Power distribution network multi-equipment fault detection method and device, terminal equipment and storage medium

Abstract

The invention discloses a power distribution network multi-equipment fault detection method and device, terminal equipment and a storage medium, and belongs to the field of power systems. The method comprises the steps of obtaining time sequence data of a plurality of devices of a to-be-detected power distribution network; performing fast Fourier transform and frequency division according to the time sequence data to obtain a plurality of initial frequency patches; according to the initial frequency patch and a preset learning weight matrix, generating a mask under the constraint of a preset clustering loss function and a regularization loss function to obtain a binary mask matrix; performing channel filtering and information fusion processing on the initial frequency patch according to the binary mask matrix to obtain a target frequency patch, and performing frequency domain reconstruction and time domain reconstruction under the constraint of a preset frequency domain loss function and a time domain loss function to obtain a frequency spectrum reconstruction signal; and inputting the frequency spectrum reconstruction signal into a preset fault detection model to obtain an equipment fault detection result of the to-be-detected power distribution network. According to the invention, the problem of low fault detection accuracy in the prior art can be solved.

Description

Technical Field

[0001] This invention relates to the field of power system technology, and in particular to a method, device, terminal equipment and storage medium for multi-device fault detection in power distribution networks. Background Technology

[0002] With the deepening of the construction of new power systems, the distribution network, as the "last mile" of power supply, is facing the dual challenges of exponential growth in equipment scale and a surge in demand for intelligent operation and maintenance. Currently, key equipment such as transformers, ring main units, pole-mounted switches, and distribution lines in the distribution network are experiencing frequent failures, making the monitoring of the operating status and fault early warning of equipment in the distribution network increasingly important.

[0003] Existing fault detection technologies for power distribution network equipment utilize neural networks to detect anomalies in power time-series data. They construct feature vectors from multi-channel sensor data and then perform classification and prediction using deep models. However, these technologies only focus on time-domain modeling and lack effective modeling of frequency-domain features. It is difficult to ensure the consistency of feature restoration and modeling between time-domain and frequency-domain features simultaneously, which leads to distortion of the overall trend and loss of local spectral details, thus affecting the accuracy of detection. Summary of the Invention

[0004] This invention provides a method, apparatus, terminal equipment, and storage medium for multi-device fault detection in power distribution networks, which can effectively solve the problem that existing technologies cannot simultaneously ensure the consistency of time-domain features and frequency-domain features in feature restoration, resulting in low accuracy of equipment fault detection.

[0005] An embodiment of the present invention provides a method for detecting faults in multiple devices in a power distribution network, comprising:

[0006] Acquire time-series data of several devices in the power distribution network to be tested; the time-series data includes: electrical operation data, physical status data, and environmental data;

[0007] Based on the time series data, a fast Fourier transform and frequency division are performed to obtain several initial frequency patches;

[0008] Based on several initial frequency patches and a preset learning weight matrix, a mask is generated to obtain a binary mask matrix; wherein, the learning weight matrix is ​​generated under the constraints of a preset clustering loss function and a preset regularization loss function.

[0009] Based on the binary mask matrix, channel filtering and channel information fusion processing are performed on several initial frequency patches to obtain several processed target frequency patches.

[0010] Based on the target frequency patch, under the constraints of a preset frequency domain loss function and a preset time domain loss function, frequency domain reconstruction and time domain reconstruction are performed to obtain the spectrum reconstruction signal.

[0011] The spectrum reconstruction signal is input into a preset fault detection model to obtain the equipment fault detection results of the power distribution network under test.

[0012] Furthermore, based on the time series data, a Fast Fourier Transform and frequency division are performed to obtain several initial frequency patches, including:

[0013] The time series data is normalized to obtain normalized time series data;

[0014] The frequency domain data is obtained by performing a fast Fourier transform on the normalized time series data.

[0015] The frequency domain data is decomposed into real frequency components and imaginary frequency components.

[0016] Frequency division is performed based on the real frequency components and the imaginary frequency components to obtain several initial frequency patches.

[0017] Furthermore, based on several initial frequency patches and a preset learning weight matrix, a mask is generated to obtain a binary mask matrix, including:

[0018] Map several initial frequency patches to query vectors, key vectors, and value vectors;

[0019] Based on the query vector, the key vector, the value vector, and the preset learning weight matrix, the channel importance score of each initial frequency patch is calculated.

[0020] Based on the channel importance score and the preset number of channels, determine whether the ranking of the channel importance score is higher than the ranking corresponding to the preset number of channels. If yes, the corresponding matrix element is 1; otherwise, the corresponding matrix element is 0, and a binary mask matrix is ​​generated.

[0021] In calculating the channel importance score, the channel attention space clustering and the number of key channels are constrained by a preset clustering loss function and a preset regularization loss function. The learning weight matrix is ​​iteratively optimized until both the clustering loss function and the regularization loss function converge, resulting in the final learning weight matrix.

[0022] Further, based on the binary mask matrix, channel filtering and channel information fusion processing are performed on several initial frequency patches to obtain several processed target frequency patches, including:

[0023] Based on the binary mask matrix, channel information filtering is performed on several initial frequency patches to obtain filtered frequency patches.

[0024] Based on the query vector, the key vector, and the value vector, the filtered frequency patch is weighted by the attention channel to obtain the weighted frequency patch;

[0025] Channel fusion is performed on the weighted frequency patches to generate several processed target frequency patches.

[0026] Furthermore, the target frequency patch includes: the real part feature matrix after channel fusion and the imaginary part feature matrix after channel fusion;

[0027] Based on the target frequency patch, under the constraints of a preset frequency domain loss function and a preset time domain loss function, frequency domain reconstruction and time domain reconstruction are performed to obtain a spectrum reconstructed signal, including:

[0028] The real part feature matrix and the imaginary part feature matrix are sequentially concatenated along the frequency band dimension to obtain the frequency domain reconstructed real part signal and frequency domain imaginary part signal;

[0029] The real part signal and the imaginary part signal in the frequency domain are converted back to time domain signals to obtain the reconstructed time domain reconstructed signal; and the real part signal and the imaginary part signal in the frequency domain are used as the frequency domain reconstructed signal.

[0030] Based on the reconstructed time-domain signal and the time series data, calculate the time-domain loss function; and based on the reconstructed frequency-domain signal and the frequency-domain data, calculate the frequency-domain loss function.

[0031] The total loss function value is calculated based on the time-domain loss function, the frequency-domain loss function, the clustering loss function, and the regularization loss function; and it is determined whether the total loss function value converges.

[0032] If the total loss function value converges, the spectrum reconstruction signal is obtained; if the total loss function value does not converge, the target frequency patch is adjusted.

[0033] Further, the reconstructed spectrum signal is input into a preset fault detection model to obtain the equipment fault detection results of the distribution network under test, including:

[0034] The reconstructed spectrum signal is input into a preset fault detection model to extract fault features; wherein, the fault features include frequency domain features and time domain features;

[0035] The fault features are enhanced based on a preset attention mechanism to obtain enhanced features;

[0036] Based on the enhanced features, a classification decision is made to obtain the prediction scores for each type of fault;

[0037] Based on the predicted score and the preset score threshold, the fault type is determined and a fault detection result is generated.

[0038] Furthermore, the training of the fault detection model includes:

[0039] Acquire historical fault data and historical operation data of the power distribution network to construct a training dataset;

[0040] Based on the training dataset, perform Fast Fourier Transform, frequency division, and spectrum reconstruction to obtain the historical spectrum reconstruction signal;

[0041] The historical spectrum reconstructed signal is input into the fault detection model to be trained to obtain the fault prediction result;

[0042] Based on the fault prediction results and the fault labels corresponding to the historical fault data, a classification loss function is calculated; and based on the classification loss function, a preset clustering loss function, a regularization loss function, a frequency domain loss function, and a time domain loss function, the overall objective function is calculated.

[0043] If the loss value of the overall objective function does not converge, update the model parameters of the fault detection model and update the fault detection model according to the updated model parameters; if the loss value of the overall objective function converges, the trained fault detection model is obtained.

[0044] As an improvement to the above solution, another embodiment of the present invention provides a multi-device fault detection device for a power distribution network, comprising:

[0045] The equipment data acquisition module is used to acquire time-series data of several devices in the power distribution network to be tested; the time-series data includes: electrical operation data, physical status data, and environmental data.

[0046] The frequency patch generation module is used to perform fast Fourier transform and frequency division based on the time series data to obtain several initial frequency patches.

[0047] The mask generation module is used to generate a binary mask matrix based on several initial frequency patches and a preset learning weight matrix; wherein the learning weight matrix is ​​generated under the constraints of a preset clustering loss function and a preset regularization loss function.

[0048] The channel fusion module is used to perform channel filtering and channel information fusion processing on several initial frequency patches according to the binary mask matrix to obtain several processed target frequency patches.

[0049] The spectrum reconstruction module is used to perform frequency domain reconstruction and time domain reconstruction based on the target frequency patch, under the constraints of a preset frequency domain loss function and a preset time domain loss function, to obtain the spectrum reconstruction signal.

[0050] The fault detection module is used to input the spectrum reconstruction signal into a preset fault detection model to obtain the equipment fault detection results of the power distribution network under test.

[0051] Another embodiment of the present invention provides a terminal device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements a multi-device fault detection method for a power distribution network as described in the above embodiments.

[0052] Another embodiment of the present invention provides a computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to execute the multi-device fault detection method for a power distribution network described in the above embodiment.

[0053] By implementing this invention, at least the following beneficial effects are achieved:

[0054] This invention provides a method, apparatus, terminal equipment, and storage medium for multi-device fault detection in power distribution networks. The method decomposes the original time-series data into initial frequency patches through Fast Fourier Transform and frequency partitioning, overcoming the shortcomings of traditional methods that neglect frequency domain fault characteristics. It also eliminates dimensional differences in sensor data from different devices, achieving a unified representation of multi-device, multi-modal data and providing accurate frequency patch data for subsequent binary mask matrix generation. Furthermore, by using clustering and regularization loss functions, it retains the frequency channels most relevant to the fault, further improving the accuracy of subsequent fault detection. Finally, it utilizes a binary mask matrix for filtering. After the channel is processed, the retained frequency patches are weighted and fused through an attention mechanism to enhance fault-sensitive features and weaken irrelevant information. This ensures that the time-domain and frequency-domain features are consistent in restoration, and that the spectrum reconstruction signal after frequency-domain reconstruction and time-domain reconstruction is accurately recovered under the constraints of preset frequency-domain and time-domain loss functions. Furthermore, through joint optimization of time-domain reconstruction loss and frequency-domain reconstruction loss, the features of the frequency-domain reconstructed signal are enhanced, thus ensuring both the overall spectrum trend and the preservation of local spectrum details. This improves the accuracy of obtaining equipment fault detection results for the distribution network under test based on the spectrum reconstruction signal. Attached Figure Description

[0055] Figure 1 This is a flowchart illustrating a method for detecting faults in multiple devices in a power distribution network according to an embodiment of the present invention.

[0056] Figure 2 This is a schematic diagram of the structure of a multi-device fault detection device for a power distribution network provided in an embodiment of the present invention. Detailed Implementation

[0057] 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.

[0058] See Figure 1 To address the problem in existing technologies where it is difficult to simultaneously ensure consistency between time-domain and frequency-domain features in feature reconstruction, leading to low accuracy in equipment fault detection, an embodiment of the present invention provides a flowchart of a multi-device fault detection method for power distribution networks, comprising:

[0059] S1. Acquire time-series data of several devices in the power distribution network to be tested; the time-series data includes: electrical operation data, physical status data, and environmental data;

[0060] Specifically, the equipment includes key components such as transformers, ring main units, pole-mounted switches, and distribution lines. The electrical operating data includes: voltage, current, active power, reactive power, power factor, power frequency deviation, and phase sequence difference; the physical state data includes: vibration parameters; and the environmental data includes: temperature and noise parameters. The purpose of data acquisition is to obtain real-time status information of several devices in the distribution network under test during operation, for subsequent processing, analysis, and fault diagnosis.

[0061] S2. Perform a fast Fourier transform and frequency division based on the time series data to obtain several initial frequency patches;

[0062] Preferably, a number of initial frequency patches are obtained by performing a Fast Fourier Transform and frequency division on the time series data, including:

[0063] The time series data is normalized to obtain normalized time series data;

[0064] The frequency domain data is obtained by performing a fast Fourier transform on the normalized time series data.

[0065] The frequency domain data is decomposed into real frequency components and imaginary frequency components.

[0066] Frequency division is performed based on the real frequency components and the imaginary frequency components to obtain several initial frequency patches.

[0067] In a preferred embodiment of the present invention, the time series data represents time-domain data from several devices. Assume the time series data is represented as X, where X∈R. N×T (N is the number of devices, T is the number of time points), normalize the time series data to obtain normalized time series data:

[0068]

[0069] in, It is normalized time series data; X (i) It is the time series data of the i-th device; parameter u i and σ i Let be the arithmetic mean and standard deviation of the time series data for the i-th device, respectively. Normalization is performed independently on the time series data for each device to eliminate differences in dimensions.

[0070] Then, the normalized time series data is subjected to a Fast Fourier Transform to obtain the frequency domain data:

[0071] X R ,X I =FFT(X) norm );

[0072] The Fast Fourier Transform (FFT) converts normalized time-series data into frequency-domain data and decomposes it into real-part frequency components X. R and the imaginary part frequency component X I .

[0073] Finally, frequency division is performed based on the real and imaginary frequency components to obtain several initial frequency patches:

[0074] P Rj =Slice(X) R ,bin j ),P Ij =Slice(X) I ,bin j ), j = 1, 2, ..., L;

[0075] Among them, P Rj ,P Ij The j-th frequency patch bin jThe corresponding real and imaginary parts of the Fourier transform result. Frequency patching divides the frequency domain signal (frequency domain data) into L local initial frequency patches to capture fine-grained features. Direct sub-band segmentation is performed on the frequency domain data, dividing the spectrum into multiple initial frequency patches for each sub-band, converting latent fault features (such as harmonics) into operable frequency domain data; patching supports fault analysis at different frequency band scales; and it provides an adapted input format for attention-based mask generation (such as Top-K channel selection).

[0076] S3. Based on several initial frequency patches and a preset learning weight matrix, a mask is generated to obtain a binary mask matrix; wherein, the learning weight matrix is ​​generated under the constraints of a preset clustering loss function and a preset regularization loss function.

[0077] Preferably, a mask is generated based on several initial frequency patches and a preset learning weight matrix to obtain a binary mask matrix, including:

[0078] Map several initial frequency patches to query vectors, key vectors, and value vectors;

[0079] Based on the query vector, the key vector, the value vector, and the preset learning weight matrix, the channel importance score of each initial frequency patch is calculated.

[0080] Based on the channel importance score and the preset number of channels, determine whether the ranking of the channel importance score is higher than the ranking corresponding to the preset number of channels. If yes, the corresponding matrix element is 1; otherwise, the corresponding matrix element is 0, and a binary mask matrix is ​​generated.

[0081] In calculating the channel importance score, the channel attention space clustering and the number of key channels are constrained by a preset clustering loss function and a preset regularization loss function. The learning weight matrix is ​​iteratively optimized until both the clustering loss function and the regularization loss function converge, resulting in the final learning weight matrix.

[0082] Specifically, the preset learning weight matrix includes the first weight matrix W corresponding to the query vector. Q The second weight matrix W corresponding to the key vector K The third weight matrix W corresponding to the value vector V And the fourth weight matrix W used to characterize the channel importance score. d W Q W K W V ∈R d×2M The attention dimension is d.

[0083] In a preferred embodiment of the present invention, several initial frequency patches are mapped to query vectors, key vectors, and value vectors:

[0084] Q j =W Q P j ,K j =W K P j V j =W V P j ; where Q j For the query vector, K j V is the key vector. j is a value vector; Pj represents the eigenvector of the j-th initial frequency patch.

[0085] Based on the query vector, the key vector, the value vector, and the preset learning weight matrix, the channel importance score D for each frequency patch is calculated. j :D j =W d P j +b d Among them, W d b is used to assess the importance of the initial frequency patch. d This is used as a bias to adjust the scoring benchmark. A channel importance score is calculated for each channel, used for the fusion of key channel information.

[0086] Based on the channel importance score and the preset number of channels, it is determined whether the ranking of the channel importance score is higher than the ranking corresponding to the preset number of channels. If so, the corresponding matrix element is 1; otherwise, the corresponding matrix element is 0, thereby generating a binary mask matrix M. j Retain the key channels in the Top-K hierarchy and achieve information fusion for critical channels. For D... j Perform a Top-K selection, retain the K channels with the highest scores, and generate a binary mask matrix M. j ∈{0,1}:M j =TopK(D j );

[0087] If a channel's importance score ranks in the top K positions in a Top-K study, then M... j The corresponding position is 1, otherwise it is 0. Sparse masking is used to suppress noise interference.

[0088] Preset clustering loss function L C By contrastive learning, sensor channels triggered by similar faults (such as overload and short circuit) are forced to cluster in the attention space, establishing causal relationships between multimodal signals such as circuit breaker action and temperature surge.

[0089] in, It is the similarity score between sample k and sample m of the learning weight matrix, and τ is a temperature coefficient that controls the smoothness of the similarity distribution.

[0090] The number of key channels represents the number of information channels that are more relevant to the initial frequency patch, with a preset regularization loss function L. R The number of critical channels is constrained to limit the number of relevant channels and prevent the mask generator from outputting a matrix of constant 1s. This also prevents potential faults caused by single-sensor anomalies (such as partial discharge ultrasonic signals) from being missed.

[0091] Where I is an all-one matrix, || || F The Frobenius norm measures the total deviation of matrix elements. Clustering loss ensures that channel selection conforms to fault semantics; regularization loss avoids selecting redundant channels and improves generalization ability.

[0092] S4. Based on the binary mask matrix, perform channel filtering and channel information fusion processing on several initial frequency patches to obtain several processed target frequency patches.

[0093] Specifically, based on the binary mask matrix, channel filtering and channel information fusion processing are performed on several initial frequency patches to obtain several processed target frequency patches, including:

[0094] Based on the binary mask matrix, channel information filtering is performed on several initial frequency patches to obtain filtered frequency patches.

[0095] Based on the query vector, the key vector, and the value vector, the filtered frequency patch is weighted by the attention channel to obtain the weighted frequency patch;

[0096] Channel fusion is performed on the weighted frequency patches to generate several processed target frequency patches.

[0097] In a preferred embodiment of the present invention, the key channel information is fused, and the calculation formula is as follows:

[0098] Where ⊙ denotes element-wise multiplication, and d is the dimension of the attention space. This is a scaling factor to prevent the Softmax gradient from vanishing.

[0099] A binary mask matrix is ​​used to filter out frequency channels unrelated to faults (channels with a mask of 0) to obtain filtered frequency patches. An attention mechanism is used to dynamically weight the retained channels (channels with a mask of 1) to highlight fault-sensitive features, resulting in weighted frequency patches. The information from multiple weighted frequency patches is fused into a compact feature representation to generate several processed target frequency patches, which facilitates subsequent time-frequency reconstruction and fault detection.

[0100] Traditional FFT processing only converts time-domain signals into frequency-domain data, but it only focuses on the overall frequency transformation stage, lacking further fine-grained extraction of local frequency features. Some equipment faults often manifest as anomalies within local frequency bands (such as local harmonic distortion or weak transient oscillations), and existing technologies struggle to efficiently capture these details through a single overall frequency domain analysis. In multi-sensor data acquisition, the data collected by each sensor often exhibits differences between devices or channels, and some sensors (such as temperature and current sensors) may introduce significant noise in certain frequency bands. Existing technologies, through simple fusion or direct splicing, are insufficient to filter out irrelevant noise and provide inadequate signal response from critical sensors, potentially leading to missed or false alarms. In multi-device data acquisition in distribution networks, due to the diverse types of sensors, different sensors may contain both critical fault information and significant noise or irrelevant information in different frequency bands. Traditional simple splicing or weighted averaging methods are insufficient to distinguish useful information from noise in the signal. This embodiment introduces an attention-based channel fusion module, which uses a dynamically generated binary mask matrix to evaluate and filter the importance of each channel, effectively eliminating irrelevant noise, strengthening key channel signals, and ensuring efficient fusion of multi-sensor data.

[0101] S5. Based on the target frequency patch, under the constraints of a preset frequency domain loss function and a preset time domain loss function, perform frequency domain reconstruction and time domain reconstruction to obtain the spectrum reconstruction signal.

[0102] Specifically, the target frequency patch includes: the real part feature matrix after channel fusion and the imaginary part feature matrix after channel fusion;

[0103] Based on the target frequency patch, under the constraints of a preset frequency domain loss function and a preset time domain loss function, frequency domain reconstruction and time domain reconstruction are performed to obtain a spectrum reconstructed signal, including:

[0104] The real part feature matrix and the imaginary part feature matrix are sequentially concatenated along the frequency band dimension to obtain the frequency domain reconstructed real part signal and frequency domain imaginary part signal;

[0105] The real part signal and the imaginary part signal in the frequency domain are converted back to time domain signals to obtain the reconstructed time domain reconstructed signal; and the real part signal and the imaginary part signal in the frequency domain are used as the frequency domain reconstructed signal.

[0106] Based on the reconstructed time-domain signal and the time series data, calculate the time-domain loss function; and based on the reconstructed frequency-domain signal and the frequency-domain data, calculate the frequency-domain loss function.

[0107] The total loss function value is calculated based on the time-domain loss function, the frequency-domain loss function, the clustering loss function, and the regularization loss function; and it is determined whether the total loss function value converges.

[0108] If the total loss function value converges, the spectrum reconstruction signal is obtained; if the total loss function value does not converge, the target frequency patch is adjusted.

[0109] Specifically, the initial frequency patch includes real and imaginary components. Therefore, after channel information fusion, the channel-fused real feature matrix and the channel-fused imaginary feature matrix are obtained. After channel fusion is completed, the processed frequency domain features (real feature matrix and imaginary feature matrix) need to be reconstructed back into complete frequency and time domain signals, and the reconstruction quality is ensured through joint optimization using multiple loss functions.

[0110] In a preferred embodiment of the present invention, the real part feature matrix and the imaginary part feature matrix are sequentially concatenated along the frequency band dimension to obtain the frequency domain reconstructed real part signal X′. R and the frequency domain imaginary part signal X′ I :

[0111]

[0112] For the j-th initial frequency patch, the real and imaginary eigenvalues ​​are obtained after channel fusion. `Concat()` represents concatenating all target frequency patches along the frequency band dimension to recover the complete spectrum; W R W I A preset learning weight matrix is ​​used to project the spliced ​​spectrum onto the original frequency domain dimension. The real part and imaginary part of the frequency domain signal are then used as the reconstructed frequency domain signal.

[0113] An iFFT decoder is used to convert the frequency domain reconstructed signal (XR′, XI′) back to the time domain reconstructed signal X′. Overall waveform consistency is constrained to suppress global offset.

[0114] X′=iFFT(X′ R ,X′ I );

[0115] Reconstruction loss functions are employed in both the time and frequency domains to enhance point-by-point and sub-sequence modeling capabilities, respectively. Based on the reconstructed time-domain signal and the time-series data, the time-domain loss function L is calculated. time :

[0116] Among them, X i,t X represents the raw time-domain data (time series data) of the i-th device at time point t. i,t′ Let T be the reconstructed time-domain data (time-domain reconstructed signal) of the i-th device at time point t. T is the total length of the time series.

[0117] Among them, X R,i X I,i X represents the real and imaginary parts of the original frequency domain signal (frequency domain data) of the i-th device, respectively. R,i′ X I,i′ d represents the real and imaginary parts of the frequency domain reconstructed signal from the i-th device, respectively. || ||1: L1 norm, measuring the absolute error of the frequency domain signal. f The dimension of each target frequency patch.

[0118] The total loss function value is calculated based on the time-domain loss function, the frequency-domain loss function, the clustering loss function, and the regularization loss function: λ1, λ2, λ3 are hyperparameters, and the balancing time-domain loss L time Frequency domain loss L freq Clustering loss L C and regularization loss L R The weights are determined. If the total loss function converges, the reconstructed spectrum signal is obtained; if the total loss function does not converge, the parameters of the iFFT decoder are updated to update the iFFT decoder.

[0119] Traditional methods mostly focus on either the time or frequency domain, making it difficult to simultaneously ensure consistency in feature reconstruction and modeling between the two. The lack of a joint optimization mechanism can lead to distortion of the overall trend and loss of local spectral details, thus affecting the accuracy of fault detection. Existing systems typically fail to consider the correlations between different devices and sensors, especially when dealing with complex faults (such as overload, short circuit, and chain reactions triggered by local abnormal signals), lacking effective cross-channel and cross-modal information fusion and relationship modeling methods. Traditional methods struggle to capture such multimodal causal relationships through single feature modeling. This embodiment designs a joint reconstruction in the time and frequency domains. By simultaneously optimizing the consistency of the time-domain waveform and the fidelity of the frequency domain spectrum during reconstruction, a dual loss function constraint is achieved, thereby ensuring overall trend stability while accurately recovering fine-grained spectral changes and improving the completeness of fault feature reconstruction. A contrastive learning strategy using clustering loss is proposed to aggregate multimodal signals of similar faults within the attention space, achieving deep-level correlation information fusion between different sensors and further improving the system's accuracy in identifying complex faults. In multi-sensor environments, individual sensors may falsely report faults due to local noise or short-term abnormal fluctuations, and existing simple fusion methods are insufficient to effectively avoid such misjudgments. Therefore, this invention introduces a regularized loss in the channel fusion module, strictly constraining the number of critical channels to ensure that a brief anomaly in a single sensor does not lead to an overall incorrect judgment, thereby improving the robustness and stability of the system.

[0120] A binary mask matrix is ​​generated using an attention mechanism to dynamically select Top-K key frequency bands related to equipment faults (such as the 50Hz power frequency and its third harmonic), isolating irrelevant noise and interference from multiple sensors such as voltage, current, and temperature, ensuring the prominent expression of key features. iFFT is used to restore the decomposed frequency domain signal to the time domain, and the time domain waveform and spectrum are jointly optimized to ensure accurate recovery of both the overall trend and local spectral information. Frequency domain reconstruction enhances the accurate recovery of harmonic distortion and high-frequency transient oscillations; time domain reconstruction constrains the overall waveform stability, avoiding local anomalies and overfitting. Joint optimization of time domain reconstruction loss and frequency domain reconstruction loss ensures that both the overall trend and local details achieve optimal modeling results, reducing the risk of overfitting. Clustering Loss (used to shorten the distance between multimodal signals of similar faults in the attention space) and regularization Loss (constraining the number of key channels to prevent misjudgment caused by noise interference) are introduced during channel fusion to further improve fault identification accuracy. Clustering loss utilizes contrastive learning to force sensor channels with similar faults to cluster in the feature space, establishing causal relationships between multimodal signals. This enables the aggregation of multimodal signals from different fault types (such as overload and short circuit) in the attention space, allowing the system to more sensitively capture signal correlations under complex faults. Regularization loss constrains the number of critical channels, preventing the mask generator from outputting a matrix of all 1s and avoiding excessive noise retention. Strict control over the number of critical channels prevents misjudgments due to anomalies in a single channel during data fusion, thus improving system robustness.

[0121] S6. Input the spectrum reconstruction signal into the preset fault detection model to obtain the equipment fault detection results of the power distribution network to be detected.

[0122] Preferably, the spectrum reconstruction signal is input into a preset fault detection model to obtain the equipment fault detection results of the distribution network under test, including:

[0123] The reconstructed spectrum signal is input into a preset fault detection model to extract fault features; wherein, the fault features include frequency domain features and time domain features;

[0124] The fault features are enhanced based on a preset attention mechanism to obtain enhanced features;

[0125] Based on the enhanced features, a classification decision is made to obtain the prediction scores for each type of fault;

[0126] Based on the predicted score and the preset score threshold, the fault type is determined and a fault detection result is generated.

[0127] Preferably, the training of the fault detection model includes:

[0128] Acquire historical fault data and historical operation data of the power distribution network to construct a training dataset;

[0129] Based on the training dataset, perform Fast Fourier Transform, frequency division, and spectrum reconstruction to obtain the historical spectrum reconstruction signal;

[0130] The historical spectrum reconstructed signal is input into the fault detection model to be trained to obtain the fault prediction result;

[0131] Based on the fault prediction results and the fault labels corresponding to the historical fault data, a classification loss function is calculated; and based on the classification loss function, a preset clustering loss function, a regularization loss function, a frequency domain loss function, and a time domain loss function, the overall objective function is calculated.

[0132] If the loss value of the overall objective function does not converge, update the model parameters of the fault detection model and update the fault detection model according to the updated model parameters; if the loss value of the overall objective function converges, the trained fault detection model is obtained.

[0133] In a preferred embodiment of the present invention, the spectrum reconstruction signal obtained after frequency division and spectrum reconstruction dual-layer optimization is fed into the fault detection model to be trained, and the prediction scores of various faults of the spectrum reconstruction signal are obtained. This indicates the likelihood that the abnormal signal is a fault, and accurately classifies the abnormal signal.

[0134] By implementing this embodiment, the original time-series data is decomposed into initial frequency patches through Fast Fourier Transform and frequency partitioning, overcoming the shortcomings of traditional methods that ignore frequency domain fault features. Simultaneously, it eliminates the dimensional differences in sensor data from different devices, achieving a unified representation of multi-device, multi-modal data and providing accurate frequency patch data for the subsequent generation of the binary mask matrix. Through clustering and regularization loss functions, the frequency channels most relevant to the fault are retained while noise channels are suppressed, further improving the accuracy of subsequent fault detection. After filtering redundant channels using the binary mask matrix, the retained frequency patches are weighted and fused using an attention mechanism, enhancing fault-sensitive features and weakening irrelevant information. This ensures consistency between time-domain and frequency-domain features in reconstruction, guaranteeing accurate recovery of the reconstructed spectrum signal after frequency and time domain reconstruction under preset frequency and time domain loss function constraints. Furthermore, through joint optimization of time and frequency domain reconstruction losses, the features of the frequency domain reconstructed signal are enhanced, thus ensuring both the overall spectral trend and the preservation of local spectral details, improving the accuracy of equipment fault detection results for the distribution network under test based on the reconstructed spectrum signal.

[0135] See Figure 2This is a schematic diagram of the structure of a multi-device fault detection device for a power distribution network according to an embodiment of the present invention, comprising:

[0136] The equipment data acquisition module is used to acquire time-series data of several devices in the power distribution network to be tested; the time-series data includes: electrical operation data, physical status data, and environmental data.

[0137] The frequency patch generation module is used to perform fast Fourier transform and frequency division based on the time series data to obtain several initial frequency patches.

[0138] The mask generation module is used to generate a binary mask matrix based on several initial frequency patches and a preset learning weight matrix; wherein the learning weight matrix is ​​generated under the constraints of a preset clustering loss function and a preset regularization loss function.

[0139] The channel fusion module is used to perform channel filtering and channel information fusion processing on several initial frequency patches according to the binary mask matrix to obtain several processed target frequency patches.

[0140] The spectrum reconstruction module is used to perform frequency domain reconstruction and time domain reconstruction based on the target frequency patch, under the constraints of a preset frequency domain loss function and a preset time domain loss function, to obtain the spectrum reconstruction signal.

[0141] The fault detection module is used to input the spectrum reconstruction signal into a preset fault detection model to obtain the equipment fault detection results of the power distribution network under test.

[0142] This invention provides a multi-device fault detection device for a power distribution network. The device acquires time-series data of several devices in the power distribution network to be detected using a data acquisition module. This time-series data includes electrical operation data, physical state data, and environmental data. A frequency patch generation module performs a Fast Fourier Transform and frequency division on the time-series data to obtain several initial frequency patches. A mask generation module generates a binary mask matrix based on the initial frequency patches and a preset learning weight matrix. The learning weight matrix is ​​generated under the constraints of a preset clustering loss function and a preset regularization loss function. A channel fusion module performs channel filtering and channel information fusion processing on the initial frequency patches using the binary mask matrix to obtain several processed target frequency patches. A spectrum reconstruction module performs frequency domain reconstruction and time domain reconstruction based on the target frequency patches, under the constraints of a preset frequency domain loss function and a preset time domain loss function, to obtain a spectrum reconstruction signal. Finally, a fault detection module inputs the spectrum reconstruction signal into a preset fault detection model to obtain the equipment fault detection results for the power distribution network to be detected.

[0143] By employing Fast Fourier Transform and frequency partitioning, the original time-series data is decomposed into initial frequency patches, overcoming the shortcomings of traditional methods that neglect frequency domain fault features. Simultaneously, it eliminates dimensional differences in sensor data from different devices, achieving a unified representation of multi-device, multi-modal data and providing accurate frequency patch data for subsequent binary mask matrix generation. Through clustering and regularization loss functions, the most fault-related frequency channels are retained while noise channels are suppressed, further improving the accuracy of subsequent fault detection. After filtering redundant channels using the binary mask matrix, an attention mechanism is used to weightedly fuse the retained frequency patches, enhancing fault-sensitive features and weakening irrelevant information. This ensures consistency between time-domain and frequency-domain features in reconstruction, guaranteeing accurate recovery of the reconstructed spectrum signal after frequency and time domain reconstruction under preset frequency and time domain loss function constraints. Furthermore, joint optimization of time and frequency domain reconstruction losses enhances the features of the frequency-domain reconstructed signal, ensuring both overall spectral trends and the preservation of local spectral details, thus improving the accuracy of fault detection results for the distribution network under test based on the reconstructed spectrum signal.

[0144] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0145] Those skilled in the art will understand that, for convenience and brevity, the specific working process of the device described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0146] Another embodiment of the present invention provides a terminal device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements a multi-device fault detection method for a power distribution network as described in the above embodiments. The terminal device may be a desktop computer, laptop, handheld computer, or cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.

[0147] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the terminal device, connecting various parts of the terminal device via various interfaces and lines.

[0148] The memory can be used to store the computer program. The processor implements various functions of the terminal device by running or executing the computer program stored in the memory and calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function, etc.; the data storage area may store data created based on the use of the mobile phone, etc. In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, RAM, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0149] Another embodiment of the present invention provides a computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to execute the multi-device fault detection method for a power distribution network described in the above embodiment.

[0150] The storage medium is a computer-readable storage medium, and the computer program is stored in the computer-readable storage medium. When executed by a processor, the computer program can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or some intermediate form. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0151] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for detecting faults in multiple devices in a power distribution network, characterized in that, include: Acquire time-series data of several devices in the power distribution network to be tested; The time-series data includes: electrical operation data, physical state data, and environmental data; Based on the time series data, a fast Fourier transform and frequency division are performed to obtain several initial frequency patches; Based on several initial frequency patches and a preset learning weight matrix, a mask is generated to obtain a binary mask matrix; wherein, the learning weight matrix is ​​generated under the constraints of a preset clustering loss function and a preset regularization loss function. Based on the binary mask matrix, channel filtering and channel information fusion processing are performed on several initial frequency patches to obtain several processed target frequency patches. Based on the target frequency patch, under the constraints of a preset frequency domain loss function and a preset time domain loss function, frequency domain reconstruction and time domain reconstruction are performed to obtain the spectrum reconstruction signal. The spectrum reconstruction signal is input into a preset fault detection model to obtain the equipment fault detection results of the power distribution network under test.

2. The method for detecting multiple equipment faults in a power distribution network as described in claim 1, characterized in that, Based on the time series data, a Fast Fourier Transform and frequency division are performed to obtain several initial frequency patches, including: The time series data is normalized to obtain normalized time series data; The frequency domain data is obtained by performing a fast Fourier transform on the normalized time series data. The frequency domain data is decomposed into real frequency components and imaginary frequency components. Frequency division is performed based on the real frequency components and the imaginary frequency components to obtain several initial frequency patches.

3. The method for detecting multiple equipment faults in a power distribution network as described in claim 2, characterized in that, Based on several initial frequency patches and a preset learning weight matrix, a mask is generated to obtain a binary mask matrix, including: Map several initial frequency patches to query vectors, key vectors, and value vectors; Based on the query vector, the key vector, the value vector, and the preset learning weight matrix, the channel importance score of each initial frequency patch is calculated. Based on the channel importance score and the preset number of channels, determine whether the ranking of the channel importance score is higher than the ranking corresponding to the preset number of channels. If yes, the corresponding matrix element is 1; otherwise, the corresponding matrix element is 0, and a binary mask matrix is ​​generated. In calculating the channel importance score, the channel attention space clustering and the number of key channels are constrained by a preset clustering loss function and a preset regularization loss function. The learning weight matrix is ​​iteratively optimized until both the clustering loss function and the regularization loss function converge, resulting in the final learning weight matrix.

4. The method for detecting multiple equipment faults in a power distribution network as described in claim 3, characterized in that, Based on the binary mask matrix, channel filtering and channel information fusion are performed on several initial frequency patches to obtain several processed target frequency patches, including: Based on the binary mask matrix, channel information filtering is performed on several initial frequency patches to obtain filtered frequency patches. Based on the query vector, the key vector, and the value vector, the filtered frequency patch is weighted by the attention channel to obtain the weighted frequency patch; Channel fusion is performed on the weighted frequency patches to generate several processed target frequency patches.

5. The method for detecting multiple equipment faults in a power distribution network as described in claim 4, characterized in that, The target frequency patch includes: the real part feature matrix after channel fusion and the imaginary part feature matrix after channel fusion; Based on the target frequency patch, under the constraints of a preset frequency domain loss function and a preset time domain loss function, frequency domain reconstruction and time domain reconstruction are performed to obtain a spectrum reconstructed signal, including: The real part feature matrix and the imaginary part feature matrix are sequentially concatenated along the frequency band dimension to obtain the frequency domain reconstructed real part signal and frequency domain imaginary part signal; The real part signal and the imaginary part signal in the frequency domain are converted back to time domain signals to obtain the reconstructed time domain reconstructed signal; and the real part signal and the imaginary part signal in the frequency domain are used as the frequency domain reconstructed signal. Based on the reconstructed time-domain signal and the time series data, calculate the time-domain loss function; and based on the reconstructed frequency-domain signal and the frequency-domain data, calculate the frequency-domain loss function. The total loss function value is calculated based on the time-domain loss function, the frequency-domain loss function, the clustering loss function, and the regularization loss function; and it is determined whether the total loss function value converges. If the total loss function value converges, the spectrum reconstruction signal is obtained; if the total loss function value does not converge, the target frequency patch is adjusted.

6. The method for detecting multiple equipment faults in a power distribution network as described in claim 5, characterized in that, The reconstructed spectrum signal is input into a preset fault detection model to obtain the equipment fault detection results of the distribution network under test, including: The reconstructed spectrum signal is input into a preset fault detection model to extract fault features; wherein, the fault features include frequency domain features and time domain features; The fault features are enhanced based on a preset attention mechanism to obtain enhanced features; Based on the enhanced features, a classification decision is made to obtain the prediction scores for each type of fault; Based on the predicted score and the preset score threshold, the fault type is determined and a fault detection result is generated.

7. The method for detecting multiple equipment faults in a power distribution network as described in claim 6, characterized in that, The training of the fault detection model includes: Acquire historical fault data and historical operation data of the power distribution network to construct a training dataset; Based on the training dataset, perform Fast Fourier Transform, frequency division, and spectrum reconstruction to obtain the historical spectrum reconstruction signal; The historical spectrum reconstructed signal is input into the fault detection model to be trained to obtain the fault prediction result; Based on the fault prediction results and the fault labels corresponding to the historical fault data, a classification loss function is calculated; and based on the classification loss function, a preset clustering loss function, a regularization loss function, a frequency domain loss function, and a time domain loss function, the overall objective function is calculated. If the loss value of the overall objective function does not converge, update the model parameters of the fault detection model and update the fault detection model according to the updated model parameters; if the loss value of the overall objective function converges, the trained fault detection model is obtained.

8. A multi-equipment fault detection device for power distribution networks, characterized in that, include: The equipment data acquisition module is used to acquire time-series data of several devices in the power distribution network to be tested; The time-series data includes: electrical operation data, physical state data, and environmental data; The frequency patch generation module is used to perform fast Fourier transform and frequency division based on the time series data to obtain several initial frequency patches. The mask generation module is used to generate a binary mask matrix based on several initial frequency patches and a preset learning weight matrix; wherein the learning weight matrix is ​​generated under the constraints of a preset clustering loss function and a preset regularization loss function. The channel fusion module is used to perform channel filtering and channel information fusion processing on several initial frequency patches according to the binary mask matrix to obtain several processed target frequency patches. The spectrum reconstruction module is used to perform frequency domain reconstruction and time domain reconstruction based on the target frequency patch, under the constraints of a preset frequency domain loss function and a preset time domain loss function, to obtain the spectrum reconstruction signal. The fault detection module is used to input the spectrum reconstruction signal into a preset fault detection model to obtain the equipment fault detection results of the power distribution network under test.

9. A terminal device, characterized in that, The method includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement a multi-device fault detection method for a power distribution network as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform a multi-device fault detection method for a power distribution network as described in any one of claims 1 to 7.

Images