Multi-turn inter-turn arc discharge grading early warning method and device based on double criteria

Abstract

The invention discloses a multi-turn inter-turn arc discharge grading early warning method and device based on double criteria. The method comprises the following steps: acquiring a monitoring signal generated by a multi-turn winding in a turn-to-turn discharge process, segmenting the monitoring signal to obtain a plurality of sliding windows, and extracting a corresponding characteristic parameter for each sliding window; inputting the characteristic parameters into a discharge stage identification model to obtain a stage probability vector, and performing probability smoothing processing on the stage probability vector to obtain a smoothing stage probability; determining a candidate stage based on the smooth stage probability, triggering a hysteresis update criterion mechanism, and outputting a stable stage state; establishing a mapping relation between the stable stage state and the grading early warning quantity to obtain a current early warning grade; and executing a persistence criterion mechanism on the current early warning level, outputting an effective early warning when a persistence criterion meets a trigger condition, otherwise, keeping the previous early warning output. According to the method, frequent hopping and instantaneous false alarm of a single-window identification result at a stage boundary are avoided.

Description

Technical Field

[0001] This invention relates to the field of graded early warning technology for inter-turn arc discharge faults in multi-turn windings of oil-paper insulated power equipment, and particularly to a graded early warning method and device for inter-turn arc discharge faults based on dual criteria. Background Technology

[0002] In oil-paper insulated power equipment, windings operate for extended periods in a complex environment characterized by high electric fields, high temperatures, and strong mechanical stress coupling. Once the inter-turn insulation deteriorates, partial discharge can easily develop into a breakdown arc fault. Such inter-turn arcs are characterized by high energy density, rapid development speed, and easy expansion of the discharge channel along adjacent turns. They can spread from single-turn breakdown to multi-turn short circuit in a very short time, causing damage to the winding structure and even triggering catastrophic failures.

[0003] Current online monitoring technologies for winding insulation status mainly focus on partial discharge, and related detection and diagnostic methods are relatively mature. However, when a fault develops to insulation breakdown and forms a continuous arc, its discharge characteristics differ fundamentally from partial discharge. Furthermore, the entire breakdown process is not instantaneous but exhibits a phased evolution from mild to severe, including near-breakdown, single-turn breakdown, double-turn breakdown, and eventually multi-turn short circuit. However, existing monitoring systems lack the ability to effectively track this critical evolutionary process and cannot provide corresponding graded early warnings for different breakdown stages.

[0004] Furthermore, in practical engineering applications, monitoring signals are often affected by factors such as electromagnetic interference, load fluctuations, and differences in sensor sensitivity, leading to fluctuations in characteristic parameters under the same fault stage. Instantaneous jumps or regressions in identification results are more likely to occur at the boundaries of different stages. If the model output of a single time window is directly used as the basis for alarms, it is highly likely to cause frequent false alarms, missed alarms, or alarm level oscillations, seriously affecting the credibility and executability of operation and maintenance decisions.

[0005] Therefore, there is an urgent need for an early warning output mechanism that can accurately identify the development stage of multi-turn inter-turn arc discharge and, on this basis, achieve stable, graded, and disturbance-resistant early warning output to support differentiated operation and maintenance strategies, thereby intervening in the early stage of faults and avoiding irreversible equipment damage. Summary of the Invention

[0006] To overcome the above problems, this invention proposes a multi-turn inter-turn arc discharge graded early warning method and device based on dual criteria.

[0007] Specifically, the object of the present invention is to provide the following aspects:

[0008] In a first aspect, a method for graded early warning of multi-turn inter-turn arc discharge is provided, the method comprising:

[0009] Step 1: Obtain the monitoring signal generated by the multi-turn winding during the inter-turn discharge process, segment the monitoring signal in a sliding window manner to obtain multiple continuous sliding windows, and extract the corresponding feature parameters for each sliding window;

[0010] Step 2: Input the feature parameters into the discharge stage identification model to obtain the stage probability vector corresponding to the nth sliding window. The stage probability vector is then subjected to probability smoothing to obtain the smoothed stage probability. ;

[0011] Step 3: Determine candidate stages based on the smoothing stage probability. This triggers the hysteresis update criterion mechanism and outputs the stable phase state. ;

[0012] Step 4, Establish a stable phase state With graded early warning volume The mapping relationship is used to obtain the current warning level;

[0013] Step 5: Execute a persistence criterion mechanism on the current warning level. When the persistence criterion meets the triggering condition, output a valid warning; otherwise, maintain the output of the previous warning.

[0014] In step 1, the monitoring signal includes one or more of the following: voltage signal, current signal, ultrasonic emission signal, ultra-high frequency electromagnetic radiation signal, and optical radiation signal.

[0015] In step 1, the characteristic parameters include maximum amplitude, average amplitude, standard deviation, kurtosis, skewness, spectral center frequency, spectral bandwidth, and pulse repetition rate.

[0016] In step 2, the stage probability vector , Let be the predicted probability that the sample in the nth window belongs to stage C.

[0017] Optionally, stage C includes a near-pre-breakdown stage, a single-turn breakdown arc stage, a double-turn breakdown arc stage, and a three-turn breakdown arc stage.

[0018] In step 2, a fully connected feedforward neural network is used as the discharge stage identification model.

[0019] In step 3, the stage with the highest probability of smoothing is selected as the candidate stage. .

[0020] Optionally, the mapping relationship is as follows: .

[0021] Secondly, a multi-turn inter-turn arc discharge graded early warning device is provided, the device comprising:

[0022] The feature acquisition module is used to acquire the monitoring signal generated by the multi-turn winding during the inter-turn discharge process, segment the monitoring signal in a sliding window manner to obtain multiple continuous sliding windows, and extract the corresponding feature parameters for each sliding window.

[0023] The probability output module is used to input the feature parameters into the discharge stage identification model to obtain the stage probability vector corresponding to the nth sliding window. The stage probability vector is then subjected to probability smoothing to obtain the smoothed stage probability. ;

[0024] The state update module is used to determine candidate stages based on the smoothing stage probability. This triggers the hysteresis update criterion mechanism and outputs the stable phase state. ;

[0025] The level mapping module is used to establish stable phase states. With graded early warning volume The mapping relationship is used to obtain the current warning level;

[0026] The alarm output module is used to perform a persistent criterion mechanism on the current warning level. When the persistent criterion meets the triggering condition, a valid warning is output; otherwise, the previous warning output is maintained.

[0027] Thirdly, a computer-readable storage medium is provided, including a stored complete computer program that, when the computer program is run, implements the method described in the first aspect.

[0028] The beneficial effects of this invention include:

[0029] (1) The present invention provides a multi-turn inter-turn arc discharge graded early warning method, which can provide executable graded early warning information during the evolution of stages such as near pre-breakdown stage, single-turn breakdown arc stage, double-turn breakdown arc stage and three-turn breakdown arc stage, reduce misjudgment and frequent alarm jumps at stage boundaries, and improve the stability and engineering availability of online early warning.

[0030] (2) The present invention provides a multi-turn inter-turn arc discharge graded early warning method, which transforms the stage identification results into graded early warnings, performs sliding average smoothing based on the model probability output, and suppresses stage boundary jitter by combining the hysteresis update criterion of the entry / exit threshold. It also outputs a stable early warning level through the continuous criterion, avoiding frequent jumps and instantaneous false alarms of the single window-level identification results at the stage boundary, and providing a promising new operation and maintenance decision-making idea for the evolution of inter-turn arcs from light to heavy. Attached Figure Description

[0031] Various other advantages and benefits of the present invention will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiments below. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. It is obvious that the drawings described below are merely some embodiments of the invention, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0032] In the attached diagram:

[0033] Figure 1 A flowchart illustrating the multi-turn inter-turn arc discharge graded early warning method of the present invention is shown;

[0034] Figure 2 The logic diagram of the dual-criteria early warning system of the present invention is shown;

[0035] Figure 3 The confusion matrix shows the accurate identification results of the discharge stage identification model in Example 1.

[0036] Figure 4 A schematic diagram of the multi-turn inter-turn arc discharge graded early warning device of the present invention is shown. Detailed Implementation

[0037] The following will refer to the appendix. Figures 1 to 4 Specific embodiments of the invention will be described in more detail below. While specific embodiments of the invention are shown in the accompanying drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the invention and to fully convey the scope of the invention to those skilled in the art.

[0038] It should be noted that certain terms are used in the specification and claims to refer to specific components. Those skilled in the art will understand that different terms may be used to refer to the same component. This specification and claims do not distinguish components based on differences in terminology, but rather on differences in function. The terms "comprising" or "including" used throughout the specification and claims are open-ended and should be interpreted as "comprising but not limited to." The following descriptions are preferred embodiments for carrying out the invention; however, these descriptions are for the purpose of understanding the general principles of the specification and are not intended to limit the scope of the invention. The scope of protection of this invention is determined by the appended claims.

[0039] In the description of this invention, it should be noted that the terms "upper," "lower," "inner," "outer," "front," and "rear," etc., indicate the orientation or positional relationship based on the orientation or positional relationship in the working state of this invention, and are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention. Furthermore, the terms "first," "second," "third," and "fourth" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0040] 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. The accompanying drawings do not constitute a limitation on the embodiments of the present invention.

[0041] On the one hand, according to the present invention, a multi-turn inter-turn arc discharge graded early warning method is provided, such as... Figures 1 to 2 As shown, the method includes:

[0042] Step 1: Obtain the monitoring signal generated by the multi-turn winding during the inter-turn discharge process, segment the monitoring signal in a sliding window manner to obtain multiple continuous sliding windows, and extract the corresponding feature parameters for each sliding window;

[0043] Step 2: Input the feature parameters into the discharge stage identification model to obtain the stage probability vector corresponding to the nth sliding window. The stage probability vector is then subjected to probability smoothing to obtain the smoothed stage probability. ;

[0044] Step 3: Determine candidate stages based on the smoothing stage probability. This triggers the hysteresis update criterion mechanism and outputs the stable phase state. ;

[0045] Step 4, Establish a stable phase state With graded early warning volume The mapping relationship is used to obtain the current warning level;

[0046] Step 5: Execute a persistence criterion mechanism on the current warning level. When the persistence criterion meets the triggering condition, output a valid warning; otherwise, maintain the output of the previous warning.

[0047] The above methods will be described in detail below.

[0048] Step 1: Obtain the monitoring signal generated during the inter-turn discharge process of the multi-turn winding, segment the monitoring signal in a sliding window manner to obtain multiple continuous sliding windows, and extract the corresponding feature parameters for each sliding window.

[0049] In step 1, inter-turn arc discharge, as a high-energy transient fault, will simultaneously excite the observed response signal in multiple physical fields such as electrical, acoustic, electromagnetic and optical fields. A single signal is easily affected by operating conditions or sensor limitations, while multi-source signal fusion can improve the robustness and comprehensiveness of fault identification.

[0050] Furthermore, the monitoring signal includes any one or a combination of the following: voltage signal, current signal, ultrasonic emission signal, ultra-high frequency electromagnetic radiation signal, and optical radiation signal. The frequency of the ultra-high frequency electromagnetic radiation signal is typically 1.5MHz-30MHz.

[0051] Preferably, the voltage signal is acquired using a resistive-capacitive voltage divider, the current signal is acquired using a Hall coil, the ultrasonic emission signal is acquired using an acoustic coupling sensor, the ultra-high frequency electromagnetic radiation signal is acquired using an ultra-high frequency antenna, and the optical radiation signal is acquired using a silicon photodiode.

[0052] In step 1, a sliding window is used to divide a long-term continuous signal into several short-time segments, so that the signal in each window is approximately regarded as a stationary process, thereby facilitating the extraction of statistically significant time-frequency domain features.

[0053] The sliding window length is typically set to 6-10 ms, and the overlap rate between adjacent windows is 30-70%. This parameter balances the timeliness and temporal resolution of feature extraction. Within each sliding window, feature parameters of the monitored signal are extracted, including maximum amplitude, average amplitude, standard deviation, kurtosis, skewness, spectral center frequency, spectral bandwidth, and pulse repetition rate, totaling eight parameters.

[0054] In step 1, the multi-turn winding used is made of metal wire, with a total of 4-8 turns and a turn-to-turn gap of 3-5 mm. This structure can simulate the weak points in the inter-turn insulation of a typical transformer winding and support experimental verification of the complete fault evolution process from near breakdown to three-turn breakdown. For example, the multi-turn winding used is made of copper wire, with a total of 4 turns and a turn-to-turn gap of 3 mm.

[0055] Step 2: Input the feature parameters into the discharge stage identification model to obtain the stage probability vector corresponding to the nth sliding window. The stage probability vector is then subjected to probability smoothing to obtain the smoothed stage probability. .

[0056] In step 2, the discharge stage identification model is trained and constructed in the following manner:

[0057] First, by simulating the inter-turn discharge process of a multi-turn winding in the near-pre-breakdown stage, single-turn breakdown arc stage, double-turn breakdown arc stage, and three-turn breakdown arc stage, at least one monitoring signal is simultaneously acquired. The monitoring signal includes voltage signal, current signal, ultrasonic emission signal, ultra-high frequency electromagnetic radiation signal, or optical radiation signal.

[0058] Secondly, extract 8-dimensional feature parameters from the collected monitoring signals according to the method described in step 1, including maximum amplitude, average amplitude, standard deviation, kurtosis, skewness, spectral center frequency, spectral bandwidth, and pulse repetition rate; and combine the feature parameters with the corresponding discharge stage labels to form a labeled sample dataset (e.g., 1 represents the near pre-breakdown stage, 2 represents the single-turn breakdown arc stage, 3 represents the double-turn breakdown arc stage, and 4 represents the three-turn breakdown arc stage).

[0059] Next, the sample dataset is divided into a training set and a validation set either chronologically or randomly. The training set is used for learning the parameters of the discharge phase recognition model; the validation set is used to verify the generalization performance of the discharge phase recognition model during training, avoiding overfitting. The training set and validation set are divided in an 8:2 or 7:3 ratio.

[0060] Finally, a discharge phase identification model is constructed, and the training set is input into the model for iterative training. An early stopping strategy is adopted, that is, training is terminated when the validation set loss does not decrease within 30 consecutive training rounds, thereby obtaining the trained discharge phase identification model.

[0061] Furthermore, a fully connected feedforward neural network is adopted as the discharge stage identification model, and its structure includes:

[0062] The input layer receives an 8-dimensional feature vector from each sliding window;

[0063] The hidden layer consists of three sequentially connected fully connected layers, with 64, 32, and 16 neurons in each layer, respectively; each layer uses the ReLU activation function to enhance nonlinear expressive power.

[0064] The output layer has a dimension of 4, which corresponds to four types of discharge stages: near breakdown arc stage, single-turn breakdown arc stage, double-turn breakdown arc stage, and triple-turn breakdown arc stage. The stage number increases with the severity of the fault.

[0065] The output layer uses the Softmax activation function to convert the network's original output into a normalized stage probability vector.

[0066] Preferably, during model training, stage labels in one-hot encoded form are used as the supervision target; cross-entropy loss function is employed; Adam optimizer is selected, and the initial learning rate is set to 1×10⁻⁶. -5 Up to 1×10-1 Training is iteratively performed in mini-batch mode, with each batch containing 30-35 samples; a Dropout mechanism is introduced between the three fully connected layers, with a uniform dropout rate of 0.2-0.4; an early stopping strategy is adopted during training, and training is terminated when the loss value on the validation set does not decrease within 30 consecutive training rounds, thus obtaining the trained discharge stage recognition model.

[0067] Furthermore, when selecting the optimal learning rate, a cross-validation grid search method is needed to determine the optimal learning rate parameter, 1×10. -5 Up to 1×10 -1 The initial learning rate covers the effective convergence range for common deep learning tasks, avoiding slow convergence due to an excessively small rate or oscillation due to an excessively large rate. In practice, initial learning rates of 0.00001, 0.0001, 0.001, 0.01, and 0.1 are used, for example. A dropout rate of 0.2-0.4 balances the model's generalization ability and information retention. This range, based on experimental verification, effectively prevents overfitting while avoiding excessive feature loss due to an excessively high dropout rate; for example, a dropout rate of 0.3 is used.

[0068] In step 2, the feature parameters obtained in step 1 are input into the trained discharge stage recognition model to obtain the stage probability vector corresponding to the nth sliding window. ,in, Let be the predicted probability that the sample in the nth window belongs to stage C. .

[0069] Furthermore, C=4 is preferred, and the four discharge stages are: near breakdown arc stage, single-turn breakdown arc stage, double-turn breakdown arc stage, and triple-turn breakdown arc stage; and the stage number increases with the severity.

[0070] In step 2, the probability smoothing process uses a moving average, satisfying: Where M is the number of sliding windows participating in the averaging, and ; Let represent the smoothed probability of the nth sliding window with stage category c. This represents the original output probability of the k-th sliding window belonging to stage category c.

[0071] In one embodiment, M is 5, which means that the original probabilities of the current sliding window and its four consecutive sliding windows are taken as an arithmetic mean. The length of each window is 8ms, and the overlap rate of adjacent windows is 50%, that is, the sliding step size is 4ms.

[0072] Step 3: Determine candidate stages based on the smoothing stage probability. This triggers the hysteresis update criterion mechanism and outputs the stable phase state. .

[0073] In step 3, for the nth sliding window, calculate the smoothing probability of each stage category. The stage with the highest probability is selected as the candidate stage: Where c is the stage category number (1-4 correspond to near breakdown, single-turn breakdown, double-turn breakdown, and triple-turn breakdown, respectively). This is the candidate stage number for the nth sliding window.

[0074] To further improve the stability of stage recognition and avoid frequent jumps caused by instantaneous noise, a hysteresis update criterion mechanism is introduced. This hysteresis update criterion mechanism utilizes an entry threshold. With exit threshold To achieve the candidate stage The lag update, in which, This is to prevent the state from fluctuating back and forth.

[0075] In a preferred embodiment Take 0.7, Take 0.55.

[0076] Specifically:

[0077] During the candidate phase And the smoothing probability of this stage Then let ;

[0078] During the candidate phase And the smoothing probability of the previous stage Then let ;

[0079] Otherwise, keeping the original state unchanged, then let .

[0080] in, The output represents the steady-state state of n-1 sliding windows. This is the output of the final stable state of the nth sliding window to be determined.

[0081] Step 4, Establish a stable phase state With graded early warning volume The mapping relationship is used to obtain the current warning level.

[0082] In step 4, the stable phase state output in step 3 is... Mapped to the corresponding tiered early warning quantity This is to enable the transformation from fault identification to actionable operational and maintenance decisions.

[0083] Since this invention defines four types of discharge stages (near pre-breakdown stage, single-turn breakdown arc stage, double-turn breakdown arc stage, and three-turn breakdown arc stage), a graded warning quantity is set. ∈{0,1,2,3}, and establish a correspondence with the stage state through the following mapping relationship: ,in, Number the stable phase of the nth sliding window ( ∈{1,2,3,4}). The corresponding warning level is denoted by n, and n is the time sequence number of the sliding window, which is consistent with n in step 3.

[0084] Furthermore, the specific meanings of each warning level are as follows:

[0085] A n =0: Normal state, corresponding to the near pre-breakdown stage, indicating that actual breakdown has not yet occurred, only insulation degradation trend exists, it is recommended to only record data and analyze long-term trends;

[0086] A n =1: Attention status, corresponding to the single-turn breakdown arc stage, indicating that the first breakdown has occurred, but the impact is limited. It is recommended to carry out a review and key monitoring.

[0087] A n =2: Warning status, corresponding to the double-turn breakdown arc stage, indicating that the fault is worsening and may trigger a chain reaction. It is recommended to arrange planned maintenance and assess the equipment operation risk.

[0088] A n =3: Emergency state, corresponding to the three-turn breakdown arc stage, indicating that the fault is serious and has approached an uncontrollable state. Immediate measures such as load reduction, shutdown, or emergency shutdown must be taken to prevent the accident from escalating.

[0089] The aforementioned tiered early warning system represents a leap from "fault detection" to "executable operation and maintenance strategies," enabling power equipment operation and maintenance personnel to develop differentiated response plans based on the early warning level, thereby improving fault prevention capabilities and system security.

[0090] Of course, the above mapping relationship It has universal applicability; when the total number of fault stages defined by the system increases to 5, 6, or more, simply assign the stability stage numbers sequentially as 1, 2, 3...C, and the corresponding graded early warning quantities can be naturally represented as 0, 1, 2...(C-1), with the mapping rule always being... .

[0091] Step 5: Execute a persistence criterion mechanism on the current warning level. When the persistence criterion meets the triggering condition, output a valid warning; otherwise, maintain the output of the previous warning.

[0092] In step 5, to further improve the reliability of the early warning results and prevent false alarms caused by instantaneous noise or signal fluctuations, a persistent criterion mechanism is introduced to adjust the graded early warning quantities generated in step 4. A secondary verification is performed, and a valid warning is only issued when the persistence condition is met.

[0093] Specifically, set the following parameters: The warning threshold represents the minimum confidence level that the probability needs to be reached in the current stage. The length of the sliding observation window is used to statistically analyze recent historical data. The minimum number of qualifying windows required to trigger the criterion, and .

[0094] Furthermore, define the criterion window statistic for the nth sliding window. for: In the formula, This is the criterion window statistic for the nth sliding window; This is an indicator variable for whether the k-th sliding window has met the criteria, and its value is determined as follows: ,in, The state of the k-th sliding window in its steady phase The probability after smoothing reflects the confidence level of this stage.

[0095] In a preferred embodiment, take Take 0.7, Take 5. The value is set to 3, meaning that if at least 3 out of the last 5 consecutive sliding windows have a stage probability of not less than 0.7, then the current warning level is considered to have sufficient persistence and credibility.

[0096] The judgment rule is as follows: if and only if If the current warning level is deemed to have sufficient persistence and credibility, a valid warning will be issued. ;when At the same time, maintain the alarm output of the previous sliding window. This mechanism effectively suppresses false alarms caused by brief disturbances while retaining the trend response capability of real fault evolution, significantly improving the robustness and practicality of the early warning system.

[0097] This invention addresses the problems of susceptibility to noise interference, frequent state transitions, and high false alarm rates in traditional discharge stage identification by introducing dual criteria. Specifically, it controls the stable stage state by updating the hysteresis criterion. The update logic, by setting entry and exit thresholds, achieves a delayed response of "high confidence for upgrades and low confidence for downgrades," effectively suppressing frequent switching of stage states at boundaries due to probability fluctuations; and verifies the graded early warning quantity through continuous criteria. The validity of the output is determined by statistical analysis of the most recent... A probability threshold is met in each sliding window. The number of qualified windows is determined only if this number is not less than a preset threshold. New warnings are only issued periodically, thus filtering out false alarms caused by transient interference. The two criteria mentioned above apply to the state recognition layer and the alarm output layer, respectively, forming a two-level protection mechanism of "stabilizing the state first, then verifying continuity." Based on this, the present invention transforms the stage recognition results into executable hierarchical warnings: smoothing is performed on the model probability output using a moving average, combined with the aforementioned dual criteria, which not only effectively suppresses stage boundary jitter but also avoids frequent jumps and transient false alarms in the critical region of a single window-level recognition result. This provides a promising new operational decision-making approach for the evolution of inter-turn arcs from light to heavy.

[0098] On the other hand, according to the present invention, a multi-turn inter-turn arc discharge graded early warning device is provided, such as... Figure 4 As shown, the device includes:

[0099] The feature acquisition module is used to acquire the monitoring signal generated by the multi-turn winding during the inter-turn discharge process, segment the monitoring signal in a sliding window manner to obtain multiple continuous sliding windows, and extract the corresponding feature parameters for each sliding window.

[0100] The probability output module is used to input the feature parameters into the discharge stage identification model to obtain the stage probability vector corresponding to the nth sliding window. The stage probability vector is then subjected to probability smoothing to obtain the smoothed stage probability. ;

[0101] The state update module is used to determine candidate stages based on the smoothing stage probability. This triggers the hysteresis update criterion mechanism and outputs the stable phase state. ;

[0102] The level mapping module is used to establish stable phase states. With graded early warning volume The mapping relationship is used to obtain the current warning level;

[0103] The alarm output module is used to perform a persistent criterion mechanism on the current warning level. When the persistent criterion meets the triggering condition, a valid warning is output; otherwise, the previous warning output is maintained.

[0104] It should be understood that the device and the method are consistent in principle, and the specific implementation of the method is applicable to the device and produces the same technical effect, which will not be elaborated here.

[0105] In another aspect, the computer-readable storage medium provided by the present invention includes a stored complete computer program that, when the computer program is run, implements the method as described in the first aspect.

[0106] Example

[0107] The present invention is further described below through specific examples; however, these examples are merely exemplary and do not constitute any limitation on the scope of protection of the present invention.

[0108] Example 1

[0109] To verify the effectiveness of this invention, a multi-turn winding inter-turn discharge simulation platform was constructed: a 4-turn copper wire winding with a 3mm inter-turn gap was used, placed in insulating oil. Voltage was gradually applied using a controllable boosting device to sequentially induce four fault stages: near-breakdown stage, single-turn breakdown arc stage, double-turn breakdown arc stage, and three-turn breakdown arc stage. Voltage signals, current signals, ultrasonic emission signals (frequency 1.5MHz-30MHz), ultra-high frequency electromagnetic radiation signals, and optical radiation signals were simultaneously acquired. The voltage signals were acquired using a resistive-capacitive voltage divider, the current signals using a Hall effect coil, the ultrasonic emission signals using an acoustic coupling sensor, the ultra-high frequency electromagnetic radiation signals using an ultra-high frequency antenna, and the optical radiation signals using a silicon photodiode.

[0110] For each stage, 200 samples were collected, and 8-dimensional features (including maximum amplitude, average amplitude, standard deviation, kurtosis, skewness, spectral center frequency, spectral bandwidth, and impulse repetition rate) were extracted to construct a dataset of 800 labeled samples. The dataset was then divided into a training set (640 samples, 160 samples per class) and a validation set (160 samples, 40 samples per class) in an 8:2 ratio.

[0111] A fully connected feedforward neural network is used as the discharge stage identification model, and its structure includes:

[0112] The input layer receives an 8-dimensional feature vector from each sliding window;

[0113] The hidden layer consists of three sequentially connected fully connected layers, with 64, 32, and 16 neurons in each layer, respectively; each layer uses the ReLU activation function to enhance nonlinear expressive power.

[0114] The output layer has a dimension of 4, which corresponds to four types of discharge stages: near pre-breakdown stage, single-turn breakdown arc stage, double-turn breakdown arc stage, and three-turn breakdown arc stage. The stage number increases with the severity of the fault.

[0115] The output layer uses the Softmax activation function to convert the network's original output into a normalized stage probability vector.

[0116] The model was trained using a training set, with one-hot encoded stage labels as the supervision target. Cross-entropy loss was employed. The Adam optimizer was selected, with an initial learning rate of 1×10⁻⁶. -3 Training is iteratively performed in mini-batch mode, with each batch containing 32 samples; a Dropout mechanism is introduced between the three fully connected layers, with a uniform dropout rate of 0.3; an early stopping strategy is adopted during training, and training is terminated when the loss value on the validation set does not decrease within 30 consecutive training rounds, thus obtaining a trained discharge stage recognition model.

[0117] The data from another 200 groups (50 groups per stage) were tested. Figure 3 The confusion matrix of the discharge stage identification model is shown, which accurately identifies 200 test samples. It can be seen that the stage identification accuracy reaches 92.5%, indicating that its classification results are good.

[0118] In the online early warning experiment, the sliding window length was set to 8ms, and the overlap rate of adjacent windows was 50% (i.e., step size 4ms); the number of probability smoothing windows was... =5; Hysteresis update criterion parameter set to 5. =0.7, =0.55; persistence criterion parameter set to =0.7, =5, =3.

[0119] To verify the effectiveness of the dual-criteria mechanism in improving the stability of early warning output, a control scheme was set up: the independent identification results of each sliding window by the fully connected feedforward neural network were directly used as the system output (referred to as "window-only output"), without introducing any state preservation, hysteresis, or persistence constraint mechanisms, to simulate the typical behavior of traditional real-time diagnostic systems. Under this control scheme and the scheme of this invention, based on the same set of independent test data (200 sets, 50 sets per stage), the following evaluation indicators were used for comparative analysis:

[0120] Window-level stage recognition accuracy: refers to the percentage of windows in the fully connected feedforward neural network whose stage prediction results for each sliding window match the true stage label, out of the total number of test windows; this metric reflects the recognition performance of the model itself and is unrelated to whether dual criteria are introduced.

[0121] Stable phase state transition count: refers to the total number of times the phase label changes between adjacent windows;

[0122] Warning level change count: refers to the number of times the warning level changes between adjacent windows;

[0123] Boundary bounce event count: refers to the phenomenon where there are more than 3 repeated transitions between two adjacent stages and the duration is less than 1 second;

[0124] Total number of alarm triggers: refers to the number of times an output with an alarm level of ≥2 occurs;

[0125] False alarm count: refers to the number of times the system outputs a level 2 alarm when no single-turn or higher breakdown event actually occurs;

[0126] Effective warning confirmation rate: refers to the proportion of alarms of level ≥2 that are successfully issued within 30 seconds before a single-turn or above breakdown event actually occurs.

[0127] The following table shows the comparison of the stability of the early warning output before and after the dual-criteria logic determination, as verified by experiments:

[0128]

[0129] The table above shows that although the window-level recognition accuracy can reach 92.5%, short-term category bounces may still occur during noise disturbances or at stage boundaries, leading to frequent alarm changes and a high number of false alarms. It is evident that the window-level recognition accuracy only reflects the model's ability to fit and classify single-window labels, but it is not directly equivalent to online alarms or the actual operating status of the equipment. By introducing hysteresis and persistence as dual criteria, without changing the window-level recognition accuracy, boundary bounces and alarm changes are significantly suppressed, the number of state transitions is significantly reduced, the effective warning confirmation rate is significantly improved, and the warning results are more stable and the decisions are more actionable.

[0130] In summary, this embodiment verifies the comprehensive advantages of the proposed method in terms of high recognition accuracy, strong anti-interference capability, and stable early warning output, and has good prospects for engineering applications.

[0131] The present invention has been described in detail above with reference to preferred embodiments and exemplary examples. However, it should be noted that these specific embodiments are merely illustrative explanations of the invention and do not constitute any limitation on the scope of protection of the invention. Various improvements, equivalent substitutions, or modifications can be made to the technical content and embodiments of the present invention without departing from the spirit and scope of protection of the invention, and all such modifications fall within the scope of protection of the present invention. The scope of protection of the present invention is defined by the appended claims.

Claims

1. A method for graded early warning of multi-turn inter-turn arc discharge, characterized in that, The method includes: Step 1: Obtain the monitoring signal generated by the multi-turn winding during the inter-turn discharge process, segment the monitoring signal in a sliding window manner to obtain multiple continuous sliding windows, and extract the corresponding feature parameters for each sliding window; Step 2: Input the feature parameters into the discharge stage identification model to obtain the stage probability vector corresponding to the nth sliding window. The stage probability vector is then subjected to probability smoothing to obtain the smoothed stage probability. ; Step 3: Determine candidate stages based on the smoothing stage probability. This triggers the hysteresis update criterion mechanism and outputs the stable phase state. ; Step 4, Establish a stable phase state With graded early warning volume The mapping relationship is used to obtain the current warning level; Step 5: Execute a persistence criterion mechanism on the current warning level. When the persistence criterion meets the triggering condition, output a valid warning; otherwise, maintain the output of the previous warning.

2. The method according to claim 1, characterized in that, Preferably, in step 1, the monitoring signal includes one or more of the following: voltage signal, current signal, ultrasonic emission signal, ultra-high frequency electromagnetic radiation signal, and optical radiation signal.

3. The method according to claim 1, characterized in that, In step 1, the characteristic parameters include maximum amplitude, average amplitude, standard deviation, kurtosis, skewness, spectral center frequency, spectral bandwidth, and pulse repetition rate.

4. The method according to claim 1, characterized in that, In step 2, the stage probability vector , Let be the predicted probability that the sample in the nth window belongs to stage C.

5. The method according to claim 4, characterized in that, Stage C includes the near-breakdown stage, the single-turn breakdown arc stage, the double-turn breakdown arc stage, and the three-turn breakdown arc stage.

6. The method according to claim 1, characterized in that, In step 2, a fully connected feedforward neural network is used as the discharge stage identification model.

7. The method according to claim 1, characterized in that, In step 3, the stage with the highest probability of smoothing is selected as the candidate stage. .

8. The method according to claim 1, characterized in that, The mapping relationship is as follows: .

9. A multi-turn inter-turn arc discharge graded early warning device, characterized in that, The device includes: The feature acquisition module is used to acquire the monitoring signal generated by the multi-turn winding during the inter-turn discharge process, segment the monitoring signal in a sliding window manner to obtain multiple continuous sliding windows, and extract the corresponding feature parameters for each sliding window. The probability output module is used to input the feature parameters into the discharge stage identification model to obtain the stage probability vector corresponding to the nth sliding window. The stage probability vector is then subjected to probability smoothing to obtain the smoothed stage probability. ; The state update module is used to determine candidate stages based on the smoothing stage probability. This triggers the hysteresis update criterion mechanism and outputs the stable phase state. ; The level mapping module is used to establish stable phase states. With graded early warning volume The mapping relationship is used to obtain the current warning level; The alarm output module is used to perform a persistent criterion mechanism on the current warning level. When the persistent criterion meets the triggering condition, a valid warning is output; otherwise, the previous warning output is maintained.

10. A computer-readable storage medium comprising a stored complete computer program that, when the computer program is run, implements the method as described in any one of claims 1-8.

Images