A circuit breaker fault diagnosis method based on an asymmetric gate convolutional network
By using asymmetric gated convolutional networks (AEG-TCN) for circuit breaker fault diagnosis, the problems of insufficient fault feature capture and noise interference in existing technologies are solved. This enables accurate temporal location and physical mechanism explanation of circuit breaker faults, supporting precise operation and maintenance of power systems.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANDONG UNIV
- Filing Date
- 2026-05-12
- Publication Date
- 2026-06-09
AI Technical Summary
Existing technologies struggle to capture the transient characteristics of faults in specific operating phases during circuit breaker fault diagnosis, and feature dilution occurs in complex electromagnetic environments, resulting in insufficient model robustness and an inability to provide physically interpretable diagnostic results.
A fault diagnosis method based on asymmetric gated convolutional network (AEG-TCN) is adopted. Through synchronous alignment of current mutation features, multi-dimensional feature representation, feature extraction of asymmetric gated temporal convolutional network, and inverse gradient weight mapping, the accurate temporal location and physical mechanism explanation of circuit breaker faults are achieved.
It enhances the ability to extract weak fault features from the transient evolution signals of circuit breakers, suppresses noise interference, provides physically interpretable diagnostic results, and supports precise operation and maintenance and full life cycle status assessment of power system switching equipment.
Smart Images
Figure CN122174054A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of circuit breaker fault diagnosis technology, specifically a circuit breaker fault diagnosis method based on asymmetric gated convolutional networks. Background Technology
[0002] In recent years, intelligent diagnostic technology based on vibration signal analysis has become an important research direction in the field of circuit breaker operation and maintenance due to its advantages of not altering the original structure of the equipment and being able to characterize the internal dynamic characteristics of the mechanism. Existing technologies typically utilize short-time Fourier transform and variational mode decomposition to extract the time-frequency fingerprint of vibration signals, and combine this with various neural network models for state discrimination. However, under actual power production conditions, this field still faces the following challenges.
[0003] First, the physical correlation between signal characteristics and action history is insufficiently explored. Circuit breaker opening and closing is a rapid, asymmetric dynamic process with a strict physical sequence. Existing technologies often treat vibration signals as a stationary sequence for overall processing, neglecting the phase evolution of signal energy over time. This lack of time-domain subdivision makes it difficult for models to capture transient characteristic shifts caused by faults in specific action phases, such as tripping, contact separation, and buffering, thus reducing sensitivity to weak early-stage faults.
[0004] Second, there is the issue of feature dilution under complex electromagnetic environments and impacts. The instantaneous operation of a circuit breaker is accompanied by strong electromagnetic pulse interference, and the broadband random vibrations generated by high-speed collisions of the mechanism can easily mask the true fault fingerprint. Traditional attention mechanisms or global feature extraction methods are prone to feature submersion when processing such strong background noise and short transient pulse signals. This makes it difficult to accurately focus on the sensitive time-domain segments beneficial for diagnosis from the non-stationary response spectrum, resulting in insufficient robustness of the model under varying operating conditions.
[0005] Third, the back-tracking step of the original time-domain signal in deep learning diagnostic decision-making is missing. Although high-performance neural networks have achieved breakthroughs in classification accuracy, in power equipment operation and maintenance practice, maintenance personnel not only focus on the fault type, but also urgently need to know the specific physical time of the fault occurrence as a basis for tracing the source of accurate maintenance. If the model cannot provide decision criteria with physical meaning, the industrial credibility of the diagnostic results will be reduced, making it difficult to meet the actual needs of intelligent operation and maintenance throughout the entire life cycle of power equipment.
[0006] To address the aforementioned challenges, there is an urgent need to explore an interpretable circuit breaker fault diagnosis scheme that can deeply integrate multi-dimensional electromechanical coupling features, possess asymmetric time-domain feature capture capabilities, and have accurate time-domain positioning capabilities during the diagnostic process. Summary of the Invention
[0007] The purpose of this invention is to provide a circuit breaker fault diagnosis method based on asymmetric gated convolutional networks, which can solve the problem that conventional diagnostic models are difficult to achieve time-domain localization and physical mechanism interpretability during the opening and closing operations of high-voltage circuit breakers in power systems due to the complex coupling of vibration and current signals, strong environmental noise interference, and lack of fault samples.
[0008] The technical solution adopted by the present invention to solve its technical problem is: a circuit breaker fault diagnosis method based on asymmetric gated convolutional network, comprising the following steps.
[0009] S1. Synchronization and alignment of electromechanical signal history based on current mutation characteristics.
[0010] S1.1 Collect the control current during the opening and closing process of the circuit breaker. Vibration spectrum throughout the entire motion process of the mechanism .
[0011] S1.2 Definition Current change at time quantification ,when continuous When a sampling point exceeds the preset sensitivity threshold, it is determined as the physical moment when the opening and closing command is issued and recorded as the starting anchor point. .
[0012] S2. Division of action phase micro-elements and multidimensional feature representation.
[0013] S2.1. The cut length is Vibration spectrum of the entire process of the mechanism's movement The time axis is divided into d infinitesimal elements at equal intervals.
[0014] S2.2. Multidimensional feature representation.
[0015] The sequences of each infinitesimal energy operator are fused with the original amplitude, kurtosis, and skewness in a multidimensional manner to construct a multidimensional feature vector reflecting the evolution of mechanical impact intensity. ;in, It is a multidimensional feature quantity formed by multidimensionally fusing the energy operator sequence of the dth infinitesimal element with the original amplitude, kurtosis and skewness.
[0016] S3. Based on the rate of change of current Adaptive physical guidance branch construction.
[0017] S3.1. Calculate the local gradient within each infinitesimal element. .
[0018] S3.2. Constructing a weight operator based on nonlinear mapping .
[0019] S3.3. Multidimensional feature quantities with weight operators Elemental level The product operation yields the reconstructed physical perception feature matrix. : ;in, It is a positive integer between 1 and d.
[0020] S4. Feature extraction based on AEG-TCN (Asymmetric gated temporal convolutional network).
[0021] AEG-TCN is used to extract cross-phase temporal evolution features and fuse multi-dimensional feature spaces within the same phase for the divided micro-elements.
[0022] S5. Action phase contribution mapping and fault time localization based on inverse gradient weights.
[0023] After the S5.1.AEG-TCN network results are output, calculate the first... Unnormalized score of fault class .
[0024] S5.2. For each fault category Perform indexing and mapping Intervals transform the strength of raw physical evidence captured in a high-dimensional feature space into mutually exclusive probability distributions. .
[0025] S5.3. Introduction A nonlinear activation operator is used to perform a second-order weighted mapping on the strength of the original phase evidence, generating a phase contribution mapping value. Using a linear stretching algorithm to Mapped to Standard interval; calculate the processed first interval. Thermodynamic intensity value of each phase element Using thermal intensity values as color indices, a pseudo-color layer is superimposed on the original vibration signal to form a heat map. The time period corresponding to the phase element of the vibration signal in the highlighted part of the heat map is the time period in which the fault occurred.
[0026] Furthermore, in step S1.1, the current transformer deployed in the circuit breaker control circuit and the vibration acceleration sensor installed in the operating mechanism are used to sample the frequency. Synchronously acquire control current during the opening and closing process Vibration spectrum throughout the entire motion process of the mechanism .
[0027] Furthermore, in step S1.2, (1); where, Let i be the current signal at time t; i(t+k) is the current signal at time t+k; k is the offset index; The length of the sliding window; denoted as the standard deviation of the current background noise.
[0028] Furthermore, in step S1.2, the method for setting the sensitivity threshold is as follows: collect the static background noise when the circuit breaker is not activated, and calculate its mean value. Preset sensitivity (2); where, This is the signal-to-noise ratio adjustment coefficient.
[0029] Furthermore, in step S2.1, the time span of a single infinitesimal element is set to... Then each infinitesimal contains 1 consecutive sampling point; mapping the original vibration signal to a signal obtained from... A sequence of infinitesimal elements (3); among which, Representing the Each action phase element.
[0030] Furthermore, in step S2.2, for a length of Action phase infinitesimal element, normalized kurtosis coefficient (4); among which, These are the vibration amplitude sampling points within the infinitesimal element; This is the mean of m sampling points in the currently processed infinitesimal element.
[0031] Furthermore, in step S2.2, the skewness operator (5); among which, This represents the signal sample value within the current phase element; This is the arithmetic mean of the signal; The standard deviation of the signal; This represents the number of sampling points within the infinitesimal element.
[0032] Furthermore, in step S2.2, the root mean square characterization of the discrete energy operator within the current action phase element... (6); among which, The current signal sampling amplitude; The amplitude of the signal sampled at the next sampling point immediately after the current moment; It represents the signal sampling amplitude of the immediately preceding sampling point at the current moment.
[0033] Furthermore, in step S3.1, (8); among which, is the current sampling step size; I(i) is the current amplitude at the i-th sampling point within the corresponding micro-element.
[0034] Furthermore, in step S3.2, (9); among which, This refers to the dynamic gain coefficient. yes The minimum value; yes The maximum value; Sigmoid is a non-linear mapping function.
[0035] Furthermore, in step S4, a lateral feature enhancement operator is introduced, and an asymmetric convolution feature fusion operation is defined: (11); among which, It is the output of asymmetric convolution; ReLU is the activation function. Represents the horizontal convolution kernel; Represents the vertical convolution kernel; It is the bias of the asymmetric convolutional layer.
[0036] Furthermore, in step S4, AEG-TCN introduces dilated convolutional layers and defines the dilated convolution operator of the e-th layer. (12); among which, The j-th column of the expanded convolution kernel; The stride is the expansion step size; j is the j-th column parameter of the convolution kernel; For the l-th infinitesimal element, the first... Characteristics of infinitesimal elements; This represents the kernel size.
[0037] Furthermore, in step S4, AEG-TCN configures a gated linear unit (GLU) after the convolutional layer, which runs parallel to the feature signal convolution process, and defines the gated feature filtering formula: (13); where G is the output of the gating layer; This is the output of the dilated convolutional layer; R is the feature extraction convolution kernel; a is the bias of the feature extraction branch; express Product; X is the gated information convolution kernel; c is the bias of the gated branch; a residual connection structure with layer normalization (LayerNorm) is introduced, and the residual state update equation is: (14); among which, This is the output of the residual layer; LayerNorm is the layer normalization function.
[0038] Furthermore, in step S5.1, (15); among which, The output representing the differential element number of the last residual block calculation is: The channel is eigenvalues; The weights of the fully connected layers determined during the pre-training process of AEG-TCN; It is the first Bias terms for fault types.
[0039] Furthermore, in step S5.2, (16); among which, The model indicates that the current sample is faulty. The final probability; is the total number of fault categories; exp is the exponential function. S θ The current sample is the one that is judged as the first. Unnormalized score for fault class.
[0040] Furthermore, in step S5.3, the output results of all residual layers are... Each eigenvalue on With the corresponding Perform point-to-point weighted summation to obtain (17); (18); among which, For the final generated first The phase element is relative to the first Contribution mapping value of the fault class; the processed value of the first fault. Thermodynamic intensity value of each phase element (19); among which, It is the minimum value among the various fault types of all the phase micro-elements generated in the end; It is the maximum value among the thermal intensity values of various fault types corresponding to all the phase micro-elements finally generated; It is a constant.
[0041] The beneficial effects of this invention are as follows: By leveraging the synergistic effect of asymmetric operators and gating mechanisms in the AEG-TCN architecture, this invention enhances the ability to extract weak fault features and suppress noise in the transient evolution signals of circuit breakers; by combining dilated convolution and residual connections, it maintains the full phase resolution of the time-domain signal while ensuring long-period physical correlation modeling, thus solving the problem of insufficient positioning accuracy in traditional diagnostic models; and finally, based on the phase contribution mapping of probabilistic backtracking, it intuitively locks the fault time period of the original signal on the original waveform, providing physically interpretable technical support for the accurate operation and maintenance and full life cycle status assessment of power system switching equipment. Attached Figure Description
[0042] Figure 1 This is a flowchart illustrating the overall concept of the present invention;
[0043] Figure 2This is a diagram of the AEG-TCN network structure of the present invention;
[0044] Figure 3 This is a flowchart of the gating mechanism. Detailed Implementation
[0045] like Figure 1 As shown, this invention first calculates the noise threshold of the original signal when the circuit breaker is not operating, and determines the starting anchor point of the electrical action by combining the current change rate; then, it extracts the action time interval and divides it into infinitesimal elements, calculating the kurtosis, skewness, and energy operators of the infinitesimal elements to form a comprehensive feature matrix; then, it inputs the formed data tensor into an asymmetric gated convolutional network (AEG-TCN) for diagnostic classification, and after obtaining the results, it uses a backtracking calculation process to deduce the fault time period in the original signal. This invention provides a circuit breaker fault diagnosis method based on an asymmetric gated convolutional network, specifically including the following steps.
[0046] S1. Synchronization and alignment of electromechanical signal history based on current mutation characteristics.
[0047] Using current transformers deployed in the circuit breaker control circuit and vibration acceleration sensors installed at key receiving points of the operating mechanism, such as the operating mechanism's support or linkage, at a sampling frequency... Synchronously acquire control current during the opening and closing process Vibration spectrum throughout the entire motion process of the mechanism .
[0048] To address the fundamental frequency shift and random noise interference present in power field applications, a sliding window energy derivative operator is used to scan the current sequence. Definition The current change at time t is characterized by: (1). Among them, Let i be the current signal at time t; i(t+k) is the current signal at time t+k; k is the offset index, representing the kth sampling point after the current time t. The length of the sliding window; This represents the standard deviation of the current background noise. continuous When a sampling point exceeds the preset sensitivity threshold, it is determined as the physical moment when the opening and closing command is issued and recorded as the starting anchor point. .
[0049] The method for setting the sensitivity threshold is as follows: collect the static background noise when the circuit breaker is not operating, and calculate its mean value. Preset sensitivity (2); where, This is the signal-to-noise ratio adjustment factor, typically ranging from 3 to 5.
[0050] S2. Division of action phase micro-elements and multidimensional feature representation.
[0051] To maintain the integrity of the entire circuit breaker operating process, starting from the anchor point The origin of the time axis, This represents the percentage of the pre-triggered time period within the selected diagnostic time period. This is the overall diagnostic time period used for selection. The time length is extracted upwards from the origin of the time axis. The pre-triggering phase signal is used to capture the very early, weak vibrations caused by the excitation of the opening and closing coils, extending downwards. As the action diagnosis time, a segment with a total length of [missing information] is extracted from the original monitoring data point set. The electromechanical coupling transient response sequence. Ensure the length is... The diagnostic signal window can construct a complete electromechanical fingerprint chain from the static pre-triggering region, the dynamic impact region to the steady-state convergence region, providing a temporal sample space for the long-range causal correlation modeling of the subsequent asymmetric gated convolutional network AEG-TCN.
[0052] S2.1. The cut length is Vibration spectrum of the entire process of the mechanism's movement The time axis is divided into d infinitesimal elements at equal intervals.
[0053] Set the time span of a single infinitesimal element as Then each infinitesimal contains Continuous sampling points. The original vibration signal is mapped to... A sequence of infinitesimal elements (3). Among them, Representing the Each action phase element.
[0054] S2.2. Multidimensional feature representation.
[0055] Kurtosis is a high-order statistic characterizing the distribution properties of a random variable and is used to measure the strength of the impulsiveness of a signal. For a length of... The action phase infinitesimal element, its normalized kurtosis coefficient for: (4). Among them, These are the vibration amplitude sampling points within the infinitesimal element; This represents the mean of m sampling points in the currently processed infinitesimal element. Kujicic is considered an impact fingerprint; unlike the mean and root mean square, kujicic is extremely sensitive to instantaneous pulses in the signal. When a circuit breaker mechanism experiences a collision, trip release, or contact closing, the vibration signal generates a strong instantaneous impact, causing the signal distribution to deviate from a normal distribution, at which point the kujicic value increases significantly. Under normal conditions, the mechanism moves smoothly, and the kujicic value is in a low, stable range; when the mechanism jams, the cam wears, or the bolts loosen, abnormal friction and collisions in the mechanism cause spikes in the signal distribution, leading to a surge in the kujicic value.
[0056] The skewness operator characterizes the degree of skewness of the circuit breaker's opening and closing waveforms relative to the mean. Under normal operating conditions, the impact vibration signal typically exhibits a relatively symmetrical distribution. However, when linkage jamming, lubricant drying, or contact spring fatigue occurs, the resistance torque of the mechanical motion changes asymmetrically, resulting in a leading or lagging tailing effect of vibration energy on the time axis. By monitoring changes in the skewness index in real time, AEG-TCN can capture this waveform distortion caused by changes in micro-friction, thereby significantly improving the early warning capability for sub-optimal conditions of the mechanism. The skewness operator obtained from the current micro-element calculation... (5). Among them, This represents the signal sample value within the current phase element; This is the arithmetic mean of the signal; The standard deviation of the signal; This represents the number of sampling points within the infinitesimal element.
[0057] The energy operator is introduced as the core dimension of the feature tensor to capture the extremely brief transient energy bursts during the opening and closing of circuit breakers. Through the nonlinear combination of the current sampling point and its preceding and following neighboring points, the energy operator can simultaneously exhibit high sensitivity to signal amplitude fluctuations and frequency jumps. Significant pulse-like abrupt changes in the energy operator value occur when the circuit breaker contacts arc, the mechanism impacts, or the trip unit actuates. By multidimensionally fusing the energy operator sequence with the original amplitude, kurtosis, and skewness, the asymmetric gated temporal convolutional network AEG-TCN identifies the starting point of potential faults with a narrower time-domain window, thus solving the time delay ambiguity problem inherent in traditional time-domain features when analyzing non-stationary, highly transient circuit breaker signals. The root mean square representation of the discrete energy operator within the current action phase micro-element is shown. (6). Among them, The current signal sampling amplitude; The amplitude of the signal sampled at the next sampling point immediately after the current moment; It represents the signal sampling amplitude of the immediately preceding sampling point at the current moment.
[0058] By traversing all infinitesimal elements and calculating multidimensional features, a multidimensional feature vector reflecting the evolution of mechanical impact intensity is constructed: (7). It is a multidimensional feature quantity formed by multidimensionally fusing the energy operator sequence of the dth infinitesimal element with the original amplitude, kurtosis and skewness.
[0059] S3. Based on the rate of change of current Adaptive physical guidance branch construction.
[0060] In circuit breaker fault diagnosis, although vibration signals characterize rich mechanical information about the current operating state of the circuit breaker, they are susceptible to interference from random environmental noise. This invention introduces the opening and closing control current to construct an adaptive physical guidance branch, aiming to dynamically weight multiple action phase micro-elements to enhance effective features and suppress background noise. The circuit breaker's control current... It is the driving force that propels the operating mechanism. The rate of change of current. It directly reflects the change in the coil excitation intensity and the motion state of the electromagnet armature. When When the vibration is large, it usually corresponds to the tripping trigger of the mechanism or the period of electromagnetic force burst. The vibration signal generated at this time has extremely high diagnostic value. When the value approaches zero and the action is completed or has not yet started, the sensor mainly collects electromagnetic noise or environmental interference, corresponding to the background noise range. This is for the action phase micro-element sequence divided in step S2. Simultaneously calculate the current change rate characteristics within the corresponding time window.
[0061] S3.1. First, calculate the local gradient within each infinitesimal element. : (8). Among them, is the current sampling step size; I(i) is the current amplitude at the i-th sampling point within the corresponding micro-element.
[0062] S3.2. To prevent excessive weight fluctuations from causing feature distortion, a weight operator based on nonlinear mapping is constructed. , (9). Among them, This is the dynamic gain coefficient, typically taken as... ; yes The minimum value; yes The maximum value; Sigmoid is a nonlinear mapping function. The weighting operator maps the dynamic intensity of the current to... Within the interval, attention factors that form physical perception are formed.
[0063] S3.3. Finally, the multi-dimensional feature quantities with weight operators Elemental level The product operation yields the reconstructed physical perception feature matrix. : (10); It is a positive integer between 1 and d.
[0064] S4. Feature extraction based on AEG-TCN (Asymmetric gated temporal convolutional network).
[0065] Obtaining the feature matrix after physical guide branch correction Subsequently, an asymmetric gated temporal convolutional network, AEG-TCN, was constructed. AEG-TCN aims to map from the original signal to high-dimensional fault features by mimicking the asymmetric physical dynamics of a circuit breaker. The AEG-TCN architecture is as follows: Figure 2 As shown, AEG-TCN consists of four parts: asymmetric convolution, dilated convolution, gated layers, and residual connections. Asymmetric convolution uses two convolutional kernels, one horizontal and one vertical, to extract cross-phase temporal evolution features from predefined micro-elements and to fuse multi-dimensional feature spaces within the same phase. Subsequently, in the dilated convolution stage, the convolution operator, with increasing layer number, captures the physical relationships between distant phase points in a jump-like manner, while relying on a padding mechanism to ensure the number of micro-elements is maintained. A mapping branch spanning the convolutional layers is constructed using residual connections, directly superimposing the original signal or previous layer feature signals onto the gated output features.
[0066] Unlike traditional neural networks that use symmetric square matrix convolution kernels, this invention utilizes asymmetric kernel decomposition techniques. Circuit breaker signals exhibit strong evolutionary causality in the time domain and strong coupling in the feature dimension. By introducing a lateral feature enhancement operator, the asymmetric convolution feature fusion operation is defined as follows: (11). Among them, It is the output of asymmetric convolution; ReLU is the activation function. Represents a transverse convolution kernel, used to capture the temporal evolution topology across phases; This represents a vertical convolution kernel, used to integrate multidimensional feature information within the same phase element; It is the bias of the asymmetric convolutional layer. This feature fusion design significantly improves the model's ability to capture the unidirectional offset features of the circuit breaker's operating phase while reducing the number of parameters and the risk of overfitting.
[0067] To perceive the causal relationships between long-distance phases throughout the entire opening and closing process of a circuit breaker, an extended convolution operator of layer e is defined. (12). Among them, The j-th column of the expanded convolution kernel; The stride is the expansion step size; j is the j-th column parameter of the convolution kernel; For the l-th infinitesimal element, the first... Characteristics of infinitesimal elements; This refers to the kernel size. As the number of layers increases... The increase in expansion factor The exponential growth rate allows the model to obtain an ultra-large effective receptive field covering the entire diagnostic time domain without losing local transient information.
[0068] To address atypical impact interference in circuit breaker vibration signals and errors caused by data padding in deep convolutional networks, AEG-TCN incorporates a gated linear unit (GLU) after the convolutional layers, operating in parallel with the feature signal convolution process. The GLU simulates a threshold-triggered mechanism in physical systems, allowing features to be passed down only when the signal energy exceeds a threshold. The GLU processing flow is as follows: Figure 3 As shown, two sets of asymmetric convolutional kernels, the feature mapping operator and the spatiotemporal gating operator, are used to extract the feature signals and gating signals from the previous layer signal, respectively. The function generates a dynamic gain coefficient from zero to one to process the characteristic signal and suppress edge effects caused by zero padding and environmental electromagnetic noise.
[0069] Define the gating feature filtering formula: (13). Where G is the output of the gating layer; This is the output of the dilated convolutional layer; R is the feature extraction convolution kernel; a is the bias of the feature extraction branch; express The product is defined as follows: X is the gated information convolution kernel; c is the bias of the gated branch. The R branch is responsible for extracting the main features, while... Branches generate physically meaningful soft masks. This gated computation method adaptively filters out residual background noise after asymmetric convolution, allowing the network to focus on high-value mechanical action features. To ensure the stability of deep network training and prevent gradient vanishing, a residual connection structure with layer normalization (LayerNorm) is introduced. The residual state update equation is: (14). Among them, This is the output of the residual layer; LayerNorm is the layer normalization function.
[0070] The above calculation process of the dilated convolutional layer, gated layer and residual normalization layer is to calculate the residual block for the eth iteration in a loop. By directly superimposing the output signal of the asymmetric convolution across layers into the output result of the residual block in each iteration, the model can form the input of the residual block in the next iteration. The model retains the original information flow of the physical guidance branch of the previous layer, ensuring that the high-level semantic features still contain clear physical temporal references, avoiding the gradient vanishing problem, and providing a stable weight benchmark for the temporal localization of the fault time period of the original signal.
[0071] S5. Action phase contribution mapping and fault time localization based on inverse gradient weights.
[0072] S5.1. First, after the AEG-TCN network output is complete, the final matrix obtained by the network is compressed into category scores, and the first category score is calculated. Unnormalized score of fault class (15). Among them, The output representing the differential element number of the last residual block calculation is: The channel is eigenvalues; The weights of the fully connected layers determined during the pre-training process of AEG-TCN represent the channels. For the fault Sensitivity; It is the first Bias terms for fault types.
[0073] S5.2. Subsequently, for each fault category... Perform indexing and mapping Intervals aim to transform the strength of raw physical evidence captured in a high-dimensional feature space into mutually exclusive probability distributions. By amplifying the score differences between different categories through exponential operations, this process outputs the fault type determination with the highest confidence, and at the same time provides a weight gain coefficient for the phase contribution calculation in subsequent reverse backtracking. (16). Among them, The model indicates that the current sample is faulty. The final probability; is the total number of fault categories; exp is the exponential function. S θ The current sample is the one that is judged as the first. Unnormalized score for fault class.
[0074] For each individual phase element index The response values of the infinitesimal element on all feature channels Corresponding category-sensitive weights Point-to-point weighted summation is performed. This process does not perform cross-time dimension information compression, thus fully preserving the spatial distribution characteristics of the circuit breaker throughout its opening and closing process. The final generated contribution score... It represents the influence intensity of each phase element on a specific fault mode, and can quantify the degree of matching between the waveform characteristics of each element and the physical fingerprint of the target fault. (17).
[0075] S5.3. Then, the final probability is introduced. A nonlinear activation operator is used to perform a second-order weighted mapping on the strength of the original phase evidence, generating the final phase contribution mapping value. In this mapping process, the original contribution scores for each phase are first obtained using the linear rectified function ReLU. Nonlinear filtering is performed. Its physical meaning lies in retaining only the characteristic responses that positively excite the target fault mode during the complex vibration and current evolution process of the circuit breaker, while eliminating antagonistic characteristic interferences that suppress fault determination or are unrelated to the fault mechanism by forcibly clearing negative scores to zero. Then, the final probability is introduced. As a global confidence modifier. (18). Among them, For the final generated first The phase element is relative to the first Contribution mapping value for fault types.
[0076] After the model's backtracking calculation is completed, due to the order-of-magnitude difference in the absolute values of contributions between different samples, direct superposition may overwhelm the localization information of weak feature faults. Therefore, a linear stretching algorithm is used to extract the contribution values of the phase infinitesimals throughout the entire process. Mapped to Standardized intervals. This standardization process eliminates the influence of dimensions by dynamically scaling the maximum and minimum values of the target fault category across the entire time axis, allowing the diagnostic system to quantify the fault risk for each time period. The resulting thermal intensity values are used as color indexes, and a pseudo-color layer is overlaid on the original vibration signal to form a heatmap. The time period corresponding to the phase element of the vibration signal in the highlighted part of the heatmap is the time period in which the fault occurred. This processing method can highlight the fault elements that contribute the most to the diagnostic conclusion. (19). Among them, It is the minimum value among the various fault types of all the phase micro-elements generated in the end; It is the maximum value among the thermal intensity values of various fault types corresponding to all the phase micro-elements finally generated; It is a constant to prevent the denominator from being zero.
[0077] In summary, this invention enhances the model's ability to extract weak fault features and suppress noise in the transient evolution signal of circuit breakers by leveraging the synergistic effect of asymmetric operators and gating mechanisms in the AEG-TCN architecture. By combining dilated convolution and residual connections, it maintains the full phase resolution of the time-domain signal while ensuring long-term physical correlation modeling, thus solving the problem of insufficient positioning accuracy in traditional diagnostic models. Finally, based on the phase contribution mapping of probabilistic backtracking, it intuitively locks the fault time period of the original signal on the original waveform, providing physically interpretable technical support for the accurate operation and maintenance and full life cycle status assessment of power system switching equipment.
Claims
1. A circuit breaker fault diagnosis method based on asymmetric gated convolutional networks, characterized in that, Includes the following steps: S1. Synchronization Alignment of Electromechanical Signal History Based on Current Sudden Change Characteristics S1.1 Collect the control current during the opening and closing process of the circuit breaker. Vibration spectrum throughout the entire motion process of the mechanism ; S1.2 Definition Current change characterization quantity at time step ,when continuous When a sampling point exceeds the preset sensitivity threshold, it is determined as the physical moment when the opening and closing command is issued and recorded as the starting anchor point. ; S2. Division and Multidimensional Feature Representation of Action Phase Elements S2.
1. The cut length is vibration spectrum And divide it into d infinitesimal elements according to the time axis; S2.
2. Multidimensional Feature Representation The energy operator sequences of each infinitesimal element are fused with the original amplitude, kurtosis, and skewness in a multidimensional manner to construct a multidimensional feature vector reflecting the evolution of mechanical impact intensity. ;in, It is the multidimensional feature quantity of the dth infinitesimal element; S3. Based on the rate of change of current Adaptive physical guidance branch construction S3.
1. Calculate the local gradient within each infinitesimal element. ; S3.
2. Constructing a weight operator based on nonlinear mapping ; S3.
3. will and Element-level The product operation yields the reconstructed physical perception feature matrix. : ;in, It is a positive integer between 1 and d; S4. Feature extraction based on AEG-TCN (Asymmetric gated temporal convolutional network) AEG-TCN is used to extract cross-phase temporal evolution features and fuse multi-dimensional feature spaces within the same phase for the divided micro-elements; S5. Action Phase Contribution Mapping and Fault Time Location Based on Backward Gradient Weights After the S5.1.AEG-TCN network results are output, calculate the first... Unnormalized score of fault class ; S5.
2. For each fault category Perform indexing and mapping Intervals transform the strength of raw physical evidence captured in a high-dimensional feature space into mutually exclusive probability distributions. ; S5.
3. Introduction A nonlinear activation operator is used to perform a second-order weighted mapping on the strength of the original phase evidence, generating a phase contribution mapping value. Using a linear stretching algorithm to Mapped to Standard interval; calculate the processed first interval. Thermodynamic intensity value of each phase element Using thermal intensity values as color indices, a pseudo-color layer is superimposed on the original vibration signal to form a heat map. The time period corresponding to the phase element of the vibration signal in the highlighted part of the heat map is the time period in which the fault occurred.
2. The circuit breaker fault diagnosis method based on asymmetric gated convolutional networks according to claim 1, characterized in that, In step S1.1, the current transformer deployed in the circuit breaker control circuit and the vibration acceleration sensor installed in the operating mechanism are used to sample the frequency. Synchronous acquisition and .
3. The circuit breaker fault diagnosis method based on asymmetric gated convolutional networks according to claim 2, characterized in that, In step S1.2, (1); where, Let i be the current signal at time t; i(t+k) is the current signal at time t+k; k is the offset index; The length of the sliding window; This represents the standard deviation of the current background noise.
4. The circuit breaker fault diagnosis method based on asymmetric gated convolutional networks according to claim 3, characterized in that, In step S1.2, the method for setting the sensitivity threshold is as follows: collect the static background noise when the circuit breaker is not activated, and calculate its mean value. ; Preset sensitivity (2); where, This is the signal-to-noise ratio adjustment coefficient.
5. The circuit breaker fault diagnosis method based on asymmetric gated convolutional networks according to claim 4, characterized in that, In step S2.1, the time span of a single infinitesimal element is set to... Then each infinitesimal contains One continuous sampling point; The original vibration signal is mapped to... A sequence of infinitesimal elements (3); among which, Representing the Each action phase element.
6. The circuit breaker fault diagnosis method based on asymmetric gated convolutional networks according to claim 5, characterized in that, In step S2.2, for a length of Action phase infinitesimal, normalized kurtosis coefficient (4); among which, These are the vibration amplitude sampling points within the infinitesimal element; This is the mean of m sampling points in the currently processed infinitesimal element.
7. The circuit breaker fault diagnosis method based on asymmetric gated convolutional networks according to claim 6, characterized in that, In step S2.2, skewness (5); among which, This represents the signal sample value within the current phase element; This is the arithmetic mean of the signal; The standard deviation of the signal; This represents the number of sampling points within the infinitesimal element.
8. The circuit breaker fault diagnosis method based on asymmetric gated convolutional networks according to claim 7, characterized in that, In step S2.2, the root mean square characterization of the discrete energy operator within the current action phase element. (6); among which, The current signal sampling amplitude; The amplitude of the signal sampled at the next sampling point immediately after the current moment; It represents the signal sampling amplitude of the immediately preceding sampling point at the current moment.
9. The circuit breaker fault diagnosis method based on asymmetric gated convolutional networks according to claim 8, characterized in that, In step S3.1, (8); among which, is the current sampling step size; I(i) is the current amplitude at the i-th sampling point within the corresponding micro-element.
10. The circuit breaker fault diagnosis method based on asymmetric gated convolutional networks according to claim 9, characterized in that, In step S3.2, (9); among which, This refers to the dynamic gain coefficient. yes The minimum value; yes The maximum value; Sigmoid is a non-linear mapping function.
11. The circuit breaker fault diagnosis method based on asymmetric gated convolutional networks according to claim 10, characterized in that, In step S4, a lateral feature enhancement operator is introduced, and an asymmetric convolution feature fusion operation is defined: (11); among which, It is the output of asymmetric convolution; ReLU is the activation function. Represents the horizontal convolution kernel; Represents the vertical convolution kernel; It is the bias of the asymmetric convolutional layer.
12. The circuit breaker fault diagnosis method based on asymmetric gated convolutional networks according to claim 11, characterized in that, In step S4, AEG-TCN introduces dilated convolutional layers and defines the dilated convolution operator of the e-th layer. (12); among which, The j-th column of the expanded convolution kernel; The stride is the expansion step size; j is the j-th column parameter of the convolution kernel; For the l-th infinitesimal element, the first... Characteristics of infinitesimal elements; This represents the kernel size.
13. The circuit breaker fault diagnosis method based on asymmetric gated convolutional networks according to claim 12, characterized in that, In step S4, AEG-TCN configures a gated linear unit (GLU) after the convolutional layer, which runs parallel to the feature signal convolution process, and defines the gated feature filtering formula: (13); where G is the output of the gating layer; This is the output of the dilated convolutional layer; R is the feature extraction convolution kernel; a is the bias of the feature extraction branch; express Product; X is the gated information convolution kernel; c is the bias of the gated branch; a residual connection structure with layer normalization (LayerNorm) is introduced, and the residual state update equation is: (14); among which, This is the output of the residual layer; LayerNorm is the layer normalization function.
14. The circuit breaker fault diagnosis method based on asymmetric gated convolutional networks according to claim 13, characterized in that, In step S5.1, (15); among which, The output representing the differential element number of the last residual block calculation is: The channel is eigenvalues; The weights of the fully connected layers determined during the pre-training process of AEG-TCN; It is the first Bias terms for fault types.
15. A circuit breaker fault diagnosis method based on an asymmetric gated convolutional network according to claim 14, characterized in that, In step S5.2, (16); among which, The model indicates that the current sample is faulty. The final probability; is the total number of fault categories; exp is the exponential function. S θ The current sample is the one that is judged as the first. Unnormalized scores for fault classes.
16. The circuit breaker fault diagnosis method based on asymmetric gated convolutional networks according to claim 15, characterized in that, In step S5.3, With the corresponding Perform point-to-point weighted summation to obtain (17); (18); among which, For the final generated first The phase element is relative to the first Contribution mapping value of the fault class; the processed value of the first fault. Thermodynamic intensity value of each phase element (19); among which, It is the minimum value among the various fault types of all the phase micro-elements generated in the end; It is the maximum value among the thermal intensity values of various fault types corresponding to all the phase micro-elements finally generated; It is a constant.