A disconnector multi-scale power signal feature fusion and fault detection method, system, device and medium

Abstract

The application relates to the technical field of signal detection, and discloses a disconnecting switch multi-scale power signal feature fusion and fault detection method and system, which comprises the following steps: obtaining a power signal of a disconnecting switch operating mechanism and preprocessing the power signal to obtain a pretreated power signal; inputting the pretreated power signal into a pre-trained multi-scale feature extraction network to obtain multi-scale time sequence features, and extracting multi-domain auxiliary features from the pretreated power signal, wherein the multi-domain auxiliary features at least include time domain statistical features, frequency domain energy features and nonlinear dynamics features; weighting and fusing the multi-scale time sequence features and the multi-domain auxiliary features by using an attention mechanism to generate fusion features; and inputting the fusion features into a multi-task diagnosis network to output fault category prediction results and fault severity estimation results. Through multi-scale features and data enhancement, the application improves early fault identification sensitivity and model generalization capability, discriminates fault attributes in combination with multi-task learning, and realizes end-to-end diagnosis.

Description

Technical Field

[0001] This invention relates to the field of signal detection technology, and in particular to a method, system, device and medium for multi-scale power signal feature fusion and fault detection of disconnect switches. Background Technology

[0002] Gas-insulated metal-enclosed switchgear (GIS) is widely used in high-voltage, ultra-high-voltage, and extra-high-voltage power grids due to its compact structure and minimal susceptibility to environmental influences. It is also a key component of GIS, with its core function being circuit isolation and conduction. However, the fully enclosed structure of GIS makes it difficult to detect the mechanical condition of the disconnecting switches through routine inspections, resulting in highly concealed faults. Once a fault occurs, it can lead to equipment damage, power outages, or even more serious power grid accidents. Currently, the static contact condition of the contacts can be directly determined by measuring electrical parameters such as the main circuit contact resistance or loop resistance. However, these methods rely on power outage tests and cannot detect the dynamic mechanical characteristics during operation, lacking response to minor jamming caused by wear or poor lubrication. To obtain dynamic information, vibration signal analysis methods are used, which involve installing accelerometers on the equipment casing to collect mechanical vibration waves caused by operation. However, the vibration undergoes complex reflection, refraction, and attenuation as it is transmitted from the internal mechanism to the casing, resulting in severe signal distortion, reduced signal-to-noise ratio, and increased difficulty in feature extraction. Another approach is to use the drive power signal of the operating mechanism motor for diagnosis. This signal has a high signal-to-noise ratio and is easy to collect, and can indirectly reflect the mechanical state of the transmission chain.

[0003] However, existing power signal analysis methods typically focus on threshold comparisons or extracting global statistics such as average, peak, and energy values ​​from the entire power curve, paying little attention to the local transient changes implied at the instants of contact opening and closing. This results in insufficient ability to characterize early, subtle anomalies. Furthermore, most solutions fail to utilize the transient characteristics at critical events such as contact opening and closing in a refined manner, making it difficult to effectively distinguish different fault modes, especially exhibiting low sensitivity in identifying early, minor faults such as operational jamming. Simultaneously, the scarcity and uneven distribution of fault samples in industrial settings prevent data-driven diagnostic models from fully learning the distribution patterns of various fault types during training, limiting model generalization ability and making them prone to false negatives or missed positives in application. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a method and system for multi-scale power signal feature fusion and fault detection of disconnect switches to solve the problems of low sensitivity in early minor fault identification of disconnect switches and poor generalization ability of diagnostic models due to imbalance of fault samples.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for multi-scale power signal feature fusion and fault detection of disconnecting switches, comprising: acquiring the power signal of the disconnecting switch operating mechanism and performing preprocessing to obtain a preprocessed power signal; The preprocessed power signal is input into a pre-trained multi-scale feature extraction network to obtain multi-scale temporal features, and multi-domain auxiliary features are extracted from the preprocessed power signal. The multi-domain auxiliary features include at least time-domain statistical features, frequency-domain energy features, and nonlinear dynamic features. The multi-scale temporal features and the multi-domain auxiliary features are weighted and fused using an attention mechanism to generate fused features; The fused features are input into a multi-task diagnostic network, which outputs fault category prediction results and fault severity estimation results.

[0007] As a preferred embodiment of the multi-scale power signal feature fusion and fault detection method for disconnecting switches described in this invention, the multi-scale feature extraction network and the multi-task diagnostic network are trained by jointly optimizing classification loss and regression loss using a training set containing real power signal samples and synthetic power signal samples generated by a conditional generative adversarial network.

[0008] As a preferred embodiment of the multi-scale power signal feature fusion and fault detection method for disconnecting switches described in this invention, the preprocessed power signal is input into a pre-trained multi-scale feature extraction network to obtain multi-scale temporal features, including: The preprocessed power signal is input into a multi-scale feature extraction network; By using multiple residual blocks containing dilated causal convolutions and residual connections, shallow, mid-level, and deep features with different receptive fields are extracted at multiple network depths. The shallow, middle, and deep features respectively characterize the local details of the contact action transient, the trend changes in the transmission process, and the global dynamic characteristics of the entire operation process. The features from different layers are spliced ​​together to form the multi-scale temporal features.

[0009] As a preferred embodiment of the multi-scale power signal feature fusion and fault detection method for disconnecting switches described in this invention, wherein: the multi-domain auxiliary features include: The time-domain statistical characteristics include: the mean, standard deviation, skewness, and kurtosis of the power signal, which are used to reflect the overall distribution pattern of the signal; The frequency domain energy characteristics are obtained by decomposing the preprocessed power signal into multiple frequency bands and calculating the wavelet energy ratio of each frequency band, which is used to characterize the contribution of different frequency components to the signal energy. The nonlinear dynamic characteristics are obtained by calculating the entropy of the power signal to obtain complexity characteristics, which are used to capture random fluctuations in the signal.

[0010] As a preferred embodiment of the multi-scale power signal feature fusion and fault detection method for disconnecting switches described in this invention, the step of using an attention mechanism to weightedly fuse the multi-scale temporal features and the multi-domain auxiliary features to generate fused features includes: A learnable weight vector is assigned to each feature vector to be fused, and the attention weights corresponding to each feature vector are calculated by a normalized exponential function. The fused features are obtained by weighting and summing all feature vectors according to the attention weights, so that the fused features dynamically focus on the feature dimension that contributes the most to the determination of the current mechanical state of the disconnector switch.

[0011] As a preferred embodiment of the multi-scale power signal feature fusion and fault detection method for disconnecting switches described in this invention, the multi-task diagnostic network includes: a shared feature processing layer, a classification branch, and a regression branch. The classification branch takes the fusion feature as input and outputs the probability vector of the disconnect switch belonging to each preset fault category through a normalized exponential function, which serves as the fault category prediction result. The regression branch takes the fusion feature as input and outputs a value within a preset continuous interval through the S-shaped activation function, which serves as the fault severity estimation result characterizing the degree of fault deterioration.

[0012] As a preferred embodiment of the multi-scale power signal feature fusion and fault detection method for disconnecting switches described in this invention, the construction of the training set and the training of the network include: Acquire samples of the actual active power signals of the disconnecting switch under different mechanical conditions and label them with fault category tags; Using random noise and corresponding fault category labels as conditions, a generator of a conditional generative adversarial network generates simulated synthetic power signal samples, which are then distinguished as genuine or fake by a discriminator. Through adversarial training, the synthetic power signal samples generated by the generator conform to the characteristic distribution of the power curve under the corresponding fault category. The synthesized power signal samples that have undergone quality assessment are combined with the real active power signal samples to form a balanced training set in order to compensate for the data imbalance caused by the scarcity of fault samples. Based on the balanced training set, the parameters of the multi-scale feature extraction network and the multi-task diagnostic network are jointly optimized using the weighted sum of the cross-entropy loss for fault classification and the mean squared error loss for fault severity regression as the total loss.

[0013] Secondly, the present invention provides a multi-scale power signal feature fusion and fault detection system for disconnecting switches, comprising: The acquisition module is used to acquire the power signal of the disconnector switch operating mechanism and preprocess it to obtain a preprocessed power signal; The feature extraction module is used to input the preprocessed power signal into a pre-trained multi-scale feature extraction network to obtain multi-scale temporal features, and to extract multi-domain auxiliary features from the preprocessed power signal. The multi-domain auxiliary features include at least time-domain statistical features, frequency-domain energy features, and nonlinear dynamic features. The fusion module is used to perform weighted fusion of the multi-scale temporal features and the multi-domain auxiliary features using an attention mechanism to generate fused features; The fault output module is used to input the fused features into the multi-task diagnostic network and output the fault category prediction result and the fault severity estimation result.

[0014] Thirdly, the present invention provides a computer device, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, they implement the steps of a method for multi-scale power signal feature fusion and fault detection of disconnecting switches.

[0015] Fourthly, the present invention provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the steps of the method for multi-scale power signal feature fusion and fault detection of a disconnecting switch.

[0016] Compared with existing technologies, the advantages of this invention are as follows: This invention employs a multi-scale feature extraction strategy to characterize the mechanical state information contained in power signals from both macroscopic trends and microscopic transients, enabling the accurate capture of early, slight jamming features that were originally present in the global curve. Furthermore, it uses conditional generative adversarial networks to specifically expand fault samples, alleviating the constraints of imbalanced industrial field data on model training and improving the diagnostic model's recognition accuracy and generalization ability for minority-class faults. Combined with a multi-task learning framework, it provides continuous degree estimation while specifying the fault category, achieving a fine-grained quantitative description of the equipment degradation process. Overall, it reduces reliance on human experience and possesses better adaptability and generalizability. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the overall process of a method for multi-scale power signal feature fusion and fault detection of disconnecting switches according to an embodiment of the present invention. Detailed Implementation

[0019] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0020] Example 1, referring to Figure 1 As an embodiment of the present invention, a method for multi-scale power signal feature fusion and fault detection of disconnect switches is provided, comprising: S100: Obtain the power signal of the disconnector switch operating mechanism and preprocess it to obtain a preprocessed power signal; S200: Input the preprocessed power signal into a pre-trained multi-scale feature extraction network to obtain multi-scale temporal features, and extract multi-domain auxiliary features from the preprocessed power signal. The multi-domain auxiliary features include at least time-domain statistical features, frequency-domain energy features, and nonlinear dynamic features. S300: The multi-scale temporal features and the multi-domain auxiliary features are weighted and fused using an attention mechanism to generate fused features; S400: Input the fused features into the multi-task diagnostic network and output the fault category prediction result and the fault severity estimation result.

[0021] To address the existing problems in the background technology described above, this embodiment provides a method through steps S100-S400 that can automatically extract multi-scale features of power signals, namely macroscopic trends, microscopic transients, and steady-state backgrounds, and perform data augmentation through generative adversarial networks, combined with multi-task learning to achieve high-precision, fine-grained fault diagnosis.

[0022] The core concept of this method lies in multi-scale feature decoupling and event-driven diagnosis. Unlike traditional methods that analyze the entire power signal as a single object, this invention first decouples the power signal into three physically distinct and complementary scale features based on the physical mechanism of disconnector operation: Macro trend characteristics: Reflects the overall smoothness of the entire operation process, total operation time, and power consumption level.

[0023] Micro-event characteristics: Focusing on the details of signal abrupt changes within an extremely short time window near the point of contact / separation, capturing sensitive indicators such as impact, bounce, and friction at the moment of contact.

[0024] Steady-state operating characteristics: Reflects the background resistance fluctuation of the transmission chain during the stable operation phase, and can be used to judge the lubrication status and the consistency of transmission components.

[0025] By specifically modeling, extracting, and fusing the features of the above three scales, this invention can accurately capture sensitive indicators of early faults such as abnormal increase in impact at the contact point or sudden change in resistance during transmission, thereby achieving early warning and accurate positioning of the mechanical state of the disconnector switch.

[0026] Example 2, refer to Figure 1 As an embodiment of the present invention, based on the above embodiment, a method for multi-scale power signal feature fusion and fault detection of disconnect switches is provided. The method includes: S100: Obtain the power signal of the disconnector switch operating mechanism and preprocess it to obtain a preprocessed power signal; Specifically, the power signal of the disconnector switch operating mechanism can be obtained by installing a high-precision power sensor on the motor power line of the disconnector switch operating mechanism, and collecting the active power signal sequence in real time during the opening / closing operation at a sampling rate of not less than 10kHz. ,in The sequence length is given.

[0027] In one optional implementation, preprocessing may include signal normalization; to eliminate the influence of signal amplitude dimensions under different devices and operating conditions and improve the stability of model training; for example, the raw power signal is first subjected to minimum-maximum normalization. Let the original power signal sequence be... ,in Where is the sequence length. The min-max normalization method is used. in, The first in the power signal sequence Each sample value; Power signal sequence, including One sampling point; The normalized version One value; For sequence The minimum value in; :sequence The maximum value in the range; this formula is mainly used to linearly map the original power signal to the [0,1] interval while preserving the relative change characteristics of the original signal.

[0028] In another optional implementation, based on the above implementation, the preprocessing may further include: adaptive wavelet denoising; for measurement noise and electromagnetic interference present in the disconnector switch power signal, an adaptive threshold denoising method based on wavelet transform is adopted to filter out noise while preserving the steep edge features of key events such as the point of separation / combination to the maximum extent. For example, the steps are as follows: For the normalized signal Perform discrete wavelet transform to obtain wavelet coefficients. Let the wavelet transform be ,in For scale parameters, Here are the translation parameters. A soft thresholding function is used to process the wavelet coefficients. Adaptive threshold Determined adaptively based on noise level, expressed as: The noise standard deviation is robustly estimated using the median of the scale wavelet coefficients, and is expressed as: Perform inverse wavelet transform on the processed wavelet coefficients to obtain the denoised power signal, which is: in, is the soft thresholding function for wavelet coefficients; These are the wavelet coefficients after wavelet transform; The threshold is determined adaptively; The standard deviation of noise; This refers to the signal length or the number of sampling points. The noise standard deviation is estimated using the median of the wavelet coefficients; These are wavelet coefficients; This is a median function.

[0029] S200: Input the preprocessed power signal into a pre-trained multi-scale feature extraction network to obtain multi-scale temporal features, and extract multi-domain auxiliary features from the preprocessed power signal. The multi-domain auxiliary features include at least time-domain statistical features, frequency-domain energy features, and nonlinear dynamic features. In this embodiment of the application, in step S200, the preprocessed power signal is input into a pre-trained multi-scale feature extraction network to obtain multi-scale temporal features, including the following steps A1-A3: A1: Input the preprocessed power signal into the multi-scale feature extraction network; A2: By using multiple residual blocks containing dilated causal convolutions and residual connections, shallow, mid-level, and deep features with different receptive fields are extracted at multiple network depths. The shallow, middle, and deep features respectively characterize the local details of the contact action transient, the trend changes in the transmission process, and the global dynamic characteristics of the entire operation process. A3: The features from different layers are spliced ​​together to form the multi-scale temporal features.

[0030] Specifically, in A1-A3, the multi-scale feature extraction network uses a temporal convolutional network (TCN) as the core feature extractor. Its dilated causal convolution architecture ensures that the output depends only on the current and historical inputs, making it suitable for online diagnostic applications and allowing for flexible control of the receptive field size.

[0031] For example, for dilated causal convolution, suppose the input preprocessed power signal sequence is... The convolution kernel is The expansion factor is Then dilation convolution at position The output is: Among them, the expansion factor increases exponentially with the number of network layers. This ensures that the perception field expands rapidly with network depth.

[0032] in, Given an input sequence with dimension , ; The kernel is a convolution kernel with a length of . ; It is the expansion factor, which increases exponentially with the number of network layers; The position index of the output sequence; This is the dilation convolution operator.

[0033] For example, for the residual connection module, to solve the gradient vanishing problem in deep networks, each TCN residual block contains two dilated convolutions and outputs them through residual connections to avoid gradient vanishing, which can be represented as: in, This represents the transformation function within the residual block, including weight normalization, ReLU activation, and dropout regularization; This is the output of the residual block; This is the input for the residual block; This is the nonlinear transformation function within the residual block; The activation function can be ReLU.

[0034] For example, for multi-level feature extraction, the TCN network automatically extracts features by stacking multiple residual blocks. The TCN network learns three distinct levels of features with clearly defined physical meanings. The TCN network outputs feature vectors at these three levels: Shallow features Extracted from the first few layers of the network, with a small receptive field, it mainly captures local fluctuations and transient details of the power curve, corresponding to the transient characteristics of the disconnecting switch in the initial contact stage, such as impact and bounce characteristics. Mid-layer features Extracted from the middle layer of the network, with a moderate receptive field, it mainly extracts trend features at medium time scales, reflecting the stability and consistency of the mechanical transmission process. Deep features Extracted from the last few layers of the network, the receptive field coverage encodes the global temporal dependency relationship, representing the dynamic behavior of the complete closing process; The final TCN output time-series features are a concatenation of the three features: .

[0035] In this embodiment of the application, the multi-domain auxiliary features described in S200 include: The time-domain statistical characteristics include: the mean, standard deviation, skewness, and kurtosis of the power signal, which are used to reflect the overall distribution pattern of the signal; Specifically, the mean can be expressed as: ; Standard deviation can be expressed as: ; Skewness can be expressed as: ; Kurtosis can be expressed as: ; in, The mean of the power signal sequence; The standard deviation of the power signal sequence; The skewness of a power signal sequence is mainly used to measure the asymmetry of the power signal distribution. Kurtosis is the kurtosis of a power signal sequence, primarily used to measure the sharpness of the power signal distribution.

[0036] In an optional implementation, the time-domain statistical characteristics may further include a peak factor, i.e., the ratio of the signal peak value to the effective value; this reflects the impulse characteristics of the signal and can be used to identify abnormal impulses at the moment of contact action. Additionally, it may include dimensionless indicators such as a pulse factor and a margin factor, which are highly sensitive to local signal abrupt changes caused by early faults.

[0037] The frequency domain energy characteristics are obtained by decomposing the preprocessed power signal into multiple frequency bands and calculating the wavelet energy ratio of each frequency band, which is used to characterize the contribution of different frequency components to the signal energy. Specifically, the signal can be directly decomposed into different frequency bands using Discrete Wavelet Transform (DWT) in the preprocessing step, and the wavelet energy and energy proportion of each frequency band can be calculated to characterize the contribution of different frequency components. In extracting multi-scale frequency domain energy features, the following steps are taken: For scale The wavelet energy is expressed as: The energy percentage of each frequency band is calculated and expressed as follows: in, For scale The wavelet energy below; For scale Translation The wavelet coefficients at that point.

[0038] The nonlinear dynamic characteristics are obtained by calculating the entropy of the power signal to obtain complexity characteristics, which are used to capture random fluctuations in the signal. In one alternative implementation, the complexity characteristics obtained by calculating the sample entropy or approximate entropy of the power signal are used to quantify the regularity and complexity of the power signal sequence, making it sensitive to random fluctuations caused by poor lubrication, loose components, etc.

[0039] For example, the approximate entropy can be expressed as: Sample entropy can be expressed as: in, Dimensional embedding dimension; For similarity tolerance; , These represent the number of matching sequence pairs.

[0040] It should be noted that S100 and S200 combine physical mechanisms with data-driven approaches, enabling them to automatically learn global macroscopic features related to faults and capture microscopic transient impact features of rigid separation / coupling. Their information extraction depth allows for more accurate and sensitive diagnosis of minor faults such as early jamming and poor lubrication.

[0041] S300: The multi-scale temporal features and the multi-domain auxiliary features are weighted and fused using an attention mechanism to generate fused features; S400: Input the fused features into the multi-task diagnostic network and output the fault category prediction result and the fault severity estimation result.

[0042] In this embodiment of the application, the multi-scale feature extraction network and the multi-task diagnostic network in S300 and S400 are trained by jointly optimizing the classification loss and regression loss using a training set containing real power signal samples and synthetic power signal samples generated by a conditional generative adversarial network.

[0043] In this embodiment of the application, the construction of the training set and the training of the network described in S300 and S400 include the following steps B1-B4: B1: Obtain actual active power signal samples of the disconnecting switch when it is operating under different mechanical conditions, and label them with fault category tags; B2: Using random noise and the corresponding fault category label as conditions, a generator of a conditional generative adversarial network generates simulated synthetic power signal samples, which are then distinguished as true or false by a discriminator. Through adversarial training, the synthetic power signal samples generated by the generator conform to the characteristic distribution of the power curve under the corresponding fault category. Specifically, the generators mentioned above (B1-B2) Random noise that follows a standard normal distribution and fault category labels As input, the output is a synthesized denoised power curve, represented as: The generator adopts an encoder-decoder structure, and the intermediate layer introduces an attention mechanism to focus on key time segments such as the points of separation / merging.

[0044] Discriminator Receive real or generated power curves and corresponding tags Output the probability that the sample is a true sample, expressed as: Training CGAN is a minimax game process, and the objective function can be expressed as: in, Indicates a generator; This represents the discriminator network; This represents a sample of the actual power curve; This represents a random noise vector, which typically follows a normal distribution. Indicates the fault category label; Represents the actual data distribution; Indicates noise distribution; This is the expectation operator.

[0045] It should be noted that through adversarial training, the generator learns to generate power curves that are realistic and conform to the characteristics of specific fault categories, thereby effectively expanding the minority class samples.

[0046] B3: The synthesized power signal samples that have undergone quality assessment are combined with the real active power signal samples to form a balanced training set to compensate for the data imbalance caused by the scarcity of fault samples. In one alternative implementation, quality assessment, to ensure the usability of generated samples, can be conducted by comprehensively employing Fréchet distance (FID), classification consistency, and physical rationality constraints. Multidimensional evaluation metrics are used to ensure the usability of generated samples; Fréchet distance measures the difference between the actual and generated distributions; classification consistency verifies the proportion of generated samples correctly identified by the classifier; and physical rationality constraints, such as ensuring that the generated curve does not violate fundamental physical laws like energy conservation, ensure that the generated power curve conforms to the dynamic constraints of the disconnector switch.

[0047] B4: Based on the balanced training set, the parameters of the multi-scale feature extraction network and the multi-task diagnostic network are jointly optimized using the weighted sum of the cross-entropy loss for fault classification and the mean squared error loss for fault severity regression as the total loss.

[0048] Specifically, for the joint optimization objective, the training loss function of the entire model is a weighted sum of the classification loss and the regression loss, with an L2 regularization term added to suppress overfitting. in, Represents the total loss function; Represents the cross-entropy loss for a classification task; This represents the mean squared error loss of the regression task; , Represents the weighting coefficients for each type of loss; Represents the set of all parameters of the model; This represents the L2 regularization term.

[0049] It should be noted that by introducing conditional generative adversarial networks, realistic samples of rare fault types can be generated in a targeted manner, solving the data imbalance problem that is common in the field of fault diagnosis. This allows the model to fully learn the distribution characteristics of various faults during training, thereby improving the model's diagnostic accuracy and generalization ability on minority faults and avoiding missed reports due to insufficient data.

[0050] In this embodiment of the application, the step S300, which uses an attention mechanism to perform weighted fusion of the multi-scale temporal features and the multi-domain auxiliary features to generate fused features, includes the following steps C1-C2: C1: Assign a learnable weight vector to each feature vector to be fused, and calculate the attention weights corresponding to each feature vector through a normalized exponential function; C2: The fusion feature is obtained by weighting and summing all feature vectors according to the attention weights, so that the fusion feature dynamically focuses on the feature dimension that contributes the most to the current mechanical state of the disconnector.

[0051] Specifically, to adaptively fuse multi-scale features from TCN and multi-domain auxiliary features from traditional methods, an attention mechanism is designed in C1-C2, which can be represented as: Attention weights through The calculation can be expressed as: in, Indicates the first Attention weights for each feature, through The function is calculated so that the model can dynamically focus on the feature that contributes the most based on different inputs; Indicates the first The weight vector of each feature; Indicates the first 1 eigenvector; This represents the total number of features.

[0052] In this embodiment of the application, the multi-task diagnostic network described in S400 includes: a shared feature processing layer, a classification branch, and a regression branch; The classification branch takes the fusion feature as input and outputs the probability vector of the disconnect switch belonging to each preset fault category through a normalized exponential function, which serves as the fault category prediction result. The regression branch takes the fusion feature as input and outputs a value within a preset continuous interval through the S-shaped activation function, which serves as the fault severity estimation result characterizing the degree of fault deterioration.

[0053] Specifically, a multi-task learning framework with a shared underlying feature extraction layer is constructed, which simultaneously outputs the results of two tasks: fault classification and fault severity estimation.

[0054] Fault classification can use a softmax classifier to output the probability vector of a sample belonging to each fault category. The fault severity regression task uses a fully connected layer with a sigmoid activation function to output a continuous value in the interval [0,1], representing the continuous value of the fault severity. in, This is the sigmoid function, which restricts the output to the interval [0,1].

[0055] It should be noted that the multi-task learning framework can simultaneously output fault category and fault severity. Compared with traditional methods that can only provide binary or limited classification results of normal or abnormal, the continuous degree quantitative information provided by this solution can more precisely characterize the degradation process of equipment, providing a more direct basis for condition-based maintenance and precise maintenance decisions.

[0056] Overall, it can achieve end-to-end automated processing from signal acquisition and preprocessing, feature extraction to fault diagnosis; the model can learn autonomously without relying on expert experience to set thresholds, and can quickly adapt through continuous learning or transfer learning after equipment aging or changes in operating conditions, thus having better robustness and practicality.

[0057] Example 3 illustrates a schematic scheme for a method of multi-scale power signal feature fusion and fault detection for disconnecting switches. It should be noted that the technical solution of this system for multi-scale power signal feature fusion and fault detection for disconnecting switches is based on the same concept as the technical solution of the aforementioned method for multi-scale power signal feature fusion and fault detection for disconnecting switches. Details not described in detail in this embodiment can be found in the description of the aforementioned method for multi-scale power signal feature fusion and fault detection for disconnecting switches.

[0058] This embodiment provides a multi-scale power signal feature fusion and fault detection system for disconnecting switches, including: The acquisition module is used to acquire the power signal of the disconnector switch operating mechanism and preprocess it to obtain a preprocessed power signal; The feature extraction module is used to input the preprocessed power signal into a pre-trained multi-scale feature extraction network to obtain multi-scale temporal features, and to extract multi-domain auxiliary features from the preprocessed power signal. The multi-domain auxiliary features include at least time-domain statistical features, frequency-domain energy features, and nonlinear dynamic features. The fusion module is used to perform weighted fusion of the multi-scale temporal features and the multi-domain auxiliary features using an attention mechanism to generate fused features; The fault output module is used to input the fused features into the multi-task diagnostic network and output the fault category prediction result and the fault severity estimation result.

[0059] This embodiment also provides a computer device applicable to the case of multi-scale power signal feature fusion and fault detection of disconnect switches, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the method for multi-scale power signal feature fusion and fault detection of disconnect switches as proposed in the above embodiment.

[0060] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements a method for multi-scale power signal feature fusion and fault detection of disconnect switches as proposed in the above embodiments.

[0061] The storage medium proposed in this embodiment and the method for implementing multi-scale power signal feature fusion and fault detection of an isolating switch proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0062] Based on the above description of the implementation methods, those skilled in the art will clearly understand that the present invention can be implemented using software and necessary general-purpose hardware. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0063] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for multi-scale power signal feature fusion and fault detection of disconnecting switches, characterized in that, include: The power signal of the disconnector switch operating mechanism is acquired and preprocessed to obtain a preprocessed power signal; The preprocessed power signal is input into a pre-trained multi-scale feature extraction network to obtain multi-scale temporal features, and multi-domain auxiliary features are extracted from the preprocessed power signal. The multi-domain auxiliary features include at least time-domain statistical features, frequency-domain energy features, and nonlinear dynamic features. The multi-scale temporal features and the multi-domain auxiliary features are weighted and fused using an attention mechanism to generate fused features; The fused features are input into a multi-task diagnostic network, which outputs fault category prediction results and fault severity estimation results.

2. The method for multi-scale power signal feature fusion and fault detection of disconnecting switches as described in claim 1, characterized in that, The multi-scale feature extraction network and the multi-task diagnostic network are trained using a training set containing real power signal samples and synthetic power signal samples generated by a conditional generative adversarial network, by jointly optimizing the classification loss and regression loss.

3. The method for multi-scale power signal feature fusion and fault detection of disconnecting switches as described in claim 2, characterized in that, The preprocessed power signal is input into a pre-trained multi-scale feature extraction network to obtain multi-scale temporal features, including: The preprocessed power signal is input into a multi-scale feature extraction network; By using multiple residual blocks containing dilated causal convolutions and residual connections, shallow, mid-level, and deep features with different receptive fields are extracted at multiple network depths. The shallow, middle, and deep features respectively characterize the local details of the contact action transient, the trend changes in the transmission process, and the global dynamic characteristics of the entire operation process. The features from different layers are spliced ​​together to form the multi-scale temporal features.

4. The method for multi-scale power signal feature fusion and fault detection of disconnecting switches as described in claim 2 or 3, characterized in that, In the multi-domain auxiliary features: The time-domain statistical characteristics include: the mean, standard deviation, skewness, and kurtosis of the power signal, which are used to reflect the overall distribution pattern of the signal; The frequency domain energy characteristics are obtained by decomposing the preprocessed power signal into multiple frequency bands and calculating the wavelet energy ratio of each frequency band, which is used to characterize the contribution of different frequency components to the signal energy. The nonlinear dynamic characteristics are obtained by calculating the entropy of the power signal to obtain complexity characteristics, which are used to capture random fluctuations in the signal.

5. The method for multi-scale power signal feature fusion and fault detection of disconnecting switches as described in claim 4, characterized in that, The step of using an attention mechanism to weightedly fuse the multi-scale temporal features and the multi-domain auxiliary features to generate fused features includes: A learnable weight vector is assigned to each feature vector to be fused, and the attention weights corresponding to each feature vector are calculated by a normalized exponential function. The fused features are obtained by weighting and summing all feature vectors according to the attention weights, so that the fused features dynamically focus on the feature dimension that contributes the most to the determination of the current mechanical state of the disconnector switch.

6. The method for multi-scale power signal feature fusion and fault detection of disconnecting switches as described in claim 5, characterized in that, The multi-task diagnostic network includes: a shared feature processing layer, a classification branch, and a regression branch; The classification branch takes the fusion feature as input and outputs the probability vector of the disconnect switch belonging to each preset fault category through a normalized exponential function, which serves as the fault category prediction result. The regression branch takes the fusion feature as input and outputs a value within a preset continuous interval through the S-shaped activation function, which serves as the fault severity estimation result characterizing the degree of fault deterioration.

7. A method for multi-scale power signal feature fusion and fault detection of disconnecting switches as described in claim 2 or 6, characterized in that, The construction of the training set and the training of the network include: Acquire samples of the actual active power signals of the disconnecting switch under different mechanical conditions and label them with fault category tags; Using random noise and corresponding fault category labels as conditions, a generator of a conditional generative adversarial network generates simulated synthetic power signal samples, which are then distinguished as genuine or fake by a discriminator. Through adversarial training, the synthetic power signal samples generated by the generator conform to the characteristic distribution of the power curve under the corresponding fault category. The synthesized power signal samples that have undergone quality assessment are combined with the real active power signal samples to form a balanced training set in order to compensate for the data imbalance caused by the scarcity of fault samples. Based on the balanced training set, the parameters of the multi-scale feature extraction network and the multi-task diagnostic network are jointly optimized using the weighted sum of the cross-entropy loss for fault classification and the mean squared error loss for fault severity regression as the total loss.

8. A multi-scale power signal feature fusion and fault detection system for disconnecting switches, using the method described in any one of claims 1-7, characterized in that, include: The acquisition module is used to acquire the power signal of the disconnector switch operating mechanism and preprocess it to obtain a preprocessed power signal; The feature extraction module is used to input the preprocessed power signal into a pre-trained multi-scale feature extraction network to obtain multi-scale temporal features, and to extract multi-domain auxiliary features from the preprocessed power signal. The multi-domain auxiliary features include at least time-domain statistical features, frequency-domain energy features, and nonlinear dynamic features. The fusion module is used to perform weighted fusion of the multi-scale temporal features and the multi-domain auxiliary features using an attention mechanism to generate fused features; The fault output module is used to input the fused features into the multi-task diagnostic network and output the fault category prediction result and the fault severity estimation result.

9. A computer device, characterized in that, include: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, they implement the steps of the method for multi-scale power signal feature fusion and fault detection of disconnecting switches as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores computer-executable instructions, which, when executed by a processor, implement the steps of the method for multi-scale power signal feature fusion and fault detection of a disconnecting switch as described in any one of claims 1 to 7.

Images