, intermittent arcing) are identified through manual annotation or expert systems to form an actual operating data set.

[0071] The extended mapping library is constructed based on the correlation between the fused feature vector generated in step 2 and the discharge mode. This library integrates laboratory simulation data (70%) and actual operating data (30%). The 7:3 ratio is chosen to balance the controllability of laboratory data (clear discharge patterns and stable features) with the authenticity of actual data (covering dynamic operating conditions and complex features): 70% of the laboratory data fully enables learning of basic discharge patterns, while the 30% of the actual data supplements the feature diversity under complex operating conditions. Each data entry consists of four components: the fused feature vector generated in step 2 (6-dimensional, e.g., [H2 = 50 ppm, C2H2 = 10 ppm, total hydrocarbons = 80 ppm, CO = 200 ppm, CO2 = 1500 ppm, normalized discharge energy = 1.2]), the discharge mode (partial discharge, spark discharge, arc discharge, mixed discharge, intermittent arc), the discharge severity (mild: partial discharge; moderate: spark discharge; severe: arc discharge / mixed discharge), and the operating condition label (e.g., "load sudden change," "vibration interference," "high temperature and high humidity"). Among them, the working condition label is used to mark the specific scenarios of data generation (such as the scenario where "load mutation" corresponds to a voltage fluctuation of ±15%), helping the model learn the discharge rules under different working conditions.

[0072] Specifically, the neural network-based power transformer discharge fault diagnosis method of the present invention serves as the core input processing module for model training in step 4. The convolutional neural network (CNN) branch inputs the temporal features of the partial discharge signal in the expanded mapping relationship library in step 3 (pulse amplitude average, frequency distribution variance, number of discharges, which are components of the fused feature vector in step 2). Local features (such as local patterns of the pulse waveform) are extracted through one convolutional layer, and the dimension is reduced (key features are retained) through one maximum pooling layer to capture the short-term patterns of the partial discharge signal in the fused feature vector in step 2 (such as changes in pulse sequence density).

[0073] As another core input processing module for model training in step 4, the long short-term memory (LSTM) branch inputs the time series data of characteristic gases in oil (hydrogen, acetylene, total hydrocarbons, carbon monoxide, and carbon dioxide volume fractions, which are components of the fused feature vector in step 2) from the expanded mapping relationship library in step 3. Through one LSTM layer, it captures the long-term variation trend of gas content (such as the accumulation rate of acetylene content over time) and extracts the temporal correlation features of the characteristic gases in oil from the fused feature vector in step 2.

[0074] Based on the multimodal 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 (integrating short-term discharge patterns and long-term gas trends), and the discharge severity probability distribution (mild: partial discharge; moderate: spark discharge; severe: arc discharge / mixed discharge) is output through the Softmax activation function, providing a classification basis for multi-dimensional feature fusion for the diagnostic model training in step 4.

[0075] The construction and training process of the hybrid neural network in step 4 of the present invention uses a multimodal branch design of a convolutional neural network (CNN) and a long short-term memory network (LSTM) to extract the short-term pattern of the partial discharge signal and the long-term trend of the characteristic gas in the oil. After feature fusion, the probability distribution of the discharge severity is output, providing a multi-dimensional classification basis for diagnostic model training.

[0076] The convolutional neural network (CNN) branch serves as the short-term pattern extraction module for PD signals. Its input is the temporal features of the PD signal from the expanded mapping library in step 3. Specifically, these include the average pulse amplitude (reflecting the average level of discharge intensity), the frequency distribution variance (reflecting the concentration of discharge pulses along the time axis), and the number of discharges per unit time (reflecting the frequency of discharge activity) from the fused feature vector in step 2. This branch uses a single convolutional layer (e.g., a 3×3 convolutional kernel) to extract local features from the input temporal features. The convolutional kernel slides over the input sequence, capturing local correlations between features at adjacent time points (e.g., increasing or decreasing patterns of pulse amplitude within multiple consecutive time windows) through a weighted summation operation. The output is a feature map containing the short-term patterns of PD. Subsequently, a single max pooling layer (e.g., a 2×2 pooling kernel) downsamples the feature map. By selecting the maximum value in a local region, the feature map retains key features (e.g., the distribution of high-amplitude pulses) while reducing the feature dimensionality and computational complexity. Through this process, the CNN branch can effectively capture the short-term patterns of the local discharge signal in the fused feature vector in 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 characteristic information reflecting the instantaneous state of the discharge.

[0077] The long-term short-term memory (LSTM) branch serves as a module for extracting long-term trends in characteristic gases in oil. Its input is the time series data of characteristic gases in oil from the expanded mapping library in step 3, specifically the volume fractions of hydrogen (H2), acetylene (C2H2), total hydrocarbons, carbon monoxide (CO), and carbon dioxide (CO2) in the fused feature vector from step 2. The LSTM layer implements sequence modeling on this time series data through an internal forget gate, input gate, and output gate mechanism. The forget gate determines whether to retain or discard historical information (such as the acetylene content in the previous five time windows); the input gate updates 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 captures long-term trends in characteristic gases in oil (such as the continuous increase in acetylene content over time, reflecting the continuous thermal decomposition of insulation materials; the periodic fluctuations in carbon dioxide content, reflecting the impact of ambient temperature and humidity changes on gas diffusion), providing the model with temporal correlation features that reflect the progress of insulation degradation.

[0078] To facilitate multimodal feature fusion, the outputs of the CNN and LSTM branches are integrated through a fully connected layer, and a softmax activation function is used to output a probability distribution of discharge severity. Specifically, the short-term pattern features (e.g., sudden changes in pulse density) output by the CNN branch and the long-term trend features (e.g., sustained growth of acetylene) output by the LSTM branch are input into a fully connected layer (e.g., 32 neurons). Linear transformations and nonlinear activations (e.g., ReLU functions) are used to achieve cross-modal feature fusion and extract correlations between the two types of features (e.g., a sudden change in pulse density accompanied by a rapid growth of acetylene may indicate an arc discharge). Finally, a softmax activation function maps the fused features to the [0, 1] interval, outputting a probability distribution of three discharge severity categories: mild (partial discharge, probability 0.8); moderate (spark discharge, probability 0.15); and severe (arcing / mixed discharge, probability 0.05), corresponding to the early, mid, and late stages of insulation degradation, respectively.

[0079] Specifically, the neural network-based power transformer discharge fault diagnosis method described in the present invention is a specific implementation of "deploying the trained diagnostic model on the edge computing device" in step 5. The diagnostic model trained in step 4 is deployed on the substation edge computing device, which receives the fused feature vector generated in real time in step 2 (which is the real-time output of step 2 "generating a fused feature vector for subsequent analysis") and outputs the discharge severity probability distribution and the corresponding operating condition label (such as "load mutation scenario").

[0080] As a specific implementation of "updating model parameters through an online learning mechanism" in step 5, the online learning mechanism accumulates 100 sets of new fused feature vectors generated in step 2 through edge computing devices (including complex working condition data such as mixed discharge and intermittent arc, which are new data in step 2 "generating fused feature vectors for subsequent analysis"), and uses the stochastic gradient descent algorithm to fine-tune the weight parameters of the fully connected layer and output layer of the hybrid neural network described in step 5 to maintain the adaptability of the diagnostic model in step 5 to the dynamic operating environment.

[0081] The deployment and online learning mechanism of the diagnostic model in step 5 of the present invention realizes the online application and dynamic optimization of the diagnostic model through the real-time data processing capabilities and incremental learning algorithms of edge computing devices, solving the problem of decreased generalization ability of traditional models due to changes in the operating environment.

[0082] The diagnostic model is deployed to achieve real-time fault diagnosis. The hybrid neural network model trained in Step 4 is deployed on a substation edge computing device (such as the edge computing unit of an integrated automation system). This device, typically located in a control cabinet within the substation, features low-latency data processing (response time ≤ 100ms) and supports real-time communication with sensors and SCADA systems (via Modbus or MQTT). The deployment process includes model parameter loading (transferring trained weight parameters from the training server to the edge computing device), interface configuration (defining the input format of the fused feature vector and the output protocol for diagnostic results), and runtime environment initialization (installing a deep learning framework such as TensorFlow Lite to optimize model inference speed). After the deployment is completed, the edge computing device receives the fused feature vector generated in step 2 in real time (transmitted via industrial Ethernet, with a frequency consistent with the feature generation cycle in step 2, that is, one group every 5 minutes), calls the model inference interface to output the discharge severity probability distribution (such as "partial discharge probability 85%, spark discharge probability 10%, arc discharge / mixed discharge probability 5%") and the corresponding working condition label (such as "load mutation scenario" and "vibration interference scenario"), and displays the diagnosis results through the human-machine interface (HMI) for reference by operation and maintenance personnel.

[0083] The online learning mechanism aims to maintain the model's adaptability to dynamic operating environments, enabling incremental updates of model parameters using the local computing resources of the edge computing device. Specifically, the edge computing device has a built-in data cache module. Upon receiving each new fused feature vector generated in step 2 (including data from complex operating conditions such as mixed discharges and intermittent arcs), it performs validity checks (e.g., eliminating missing data or outliers) and stores valid data in the cache. When the cache accumulates 100 sets of valid data, the online learning process is triggered: the stochastic gradient descent (SGD) algorithm is used to fine-tune the weight parameters of the fully connected and output layers of the hybrid neural network (with a learning rate set to 0.001 to avoid overfitting). Fine-tuning the fully connected and output layers is chosen because these layers are directly responsible for the fusion of multimodal features and classification decisions, and are more sensitive to adaptive adjustments to new operating conditions. The underlying feature extraction layers (convolutional and LSTM layers) of the CNN and LSTM branches, which have already learned basic discharge pattern characteristics, do not require frequent adjustments. Through this fine-tuning operation, the model can gradually learn the characteristic patterns under new operating conditions (such as the synergistic growth law of the volume fractions of hydrogen and acetylene in high temperature and high humidity environments), and improve its adaptability to dynamic operating environments.

[0084] Specifically, the neural network-based power transformer discharge fault diagnosis method described in the present invention is a specific implementation of the "dynamic verification of model diagnostic performance" in step 5. The dynamic verification process extracts 100 sets of fused feature vectors corresponding to complex operating conditions (mixed discharge, intermittent arc) from the newly generated fused feature vectors in step 2 (real-time data from step 2, "generating fused feature vectors for subsequent analysis") every month as a test set. These fused feature vectors are input into the diagnostic model deployed on the edge computing device in step 5 to evaluate the accuracy of discharge severity diagnosis.

[0085] Based on the goal of "maintaining the model's adaptability to dynamic operating environments" in step 5, if the accuracy rate does not reach the preset threshold for two consecutive months, re-execute steps 3 (expanding the mapping relationship library, supplementing the fusion feature vector and discharge pattern association data of the new operating condition, and providing more comprehensive training samples for step 4) and step 4 (adjusting the hybrid neural network structure, such as increasing the number of LSTM hidden layer nodes, to improve the model's feature extraction capability for the new operating condition) to generate an updated diagnostic model and replace the old model.

[0086] 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, thereby solving the problem of diagnostic performance degradation of traditional models caused by changes in operating conditions.

[0087] The dynamic validation process focuses on evaluating the model's generalization capabilities for complex operating conditions. This is achieved through the following steps: Each month, from the newly generated fused feature vectors in step 2 (i.e., 6-dimensional feature data collected and processed in real time, including hydrogen and acetylene volume fractions and normalized discharge energy), complex operating condition data (e.g., samples labeled "mixed discharge" or "intermittent arc") are filtered by operating condition label. A random sampling method is used to select 100 groups as the test set. The selection of 100 groups is based on statistical significance considerations (a sample size of ≥30 satisfies 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. The model's inference interface is called to output a probability distribution of discharge severity (e.g., "arc discharge / mixed discharge probability 90%)." This is then compared with the manually labeled true severity (determined by experts based on discharge characteristics and historical cases). The diagnostic accuracy (number of correctly classified samples divided by total number of samples) is calculated. This accuracy reflects the model's ability to identify recent complex operating conditions and is a key metric for evaluating model adaptability.

[0088] If the diagnostic accuracy rate falls below a preset threshold (such as the 90% required by industry standards or a benchmark value determined based on historical training data) for two consecutive months, 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 operating transformers, discharge data under new operating conditions (such as the newly added "high-frequency vibration + sudden load increase" composite scenario) is collected. The corresponding fused feature vectors (such as abnormal fluctuations in the hydrogen gas integral fraction under vibration scenarios and high-frequency pulses in the standardized discharge energy) are simultaneously generated. The new discharge mode (such as "vibration-induced mixed discharge") and severity (such as "severe") are annotated. The new data is then added to the expanded mapping relationship library in a 3:7 ratio (actual operation data: laboratory simulation data).

[0089] Then, re-execute step 4, adjust the hybrid neural network structure, and retrain the model. The specific method of structural adjustment is determined by the new operating conditions. If the new data shows more complex long-term trends in characteristic gases in the oil (e.g., a multi-stage increase in acetylene content), increase the number of hidden nodes in the LSTM layer (e.g., from 16 to 24) to improve the ability to capture long-term dependencies. If the short-term patterns of the partial discharge signal become more discrete (e.g., an increase in the frequency of sudden changes in pulse density), increase the number of convolutional kernels in the CNN layer (e.g., from 16 to 32) to enhance the ability to extract local features. The adjusted network structure is retrained using the expanded mapping library (training set: validation set = 8:2), using the same optimization strategy as step 4 (cross-entropy loss function, gradient descent, and dropout to suppress overfitting) until the validation set error rate stabilizes.

[0090] Finally, the trained new model is deployed to the edge computing device, replacing the old model. Before the replacement, the new model's diagnostic accuracy on both the historical test set and the new operating condition test set must be verified through offline testing (both must be ≥ the preset threshold).

[0091] This invention addresses the low efficiency of data fusion for partial discharge patterns, oil chromatograms, and electrical quantities through the collaborative collection and feature standardization of multi-source data. Step 1: Deploy a high-frequency current coil transformer, a resistor-capacitor voltage divider, an online gas sensor, and a SCADA system 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 operating parameters (load current, ambient temperature and humidity). GPS or NTP protocols are used to unify timestamps and achieve time alignment of the multi-source data. Step 2: Using the sampling period of the characteristic gas data in oil as a window, sliding window statistics are performed on the partial discharge signal (extracting statistics such as the pulse amplitude mean and frequency distribution variance), converting the high-frequency transient signal into low-frequency statistics. Simultaneously, the Z-score method is used to normalize the electrical quantity parameter (load current) and the discharge characteristic parameter (discharge energy) to eliminate dimensional differences. This generates a six-dimensional fusion feature vector containing the gas volume fraction and the normalized discharge energy, providing structured input for multi-source data fusion.

[0092] Explanation of the technical features of the present invention

[0093] Partial discharge signals: Electrical signals such as high-frequency current and voltage fluctuations are generated when the insulation material inside a power transformer partially breaks down due to factors such as aging, electric field distortion, or mechanical stress. These signals are early signs of insulation degradation and directly reflect the instantaneous characteristics of the discharge (such as intensity and frequency), making them the core electrical basis for fault diagnosis.

[0094] Characteristic gas data in oil: This data measures the volume fraction of gases generated by the degradation of insulating materials (such as oil-paper) during transformer oil discharge or thermal decomposition. These gases primarily include hydrogen (H2), acetylene (C2H2), total hydrocarbons (the 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 degradation and serves as non-electrical signature for determining the type and severity of discharge.

[0095] Power supply system operating parameters: These parameters reflect the external environment and load data of the transformer's operating status, including load current (reflecting the power load), voltage amplitude (reflecting system voltage stability), and ambient temperature and humidity (affecting insulation material performance). These parameters are used to analyze the correlation between discharge faults and system operating status and assist in locating fault scenarios.

[0096] Timestamp: A unified time identifier (e.g., "2025-05-21 10:00:00") is used to label multi-source data (partial discharge signals, characteristic gas data in oil, and operating parameters). Synchronizing sensor clocks via GPS or network time protocol aligns different types of data in time, facilitating subsequent analysis of multi-dimensional information within the same time period.

[0097] Statistical aggregation: For high-frequency transient partial discharge signals (such as high-frequency current and voltage pulses), the sampling period (e.g., 5 minutes) of the characteristic gas-in-oil data is used as a window to extract statistics such as the average pulse amplitude (reflecting the overall level of discharge intensity), the frequency distribution variance (reflecting the density of discharges), and the number of discharges per unit time (reflecting the frequency of discharge activity). This operation converts high-frequency signals into low-frequency statistics that match the time scale of the gas-in-oil data, resolving the time scale mismatch problem of multi-source data.

[0098] Normalization: The process of eliminating dimensional differences between electrical parameters (such as load current in amperes) and discharge characteristic parameters (such as discharge energy in joules) of different physical dimensions using methods such as the Z-score. For example, load current and discharge energy are converted to dimensionless values ​​with a mean of 0 and a standard deviation of 1, making the two types of data comparable and facilitating subsequent feature fusion.

[0099] Fusion feature vectors: Structured data (e.g., [H2, C2H2, total hydrocarbons, CO, CO2, normalized discharge energy]) is formed by combining characteristic gas data in oil (e.g., hydrogen and acetylene volume fractions) with standardized discharge characteristic parameters (e.g., normalized discharge energy) in a fixed order. This vector integrates the physical meaning (gas content, discharge intensity) and numerical characteristics (time alignment, dimensionality uniformity) of multi-source data. It serves as input to the neural network model, establishing a correlation between features and discharge patterns.

[0100] Extended Mapping Relationship Library: This sample set consists of a mixture of laboratory simulation data and actual operating data. Laboratory simulation data is acquired using devices such as a pin-plate electrode model (generating partial discharge patterns and spark discharge data) and a suspended electrode model (generating mixed discharge data). Actual operating data is collected using temporary monitoring devices, including complex discharge data (such as intermittent arcing) under scenarios such as sudden load changes and drastic temperature and humidity fluctuations. Each data entry in the library includes a fused feature vector, discharge pattern (such as partial discharge pattern or mixed discharge), discharge severity (mild / moderate / severe), and operating condition label (such as "load sudden change"), providing multi-scenario sample support for model training.

[0101] Convolutional Neural Network (CNN) branch: This neural network module is used to extract short-term patterns in partial discharge signals. It takes as input the temporal characteristics of the partial discharge signal (e.g., the average pulse amplitude and the variance of the frequency distribution). The convolutional layer captures local features (e.g., changes in the density of the pulse waveform). After dimensionality reduction by the pooling layer, the output reflects the instantaneous state of the discharge (e.g., a sudden change from a sparse to a dense pulse sequence).

[0102] Long Short-Term Memory (LSTM) branch: This neural network module is designed to capture long-term trends in characteristic gases in oil. It takes time series data of characteristic gases in oil (such as acetylene volume fraction) as input and processes this data through internal gating mechanisms (forget gate, input gate, and output gate). It extracts temporal correlation features of gas content (such as the accumulation rate of acetylene over time) to reflect the progress of insulation degradation.

[0103] Hybrid neural network architecture: This model consists of a CNN and LSTM branches. The CNN extracts short-term patterns in partial discharge signals (such as changes in pulse density), while the LSTM captures long-term trends in oil-gas emissions (such as the continued growth of acetylene). The outputs of the two branches are fused through a fully connected layer and then activated using a softmax function to output a probability distribution of discharge severity (e.g., "mild: 85% probability of partial discharge"), enabling collaborative learning of multimodal features.

[0104] Edge computing devices: Low-latency data processing equipment deployed within substations (such as edge computing units in integrated automation systems) receives fused feature vectors and runs diagnostic models in real time. This reduces data transmission latency and supports real-time output of discharge severity and operating condition labels (such as sudden load changes), meeting the power system's rapid response requirements.

[0105] Online learning mechanism: The edge computing device accumulates newly collected fused feature vectors (such as mixed discharge and intermittent arc data) and uses a stochastic gradient descent algorithm to fine-tune the weight parameters of the hybrid neural network's fully connected and output layers. This mechanism enables the model to gradually learn new operating conditions (such as the synergistic growth of hydrogen and acetylene in high-temperature and high-humidity environments), maintaining adaptability to dynamic operating environments.

[0106] Dynamic verification process: Complex working condition data (such as mixed discharge and intermittent arc) are extracted from the newly generated fusion feature vector every month as a test set and input into the diagnostic model to evaluate the accuracy. If the preset threshold (such as 90%) is not reached for two consecutive months, the extended mapping relationship library is rebuilt (supplemented with new working condition data) and the network structure is adjusted (such as increasing the number of LSTM hidden layer nodes). A new version of the model is generated to replace the old model, so that the model can continuously adapt to environmental changes.

[0107] The specific implementation of the present invention focuses on multi-source data fusion, model generalization capability improvement, and dynamic adaptability optimization for power transformer discharge fault diagnosis. Combined with the actual operation scenarios of the power supply system, the technical solutions described in the claims are achieved through the following technical means:

[0108] For multi-source data collaborative collection, sensors are deployed throughout the power transformer and along its power supply lines. A high-frequency current coil transformer is installed at the transformer bushing's end shield grounding wire, the primary path for partial discharge current to discharge to the ground. This location effectively captures the high-frequency current signal (frequency range 100kHz-10MHz) generated by partial discharge. A resistor-capacitor voltage divider is connected in parallel to the high-voltage winding's outlet, where the electric field is concentrated. Partial discharge can induce transient voltage changes. The resistor-capacitor voltage divider simultaneously measures the voltage fluctuation signal during the discharge process, complementing the high-frequency current signal. Oil characteristic gas data is collected by an online gas sensor integrated into the transformer's oil flow circuit (e.g., the oil flow pipeline connecting the oil pillow to the transformer). Photoacoustic spectroscopy is used to continuously monitor the volume fractions of five characteristic gases: hydrogen (H2), acetylene (C2H2), total hydrocarbons (the sum of methane, ethylene, and other hydrocarbons), carbon monoxide (CO), and carbon dioxide (CO2). This reflects the dynamic changes in these gases during insulation material degradation. Power supply system operating parameters (load current, voltage amplitude, ambient temperature and humidity) are accessed through the SCADA system via the Modbus protocol. The Modbus protocol supports real-time communication between multiple devices, ensuring stable acquisition of system operating status data. All sensors and the SCADA system synchronize their clocks using GPS (without a network, time error ≤ 1μs) or NTP (with a network, time error ≤ 10ms). This ensures that partial discharge signals (sampling frequency 1MHz), characteristic gas data in oil (sampling period 5 minutes), and operating parameters (sampling period 1 minute) are marked with unified timestamps, providing a time reference for subsequent data alignment.

[0109] In the multi-source data feature fusion phase, the high-frequency transient characteristics of partial discharge signals (such as high-frequency current and voltage pulses) are statistically aggregated using the 5-minute sampling period of the oil-gas characteristic data as a sliding window. Within each 5-minute window, the average pulse amplitude (reflecting the overall level of discharge intensity within that 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 within the window divided by 5 minutes, reflecting the frequency of discharge activity) are calculated. This converts the high-frequency transient signals into low-frequency statistics consistent with the time scale of the oil-gas characteristic data. Z-score normalization is used to eliminate dimensional differences in the electrical parameter (load current obtained from the SCADA system, in amperes) and the discharge characteristic parameter (discharge energy calculated by integrating the voltage and current signals, in joules). The mean and standard deviation of the load current and discharge energy are calculated, respectively, and the raw values ​​are converted to a dimensionless form (raw value - mean) / standard deviation) to make the two parameters comparable. Finally, the volume fractions of the characteristic gases in the five oils were combined with the normalized discharge energy to form a six-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 served as the structured input data for the subsequent mapping library construction.

[0110] In the construction phase of the extended mapping relationship library, laboratory simulation data is generated through the needle-plate electrode model, the floating electrode model and the vibration interference device: the needle-plate electrode model uses a pure tungsten needle electrode and a brass plate electrode to form a fixed gap (such as 2mm) to simulate local 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 to the needle-plate electrode (simulating the looseness of the winding pad), and adjusts the distance between the floating electrode and the main electrode (such as 1-3mm) to generate mixed discharge data of the local discharge superimposed on the floating discharge model (large fluctuations in pulse amplitude and irregular frequency); the vibration interference device applies 0.5g-2g mechanical vibration (g is the acceleration of gravity) through an electromagnetic exciter to simulate the discontinuous phenomenon of the discharge channel caused by mechanical stress, and generates intermittent arc data (the discharge channel exists for 2-8 seconds and the energy fluctuates periodically). Actual operating data was collected through temporary monitoring devices under scenarios involving sudden load changes (such as ±15% voltage fluctuations caused by transformer paralleling / de-coupling) and dramatic changes in ambient temperature and humidity (temperature from 25°C to 85°C, humidity from 30% to 90%). Fusion feature vectors and corresponding complex patterns such as mixed discharge and intermittent arcing were simultaneously recorded. An extended mapping database, comprised of 70% laboratory data and 30% field data, contains each data entry containing a fusion feature vector, discharge mode (partial discharge, spark discharge, arc discharge, mixed discharge, intermittent arc), discharge severity (mild: partial discharge; moderate: spark discharge; severe: arc discharge / mixed discharge), and operating condition labels (such as "sudden load change," "vibration interference," and "high temperature and high humidity"), providing a multi-scenario sample set for model training.

[0111] The hybrid neural network model training phase utilizes a hybrid architecture of a convolutional neural network (CNN) and a long short-term memory (LSTM) network. The CNN branch inputs the temporal characteristics of the PD signal (mean pulse amplitude, frequency distribution variance, and number of discharges per unit time). A 3×3 convolutional layer extracts local features (such as the increasing or decreasing pattern of pulse amplitude within a continuous time window). A 2×2 max pooling layer reduces the dimensionality (preserving key features such as the distribution location of high-amplitude pulses) to capture short-term patterns in the PD signal (such as sudden changes from sparse to dense pulse sequences). The LSTM branch inputs time series data of characteristic gases in the oil (H2, C2H2, total hydrocarbons, CO, and CO2 volume fractions). Using a single LSTM layer with 16 hidden nodes, the LSTM layer employs a forget gate (retaining historical gas content information), an input gate (updating the current gas increment), and an output gate (integrating historical and current information) to capture long-term trends in gas content (such as the continuous increase in acetylene content over time, reflecting the ongoing thermal decomposition of the insulating material). The outputs of the two branches are fused through a fully connected layer (32 nodes) and then activated using a softmax function to output a probability distribution of discharge severity (mild, moderate, severe). During training, the weight parameters are optimized using a cross-entropy loss function combined with gradient descent. A dropout layer (20% dropout rate) is introduced to mitigate overfitting. Early stopping is triggered when the validation set error rate does not decrease after five consecutive iterations.

[0112] During the model deployment and dynamic optimization phase, the trained hybrid neural network model is deployed on the substation's edge computing device (such as the edge computing unit of the integrated automation system). This device supports Modbus / MQTT protocols for real-time communication with sensors and SCADA systems (response time ≤ 100ms). It receives real-time fused feature vectors (one set every 5 minutes) and calls the model inference interface to output discharge severity probability distributions (e.g., "85% probability of partial discharge, 10% probability of spark discharge, 5% probability of arc discharge / mixed discharge") and corresponding operating condition labels (e.g., "load sudden change scenario"). To maintain the model's adaptability to dynamic environments, the edge computing device has a built-in data cache module. Each new fused feature vector (including data for complex operating conditions such as mixed discharge and intermittent arcing) is validated to eliminate missing data or outliers. When 100 valid data sets have been cached, the weight parameters of the fully connected and output layers are fine-tuned using the stochastic gradient descent algorithm (learning rate 0.001). (The underlying convolutional and LSTM layers, as they have already learned basic features, do not require frequent adjustments.) Every month, complex working condition data (mixed discharge, intermittent arc) are filtered from the newly generated fusion feature vectors according to the working condition labels, and 100 groups are randomly selected as test sets. The diagnostic accuracy is input into the model to evaluate (target ≥90%). If the threshold is not reached for two consecutive months, the mapping library construction (supplemented with new working condition data such as "high-frequency vibration + sudden load increase") and model training (such as increasing the number of LSTM hidden layer nodes to 24 to improve the ability to capture long-term trends) are re-executed to generate a new version of the model and replace the old model.

[0113] Through the above implementation methods, the present invention achieves time alignment and dimensional unification of multi-source data, solving the problem of low efficiency in data fusion of local discharge patterns, oil chromatograms, electrical quantities, etc.; by expanding the mapping relationship library to cover multi-scenario discharge patterns, and combining hybrid neural networks to fuse short-term discharge patterns with long-term gas trends, the model's generalization ability for 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 the power transformer, ultimately improving the operational stability of the power supply system.

[0114] To address the problem of insufficient generalization of neural network models for complex discharge patterns, the present invention improves the model's learning capabilities by expanding the mapping relationship library and designing a multimodal network structure. Step 3 constructs an expanded mapping relationship library, integrating laboratory simulation data (local discharge patterns and spark discharge data generated by the needle-plate electrode model, and mixed discharge data generated by the suspended electrode model) with actual operating data (intermittent arc data under load mutations and drastic changes in ambient temperature and humidity), covering five types of patterns, including local discharge patterns and mixed discharges, and annotating discharge severity and operating condition labels to provide a multi-scenario sample set for model training. Step 4 adopts a hybrid neural network structure: the CNN branch extracts short-term patterns of local discharge signals (such as changes in pulse density), and the LSTM branch captures long-term trends of characteristic gases in the oil (such as the accumulation rate of acetylene content). The outputs of the two branches are fused through a fully connected layer and then output through Softmax to achieve collaborative learning of multimodal features and enhance the model's ability to recognize complex patterns.

[0115] 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. Step 5 deploys the trained model on the edge computing device of the substation, receives the fused feature vector in real time and outputs the diagnostic results; every time 100 sets of new fused feature vectors (including complex working condition data) are accumulated, the stochastic gradient descent algorithm is used to fine-tune the weight parameters of the fully connected layer and the output layer to achieve incremental updates of the model parameters. 100 sets of complex working condition data are extracted each month as a test set to evaluate the diagnostic accuracy. If the threshold is not reached for two consecutive months, step 3 (expanding the mapping relationship library) and step 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 new working condition features, avoid the degradation of diagnostic performance due to environmental changes, and ultimately improve the operational stability of the power supply system.

Claims

1. A method for diagnosing power transformer discharge faults based on a neural network, characterized in that: include: Step 1: Using a sensor group deployed on the power transformer and power supply lines, partial discharge signals, characteristic gas data in oil, and power supply system operating parameters are collected in real time. Discharge characteristic parameters are calculated based on the partial discharge signals and timestamped. The discharge characteristic parameters include pulse amplitude, frequency distribution, and the 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 standardize the power supply system operating 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, which includes fused feature vectors corresponding to partial discharge mode, spark discharge mode, and mixed discharge mode. Collect actual operation data through a temporary monitoring device of an actual operating transformer. The actual operation data includes fused feature vectors corresponding to intermittent arc mode under scenarios of sudden load change and drastic changes in temperature and humidity. Based on the fused 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 association data between the fused feature vectors and discharge modes, including mixed discharge and intermittent arc. Step 4: using the fused feature vector in the extended mapping relationship library constructed in Step 3 as input, a hybrid neural network structure including a convolutional neural network branch and a long short-term memory network branch is used to train a diagnostic model, wherein the diagnostic model outputs a probability distribution of discharge severity; Step 5: Deploy the diagnostic model trained in step 4 on the edge computing device, use the fused feature vector generated in real time in step 2 to diagnose the discharge fault, and update the parameters of the diagnostic model trained in step 4 through the online learning mechanism; The sensor group includes a high-frequency current coil transformer for collecting partial discharge signals, a resistor-capacitor voltage divider for measuring discharge voltage fluctuations, an online gas sensor for monitoring characteristic gas data in oil, and a SCADA system for obtaining power supply system operating parameters; The high-frequency current coil mutual inductor 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 RC voltage divider is connected in parallel to the high voltage winding output terminal, and the voltage fluctuation signal during the discharge process is measured to supplement the collection of partial discharge signals. The online gas sensor is integrated into the transformer oil flow loop to acquire the characteristic gas data in the oil in step 1 by monitoring the volume fractions of hydrogen, acetylene, total hydrocarbons, carbon monoxide, and carbon dioxide. The SCADA system is connected via the Modbus protocol and acquires the power supply system operating parameters in step 1 by obtaining the load current, voltage amplitude, and ambient temperature and humidity. The sensor group and SCADA system are synchronized through GPS or network time protocol to ensure that the timestamps of partial discharge signals, characteristic gas data in oil and operating parameters are consistent.

2. The method for diagnosing power transformer discharge faults based on a neural network according to claim 1, characterized in that: Also includes: The statistical aggregation of partial discharge signals uses the sampling period of the characteristic gas data in oil as a window, performs sliding window statistics on the high-frequency current and voltage signals collected by the high-frequency current coil mutual inductor and the resistance-capacitance voltage divider, extracts the pulse amplitude mean, frequency distribution variance, and number of discharges per unit time within the window, and converts the high-frequency instantaneous partial discharge signal in step 1 into a statistic that matches the time scale of the characteristic gas data in oil. For the electrical quantity parameters and discharge characteristic parameters collected in step 1, the load current obtained by the SCADA system and the discharge energy calculated based on the voltage and current signals are standardized using the Z-score method to eliminate the dimensional differences between current and energy; The characteristic gas data in the oil in step 1 is combined 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 online gas sensor and the standardized discharge energy, which serves as the structured input data for the associated discharge mode in the expanded mapping relationship library in step 3.

3. The method for diagnosing power transformer discharge faults based on a neural network according to claim 2, characterized in that: Also includes: Laboratory simulation data is generated using a needle-plate electrode model, a suspended electrode model, and a vibration interference device. The laboratory simulation data includes fused feature vectors and corresponding partial discharge modes, spark discharge modes, arc discharge modes, and mixed discharge data of partial discharge and suspended discharge models. Actual operation data is collected through temporary monitoring devices under load mutation and drastic changes in ambient temperature and humidity. The discharge data includes the fused feature vector generated in step 2 and the corresponding mixed discharge and intermittent arc discharge modes. Based on the association requirements between the fused feature vector generated in step 2 and the discharge mode, the extended mapping relationship library is composed of a mixture of 70% laboratory simulation data and 30% actual operation data. Each piece of data includes the fused feature vector generated in step 2, the discharge mode, the discharge severity and the operating 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.

4. The method for diagnosing power transformer discharge faults based on a neural network according to claim 3, characterized in that: Also includes: The convolutional neural network branch inputs the temporal features of the partial discharge signal in the expanded mapping relationship library in step 3, extracts local features through a layer of convolutional layer, reduces the dimension through a layer of maximum pooling layer, and captures the short-term pattern of the partial discharge signal in the fused feature vector in step 2; The long short-term memory network branch inputs the time series data of characteristic gases in oil from the expanded mapping relationship library in step 3, captures the long-term variation trend of gas content through a single LSTM layer, and extracts the temporal correlation features of characteristic gases in oil from the fused feature vector in step 2. Based on the multimodal 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 the 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.

5. The method for diagnosing power transformer discharge faults based on neural network according to claim 4, characterized in that: Also includes: The diagnostic model trained in step 4 is deployed on the substation edge computing device, which receives the fused feature vector generated in real time in step 2 and outputs the probability distribution of discharge severity and the corresponding working condition label. The online learning mechanism accumulates 100 sets of new fused 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 output layer of the hybrid neural network in step 5.

6. The method for diagnosing power transformer discharge faults based on neural network according to claim 5, characterized in that: Also includes: The dynamic verification process extracts 100 sets of fused feature vectors corresponding to complex working conditions from the newly generated fused feature vectors in step 2 every month as a test set. These fused feature vectors are input into the diagnostic model deployed on the edge computing device in step 5 to evaluate the accuracy of discharge severity diagnosis. If the accuracy rate does not reach the preset threshold for two consecutive months, re-execute steps 3 and 4 to generate an updated diagnostic model and replace the old model.

Images