A method and apparatus for detecting abnormalities in a converter transformer, and a storage medium.

Abstract

This invention belongs to the field of converter transformer detection technology, specifically relating to a method, device, and storage medium for detecting converter transformer anomalies. The method involves acquiring monitoring data of the converter transformer, including external driving parameters and internal state response parameters. For each external driving parameter, time-domain correlation analysis is used to analyze its optimal hysteresis time with each internal state response parameter. Using the optimal hysteresis time, each internal state response parameter is feature-aligned with the external driving parameter to construct a time-domain aligned multidimensional feature matrix. The time-domain aligned multidimensional feature matrices corresponding to all external driving parameters are input into a trained health baseline prediction model to predict the dynamic health baseline corresponding to each internal state response parameter. Anomaly detection of the converter transformer is then performed based on the prediction results. This invention achieves a leap from "post-event alarm" to "very early warning," significantly advancing the warning time window.

Description

Technical Field

[0001] This invention belongs to the field of converter transformer detection technology, specifically relating to a converter transformer anomaly detection method and device, and storage medium. Background Technology

[0002] The current online monitoring and fault diagnosis technology for converter transformers faces several key bottlenecks in its practical application in engineering.

[0003] At the early warning mechanism level, traditional methods are highly susceptible to interference from system operating conditions. Traditional methods rely heavily on static absolute value thresholds or gas generation rate thresholds set by standards such as the "Guidelines for Analysis and Judgment of Dissolved Gases in Transformer Oil" for alarm purposes. However, in actual operation, the thermal decomposition process of the internal insulating oil of converter transformers is directly and strongly affected by fluctuations in the grid's operating load (operating power). Normal thermal stress gas generation under high load conditions is often misjudged as an early fault (false alarm), while abnormally small gas generation under low load conditions may not reach the static threshold, resulting in missed alarms. Traditional static threshold alarms are easily affected by grid load fluctuations and exhibit significant lag.

[0004] At the data analysis and feature extraction level, existing models adhere to the approach of independent monitoring of single variables. Existing data-driven methods (such as single neural networks, support vector machines, etc.) predict parameters such as oil chromatography gas concentration, core grounding current, and SF6 gas state as isolated random time series. Chinese invention patent application CN118501775A, published on August 16, 2024, discloses a method and apparatus for identifying converter transformer faults based on vibration signals. This method first uses sensors to collect vibration signals, determines the comparison vibration signal and the target analysis vibration signal, extracts the DC-side vibration signal and determines its characteristic distribution range, and then judges the transformer fault and determines the location of the fault. This patent can determine whether the characteristic value of the separated DC-side vibration signal is within the characteristic distribution range. If the characteristic value is within the distribution range, it determines that the winding and converter valve side of the converter transformer under analysis are operating normally; if the characteristic value is outside the distribution range, a fault has occurred at the winding and converter valve side of the converter transformer, achieving accurate judgment of the fault location of the converter transformer. These methods neglect the thermodynamic hysteresis coupling effect between "external driving force (operating power)" and "internal state response (specific gas production, leakage current abrupt change)" under multiphysics conditions. The abrupt change in converter transformer load propagates to the rise in oil temperature, and then to the decomposition of insulating oil producing characteristic gases (such as hydrogen, acetylene, total hydrocarbons, etc.) which diffuse to the sensors, resulting in a significant physical time delay. The lack of decoupling and exploration of this temporal evolution characteristic leads to low model prediction accuracy and poor physical interpretability.

[0005] At the level of anomaly detection and decision-making, mainstream methods lack the ability to dynamically track the "personalized" health baseline of equipment, and cannot accurately intercept equipment defects in the very early stage of their evolution from the incubation period to the sudden change period, making it difficult to meet the predictive maintenance needs of power equipment under high-dimensional and complex operating conditions.

[0006] The aforementioned systemic shortcomings collectively hinder the leap from "post-event alarm" to "very early and accurate warning" in converter transformer condition assessment technology. Summary of the Invention

[0007] The purpose of this invention is to provide a method, device, and storage medium for detecting anomalies in converter transformers, in order to solve the problems in the prior art where traditional static threshold alarms are easily affected by power grid load fluctuations and have serious lag, as well as the lack of decoupling and mining of time-series evolution characteristics, resulting in low model prediction accuracy.

[0008] To address the aforementioned technical problems, the first aspect of this invention provides a method for detecting anomalies in a converter transformer, the method comprising:

[0009] 1) Acquire monitoring data of the converter transformer. The monitoring data includes external driving parameters and internal state response parameters. External driving parameters can cause delay changes in internal state response parameters.

[0010] 2) For an external driving parameter, the optimal hysteresis time between it and each internal state response parameter is analyzed using the time-domain correlation analysis method. The optimal hysteresis time is used to align the features of each internal state response parameter with the external driving parameter to construct a time-domain aligned multidimensional feature matrix.

[0011] 3) Input the time-domain aligned multidimensional feature matrix corresponding to all external driving parameters into the trained health baseline prediction model to predict the dynamic health baseline corresponding to each internal state response parameter. The dynamic health baseline is the normal level that the internal state response parameter should be in a healthy operating state under the current conditions; and perform converter transformer anomaly detection based on the prediction results.

[0012] In one possible approach, the optimal hysteresis time is analyzed using a cross-correlation function.

[0013] In one possible implementation, the health baseline prediction model includes an LSTM temporal feature extraction layer, an attention weighting layer, and an output layer; the LSTM temporal feature extraction layer is used to extract long-term temporal dependencies of time series using LSTM; the attention weighting layer is used to weight the importance of the hidden states at each time step output by the LSTM temporal feature extraction layer to obtain a context vector; and the output layer is used to output a dynamic health baseline based on the context vector.

[0014] In one possible implementation, the output layer is a deep neural network.

[0015] In one possible implementation, the method for detecting converter transformer anomalies based on prediction results is as follows:

[0016] Collect historical data confirming that the converter transformer is in a healthy state and calculate the state residuals accordingly to obtain the state residual sequence. The state residuals refer to the residuals between the actual values ​​of each monitored internal state response parameter and the corresponding dynamic health baseline.

[0017] The probability density function of the fitted state residual sequence is integrated, and the confidence quantile in the integration result is determined based on the set confidence interval. Then, the upper and lower limit thresholds of the residual corresponding to the confidence quantile are obtained in reverse.

[0018] Obtain the actual state residual between each internal state response parameter monitored online and the corresponding dynamic health baseline. If the actual state residual of N consecutive time steps is not within the range of the upper and lower limit thresholds of the residual, it is determined that the converter transformer equipment deviates from the dynamic health baseline.

[0019] In one possible implementation, when analyzing the transformer oil-paper insulation system of a converter transformer, the external driving parameters include operating power, and the internal state response parameters include total hydrocarbons in the body oil chromatogram and single hydrogen in the body riser seat; when analyzing the core clamping system of a converter transformer, the external driving parameters include motion power and harmonic characteristics, and the internal state response parameters include the grounding current of the body core clamping system; when analyzing the bushing sealing system of a converter transformer, the external driving parameters include operating power and ambient temperature, and the internal state response parameters include the SF6 gas density / pressure of the valve-side bushing.

[0020] In one possible implementation, a nonparametric kernel density estimation method is used to fit the probability density function of the state residual sequence.

[0021] In one possible implementation, 1) also includes using a normalization method to make the acquired monitoring data dimensionless and aligning the acquired monitoring data to a uniform time resolution.

[0022] To address the aforementioned technical problems, a second aspect of the present invention provides a converter transformer anomaly detection device, comprising a processor, the processor being configured to execute a computer program to implement the steps of the method in any possible implementation of the first aspect of the present invention.

[0023] To address the aforementioned technical problems, a third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method in any possible implementation of the first aspect of the present invention.

[0024] The beneficial effects of this invention are as follows: Breaking away from conventional single-variable monitoring approaches, this invention divides the monitoring data of the converter transformer into external driving parameters and internal state response parameters. It employs time-domain correlation analysis to analyze the thermodynamic hysteresis coupling effect between these two parameters, taking into account the physical time delay. This eliminates temporal misalignments caused by factors such as physical propagation paths and sensor response differences when constructing the time-domain aligned multidimensional feature matrix. This ensures the constructed time-domain aligned multidimensional feature matrix truly reflects causal relationships rather than simple correlations, enhancing the model's interpretability and generalization ability. Furthermore, a trained health baseline prediction model is used for dynamic health baseline prediction. This makes the dynamic health baseline no longer a rigid numerical value, but a "health baseline" that fluctuates with the external driving parameters, significantly improving the robustness of anomaly detection. This invention achieves a leap from "post-event alarm" to "very early warning," significantly advancing the warning time window and securing valuable time for predictive maintenance and power outage repair plans for the converter transformer, demonstrating excellent engineering practicality. Attached Figure Description

[0025] Figure 1 This is a flowchart of the converter transformer anomaly detection method of the present invention;

[0026] Figure 2 This is a block diagram of the multidimensional heterogeneous data preprocessing and feature extraction module that takes into account thermal hysteresis effect of the present invention;

[0027] Figure 3 This is a diagram of the dynamic baseline neural network structure of the fusion attention mechanism of the present invention;

[0028] Figure 4 This is a diagram illustrating the verification and early warning triggering effect of measured data of a converter transformer based on the method of this invention;

[0029] Figure 5 This is a simplified flowchart of the processing flow of the present invention;

[0030] Figure 6 This is a structural diagram of the converter transformer abnormality detection device of the present invention. Detailed Implementation

[0031] This invention breaks with conventional univariate monitoring approaches by combining external driving parameters and internal state response parameters. It utilizes time-domain correlation analysis to construct multi-scale feature engineering to deeply mine the hysteresis coupling effect driven by thermodynamics, and based on this, predicts the dynamic health baseline. To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings.

[0032] Implementation method 1 of a converter transformer anomaly detection method:

[0033] The present invention provides a method for detecting abnormalities in a converter transformer, the process of which is as follows:

[0034] Step 1: Obtain monitoring data of the converter transformer. The monitoring data includes external driving parameters and internal state response parameters. External driving parameters can cause delay changes in internal state response parameters.

[0035] When analyzing different subsystems of a converter transformer, the collected external driving parameters and internal state response parameters differ, reflecting a differentiated design based on the subsystem's physical characteristics and excitation-response mechanism. For example, when analyzing the transformer oil-paper insulation system, external driving parameters include operating power, while internal state response parameters include total hydrocarbons in the body oil chromatogram and single hydrogen atoms in the body riser. When analyzing the core clamping system, external driving parameters include dynamic power and harmonic characteristics, while internal state response parameters include the core clamping grounding current. Similarly, when analyzing the bushing sealing system, external driving parameters include operating power and ambient temperature, while internal state response parameters include the SF6 gas density / pressure in the valve-side bushing. This differentiated design allows for a more precise characterization of the local hysteresis coupling behavior of each subsystem, thereby optimizing the optimal hysteresis time and constructing a more physically meaningful feature matrix.

[0036] Furthermore, considering the differences in the dimensions of various types of data, this embodiment uses a normalization method to perform dimensionless processing on various feature data; at the same time, to address the problem of inconsistent sampling frequencies of different sensors, a mean filtering and linear interpolation algorithm are combined to align the resampling of multidimensional heterogeneous sequences to a unified time resolution.

[0037] Step 2: For an external driving parameter, use time-domain correlation analysis to analyze its optimal hysteresis time with each internal state response parameter; use the optimal hysteresis time to align the features of each internal state response parameter with the external driving parameter to construct a time-domain aligned multidimensional feature matrix.

[0038] In this embodiment, the optimal hysteresis time is analyzed using the cross-correlation function.

[0039] For example, the external drive parameter is selected as operating power, the internal state response parameters are selected as total hydrocarbons in bulk oil chromatography and single hydrogen in bulk riser, and the historical time step window is set to... The optimal lag time, obtained based on the total hydrocarbons from bulk oil chromatography and the operating power, is: The optimal hysteresis time, calculated based on the single hydrogen atom of the bulk riser and the operating power, is: At any time Extract past time periods The operating power sequence within the time frame is used as the driving feature input to extract the lag time period. State response sequence and hysteresis time of total hydrocarbons in bulk oil chromatography The state response sequence of the single hydrogen atom in the internal body is used as the state feature input, and the whole matrix constitutes a time-domain aligned multidimensional feature matrix corresponding to the operating power. Of course, if there are other external driving parameters (such as ambient temperature), a time-domain aligned multidimensional feature matrix is ​​calculated using the same method as above. The time-domain aligned multidimensional feature matrices corresponding to all external driving parameters are used as inputs to the subsequent health baseline prediction model.

[0040] Step 3: Input the time-domain aligned multidimensional feature matrix corresponding to all external driving parameters into the trained health baseline prediction model to predict the dynamic health baseline corresponding to each internal state response parameter. The dynamic health baseline is the normal level that the internal state response parameter should be in a healthy operating state under the current conditions.

[0041] In this embodiment, the health baseline prediction model includes an LSTM temporal feature extraction layer, an attention weighting layer, and an output layer. The LSTM temporal feature extraction layer is used to extract long-term temporal dependencies of time series using LSTM. The attention weighting layer is used to weight the importance of the hidden states at each time step output by the LSTM temporal feature extraction layer to obtain a context vector. The output layer is used to output a dynamic health baseline based on the context vector. As one embodiment of the output layer, the output layer is a deep neural network (DNN); as one embodiment of the attention weighting layer, the attention weighting layer is additive attention.

[0042] As a better approach, population-based metaheuristic optimization algorithms, such as genetic algorithms, can be used to adaptively optimize the model parameters of the health baseline prediction model.

[0043] The health baseline is dynamic, fluctuating with the load. It can adaptively raise the baseline when there is normal gas / current fluctuation under high load, and sensitively detect abnormalities when there is abnormal gas production under low load, thus greatly improving the robustness of the early warning system.

[0044] Step 4: Perform converter transformer anomaly detection based on the predicted dynamic health baseline.

[0045] In this embodiment, the method for detecting converter transformer anomalies based on the predicted dynamic health baseline is as follows:

[0046] 1. Collect historical data confirming the health status of the converter transformer and calculate the corresponding state residuals to obtain the state residual sequence. The state residuals refer to the actual values ​​of the monitored internal state response parameters. With the corresponding dynamic health baseline residuals between , .

[0047] 2. Fit the probability density function of the state residual sequence, integrate the probability density function, determine the confidence quantile in the integration result based on the set confidence interval, and then inversely calculate the upper and lower limits of the residual corresponding to the confidence quantile. and Furthermore, a nonparametric kernel density estimation method (KDE) can be selected to fit the probability density function. For example, when the centroid interval is set to 99%, the corresponding confidence quantiles are 1% and 99%.

[0048] 3. Obtain the actual state residuals between each internal state response parameter monitored online and the corresponding dynamic health baseline. If the actual state residuals for N consecutive time steps are not within the range of the baseline, the remaining data will be considered as follows: and Within the range of components, it is determined that the converter transformer equipment deviates from the dynamic health baseline.

[0049] This completes the converter transformer anomaly detection method of the present invention.

[0050] Implementation method 2 of a converter transformer anomaly detection method:

[0051] To address the shortcomings of traditional static threshold alarms for converter transformers, which are susceptible to grid load fluctuations and exhibit significant lag, this invention proposes a dynamic early warning method for early faults based on multi-source heterogeneous parameter fusion and temporal feature decoupling—a method for detecting anomalies in converter transformers. This approach breaks away from conventional single-variable monitoring, treating "operating power" as an external physical driving force and considering "online oil chromatography," "grounding current," and "shelter SF6 status" as internal multi-dimensional state responses. By constructing multi-scale feature engineering to deeply mine the hysteresis coupling effect driven by thermodynamics, a deep learning model with an attention mechanism is used to predict the expected state of the equipment under specific load conditions, accurately constructing a "dynamic health baseline" that adaptively fluctuates with system load. Finally, by calculating the state residual between multi-dimensional measured values ​​and this baseline in real time, and combining it with kernel density estimation methods to generate an adaptive residual trigger threshold, accurate interception and early warning are achieved at the very early stage of equipment defects evolving from the incubation period to the abrupt change period. This provides a highly interpretable and practically valuable intelligent solution for predictive maintenance of power equipment under high-dimensional and complex operating conditions. The following section will elaborate on this approach. Figure 1 , Figure 2 , Figure 3 and Figure 5 A detailed explanation of the entire method and process is provided below:

[0052] Step 1: Sensing and preprocessing of multidimensional heterogeneous state parameters of converter transformers.

[0053] Collect multi-source heterogeneous monitoring data of the converter transformer, specifically including external drive parameters and internal state response parameters.

[0054] External driving parameters Figure 1 The "equipment load data" specifically includes operating power and ambient temperature. Internal state response parameters are... Figure 1 The "equipment status data" specifically includes the total hydrocarbon / acetylene content in the online oil chromatography of the main body, the online single hydrogen content of the main body riser, the grounding current of the main body core clamp, and the SF6 gas density / pressure in the valve side bushing. External drive parameters and internal state response parameters together constitute the dynamic baseline modeling input for "external drive-internal response" in subsequent steps.

[0055] The min-max normalization method is used to perform dimensionless processing on various feature data to eliminate the dimensional differences between electrical and chemical quantities. To address the issue of inconsistent sampling frequencies among different sensors, a mean filtering and linear interpolation algorithm are combined to align the resampling of multidimensional heterogeneous sequences to a uniform time resolution (e.g., 1 hour / sample).

[0056] Step 2: Decoupling of multi-scale features and construction of a sliding window for thermal hysteresis effect.

[0057] The whole process is as follows Figure 2 As shown.

[0058] Define the operating power sequence as follows:

[0059]

[0060] In the formula, This represents the operating power at each time point.

[0061] Define the state response sequence as follows:

[0062]

[0063] In the formula, This represents the values ​​of a certain type of state response parameter at continuous time points, such as a total hydrocarbon concentration sequence. When multiple state response parameters exist, each parameter forms its own state response sequence, and hysteresis correlation analysis and dynamic baseline modeling are performed separately with external driving parameters.

[0064] Due to the time lag in insulating oil decomposition caused by load-generated heat, the optimal lag window between operating power and state response is analyzed using the cross-correlation function. (i.e., optimal hysteresis time):

[0065]

[0066] Optimization makes The largest The value serves as the time delay constant for feature alignment, i.e., the optimal hysteresis time. A fusion feature matrix incorporating the hysteresis effect is constructed. (Also known as time-domain aligned multidimensional features), which includes power features (such as mean, difference slope) and state features within the current and historical sliding windows. Specifically, the historical time step window is set as... At any time Extract past time periods The operating power sequence within the time frame is used as the driving feature input, and the lag time period is extracted. The state response sequence (such as a specific gas concentration) within the system is used as the state feature input to form a multidimensional time series feature pair.

[0067] Step 3: Dynamic health baseline modeling based on Attention-LSTM.

[0068] Construct an LSTM deep network with an attention mechanism (Attention-LSTM). Its input is a multi-dimensional sliding window feature matrix that takes into account the hysteresis effect. This includes historical operating power sequences, power change characteristics, ambient temperature characteristics, and historical sequences of corresponding state response parameters.

[0069] The constructed network structure is as follows: The Attention-LSTM model consists of a sequentially connected LSTM temporal feature extraction layer, an Attention weighting layer, and a Deep Neural Network (DNN). The LSTM temporal feature extraction layer is used to extract long-term temporal dependencies of the time series using LSTM. The Attention weighting layer is used to weight the importance of the hidden states at each time step output by the LSTM temporal feature extraction layer to obtain a context vector. The DNN is used to output a dynamic health baseline based on this context vector. Alternatively, multiple state parameters can be combined into a multi-dimensional state-response matrix, but in the anomaly detection stage, the prediction baseline and residual are still calculated separately for each state parameter.

[0070] The LSTM temporal feature extraction layer is used to extract long-term temporal dependencies in time series. The core formulas for forget gate, input gate, output gate, cell state and hidden state updates are as follows:

[0071]

[0072]

[0073]

[0074]

[0075]

[0076]

[0077] In the formula, For the output of the forget gate, For the sigmoid function, This is the weight matrix. This indicates the concatenation of the previous hidden state with the current input. For bias, For input gate output, This is the weight matrix. For bias, Candidate cell state, This is the weight matrix. For bias, , These represent the current cell state and the previous cell state, respectively. For the output of the output gate, This is the weight matrix. For bias, It is currently hidden.

[0078] Introducing an attention mechanism to the output of the LSTM hidden layer Dynamic weights are assigned to amplify the impact of key high-load periods on gas production behavior:

[0079]

[0080]

[0081]

[0082] Figure 3 The leftmost "attention mechanism" diagram corresponds to the process of weighting the importance of the hidden states at each time step of the LSTM output. Specifically, the LSTM first outputs a sequence of hidden states. , The attention mechanism represents the temporal features at time step t through the weight matrix. Bias and score vector The importance score for each time step was calculated. The attention weights are then obtained through Softmax normalization. , This represents the contribution of the t-th time step to the current prediction result; subsequently, the hidden states are weighted and summed to obtain the context vector C, which is then input into the subsequent DNN output to create a dynamic health baseline. , and These are the parameters obtained through adaptive learning during model training. The larger the value, the more significant the impact of load or state changes at that moment on the current baseline forecast.

[0083] The predicted value is output through the fully connected layer in the DNN. This output is not an absolute prediction of the future, but rather represents the "dynamic health baseline" that the converter transformer should exhibit under the current input power load timing characteristics.

[0084] Furthermore, a genetic algorithm is used to adaptively optimize the model parameters of the health baseline prediction model.

[0085] Step 4: Adaptive residual anomaly detection based on kernel density estimation (KDE).

[0086] Calculate the actual online monitoring status value Dynamic health baseline output by the model State residuals between:

[0087]

[0088] Within the historical intervals during which the equipment was confirmed to be in a healthy state, residual sequences were collected, and the probability density function of the residuals was fitted using nonparametric kernel density estimation (KDE). ,by To smooth bandwidth, Let the kernel function (Gaussian kernel) be used, and its probability density distribution be estimated as follows:

[0089]

[0090] By setting a confidence interval (e.g., 99%), the cumulative distribution function (CDF) is calculated by integrating the probability density function. Then, the upper limit threshold of the residual that adapts to the system load fluctuation is obtained by inversely calculating the residuals at the 1% and 99% confidence quantiles. With lower threshold .

[0091] Step 5: Multi-physics collaborative early warning decision-making and very early defect interception.

[0092] This invention extends the aforementioned dynamic baseline method to the three core subsystems of converter transformers, enabling full utilization of online monitoring data and multi-source collaborative early warning:

[0093] 1) Oil-paper insulation system: Combined analysis of "total hydrocarbons in the bulk oil chromatography" and "single hydrogen in the bulk riser seat", utilizing the spatial diffusion lag caused by oil flow circulation to locate the internal latent overheating / discharge gas generation source.

[0094] 2) Core clamping system: Driven by "operating power and harmonic characteristics", the system uses "core clamping grounding current" as a response to construct a normal leakage current baseline and accurately capture weak multi-point grounding faults.

[0095] 3) Casing sealing system: Driven by ambient temperature and operating power, it takes "SF6 gas in the valve side casing" as the response, eliminates the regular fluctuations in SF6 pressure / density caused by sudden changes in the external environment, and accurately warns of micro-leakage or moisture defects in the casing.

[0096] During online monitoring, real-time residuals are continuously calculated. Design a sliding window judgment logic: when N consecutive time steps meet the condition... The system determines that the equipment deviates from the dynamic health baseline and triggers an early fault warning signal.

[0097] To verify the effectiveness of the method of the present invention, the actual operation monitoring data of the low YDC converter transformer at pole 2 of a certain converter station from March to April 2024 were used as a sample for testing.

[0098] As attached Figure 4 As shown in (a), this figure illustrates the state monitoring and dynamic baseline fitting curves considering external load-driven conditions. During the normal operation phase of the equipment in early to mid-March, the measured oil chromatography state response (solid line in the figure) fluctuated with changes in external operating conditions such as operating power. The Attention-LSTM dynamic health baseline (dashed line in the figure) proposed in this invention can accurately track and encapsulate this normal fluctuation trajectory, effectively avoiding false alarms caused by normal thermal stress gas production due to sudden load increases in traditional fixed thresholds. However, in late March, the measured gas concentration showed an abnormal increase that deviated from the load pattern. At this time, the dynamic baseline still maintained the healthy gas production level that the current load should maintain. The significant divergence between the two trajectories intuitively revealed the occurrence of latent defects.

[0099] As attached Figure 4 As shown in (b) of the figure, this diagram illustrates the adaptive anomaly detection and early warning triggering process based on KDE. During the healthy operation phase, the state residual (measured value minus predicted baseline) remains stably within the adaptive threshold band generated by KDE (an envelope consisting of 99% and 1% confidence levels), demonstrating strong anti-interference capability. When internal defects in the equipment enter a slow development phase, gas production characteristics change, and the state residual increases dramatically. For example... Figure 4(b) clearly indicates that on March 30, 2024, the measured residuals experienced a step change and continuously exceeded the 99% confidence level adaptive upper limit threshold set based on KDE. At this time, although the absolute concentration of the gas often had not yet reached the static alarm threshold specified in traditional guidelines, this system had already successfully triggered an alarm accurately at the "early warning interception point (2024-03-30)" marked in the figure, thanks to its high sensitivity to the residual distribution characteristics.

[0100] These examples fully demonstrate that the method proposed in this invention achieves a leap from "post-event alarm" to "very early warning", significantly advancing the warning time window and gaining valuable golden time for predictive maintenance and power outage repair plans of converter transformers, demonstrating excellent engineering practical effectiveness.

[0101] An implementation method for a converter transformer anomaly detection device:

[0102] The present invention provides a converter transformer anomaly detection device, such as... Figure 6 As shown, the system includes a memory, a processor, a system bus, and a computer program stored in the memory. The processor and memory communicate and exchange data with each other via the system bus. The processor executes the computer program to implement a converter transformer anomaly detection method according to the present invention.

[0103] The processor can be a general-purpose CPU (such as Intel Xeon or ARM Cortex-A series) for software deployment; a GPU (such as NVIDIA Tesla) can be used to accelerate deep learning forward inference; or an FPGA / ASIC (such as Xilinx UltraScale+ or dedicated neural network acceleration chips) can be used to achieve hardware-level low-latency inference. Memory can be dynamic random access memory (RAM) for high-speed read and write, NVRAM / SSD for persistent model weights and historical monitoring data, or tape library / cloud object storage for long-term archiving and offline retraining.

[0104] The device can be deployed in a field rack configuration, where it is installed in a transformer control cabinet or a power distribution room. In this configuration, data acquisition and processing are completed locally, resulting in extremely low network latency and millisecond-level real-time anomaly alarms; it also poses less risk to network security. However, rack space is limited, requiring high levels of heat dissipation and power redundancy, and hardware upgrades necessitate on-site shutdown, leading to relatively high maintenance costs. Alternatively, the device can be deployed as an edge computing box / industrial PC, placing it in a substation power distribution room or as an independent edge server near an HVDC station. This deployment offers greater computing resources (CPU+GPU / FPGA combination) than a rack configuration, supporting model iteration and online learning while maintaining low transmission latency. Redundant power supply boxes can be used to improve reliability. However, it is limited by on-site electromagnetic compatibility (EMC) and ambient temperature, requiring additional protective measures, and may lose cloud backup if network connectivity is poor. Furthermore, a centralized cloud / data center deployment is also possible, where the device's processing functions are encapsulated as containers or microservices and run on a cloud platform or in a company data center. This deployment method offers elastic scaling of computing power, enabling rapid deployment of the latest deep learning models. It facilitates unified management and cross-site business analysis, with hardware maintenance handled by the service provider and minimal on-site facility requirements. However, it relies on reliable industrial private networks or 5G backhaul links, and transmission latency and bandwidth fluctuations may affect real-time performance. Data security and privacy compliance require additional encryption and authentication measures.

[0105] One embodiment of a computer-readable storage medium:

[0106] This invention discloses a computer-readable storage medium comprising a 256GB solid-state drive (SSD) storing a set of instructions applied to a processor (i.e., the program code of the converter transformer anomaly detection method of this invention) and key data files for model inference, including trained Attention-LSTM network weights, kernel density estimation (KDE) threshold parameters, and sampling templates for multi-source time-series data such as historical operating power, oil chromatography, and grounding current. In other words, when the computer program stored on the computer-readable storage medium is executed by the processor, it implements the converter transformer anomaly detection method of this invention.

[0107] This SSD can be integrated into industrial control servers in substations (rack-mounted deployment), or mounted in portable edge computing boxes and connected to the main control unit via a high-speed USB-3.0 interface, or mounted on cloud servers in the company's data center via network storage protocols (such as NFS / SMB) to achieve remote software upgrades and unified management. The advantages of using SSDs as storage media include high read / write speeds (sequential read / write up to 550MB / s, random read / write up to 100kIOPS), meeting the needs of real-time anomaly detection for model loading and rapid access to large-scale time-series data. Compared to disposable media such as optical discs or magnetic tapes, SSDs are rewritable, have a long lifespan, are small in size, and are highly vibration-resistant, making them suitable for field scenarios with complex electromagnetic environments and large temperature variations. Network mounting further enhances storage scalability and maintenance convenience, enabling multi-site transformer monitoring systems to share a unified model library and historical database, reducing overall system operation and maintenance costs.

[0108] This invention has the following characteristics:

[0109] (1) Overcoming the defects of false alarms / false alarms of static thresholds and achieving highly robust early warning. It breaks through the technical bottleneck that traditional fixed threshold standards cannot adapt to drastic load fluctuations. By using "operating power" as an external driving force input, the model no longer predicts rigid values, but a "healthy baseline" that fluctuates dynamically with the load. It can adaptively raise the baseline when there is normal gas production / current fluctuation under high load. Tests have shown that it can reduce the false alarm rate of the system under complex operating conditions by more than 85%. It can sensitively detect anomalies when there is abnormal gas production under low load, which greatly improves the robustness of the early warning system.

[0110] (2) Multi-physics panoramic defense, fully exploring the value of heterogeneous data. The solution breaks through the limitations of single oil chromatography monitoring and creatively incorporates heterogeneous data such as single hydrogen in the riser, grounding current of the iron core clamp, and SF6 gas in the bushing into a unified "external drive-internal response" dynamic baseline architecture. It realizes full-coverage collaborative diagnosis of the three core subsystems of insulation, iron core, and bushing, and builds a tight technical barrier.

[0111] (3) Integrating physical mechanisms with global optimization mechanisms enhances model interpretability and generalization ability. A multi-scale sliding window feature engineering approach based on the "thermodynamic hysteresis coupling effect" is innovatively introduced, and a genetic algorithm is used to achieve adaptive optimization of the network structure. This enables the AI ​​model to accurately learn the physical causal laws of "heat-driven - delayed response," significantly improving prediction accuracy and providing a physically interpretable industrial-grade solution.

[0112] (4) Achieving very early detection of fault evolution, enabling predictive maintenance. A non-parametric adaptive trigger threshold is generated using kernel density estimation (KDE), exhibiting extremely high sensitivity to minute state residuals. This method can detect abnormal trajectory deviations 15 to 30 days or even earlier than traditional absolute value alarms, during the early "quantitative change" stage of equipment defect evolution from latency to abrupt change. This transforms emergency repairs into calm, planned maintenance, demonstrating high engineering practical value and economic benefits.

Claims

1. A method for detecting anomalies in a converter transformer, characterized in that, The method includes: 1) Acquire monitoring data of the converter transformer. The monitoring data includes external driving parameters and internal state response parameters. External driving parameters can cause delay changes in internal state response parameters. 2) For an external driving parameter, the optimal hysteresis time between it and each internal state response parameter is analyzed using the time-domain correlation analysis method. The optimal hysteresis time is used to align the features of each internal state response parameter with the external driving parameter to construct a time-domain aligned multidimensional feature matrix. 3) Input the time-domain aligned multidimensional feature matrix corresponding to all external driving parameters into the trained health baseline prediction model to predict the dynamic health baseline corresponding to each internal state response parameter. The dynamic health baseline is the normal level that the internal state response parameter should be in a healthy operating state under the current conditions; and perform converter transformer anomaly detection based on the prediction results.

2. The converter transformer anomaly detection method according to claim 1, characterized in that, The optimal hysteresis time is analyzed using the cross-correlation function.

3. The converter transformer anomaly detection method according to claim 1, characterized in that, The health baseline prediction model includes an LSTM temporal feature extraction layer, an attention weighting layer, and an output layer; the LSTM temporal feature extraction layer is used to extract long-term temporal dependencies of time series using LSTM; The attention weighting layer is used to weight the hidden states at each time step of the LSTM temporal feature extraction layer to obtain a context vector; the output layer is used to output a dynamic health baseline based on the context vector.

4. The converter transformer anomaly detection method according to claim 3, characterized in that, The output layer is a deep neural network.

5. The converter transformer anomaly detection method according to claim 1, characterized in that, The method for detecting converter transformer anomalies based on prediction results is as follows: Collect historical data confirming that the converter transformer is in a healthy state and calculate the state residuals accordingly to obtain the state residual sequence. The state residuals refer to the residuals between the actual values ​​of each monitored internal state response parameter and the corresponding dynamic health baseline. The probability density function of the fitted state residual sequence is integrated, and the confidence quantile in the integration result is determined based on the set confidence interval. Then, the upper and lower limit thresholds of the residual corresponding to the confidence quantile are obtained in reverse. Obtain the actual state residual between each internal state response parameter monitored online and the corresponding dynamic health baseline. If the actual state residual of N consecutive time steps is not within the range of the upper and lower limit thresholds of the residual, it is determined that the converter transformer equipment deviates from the dynamic health baseline.

6. The converter transformer anomaly detection method according to claim 1, characterized in that, When analyzing the transformer oil-paper insulation system of a converter transformer, external driving parameters include operating power, and internal state response parameters include total hydrocarbons in the body oil chromatogram and single hydrogen in the body riser seat. When analyzing the core clamping system of a converter transformer, external driving parameters include motion power and harmonic characteristics, and internal state response parameters include the grounding current of the body core clamping system. When analyzing the bushing sealing system of a converter transformer, external driving parameters include operating power and ambient temperature, and internal state response parameters include the SF6 gas density / pressure in the valve-side bushing.

7. The converter transformer anomaly detection method according to claim 5, characterized in that, The probability density function of the state residual sequence is fitted using a nonparametric kernel density estimation method.

8. The converter transformer anomaly detection method according to any one of claims 1 to 7, characterized in that, 1) also includes using a normalization method to perform dimensionless processing on the acquired monitoring data and aligning the acquired monitoring data to a uniform time resolution.

9. A converter transformer anomaly detection device, comprising a processor, characterized in that, The processor executes a computer program to implement the steps of the method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 8.

Images