Direct-current insulation fault accurate positioning method fusing break variable and trend analysis

Abstract

The invention belongs to the technical field of fault positioning, and provides a direct current insulation fault accurate positioning method fusing break variable and trend analysis. The method comprises the following steps: recording and storing current waveform data of a fault branch within a set time period before and after a fault event after triggering break variable recording; continuously monitoring the insulation parameters of each branch in parallel, and generating steady-state trend data of the insulation parameters changing along with time; multi-dimensional features are extracted from the steady state trend data, a dynamic first step length and a dynamic second step length are obtained based on analysis of the multi-dimensional features, and transient features and trend change features are extracted according to the dynamic first step length and the dynamic second step length; processing the transient characteristics and the trend change characteristics by using a confidence coefficient calculation model to obtain a confidence coefficient score; and when the confidence score is higher than an alarm threshold value, sending out a highest-level alarm and executing fault branch positioning. According to the invention, instantaneous disturbance and permanent insulation degradation can be effectively distinguished, and the false alarm rate and the missing report rate are obviously reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fault location, in particular to a direct current insulation fault precise positioning method fusing mutation variable and trend analysis. BACKGROUND

[0002] The direct current system is the heart of many key fields such as power, rail transit, data center, etc., and the continuity and reliability of its power supply are crucial. The insulation condition of the direct current system to the ground is directly related to the safe and stable operation of the entire system. Once an insulation fault occurs, it may not only cause power interruption, but also lead to equipment damage and even serious accidents such as fire. Therefore, achieving rapid and accurate positioning of the insulation fault of the direct current system has always been the core issue of technical research and product development in this field.

[0003] At present, the mainstream direct current insulation fault positioning method relies on the steady-state trend analysis principle. This method continuously monitors the voltage and current signals of the positive and negative busbars to the ground through high-precision sensors, and analyzes the trend of the insulation parameters of each branch with time based on an insulation resistance calculation model. By identifying the continuous and slow downward trend of the insulation resistance of a specific branch, early or gradual insulation degradation problems can be warned and located. This method has certain detection effect on chronic and permanent insulation defects.

[0004] However, in actual operation conditions, transient disturbances often occur during the operation of the direct current system, such as switching of large-capacity loads, switching operations or external electromagnetic interference, etc. These events can cause transient mutations in branch currents, which may appear as a short-lived spike or burr on the steady-state trend curve. The existing method cannot effectively distinguish between transient disturbances caused by operational disturbances and the initial stage of permanent insulation degradation that truly indicates insulation breakdown. This can easily lead to false alarms, increasing the troubleshooting burden of maintenance personnel.

[0005] In addition, when a fault occurs, especially a transient fault, its key transient characteristics (such as mutation amplitude, waveform, duration, etc.) will quickly disappear. Traditional steady-state monitoring methods cannot capture and record these transient and rich fault characteristic information, resulting in the absence of fault fingerprints that can be used for subsequent analysis, which is not conducive to in-depth diagnosis and analysis of the fault cause.

[0006] In summary, the existing insulation fault positioning technology based on single steady-state trend analysis has deficiencies in anti-interference ability, response speed and fault information integrity, and cannot meet the growing demand for precise, reliable and rapid positioning of direct current insulation faults.

[0007] Therefore, there is an urgent need for a method that can comprehensively analyze transient and steady-state information to accurately distinguish and locate the type of insulation fault. SUMMARY

[0008] Therefore, the present application aims to overcome the above-mentioned defects of the prior art, and proposes a DC insulation fault precise positioning method fusing mutation variable and trend analysis.

[0009] The present application provides a DC insulation fault precise positioning method fusing mutation variable and trend analysis, comprising the following steps: Real-time monitoring of the current signal of each branch of the DC system, when the current change or the current change rate of any branch is detected to exceed the preset mutation variable threshold, triggering the mutation variable recording; After triggering the mutation variable recording, recording and saving the current waveform data of the fault branch within a set time period before and after the fault event; and, in parallel, continuously monitoring the insulation parameters of each branch, and generating the steady-state trend data of the insulation parameters changing with time; From the steady-state trend data, a multi-dimensional feature is extracted, based on which a dynamic first step and a second step are analyzed, based on the first step, a transient feature is extracted from the current waveform data, and based on the second step, a trend change feature is extracted from the steady-state trend data; Using a confidence calculation model to process the transient feature and the trend change feature, to obtain a confidence score for characterizing the event as a permanent insulation fault; When the confidence score is higher than the alarm threshold, issuing a highest-level alarm and performing fault branch positioning.

[0010] The beneficial technical effects of the present application are at least: The present application triggers recording by real-time monitoring of the current mutation variable, and constitutes a double criterion with steady-state trend analysis, uses the self-adaptive step dynamically analyzed from the steady-state data to accurately extract transient and trend features, and then fuses and analyzes through a confidence model, to finally realize precise positioning of the DC system insulation fault, effectively distinguishes transient disturbance and permanent insulation deterioration, significantly reduces the false alarm rate and the missed alarm rate, and improves the reliability, accuracy and intelligent level of fault positioning. BRIEF DESCRIPTION OF DRAWINGS

[0011] Figure 1 A flowchart of a DC insulation fault precise positioning method fusing mutation variable and trend analysis disclosed by the embodiments of the present application; Figure 2 A structure diagram of a step mapping model disclosed by the embodiments of the present application. DETAILED DESCRIPTION

[0012] The technical solutions of the present application will be further described in detail below through the drawings and embodiments.

[0013] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] To facilitate understanding of the embodiments of the present invention, further explanations and descriptions will be provided below with reference to the accompanying drawings and specific embodiments. These embodiments do not constitute a limitation on the embodiments of the present invention.

[0015] like Figure 1 As shown in the figure, this invention discloses a method 100 for accurate DC insulation fault location by integrating abrupt change and trend analysis, comprising the following steps: Step 10: Monitor the current signal of each branch of the DC system in real time. When the current change or rate of change of any branch exceeds the preset threshold for sudden change, trigger the sudden change waveform recording. In this step, the solution of the present invention is applied to a DC insulation fault monitoring system. The system uses high-precision current sensors (such as Hall sensors) deployed in each branch to continuously collect and monitor the current signals of all branches of the DC system in real time.

[0016] When the change in current or rate of change in any branch of a DC system exceeds a preset sudden change threshold within a very short time (e.g., milliseconds), an abnormal current event is immediately determined to have occurred in that branch, and the sudden change recording function is automatically triggered. This mechanism ensures extremely high sensitivity to sudden faults. It is understood that the sudden change threshold is a safety margin value set based on statistical analysis of the maximum possible disturbance current under normal operating conditions of the DC system; for example, it can be set to 3 to 5 times the amplitude of normal current fluctuations.

[0017] Step 20: After triggering the sudden change waveform recording, record and save the current waveform data of the faulty branch within a set time period before and after the fault event; and monitor the insulation parameters of each branch in parallel and continuously, and generate steady-state trend data of the insulation parameters changing over time. In this step, the solution of the present invention includes two parallel and independent data streams, as follows: The first data stream (i.e., transient data capture): After triggering the sudden change waveform recording in step 10, the complete current waveform data of the faulty branch is recorded and saved at a high sampling rate (e.g., not less than 10kHz) for a set period of time before, during, and after the fault event. This current waveform data includes at least rich transient characteristic information such as sudden change peak value, waveform shape, oscillation frequency, and duration.

[0018] The second data stream (steady-state data accumulation): The steady-state monitoring thread is running independently and continuously at the same time, which collects the current of each branch and the voltage of the bus to ground at a lower frequency (for example, once per second), and calculates the insulation parameters (the most core is the insulation resistance value) of each branch through the insulation resistance calculation model (such as the balanced bridge method). These insulation parameters are recorded in chronological order, thereby generating steady-state trend data that can reflect its long-term and slow change rule.

[0019] Step 30, a plurality of dimensional features are extracted from the steady-state trend data, and a dynamic first step length and a second step length are obtained based on the analysis of the plurality of dimensional features, a transient feature is extracted from the current waveform data based on the first step length, and a trend change feature is extracted from the steady-state trend data based on the second step length; In this step, the steady-state trend data generated in step 20 is analyzed in depth, and a plurality of dimensional features that can represent the current running state of the DC system are extracted, including but not limited to: the average value, variance (volatility) of the recent insulation resistance, the decline slope within a certain time window, and the autocorrelation of historical data, etc.

[0020] Based on the above-mentioned plurality of dimensional features, through a pre-defined decision logic or mapping model, the following two key analysis step lengths are dynamically analyzed and obtained: The first step length: used to extract the transient feature from the current waveform data. When the DC system stability is poor (such as large variance) or there are early signs of degradation (such as negative slope), the first step length can be shortened to capture finer transient details at a higher time resolution, avoiding missing small fault signs.

[0021] The second step length: used to extract the trend change feature from the steady-state trend data. When a long-term and slow degradation trend needs to be confirmed, the second step length can be lengthened to filter out short-term fluctuation noise, so as to more accurately represent the macroscopic change rule of the insulation parameter.

[0022] Using the two dynamically adjusted step lengths, transient features (such as sudden change energy, main oscillation frequency) are extracted from high-resolution current waveform data, and trend change features (such as resistance decay amount within a certain time window) are extracted from steady-state trend data.

[0023] It can be understood that this adaptive multi-step feature extraction mechanism adopted in this step can ensure that the most discriminative feature information can be obtained under different DC system states.

[0024] Step 40, using a confidence calculation model to process the transient feature and the trend change feature to obtain a confidence score for representing that the event is a permanent insulation fault; In this step, a confidence calculation model is pre-built and trained, which can be built based on machine learning classifiers (such as support vector machines or random forests).

[0025] The confidence calculation model receives the transient and trend change features extracted in step 30 as input. It then performs weighted fusion and comprehensive analysis on these features from different time scales and with different physical meanings, outputting a quantitative confidence score. This confidence score is a value in the range [0,1], used to comprehensively characterize the probability that the current event is a permanent insulation failure. A higher score indicates stronger consistency between the transient abrupt change and steady-state degradation evidence, and a greater certainty of classifying it as a permanent failure.

[0026] Step 50: When the confidence score is higher than the alarm threshold, issue the highest level alarm and perform fault branch location.

[0027] In this step, the confidence score calculated in step 40 is compared with a preset alarm threshold. This alarm threshold is set after verification using a large amount of historical fault data, and it can balance the false alarm rate and the missed alarm rate.

[0028] If the confidence score is higher than the alarm threshold, the event is determined to be a permanent insulation fault, and the highest level alarm is immediately triggered (such as flashing red on the monitoring interface, sounding an alarm, or sending a text message). The specific branch number where the fault occurred is accurately located, providing maintenance personnel with a clear and specific handling target.

[0029] If the confidence score does not exceed the alarm threshold, it is judged as a transient disturbance or suspected fault, and recorded or a low-level warning is issued, thereby effectively reducing the false alarm rate and improving the reliability and credibility of the DC insulation fault monitoring system.

[0030] This invention uses real-time monitoring of sudden current changes to trigger waveform recording and establishes a dual criterion with steady-state trend analysis. It utilizes an adaptive step size dynamically analyzed from steady-state data to accurately extract transient and trend features, and then integrates and judges them through a confidence model. Ultimately, it achieves accurate location of insulation faults in DC systems, effectively distinguishes between transient disturbances and permanent insulation degradation, significantly reduces false alarm and missed alarm rates, and improves the reliability, accuracy, and intelligence level of fault location.

[0031] As an example, the step of triggering a sudden change waveform recording when the current change or rate of change in any branch exceeds a preset sudden change threshold includes: The instantaneous rate of change and the amount of change within a set time window of the current in each branch are calculated in real time. When the absolute value of at least one of the instantaneous rate of change and the amount of change of any branch exceeds the preset rate threshold and quantity threshold respectively, the sudden change waveform of the faulty branch is immediately triggered.

[0032] In this embodiment, the instantaneous rate of change (di / dt) and the amount of change (ΔI) within a set time window are calculated in parallel and in real time. The instantaneous rate of change reflects the drasticness and speed of current change, and is particularly sensitive to capturing extremely fast transient processes on the order of nanoseconds to microseconds, such as insulation breakdown; while the amount of change reflects the cumulative deviation of the current within a slightly longer but fixed time window (e.g., 1-10 milliseconds), and is more effective in identifying a sustained increase in leakage current with a slightly slower rate of change but a larger magnitude.

[0033] Simultaneously, preset threshold values ​​for the instantaneous rate of change and the amount of change are established for rate and amount, respectively. If either the instantaneous rate of change or the cumulative amount of change of the current exceeds its corresponding threshold, it is determined to be a definite fault disturbance signal. It is understood that these two threshold values ​​are safety margin values ​​independently set based on statistical analysis of various disturbances (such as load switching) under normal DC system operation. Once the above conditions are met, the waveform recording function for that branch is immediately triggered, and the action target is clearly specified.

[0034] As an example, after triggering the sudden change waveform recording, the current waveform data of the faulty branch within a set time period before and after the fault event is recorded and saved, including: Record and save the current waveform data of the faulty branch within a set time period before and after the fault event; wherein, the set time period is a dynamic time period, and its specific duration is adaptively adjusted based on the characteristics of the current mutation that triggers this waveform recording.

[0035] In this embodiment, the specific duration of the set time period for acquiring current waveform data is dynamic. Specifically, the current surge characteristics that trigger this waveform recording are extracted, which may include the amplitude of the instantaneous rate of change (di / dt) calculated in step 10, the amplitude of the change within the set time window (ΔI), or a comprehensive index composed of both. For example, when the absolute value of the detected di / dt or ΔI far exceeds the alarm threshold, it indicates that it may be a rapidly developing and high-energy serious fault. In this case, the dynamic time period is automatically extended to capture a more complete fault development process and decaying oscillation waveform, providing richer data for in-depth fault analysis. Conversely, for slight disturbances that slightly exceed the alarm threshold, a shorter dynamic time period can be used, which can effectively save storage space while ensuring that key information is not lost.

[0036] It is understandable that the above adaptive adjustment can be implemented based on comparison data or an experience-based conversion function, without any specific limitations.

[0037] The adaptive mechanism adopted in this embodiment can ensure that fault data for a sufficiently long time window can be saved for major faults, while avoiding the invalid occupation of storage space for minor disturbances, thereby achieving optimized configuration of data recording resources.

[0038] As an example, multi-dimensional features are extracted from the steady-state trend data, and dynamic first-step length and second-step length are derived based on the analysis of these multi-dimensional features, including: Step 301: Perform statistical analysis on the steady-state trend data at multiple preset continuous time scales to extract multi-dimensional features, including short-term volatility, medium-term change slope, and long-term autocorrelation. In this step, to ensure the timeliness of feature extraction and meet the rapid response requirements of DC insulation fault monitoring, this embodiment selects steady-state trend data within the most recent short time window for analysis. The setting of this time window fully considers the real-time requirements of fault location; for example, it can be set to data within the most recent 30 minutes. This duration can capture the dynamic changes in the system state while avoiding response delays caused by excessively long historical data. Based on the steady-state trend data within this short time window, the following key features are extracted using a multi-timescale analysis method: Short-term volatility: This is obtained by calculating the standard deviation of insulation parameters (such as insulation resistance) over the most recent few minutes. Understandably, this characteristic reflects the current instantaneous stable state of the DC system. High short-term volatility indicates that the DC system is experiencing transient disturbances or has an unstable grounding condition.

[0039] Mid-term change slope: The slope of the trend line is obtained by linearly fitting the insulation parameters over the most recent 10-30 minutes. This characteristic represents the recent rapid change trend of the insulation parameters; a negative slope with a large absolute value indicates rapid insulation degradation, while a slope close to zero indicates relative stability.

[0040] Long-term autocorrelation: By calculating the autocorrelation function of data within this time window at different time lags, the persistence characteristics of the data are analyzed. It can be understood that this feature characterizes the persistence and regularity of insulation state changes; a strong positive autocorrelation indicates that insulation degradation is a continuous process.

[0041] Step 302: Input the multi-dimensional features into a predefined step size mapping model to obtain a first step size and a second step size that match the current state of the DC system; wherein, the first step size is negatively correlated with the features characterizing the stability of the DC system, and the second step size is positively correlated with the features characterizing the insulation degradation trend.

[0042] In this step, the multi-dimensional features extracted from the short-term steady-state trend data in step 301 are input into a predefined step size mapping model. This step size mapping model can be a lightweight regression model, which is trained based on a large amount of historical running data and can deeply understand the complex coupling relationship between various features and their comprehensive impact on the analysis step size.

[0043] The step-size mapping model outputs the corresponding analysis step size according to the following adaptive logic: (1) The first step length is negatively correlated with the comprehensive characterization of DC system stability. When short-term volatility is high, it indicates that there is a significant transient disturbance in the DC system. In this case, long-term autocorrelation plays an important auxiliary role in judgment: if the long-term autocorrelation is also low, it indicates that the fluctuation is an occasional disturbance, and the step-size mapping model should appropriately shorten the first step length; if the long-term autocorrelation is high, it indicates that the DC system is in a continuous unstable state, and the step-size mapping model will output a shorter first step length (e.g., 10-50 milliseconds) to capture potential transient fault characteristics with the highest time resolution.

[0044] (2) The second step length is positively correlated with the comprehensive characterization of insulation degradation trend. When the negative value of the slope of the medium-term change increases, it indicates that the insulation is showing a deterioration trend. At this time, long-term autocorrelation provides key verification: if the long-term autocorrelation is high, it confirms that the deterioration trend is persistent, and the step-size mapping model will output a longer second step (such as 1-5 minutes) to accurately grasp the macro-change pattern; if the long-term autocorrelation is low, it may be judged as a temporary fluctuation, and the step-size mapping model will adopt a relatively conservative step-size setting.

[0045] (3) Synergistic mechanism among multi-dimensional features The three features mentioned above form a complete judgment chain within the step-size mapping model: short-term volatility reflects the immediate state, the slope of medium-term changes indicates the development trend, and long-term autocorrelation provides historical evidence. The model achieves this by using deep learning to understand the intrinsic relationship between the three features: when short-term volatility is high and long-term autocorrelation is low, it is identified as an instantaneous disturbance; when the slope of medium-term changes is large and long-term autocorrelation is high, it confirms continuous deterioration; when the three features show inconsistency, a cautious step-size strategy is adopted.

[0046] This implementation method dynamically extracts multi-dimensional features based on short-time steady-state data and adaptively adjusts the analysis step size using a lightweight regression model. This allows the time scale for transient feature capture and trend analysis to accurately match the real-time operating status of the DC system. As a result, while ensuring rapid fault response, it can also significantly improve the accuracy and reliability of transient fault identification and insulation degradation trend judgment.

[0047] As an example, such as Figure 2As shown, the step size mapping model includes a feature preprocessing layer, a feature fusion layer, and a step size regression layer connected in sequence; The feature preprocessing layer is used to standardize the input multi-dimensional features and eliminate the influence of units. The feature fusion layer adopts a fully connected neural network structure to learn the weight relationship of the influence of different features on the step size, thereby achieving deep feature fusion. The step-size regression layer uses a linear activation function to map the fused features into continuous first and second step-size values.

[0048] In this embodiment, the feature preprocessing layer serves as the data input interface for the step-size mapping model. It is responsible for receiving the raw multi-dimensional features extracted from the steady-state trend data, including short-term volatility, medium-term slope of change, and long-term autocorrelation. Since these features have different physical meanings and dimensions, this layer uses the Z-score normalization method to process the input features, so that the values ​​of each feature are normalized to a similar scale range.

[0049] The feature fusion layer is the core computational layer of the step-size mapping model. It adopts a fully connected neural network structure that can simultaneously process and deeply fuse information from three dimensions: short-term volatility, medium-term change slope, and long-term autocorrelation.

[0050] This layer automatically learns and assigns appropriate decision weights to long-term autocorrelation through nonlinear transformation. For example, when long-term autocorrelation is high, the step-size mapping model interprets the current system state as having strong historical inertia. In this case, if the intermediate-term slope also shows negative deterioration, the step-size mapping model gains support from the historical persistence of this trend and outputs a longer second step with more confidence to confirm this macro trend; conversely, if long-term autocorrelation is low, the step-size mapping model will be more cautious about changes in the intermediate-term slope.

[0051] Similarly, long-term autocorrelation works in conjunction with short-term volatility to influence the decision on the first step length. High volatility under high long-term autocorrelation may indicate persistent instability in the DC system, thus driving the step-size mapping model to output a shorter first step length.

[0052] Through this mechanism, the feature fusion layer achieves intelligent weighting and comprehensive judgment of multi-dimensional features, rather than simple linear judgment.

[0053] The step-size regression layer receives the deeply fused high-dimensional feature vectors output from the feature fusion layer. This layer uses a linear activation function to precisely map the complex system state represented by the fused features into two continuous actual control parameters, namely the first step size and the second step size, through a linear transformation of the weight matrix and the bias vector.

[0054] As an example, a confidence score is obtained by processing the transient features and the trend change features using a confidence calculation model to characterize the event as a permanent insulation failure, including: Step 401: Concatenate the transient features and the trend change features. During the concatenation process, random interference features are inserted according to a dynamically determined insertion interval to form an anti-interference fusion feature vector. In this step, under actual DC system operating conditions, monitoring signals are inevitably affected by various electromagnetic interferences, measurement noise, and environmental factors, leading to varying degrees of disturbance and bias in the extracted features. If the extracted features are directly input into the confidence calculation model, the model may be overly sensitive to these minor disturbances, causing fluctuations in diagnostic results and affecting the accuracy of fault diagnosis. To address this, this embodiment inserts random interference features during the feature cascading process to simulate the noise impact in the real environment. This allows the model to focus on more discriminative feature patterns rather than overly relying on the precise numerical values ​​of the features when making decisions, thereby improving stability in complex real-world environments.

[0055] Specifically, the transient features (such as mutation energy and main oscillation frequency) extracted in step 30 are connected with trend change features (such as resistance decay within a specific time window) in a predetermined order to form a comprehensive feature vector characterizing the fault event. Simultaneously, during the cascading process, random disturbance features are inserted at specific positions on the feature vector according to a dynamically determined insertion interval. It can be understood that the random disturbance features are scalars randomly generated within a preset small numerical range (e.g., [-0.1, 0.1]), which can simulate the effects of real noise without destroying the physical meaning of the original features.

[0056] After the above processing, an anti-interference fusion feature vector is obtained, which contains both the original transient and trend information and structured random noise, and serves as the input to the confidence calculation model.

[0057] Step 402: Input the anti-interference fusion feature vector into the confidence calculation model and output the confidence score. The confidence score comprehensively represents the probability of permanent insulation fault occurrence based on the joint judgment of transient change and steady-state trend.

[0058] In this step, the confidence calculation model is a pre-trained machine learning model that has been trained on a large amount of historical fault data (including various transient disturbances and permanent fault cases) and has established an accurate mapping relationship from feature vectors to fault types.

[0059] The anti-interference fusion feature vector formed in step 401 is input into the confidence calculation model. Through its inherent nonlinear structure, each feature in the vector (including the original feature and the inserted interference feature) is weighted, combined, and judged. The model focuses on evaluating the consistency between the abrupt change of transient features and the persistence of trend change features, ultimately outputting a confidence score. It can be understood that this confidence score is a continuous probability value. A higher score indicates a stronger consistency between the two pieces of evidence: transient abrupt change and steady-state degradation, and a greater probability that the current event is a permanent insulation failure. A lower score indicates that it is more likely a transient disturbance or noise event.

[0060] This embodiment constructs an anti-interference fusion feature vector by dynamically inserting random interference features into the feature cascade, enabling the confidence calculation model to effectively simulate feature perturbations in a real noise environment during the inference stage. This significantly improves the model's robustness to minor feature changes and enhances the stability and reliability of fault diagnosis results.

[0061] As an example, the insertion interval length is dynamically adjusted based on the signal-to-noise ratio of the current DC system; the lower the signal-to-noise ratio, the shorter the insertion interval, and vice versa.

[0062] In this embodiment, the electromagnetic interference level of the operating environment is dynamically changing during DC system operation. Using a fixed insertion interval will not be suitable for the varied operating conditions. Specifically, in a strong interference environment, a fixed interference feature insertion density may not be sufficient to fully simulate the impact of real noise; while in a clean signal environment, excessively dense interference features may introduce unnecessary noise interference. Therefore, this invention establishes a dynamic adjustment mechanism based on the signal-to-noise ratio, enabling the interference feature insertion density to intelligently match the current noise level of the environment, achieving optimal matching between noise simulation and the real environment.

[0063] Specifically, by analyzing steady-state monitoring data in real time, the energy ratio of the effective component to the noise component in the insulation parameter signal is calculated to accurately reflect the complexity of the current electromagnetic environment and the signal quality. When the signal-to-noise ratio is low, it indicates severe environmental electromagnetic interference. In this case, the insertion interval should be shortened (e.g., inserting an interference feature every 2-3 features) to increase the interference feature density, simulate a strong noise environment, and improve the model's anti-interference capability under complex conditions. When the signal-to-noise ratio is high, it indicates good signal quality. In this case, the insertion interval should be extended (e.g., inserting an interference feature every 8-10 features) to reduce the interference feature density, maintaining the necessary regularization effect while avoiding the impact of excessive interference on the learning of effective features.

[0064] This embodiment dynamically adjusts the insertion interval of random interference features based on the real-time signal-to-noise ratio of the DC system, enabling the noise injection intensity during feature cascading to accurately match the complexity of the current electromagnetic environment. This enhances the robustness of the model under strong interference and maintains recognition accuracy under weak interference, thus optimizing fault diagnosis performance under different operating conditions.

[0065] As an example, the confidence calculation model is a pre-trained gradient boosting decision tree model.

[0066] In this embodiment, in the DC insulation fault diagnosis scenario, it is necessary to combine transient change characteristics and steady-state trend characteristics for joint judgment. These characteristics have complex nonlinear relationships, and Gradient Boosting Decision Tree (GBDT) is good at handling such problems. Therefore, it is preferable to set the confidence calculation model as a pre-trained gradient boosting decision tree model.

[0067] As an example, the process of locating the faulty branch includes: Step 501: Based on the comparison results of the confidence score and the alarm threshold, and combined with the original branch number that triggered this diagnostic process, determine the location of the permanent insulation fault.

[0068] In this embodiment, the confidence score calculated in step 40 is compared with a preset alarm threshold. When the confidence score is higher than the alarm threshold, a permanent insulation fault event is determined to have occurred. Next, this diagnostic conclusion is precisely linked to the original branch number that triggered the diagnostic process, where the original branch number is a unique identifier already recorded when the abrupt change waveform was triggered in step 10. Through this association, the fault can be clearly located in a specific feeder branch of the DC system, such as "DC bus section I - feeder branch 05".

[0069] As an example, the fault branch localization process further includes: Step 502: Generate a complete fault report including the fault branch number, confidence score, fault timestamp, and path to the associated waveform data file; push the fault location results to the monitoring backend in real time and update the branch insulation status database synchronously.

[0070] In this embodiment, a structured fault report is automatically generated as the basis for subsequent accident analysis and accountability. The report includes at least: the fault branch number, used to identify the specific location of the fault; a confidence score, used to provide the reliability of this diagnosis for maintenance personnel's decision-making reference; a fault timestamp, used to accurately record the time of the fault occurrence; and a path to the associated waveform data file, used to provide a link to the current waveform data saved in step 20, facilitating maintenance personnel to retrieve and conduct in-depth analysis of the fault transient process.

[0071] Next, the simplified fault location results (such as branch number and alarm level) are pushed to the monitoring backend in real time via the system bus (such as IEC 61850 MMS), triggering audible and visual alarms to remind operators to handle the situation promptly. Simultaneously, the branch insulation status database is updated, marking the branch's status as "insulation fault" or a similar warning status.

[0072] Although the invention has been specifically shown and described with reference to preferred embodiments, those skilled in the art will understand that various modifications in form and detail may be made without departing from the spirit and scope of the invention. Accordingly, the disclosed invention should be considered merely illustrative and limited only by the scope specified in the appended claims.

Claims

1. A method for precise location of DC insulation faults integrating abrupt change and trend analysis, characterized in that, Includes the following steps: Real-time monitoring of current signals in each branch of the DC system; when the current change or rate of change in any branch exceeds the preset threshold for sudden change, sudden change waveform recording is triggered. After triggering the sudden change waveform recording, the current waveform data of the faulty branch within a set time period before and after the fault event is recorded and saved. In addition, the insulation parameters of each branch are monitored in parallel and continuously, and steady-state trend data of the insulation parameters over time are generated; Multi-dimensional features are extracted from the steady-state trend data. Based on the multi-dimensional features, the dynamic first step length and second step length are analyzed. Transient features are extracted from the current waveform data based on the first step length. Trend change features are extracted from the steady-state trend data based on the second step length. The transient features and trend change features are processed using a confidence calculation model to obtain a confidence score that characterizes the event as a permanent insulation fault. When the confidence score is higher than the alarm threshold, the highest level alarm is issued and fault branch location is performed.

2. The method for precise location of DC insulation faults by integrating abrupt change and trend analysis according to claim 1, characterized in that: When the current change or rate of change in any branch exceeds a preset abrupt change threshold, abrupt change recording is triggered, including: The instantaneous rate of change and the amount of change within a set time window of the current in each branch are calculated in real time. When the absolute value of at least one of the instantaneous rate of change and the amount of change of any branch exceeds the preset rate threshold and quantity threshold respectively, the sudden change waveform of the faulty branch is immediately triggered.

3. The method for precise location of DC insulation faults by integrating abrupt change and trend analysis according to claim 1, characterized in that: After triggering the sudden change waveform recording, the current waveform data of the faulty branch within a set time period before and after the fault event is recorded and saved, including: Record and save the current waveform data of the faulty branch within a set time period before and after the fault event; wherein, the set time period is a dynamic time period, and its specific duration is adaptively adjusted based on the characteristics of the current mutation that triggers this waveform recording.

4. The method for precise location of DC insulation faults by integrating abrupt change and trend analysis according to claim 1, characterized in that: Multi-dimensional features are extracted from the steady-state trend data, and the dynamic first step length and second step length are derived based on the analysis of these multi-dimensional features, including: Statistical analysis is performed on the steady-state trend data at multiple preset continuous time scales to extract multi-dimensional features, including short-term volatility, medium-term change slope, and long-term autocorrelation. The multi-dimensional features are input into a predefined step size mapping model to obtain a first step size and a second step size that match the current state of the DC system; wherein, the first step size is negatively correlated with the features characterizing system stability, and the second step size is positively correlated with the features characterizing insulation degradation trend.

5. The method for precise location of DC insulation faults by integrating abrupt change and trend analysis according to claim 4, characterized in that: The step size mapping model includes a feature preprocessing layer, a feature fusion layer, and a step size regression layer connected in sequence. The feature preprocessing layer is used to standardize the input multi-dimensional features and eliminate the influence of units. The feature fusion layer adopts a fully connected neural network structure to learn the weight relationship of the influence of different features on the step size, thereby achieving deep feature fusion. The step-size regression layer uses a linear activation function to map the fused features into continuous first and second step-size values.

6. The method for precise location of DC insulation faults by integrating abrupt change and trend analysis according to claim 5, characterized in that: The transient characteristics and trend change characteristics are processed using a confidence calculation model to obtain a confidence score characterizing the event as a permanent insulation failure, including: The transient features and the trend change features are concatenated. During the concatenation process, random interference features are inserted according to a dynamically determined insertion interval to form an anti-interference fusion feature vector. The anti-interference fusion feature vector is input into the confidence calculation model, and a confidence score is output. This confidence score comprehensively represents the probability of permanent insulation failure based on the joint judgment of transient changes and steady-state trends.

7. The method for precise location of DC insulation faults by integrating abrupt change and trend analysis according to claim 6, characterized in that: The insertion interval length is dynamically adjusted based on the signal-to-noise ratio of the current DC system. The lower the signal-to-noise ratio, the shorter the insertion interval, and vice versa.

8. A method for precise location of DC insulation faults by integrating abrupt change and trend analysis according to claim 6 or 7, characterized in that: The confidence calculation model is a pre-trained gradient boosting decision tree model.

9. The method for precise location of DC insulation faults by integrating abrupt change and trend analysis according to claim 1, characterized in that: The fault branch location process includes: Based on the comparison results of the confidence score and the alarm threshold, combined with the original branch number that triggered this diagnostic process, the location of the permanent insulation fault is determined.

10. The method for precise location of DC insulation faults by integrating abrupt change and trend analysis according to claim 9, characterized in that: The fault branch location process also includes: Generate a complete fault report including the fault branch number, confidence score, fault timestamp, and path to the associated waveform data file; push the fault location results to the monitoring backend in real time and update the branch insulation status database synchronously.

Images