Fault Identification Method and System Based on PSO Dual Frequency Domain Weighted Time-Frequency Analysis

Abstract

This invention discloses a fault identification method and system based on PSO (Programmable Array Sort) dual-frequency domain weighted time-frequency analysis, comprising the following steps: acquiring the transient wave signal after a fault and performing continuous wavelet transform (CWT) and fast Fourier transform (FFT) on it respectively, then inputting it into a trained PSO-CNN joint iterative optimization model, and outputting the fault type. This invention provides a time-frequency analysis method that adaptively senses the distribution of fault features, effectively suppresses background noise, and significantly improves the energy concentration of fault features, thus meeting the needs for accurate capture of weak faults in the actual operation of distribution networks in new power systems.

Description

Technical Field

[0001] This invention relates to the field of power system technology, and specifically to a fault identification method and system based on PSO dual frequency domain weighted time-frequency analysis. Background Technology

[0002] In new power systems, the distribution network structure is becoming increasingly complex, with numerous branch lines, resulting in fault traveling wave signals often accompanied by strong background noise and various interferences. When a single-phase ground fault occurs in the distribution network, especially a high-resistance ground fault, the resulting traveling wave signal is extremely weak, only at the millivolt level, and is easily submerged by random noise in the line and the boundary effects of the signal processing algorithm itself, resulting in a "feature masking" phenomenon.

[0003] Existing methods for fault identification in distribution networks mainly rely on the traveling wave method, which utilizes time-frequency transformation technology to construct a panoramic view of the fault transient state. While the Fast Fourier Transform (FFT) can accurately capture the global frequency distribution of the signal, as a one-dimensional signal representation, it lacks time dimension information and cannot construct a time-frequency correspondence. Although the Continuous Wavelet Transform (CWT) establishes a good time-frequency mapping relationship, in low signal-to-noise ratio environments, its "magnifying glass" effect amplifies both background noise and fault features simultaneously, and it is prone to generating high-energy boundary artifacts in the low-frequency band, resulting in blurred effective features in the time-frequency image.

[0004] Simply superimposing FFT and CWT linearly cannot solve the aforementioned problems of physical dimension misalignment and noise accumulation. Furthermore, different types of faults (such as A-phase, B-phase, and C-phase grounding) and disturbances exhibit different energy distributions in the frequency domain, making it difficult for traditional methods to manually set uniform values ​​to adapt to all fault scenarios. Therefore, there is an urgent need for a time-frequency analysis method that can adaptively sense fault feature distribution, effectively suppress background noise, and significantly improve the energy concentration of fault features to meet the requirements for accurate capture of weak faults in the actual operation of new power system distribution networks. Summary of the Invention

[0005] In view of this, the present invention provides a fault identification method and system based on PSO dual frequency domain weighted time-frequency analysis, which can at least solve the problem that the distribution network of new power system is difficult to capture weak faults in the prior art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: The fault identification method based on PSO dual frequency domain weighted time-frequency analysis includes the following steps: The transient waveform signal after the fault is acquired, and after performing continuous wavelet transform (CWT) and fast fourier transform (FFT) respectively, it is input into a trained PSO-CNN joint iterative optimization model to output the fault type. The training process includes: S1: Acquire transient waveform signal after fault Samples, establish a database; S2: Will The time-frequency matrix generated by performing CWT and FFT respectively and global spectrum ; S3: In the PSO-CNN joint iterative optimization model, the PSO algorithm is used to calculate... and The optimal weights for the time-frequency images are obtained by weighted fusion. and The input is fed into the CNN network for forward propagation and fast training. After the fast training of the CNN network converges, it outputs the fitness function value as the classification accuracy and the corresponding fault type judgment result. S4: Based on the fitness function value, the PSO algorithm determines whether it has converged. If not, the PSO algorithm updates the current optimal weight and repeats S3 until the PSO algorithm converges. If it has converged, the iteration ends, the global optimal weight is output, and the training is completed.

[0007] Preferably, the specific content of S1 includes: For any distribution network, a corresponding simulation model is built, with n fault points set in each branch, and different fault conditions are set for each fault point; After a fault occurs, the transient signals extracted by the detection devices at each end are collected, and then the linear modulus components are obtained; Arbitrary linear mode components are selected and noise is injected to form a transient wave signal. ,by Centered on the point of abrupt change in amplitude, data within any preset time window before and after the time of the fault occurrence are extracted as input samples for a single analysis. Repeatedly collect input samples to build a database.

[0008] Preferably, The specific content of performing Continuous Wavelet Transform (CWT) includes: Transient wave signal Record : ; in, For the actual fault traveling wave component, It is Gaussian white noise that follows a normal distribution. This refers to the non-periodic DC attenuation and boundary effect components. For time; right Perform continuous wavelet transform to obtain : ; in, The scaling parameters for CWT, For the translation parameters of CWT, Characterized at time t, after scaling transformation of scale a and translation to The wavelet basis function after the position; the integral process represents the scale parameter. With translation parameters The degree of local correlation between the wavelet basis and the original signal x(t).

[0009] Preferably, The specific content of performing the Fast Fourier Transform (FFT) includes: For transient wave signals The corresponding one-dimensional time-domain discrete sampling signal sequence Perform a Fast Fourier Transform to obtain the frequency domain discrete spectrum sequence after the transformation. : ; ; in, and These are the frequency domain index and the time domain index, respectively. N is the total number of sampling points within the sampling window, where N = sampling frequency × sampling window length. Let be the rotation factor of the discrete Fourier transform, and represent the root of unit in the complex plane. The imaginary unit; Mapping the frequency domain of the FFT to the scale domain of the CWT yields the mapped global spectrum. : ; in, It is a linear interpolation mapping operator. The scaling parameter for CWT.

[0010] Preferably, the specific content of the time-frequency image obtained by weighted fusion includes: Time-frequency image obtained by weighted fusion for: ; in, scale parameter Translation parameters The time-frequency matrix below, for Feature enhancement weights, scale parameter The global spectrum below, for Global frequency weights; For the dimension expansion operation, a one-dimensional vector is... Copy to extend with The same two-dimensional scale.

[0011] Preferably, calculation is performed using the PSO algorithm. and The optimal weights for the time-frequency images are obtained by weighted fusion. and The specific content includes: Each particle is defined with a feature weighting vector. : ; in, For the first Each particle contains a complete spatial position vector. The total number of scale decompositions. =1,2,3…… , For scale parameters; Will be via weight vector The reconstructed time-frequency image is denoted as ,Will Input the CNN network for forward propagation and fast training, and evaluate the function according to the deep convolutional network. Perform fault classification and assessment; Calculate the fitness function value : ; in, Let i be the fitness score of the current weight strategy for the i-th particle. This indicates the use of the weight vector of the i-th particle. The reconstructed double-weighted time-frequency image, This represents the evaluation function for deep convolutional networks.

[0012] Preferably, the specific content of the PSO algorithm updating the current optimal weight includes: A global search is performed in the N-dimensional feature space, and the particle update method is as follows: ; ; in, The first The particle in the first Second and third Optimization speed in +1 iterations; The weighting parameter controls the tendency of particles to maintain their original motion state; and Individual learning factors and social learning factors control the step size by which particles approach their own optimal value and the group's optimal value. and The random numbers are uniformly distributed between [0,1], which increases the randomness of the search and the ability to escape local optima; For the first The optimal position of an individual particle that reaches the highest fitness function value in historical iterations; and Don't be the first The particle in the first Second and third The feature weighted vector in +1 iterations; This represents the globally optimal position where the entire particle population achieves the highest fitness function value in historical iterations.

[0013] The fault identification system based on PSO dual frequency domain weighted time-frequency analysis includes a processor, a memory, and a computer program stored in the memory and executable by the processor. When the processor runs the computer program, it implements the above-mentioned fault identification method based on PSO dual frequency domain weighted time-frequency analysis.

[0014] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a fault identification method and system based on PSO dual frequency domain weighted time-frequency analysis, which has the following beneficial effects: 1. This invention solves the problem of weak signals being masked by noise and improves the signal-to-noise ratio: This invention introduces FFT as a global frequency gate and uses the PSO algorithm to optimize and suppress the spectral leakage effect in CWT, which significantly improves the feature saliency of weak fault panoramic waves in low signal-to-noise ratio environments.

[0015] 2. This invention can improve energy concentration and enhance recognition accuracy: Through adaptive dual weighting, time-frequency energy is highly concentrated from the diffuse state to the local time-frequency region where the fault features are located, which greatly reduces the entropy value of the time-frequency image, enabling the back-end CNN to easily capture the subtle texture differences of different fault types (A / B / C phases) and distinguish between normal disturbance signals and fault signals. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. 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.

[0017] Figure 1 The spectral leakage effect diagram provided for the embodiments of the present invention is shown in the figure. Figure 1 (a) shows the spectral leakage effect before the introduction of FFT. Figure 1 (b) shows the effect after FFT suppression; Figure 2 The flowchart of PSO-CNN joint iterative optimization in the fault identification method based on dual frequency domain weighted time-frequency analysis provided by the present invention is shown. Figure 3 The flowchart of the fault location method based on PSO dual frequency domain weighted time-frequency analysis provided by the present invention is shown below. Figure 4 This is a comparison chart of the PSO-CNN joint iterative optimization performance provided in an embodiment of the present invention, wherein... Figure 4 (a) shows the effect of pure CWT transformation for AG faults; Figure 4 (b) is a diagram showing the effect of AG fault PSO iteration; Figure 5 This is a diagram of the IEEE 14-node distribution network topology provided in an embodiment of the present invention. Figure 6 The accuracy rates corresponding to different initial angles and different transition resistances provided in the embodiments of the present invention; Figure 7 This is a fault clustering diagram with an average accuracy of 95% provided in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] This invention provides a fault identification method based on PSO dual frequency domain weighted time-frequency analysis, comprising the following steps: The transient waveform signal after the fault is acquired, and after performing continuous wavelet transform (CWT) and fast fourier transform (FFT) respectively, it is input into a trained PSO-CNN joint iterative optimization model to output the fault type. The training process includes: S1: Acquire transient waveform signal after fault Samples, establish a database; S2: Will The time-frequency matrix generated by performing CWT and FFT respectively and global spectrum ; S3: In the PSO-CNN joint iterative optimization model, the PSO algorithm is used to calculate... and The optimal weights for the time-frequency images are obtained by weighted fusion. and The input is fed into the CNN network for forward propagation and fast training. After the fast training of the CNN network converges, it outputs the fitness function value as the classification accuracy and the corresponding fault type judgment result. S4: Based on the fitness function value, the PSO algorithm determines whether it has converged. If not, the PSO algorithm updates the current optimal weight and repeats S3 until the PSO algorithm converges. If it has converged, the iteration ends, the global optimal weight is output, and the training is completed.

[0020] To further implement the above technical solution, the specific content of S1 includes: For any distribution network, a corresponding simulation model is built, with n fault points set in each branch, and different fault conditions are set for each fault point; After a fault occurs, the transient signals extracted by the detection devices at each end are collected, and then the linear modulus components are obtained; Arbitrary linear mode components are selected and noise is injected to form a transient wave signal. ,by Centered on the point of abrupt change in amplitude, data within any preset time window before and after the time of the fault occurrence are extracted as input samples for a single analysis. Repeatedly collect input samples to build a database.

[0021] It should be noted that: For a specific distribution network, a simulation model with identical line parameters and topology was built in PSCAD / EMTDC to construct a 10kV radial distribution network model. To improve the generalization ability of the trained model, multiple fault points were set for each branch, and a database was established. Furthermore, each fault point could be configured with multiple fault scenarios. For example, for each fault point, three fault types—A-phase grounding, B-phase grounding, and C-phase grounding—were simulated. Six different initial phase angles were configured: 1.5°, 30°, 60°, 90°, 180°, and 270°; and seven different transition resistances were configured: 0Ω, 100Ω, 200Ω, 500Ω, 1000Ω, 2000Ω, and 2500Ω. In practical applications, fault scenarios can be set according to requirements.

[0022] In this embodiment, the linear mode components are obtained through Kelvin transform: ; In the formula, , and These are the transient signals of phases A, B, and C, respectively. The transformed zero-mode component, and These are the two line mode components after transformation; in this embodiment, components with strong anti-attenuation capability and stable wave velocity are selected. The transient wave signal Ewave is used as the input sequence for subsequent time-frequency analysis. The preset time window is set to 400 sampling points.

[0023] To further implement the above technical solution, The specific content of performing Continuous Wavelet Transform (CWT) includes: Transient wave signal Record : ; in, For the actual fault traveling wave component, It is Gaussian white noise that follows a normal distribution. This refers to the non-periodic DC attenuation and boundary effect components. For time; right Perform continuous wavelet transform to obtain : ; in, The scaling parameters for CWT, For the translation parameters of CWT, Characterized at time t, after scaling transformation of scale a and translation to The wavelet basis function after the position; the integral process represents the scale parameter. With translation parameters The degree of local correlation between the wavelet basis and the original signal x(t).

[0024] It should be noted that: The coefficient matrix obtained after continuous wavelet transform The corresponding decomposition terms are: ; in, : Effective wavelet coefficients corresponding to the actual fault traveling wave; : Wavelet coefficients corresponding to broadband random noise; : Corresponds to the high-energy artifact coefficients caused by boundary effects and spectral leakage; For trend items When the scale 'a' is large, i.e., at low frequencies, the wavelet basis... The support interval becomes very wide, even exceeding the observation window. This leads to boundary effects, which are essentially caused by the divergence of wavelet coefficients due to the sudden truncation of data. At the moment of high-resistivity fault, the traveling wave undergoes a drastic change. When wavelet bases deal with such drastic changes, spectral leakage can resemble edge effects.

[0025] WE is defined as wavelet entropy, which is a metric for measuring the complexity of the energy distribution of a signal in the time-frequency domain. ; Where is the relative energy at the j-th scale. This represents the energy distribution, where K is the total number of scale layers divided by the wavelet transform. WE This represents the wavelet entropy value. Distribution. The more uniform the surface, the more dispersed the energy, leading to spectral leakage. WE The larger the value, the more concentrated the energy is at a certain local high-frequency scale (ideal fault characteristics). WE The smaller the value, the better. Ideally, the high-frequency scale should contain the vast majority of the energy. However, due to the spectral leakage characteristics of the CWT, a large amount of energy is incorrectly projected onto the low-frequency scale, and this leakage affects the energy distribution. The distribution becomes flatter, rather than concentrated on a few high-frequency scales. According to the definition of entropy, the more uniform the distribution, the greater the entropy value. Therefore, a fault signal that should have obvious characteristics (low entropy) may be calculated to have a high entropy value due to class boundary effects, i.e., an artifact.

[0026] When analyzing traveling waves of weak faults in distribution networks, traditional continuous wavelet transform (CWT) is prone to spectral leakage and boundary effects. From the physical perspective of information distribution, the ideal transient characteristics of a fault should exhibit a high concentration of energy at local high-frequency scales, corresponding to extremely low wavelet entropy. However, due to the unreliability of traditional CWT in handling transient changes in high-resistivity faults, a large amount of energy is incorrectly projected to low-frequency scales, leading to an abnormally high wavelet entropy in the low-frequency region, exhibiting an energy dispersion state. This high-entropy boundary artifact severely masks the originally weak low-entropy fault characteristics, resulting in a 'feature masking' phenomenon.

[0027] Furthermore, due to the highly variable transition resistance and initial phase angle of faults in actual distribution networks, the distribution of these high-entropy artifacts exhibits a high degree of dynamic randomness. Traditional methods that attempt to forcibly filter out artifacts by manually calculating and setting a fixed 'wavelet entropy truncation threshold' are not only difficult to tune parameters for, but also prone to the accidental deletion of valid fault features or the omission of strong noise.

[0028] Given the inherent defects of traditional fixed threshold schemes, the scheme disclosed in this application abandons the rigid step of 'manual judgment and artifact removal' and directly introduces a dual frequency domain weighting and PSO-CNN closed-loop joint iterative architecture, transforming the task of overcoming the high-entropy spectrum leakage effect into an adaptive weight dimensionality reduction optimization process driven by deep learning accuracy.

[0029] To further implement the above technical solution, The specific content of performing the Fast Fourier Transform (FFT) includes: For transient wave signals The corresponding one-dimensional time-domain discrete sampling signal sequence Perform a Fast Fourier Transform to obtain the frequency domain discrete spectrum sequence after the transformation. : ; ; in, and These are the frequency domain index and the time domain index, respectively. N is the total number of sampling points within the sampling window, where N = sampling frequency × sampling window length. Let be the rotation factor of the discrete Fourier transform, and represent the root of unit in the complex plane. The imaginary unit; Mapping the frequency domain of the FFT to the scale domain of the CWT yields the mapped global spectrum. : ; in, It is a linear interpolation mapping operator. The scaling parameter for CWT.

[0030] It should be noted that: To address the spectral leakage issue in CWT, the CWT transformation is performed simultaneously with... Perform FFT transformation.

[0031] For the conversion formula based on center frequency Nonlinear interpolation mapping operator. This is a one-dimensional vector after mapping and alignment. It represents the true global frequency domain energy level of the signal at each wavelet scale 'a'.

[0032] The mapping relationship follows The array indices of FFT and the scale indices of CWT cannot be directly aligned; interpolation using frequency values ​​as an intermediary is required. It is a time-invariant vector representing the global energy level at scale *a*. The FFT can accurately capture the global frequency distribution and suppress spectral leakage to some extent. However, its output cannot directly modify the CWT result; a medium is needed. Therefore, PSO (Particle Swarm Optimization) is introduced to weaken the spectral leakage effect through weight control. The weakening effect is shown in the figure below. Figure 1 As shown.

[0033] To further implement the above technical solution, the specific content of the time-frequency image obtained by weighted fusion includes: Time-frequency image obtained by weighted fusion for: ; in, scale parameter Translation parameters The time-frequency matrix below, for Feature enhancement weights, scale parameter The global spectrum below, for Global frequency weights; For the dimension expansion operation, a one-dimensional vector is... Copy to extend with The same two-dimensional scale.

[0034] It should be noted that: This is the final two-dimensional time-frequency feature matrix reconstructed after weighted fusion (i.e., the image input to the CNN). The feature enhancement weights for the continuous wavelet transform (CWT) to be optimized at scale a are given. The global frequency weights of the Fast Fourier Transform (FFT) to be optimized at scale a. For the dimension expansion operation, a one-dimensional vector is... Copy to extend with The same two-dimensional scale.

[0035] When 'a' is in the low-frequency artifact region: It's huge (an illusion), but Smaller, by making Optimization can achieve mathematical removal of artifacts. When 'a' is in the noise region, let... This can achieve noise reduction. When 'a' is in the fault characteristic region, the wavelet entropy WE shows a significantly low value, making... Enlarging the size enhances the features.

[0036] To further implement the above technical solution, the PSO algorithm is used for calculation. and The optimal weights for the time-frequency images are obtained by weighted fusion. and The specific content includes: Each particle is defined with a feature weighting vector. : ; in, For the first Each particle contains a complete spatial position vector. The total number of scale decompositions. =1,2,3…… , For scale parameters; Will be via weight vector The reconstructed time-frequency image is denoted as ,Will Input the CNN network for forward propagation and fast training, and evaluate the function according to the deep convolutional network. Perform fault classification and assessment; Calculate the fitness function value : ; in, Let i be the fitness score of the current weight strategy for the i-th particle. This indicates the use of the weight vector of the i-th particle. The reconstructed double-weighted time-frequency image, This represents the evaluation function for deep convolutional networks.

[0037] It should be noted that: The specific meaning of this operation is: to reconstruct the image. The input sample is forward-propagated through the constructed LeNet-5 network, and the network outputs the fault classification on the test / validation set.

[0038] It should be noted that: In actual power distribution networks, the artifact distribution and characteristic frequency bands corresponding to different transition resistances, initial phase angles, and fault types (such as A / B / C phase grounding) are dynamically changing and cannot be fixed by manual presets. and Parameters. To implement the above ideal feature reconstruction model in engineering, it is necessary to find the optimal parameters across the entire frequency band. and In the combination process, wavelet transform (CWT) divides the frequency into 64 scales, each scale corresponding to a specific frequency. , There are a total of 128 unknown weight parameters. Moreover, the impact of these 128 parameters on the final classification accuracy is not a simple linear relationship, but rather exhibits multiple local values, transforming it into a parameter optimization problem in a high-dimensional space.

[0039] The high-dimensional non-convex optimization problem is to find a set of optimal nonlinear weighted strategies in the feature space to maximize the separability of fault types. In the PSO algorithm, each particle represents a potential feature reconstruction strategy.

[0040] in, Let i be the complete spatial position orientation contained in the i-th particle. K This represents the total number of scale decompositions. The vector has a dimension of 2. K It essentially includes the CWT weights corresponding to all scales. With FFT weights The weight vector evolves iteratively during the optimization process and is not pre-defined. A closed-loop feedback mechanism is established to evaluate the effectiveness of the feature reconstruction strategy by defining the objective function as maximizing classification accuracy. This is achieved through the weight vector. The reconstructed time-frequency images are input into the CNN (LeNet-5) for rapid training and evaluation.

[0041] The objective function directly quantifies the ability of the current weights to refine fault features through feedback from the CNN (LeNet-5). Due to the nonlinear nature of the CNN model, the objective function exhibits high nonconvexity in the parameter space, meaning it has a large number of local extrema. The PSO algorithm performs a global search in the N-dimensional feature space through individual cognition and social cooperation among particles.

[0042] To further implement the above technical solution, the PSO algorithm updates the current optimal weights in the following ways: A global search is performed in the N-dimensional feature space, and the particle update method is as follows: ; ; in, and The first The particle in the first Second and third Optimization speed in +1 iterations; The weighting parameter controls the tendency of particles to maintain their original motion state; and Individual learning factors and social learning factors control the step size by which particles approach their own optimal value and the group's optimal value. and The random numbers are uniformly distributed between [0,1], which increases the randomness of the search and the ability to escape local optima; For the first The optimal position of an individual particle that reaches the highest fitness function value in historical iterations; and Don't be the first The particle in the first Second and third The feature weighted vector in +1 iterations; This represents the globally optimal position where the entire particle population achieves the highest fitness function value in historical iterations.

[0043] It should be noted that: Through iteration, the algorithm automatically suppresses frequency band weights containing noisy or redundant information and enhances frequency band weights containing key features of fault traveling waves. Mathematically, this is equivalent to finding the optimal nonlinear projection operator. By adaptively stretching or compressing various dimensions of the feature space, modules such as... Figure 2 As shown, this module can maximize the inter-class distance between different fault types in the reconstructed feature space.

[0044] Throughout the actual model training process, the fault traveling wave sample database established by S1 was randomly divided into training and test sets according to a preset ratio. The original fault data in the training set underwent standardized preprocessing to eliminate the influence of dimensions, and a certain proportion of Gaussian white noise was artificially injected to construct a highly challenging noise environment and enhance the model's feature extraction capabilities.

[0045] In each iteration, the training data input is transformed. Using the weight vector of the current particle, the FFT spectral features and CWT time-frequency features of the original signal are calculated respectively, achieving adaptive attenuation of the noise frequency band and targeted enhancement of the fault feature frequency band. The weighted dual features are then reconstructed into the final time-frequency image M. weighted The CNN network (LeNet-5 is used in this embodiment) is input for fast training, its classification accuracy on the validation set is calculated, and the classification accuracy Acc is used as the fitness function to feed back to the PSO algorithm.

[0046] Based on feedback values, the algorithm uses a velocity-position update formula to drive the particle swarm to move in a high-dimensional non-convex weight space. If an increase in the weight of a certain frequency band leads to an increase in Acc, the particle swarm will gather in that direction (feature enhancement); if it leads to a decrease in Acc, the weight will be suppressed in the opposite direction (noise filtering). During this process, a limiting mechanism is introduced to prevent weight divergence.

[0047] Repeat the above steps until the preset maximum number of iterations or the fitness value converges. At this point, the algorithm stops iterating and outputs the globally optimal dual-feature weighted vector W. best And based on this optimal weight, the CNN fault diagnosis model has been trained, completing the offline training phase. The complete flowchart is as follows: Figure 3 As shown.

[0048] The fault identification system based on PSO dual frequency domain weighted time-frequency analysis includes a processor, a memory, and a computer program stored in the memory and executable by the processor. When the processor runs the computer program, it implements the above-mentioned fault identification method based on PSO dual frequency domain weighted time-frequency analysis.

[0049] This invention addresses the engineering challenges of complex distribution network structures in new power systems, where the traveling wave signals from high-impedance grounding faults are weak and easily masked by strong background noise and algorithm boundary effects. To amplify and highlight the traveling wave signal, a fault identification method and system based on PSO (Programmable Array Signal Processing) with dual frequency domain weighted time-frequency analysis is proposed. To overcome the shortcomings of existing technologies that rely solely on linear superposition, which cannot address the physical dimension misalignment and noise accumulation between FFT (Fault-to-Fault) and CWT (Concurrent-to-Wave) and lead to the simultaneous amplification of background noise and artifacts in low signal-to-noise ratio environments, this invention constructs a signal enhancement model based on dual weighting of the FFT spectrum and CWT scale space. This model establishes an adaptive feature reconstruction mechanism by analyzing the energy distribution differences between the fault traveling wave and background noise in the frequency domain and scale space. Utilizing the global search capability of evolutionary algorithms in the parameter space, it automatically adapts to the spectral characteristics under different fault types and disturbance scenarios. This mechanism can accurately identify and suppress the spectral leakage effect and line random noise generated by CWT, while directionally enhancing the key texture features of weak millivolt-level fault traveling waves, thus solving the "feature masking" problem of weak fault signals in strong interference environments from the source of signal processing.

[0050] To visually demonstrate the feature enhancement effect of the dual frequency domain weighting mechanism of this invention, the time-frequency output results before and after feature reconstruction were extracted and compared, such as... Figure 4 As shown. Figure 4 (a) shows the time-frequency feature map of the AG fault extracted using pure continuous wavelet transform (CWT). It is clearly visible that a large number of high-energy boundary artifacts appear in the low-frequency region, severely masking the truly effective high-frequency transient change features. Figure 4 (b) is the weighted time-frequency effect after PSO-CNN joint iterative optimization of the present invention. Low-frequency artifacts are precisely suppressed by mathematical mechanisms, and time-frequency energy is highly focused at the fault mutation point, presenting an extremely sharp and clear effective feature peak, thus solving the problem of 'feature masking' of weak signals from the source.

[0051] Furthermore, this invention combines a result-oriented evolutionary optimization strategy with a deep learning model capable of deep feature mining, constructing a PSO-CNN closed-loop adaptive iterative optimization architecture. By using the accuracy of the deep learning model in identifying fault types as a feedback indicator, the algorithm automatically searches for the optimal weighting strategy that maximizes the difference between fault features and noise. This closed-loop mechanism not only achieves "automatic focusing" on the effective frequency bands in weak traveling wave signals, but also effectively overcomes the generalization risk caused by incomplete feature extraction in single data-driven models under harsh conditions. It significantly improves the anti-interference capability and identification accuracy of the distribution network fault diagnosis system in extreme scenarios such as high-resistance grounding, meeting the engineering requirements for accurate capture of weak faults in the actual operation of new power systems. To further illustrate this invention, the following is a comprehensive comparison between the proposed dual-frequency domain weighting and PSO-CNN closed-loop adaptive optimization model and traditional machine learning models (SVM and decision trees based on artificial features). A simulation model is built, such as... Figure 5 As shown, more realistic fault signals are obtained and fused to form a sample database.

[0052] As shown in Table 1, the system's identification accuracy for A-phase grounding (AG), B-phase grounding (BG), and C-phase grounding (CG) faults reached 94.1%, 93.3%, and 95%, respectively, with an overall average accuracy of 94.2%.

[0053] Table 1 Recognition performance of different models ; To verify the effectiveness of the model, a complex power distribution network was constructed in PSCAD simulation software. The branches in the diagram consist of overhead lines and cables, with specific lengths determined by… Figure 5 As shown, a fault traveling wave detection device is installed at the network terminal, with the sampling rate set to 1MHz.

[0054] A sample set was constructed by iterating through the different fault parameters shown in Table 2. Waveforms were collected within 400µs after the arrival of the wavefront. A total of 3648 fault samples were generated, of which 80% were used for model training and the remaining 20% ​​were used to test the trained model.

[0055] Table 2 Fault Parameter Traversal Table for Total Sample Set ; The results are shown in Table 3, taking into account the effects of different transition resistances and initial phase angles. Since the intensity of the traveling wave signal is directly proportional to the voltage and inversely proportional to the transition resistance, it can be seen that the accuracy of fault identification is affected by the transition resistance and the initial phase angle. Furthermore, according to the wavelet frequency-scale conversion formula, the faults are concentrated in the scale 0-scale 10, i.e., 400~500kHz.

[0056] Table 3. Fault location results under different transition resistances and initial angles. ; Under high-impedance faults, the traveling wave signal is weak and difficult to judge, and is greatly affected by the initial angle, requiring high speed. In the most extreme case, i.e., 2500Ω and the voltage wave crossing zero, the fault identification accuracy of this method still reaches over 85%. In most other cases, the accuracy is higher than 95%, with an average accuracy of 95%. Simulation verification proves the feasibility of the method. The results are as follows... Figure 6 As shown.

[0057] like Figure 7 As shown, Phase A, Phase B, and Phase C represent single-phase grounding fault samples of three phases A, B, and C, respectively. Among 3648 complex test samples covering transition resistances from 0Ω to an extreme 2500Ω and initial phase angles from near zero crossing 1.5° to a peak value of 270°, fault feature points of the same phase exhibit extremely high intra-class compactness in space, i.e., points of the same color are highly clustered; while fault feature groups of different phases form extremely clear and broad inter-class decision boundaries, with the three feature clusters being independent of each other.

[0058] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A fault identification method based on PSO-based dual-frequency domain weighted time-frequency analysis, characterized in that, Includes the following steps: After acquiring the transient wave signal after the fault, perform continuous wavelet transform (CWT) and fast four-factor transform (FFT) on it respectively, and then input it into the trained PSO-CNN joint iterative optimization model to output the fault type. The training process includes: S1: Acquire transient waveform signal after fault Samples, establish a database; S2: Will The time-frequency matrix generated by performing CWT and FFT respectively and global spectrum ; S3: In the PSO-CNN joint iterative optimization model, the PSO algorithm is used to calculate... and The optimal weights for the time-frequency images are obtained by weighted fusion. and The input is fed into the CNN network for forward propagation and fast training. After the fast training of the CNN network converges, it outputs the fitness function value as the classification accuracy and the corresponding fault type judgment result. S4: Based on the fitness function value, the PSO algorithm determines whether it has converged. If not, the PSO algorithm updates the current optimal weight and repeats S3 until the PSO algorithm converges. If it has converged, the iteration ends, the global optimal weight is output, and the training is completed.

2. The fault identification method based on PSO dual-frequency domain weighted time-frequency analysis according to claim 1, characterized in that, The specific content of S1 includes: For any distribution network, a corresponding simulation model is built, with n fault points set in each branch, and different fault conditions are set for each fault point; After a fault occurs, the transient signals extracted by the detection devices at each end are collected, and then the linear modulus components are obtained; Arbitrary linear mode components are selected and noise is injected to form a transient wave signal. ,by Centered on the point of abrupt change in amplitude, data within any preset time window before and after the time of the fault occurrence are extracted as input samples for a single analysis. Repeatedly collect input samples to build a database.

3. The fault identification method based on PSO dual-frequency domain weighted time-frequency analysis according to claim 1, characterized in that, Will The specific content of performing Continuous Wavelet Transform (CWT) includes: Transient wave signal Record : ； in, For the actual fault traveling wave component, It is Gaussian white noise that follows a normal distribution. This refers to the non-periodic DC attenuation and boundary effect components. For time; right Perform continuous wavelet transform to obtain : ； in, The scaling parameters for CWT, For the translation parameters of CWT, Characterized at time t, after scaling transformation of scale a and translation to The wavelet basis function after the position; the integral process represents the scale parameter. With translation parameters The degree of local correlation between the wavelet basis and the original signal x(t).

4. The fault identification method based on PSO dual-frequency domain weighted time-frequency analysis according to claim 1, characterized in that, Will The specific content of performing the Fast Fourier Transform (FFT) includes: For transient wave signals The corresponding one-dimensional time-domain discrete sampling signal sequence Perform a Fast Fourier Transform to obtain the frequency domain discrete spectrum sequence after the transformation. : ； ； in, and These are the frequency domain index and the time domain index, respectively. N is the total number of sampling points within the sampling window, where N = sampling frequency × sampling window length. Let be the rotation factor of the discrete Fourier transform, and represent the root of unit in the complex plane. The imaginary unit; Mapping the frequency domain of the FFT to the scale domain of the CWT yields the mapped global spectrum. : ； in, It is a linear interpolation mapping operator. The scaling parameter for CWT.

5. The fault identification method based on PSO dual-frequency domain weighted time-frequency analysis according to claim 1, characterized in that, The specific content of the time-frequency image obtained by weighted fusion includes: Time-frequency image obtained by weighted fusion for: ； in, scale parameter Translation parameters The time-frequency matrix below, for Feature enhancement weights, scale parameter The global spectrum below, for Global frequency weights; For the dimension expansion operation, a one-dimensional vector is... Copy to extend with The same two-dimensional scale.

6. The fault identification method based on PSO dual-frequency domain weighted time-frequency analysis according to claim 5, characterized in that, Calculated using the PSO algorithm and The optimal weights for the time-frequency images are obtained by weighted fusion. and The specific content includes: Each particle is defined with a feature weighting vector. : ； in, For the first Each particle contains a complete spatial position vector. The total number of scale decompositions. =1,2,3…… , For scale parameters; Will be via weight vector The reconstructed time-frequency image is denoted as ,Will Input the CNN network for forward propagation and fast training, and evaluate it according to the deep convolutional network evaluation function. Perform fault classification and assessment; Calculate the fitness function value : ； in, Let i be the fitness score of the current weight strategy for the i-th particle. This indicates the use of the weight vector of the i-th particle. The reconstructed double-weighted time-frequency image, This represents the evaluation function for deep convolutional networks.

7. The fault identification method based on PSO dual-frequency domain weighted time-frequency analysis according to claim 6, characterized in that, The PSO algorithm updates the current optimal weights by including: A global search is performed in the N-dimensional feature space, and the particle update method is as follows: ； ； in, The first The particle in the first Second and third Optimization speed in +1 iterations; The weighting parameter controls the tendency of particles to maintain their original motion state; and Individual learning factors and social learning factors control the step size by which particles approach their own optimal value and the group's optimal value. and The random numbers are uniformly distributed between [0,1], which increases the randomness of the search and the ability to escape local optima; For the first The optimal position of an individual particle that reaches the highest fitness function value in historical iterations; and Don't be the first The particle in the first Second and third The feature weighted vector in +1 iterations; This represents the globally optimal position where the entire particle population achieves the highest fitness function value in historical iterations.

8. A fault identification system based on PSO dual-frequency domain weighted time-frequency analysis, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and executable by the processor. When the processor executes the computer program, it implements the fault identification method based on PSO dual frequency domain weighted time-frequency analysis as described in any one of claims 1-7.

Images