A Deep Learning-Based Method for Analyzing and Determining Anomalies in Current Waveform Features

Abstract

This invention relates to the field of smart grid condition monitoring and fault diagnosis technology, specifically a deep learning-based method for anomaly analysis and discrimination of current waveform features. The method includes: extracting the power frequency and high-frequency discrete sequences of distribution line signals; fusing fundamental envelope features to construct a multimodal load addressing matrix, inputting it into a neural architecture search network, and dynamically activating matching convolutional branches in the target deep network through a soft routing gating mechanism; mapping the high-frequency sequence to a phase space trajectory matrix, performing spatiotemporal decoupling calculations and extracting distortion features through the activated branches; adjusting the bias of the fully connected layer to generate a dynamic margin discrimination bound, and outputting the anomaly confidence probability. This invention effectively filters out complex background noise and reduces model misjudgment through multidimensional feature reconstruction and adaptive dynamic routing with edge computing power, combined with a variance penalty margin compensation mechanism, thereby improving the anti-interference capability and computational efficiency of anomaly discrimination.

Description

Technical Field

[0001] This invention relates to the field of smart grid condition monitoring and fault diagnosis technology, specifically a method for analyzing and identifying anomalies in current waveform features based on deep learning. Background Technology

[0002] In modern power distribution systems, current waveforms serve as a fundamental measurement variable reflecting the operating status of electrical equipment. The analysis of waveform characteristics determines the reliability of electrical anomaly detection (such as fault arcing). Existing technologies largely rely on traditional protection devices with fixed current thresholds or employ shallow time-frequency signal analysis techniques (such as capturing zero-crossing "zero-break" phenomena and analyzing the proportion of odd harmonics). In addition, some solutions introduce Fourier transform, empirical mode decomposition, and static deep learning models for waveform feature measurement and recognition.

[0003] However, with the connection of massive nonlinear loads (such as high-frequency inverters and switching power supplies), the above measurement and analysis methods have exposed insurmountable defects in complex electromagnetic environments: First, the high-frequency harmonics injected by nonlinear devices highly overlap with the distortion characteristics of the real waveform, forming a strong "masking effect", which easily leads to false alarms and "false operation" of waveform analysis algorithms; Second, under strong background noise coupling, weak early abnormal quantities (such as when a series fault arc occurs, the effective value of the line current not only does not increase but decreases slightly) are easily submerged by environmental interference, causing traditional electrical measurement devices to miss the alarm and "refuse to operate".

[0004] Meanwhile, existing deep learning waveform analysis methods also face bottlenecks in industrial measurement applications. Conventional network models have large parameters, making it difficult to achieve fast, low-latency inference in edge-side measurement devices (such as circuit breaker MCUs) with limited storage and computing power. In addition, static models are difficult to adapt to dynamic changes in load topology, and the lack of real physical arc waveform samples leads to data imbalance, resulting in poor environmental adaptability of the models.

[0005] In summary, achieving low computational overhead, accurate anomaly feature extraction, and reliable discrimination of high-dimensional nonlinear time-series data under environments with strong noise coupling and dynamic changes has become a pressing technical challenge in this field.

[0006] To address this, a deep learning-based method for analyzing and identifying anomalies in current waveform features is proposed. Summary of the Invention

[0007] The purpose of this invention is to provide a current waveform feature anomaly analysis and discrimination method based on deep learning. By multi-dimensional feature reconstruction and adaptive neural architecture search, it solves the problems of fuzzy transient feature extraction and easy misjudgment of deep network models in complex power distribution network environments, and improves the anti-interference ability and computational efficiency of anomaly discrimination.

[0008] To achieve the above objectives, the present invention provides the following technical solution: A deep learning-based method for analyzing and identifying anomalies in current waveform features includes: acquiring discrete sampling signals of distribution lines, splitting and extracting the power frequency fundamental wave data stream and high-frequency transient discrete sequence; fusing the steady-state slow-varying envelope features extracted from the power frequency fundamental wave data stream with the power frequency starting step envelope features, and performing unsupervised density clustering analysis to construct a multimodal load addressing matrix that characterizes the current power grid load operation background features.

[0009] The multimodal load addressing matrix is ​​used as a conditional trigger. A preset differentiable neural architecture search network is input to calculate and generate an architecture probability weight matrix. Based on the architecture probability weight matrix, the target convolutional kernel weight branch in the preset target depth network that matches the background noise under the current load operation background features is dynamically activated through a soft routing gating mechanism.

[0010] The high-frequency transient discrete sequence is mapped to a two-dimensional phase space trajectory matrix and fed into the target depth network. The spatiotemporal decoupling calculation is performed by the activated target convolutional kernel weight branch to filter out the switching transient background noise and extract the dimension-reduced abnormal distortion feature vector.

[0011] Obtain the preset local smoothing constraint term corresponding to the weight branch of the target convolutional kernel, adjust the Softmax activation function bias term in the classification fully connected layer at the end of the target depth network, and generate a dynamic margin discrimination bound corresponding to the current power grid background noise intensity; input the abnormal distortion feature vector into the adjusted classification fully connected layer, and output the distribution line anomaly confidence probability through the dynamic margin discrimination bound mapping.

[0012] Preferably, the step of splitting and extracting the power frequency fundamental wave data stream and the high-frequency transient discrete sequence includes: applying a discrete wavelet packet decomposition algorithm to divide the frequency band boundary of the discrete sampling signal of the power distribution line, separating the current sampling points whose frequency range corresponds to the rated fundamental frequency of the power distribution system and the harmonics that are integer multiples of the rated frequency within a preset number, and recombining them into a power frequency fundamental wave data stream according to the time series; extracting and saving the edge detail current components whose frequency is within a preset high-frequency characteristic frequency band, and recombining them into a high-frequency transient discrete sequence according to the time series; extracting the effective value of the fundamental current and the current crest coefficient of the power frequency fundamental wave data stream within a preset number of continuous sliding time windows, and combining them to form a steady-state slowly varying envelope feature; extracting the first-order difference data of the current amplitude within a preset number of complete cycles after the zero-point level change point of the power frequency fundamental wave data stream, and combining them to form a power frequency start-up step envelope feature.

[0013] Preferably, the step of constructing a multimodal load addressing matrix by performing unsupervised density clustering analysis includes: performing a tensor concatenation operation along the data channel dimension on the steady-state slow-varying envelope feature and the power frequency starting step envelope feature, and inputting them into a density-based noise application space clustering algorithm model; calculating the Euclidean distance distribution within the neighborhood radius of the cluster core points, dividing sample points with Euclidean distance differences within a preset constant range into load equipment clusters of the same type, and calculating the spatial coordinates of the cluster centers of all load equipment clusters; extracting the spatial coordinates of all the cluster centers and the percentage of samples contained in each load equipment cluster, performing a combination operation according to a standard row and column format, and generating a multimodal load addressing matrix for characterizing the distribution status of equipment types within the distribution network.

[0014] Preferably, the step of dynamically activating the target convolutional kernel weight branch through the soft routing gating mechanism includes: the preset differentiable neural architecture search network contains a continuous operation space; the multimodal load addressing matrix is ​​input into the continuous operation space; the normalized probability values ​​of the corresponding candidate branch nodes are calculated through a fully connected layer and Softmax normalization to generate the architecture probability weight matrix; based on the numerical distribution sequence in the architecture probability weight matrix, the Gumbel distribution sampling algorithm is applied to output discrete one-hot encoded vectors; the one-hot encoded vectors are multiplied by a matrix multiplication operation with the candidate feature processing path set inside the preset target depth network; independent data paths with non-zero multiplication results are defined and activated as the target convolutional kernel weight branch, and other data paths with zero multiplication results are shut down.

[0015] Preferably, the step of mapping a high-frequency transient discrete sequence to a two-dimensional phase space trajectory matrix includes: applying a mutual information algorithm to calculate the optimal delay time constant of the high-frequency transient discrete sequence in the phase space reconstruction process; applying a false nearest neighbor algorithm to calculate the optimal embedding dimension constant of the high-frequency transient discrete sequence in the phase space reconstruction process; and performing a phase space delay coordinate reconstruction operation on the high-frequency transient discrete sequence in one-dimensional sequence format based on the optimal delay time constant and the optimal embedding dimension constant to generate the two-dimensional phase space trajectory matrix.

[0016] Preferably, the step of performing spatiotemporal decoupling calculation and extracting anomalous distortion feature vectors by the activated target convolutional kernel weight branch includes: the calculation process of the target convolutional kernel weight branch includes cascaded spatial depth separable convolution operations and temporal causal convolution operations; performing the spatial depth separable convolution operation on the two-dimensional phase space trajectory matrix, using a one-dimensional spatial convolution kernel to perform local feature mapping calculation on the two-dimensional phase space trajectory matrix along the embedding dimension, generating a first feature tensor characterizing the geometric distortion of the phase space trajectory edge; performing the temporal causal convolution operation on the first feature tensor, performing causal convolution operation according to the time series unidirectional dependency mask, suppressing isolated jump data interference, compressing the spatial dimension parameters through a global average pooling layer, and outputting the anomalous distortion feature vector with a one-dimensional structure.

[0017] Preferably, the output of the anomaly confidence probability includes: retrieving the statistical variance of the cross-entropy loss distribution calculated by the target convolutional kernel weight branch based on a preset offline historical validation set; extracting the statistical variance and constructing the preset local smoothing constraint term by combining it with a preset penalty scaling factor; subtracting the preset local smoothing constraint term as a negative compensation parameter into the standard bias parameter of the abnormal state category in the classification fully connected layer to calculate the corrected classification bias, and setting the corrected classification bias as the dynamic margin discriminant bound; performing a tensor inner product operation on the abnormal distortion feature vector and the node weights of the classification fully connected layer, fusing the dynamic margin discriminant bound, performing an exponential normalization mapping via the Softmax activation function, and outputting the anomaly confidence probability.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention separates the power frequency fundamental wave data stream from the high-frequency transient discrete sequence and constructs a multi-mode load addressing matrix using the fundamental wave characteristics, enabling the perception of the current operating background status of the distribution network. This environmental perception and feature decoupling mechanism helps distinguish between conventional high-frequency noise generated during normal nonlinear load operation or start-up and shutdown and real physical anomaly signals. In a variable and complex electromagnetic coupling environment, this scheme improves the accuracy of anomaly detection and helps reduce the probability of false triggering of protection actions.

[0019] 2. This invention designs a soft-routing gating mechanism based on a differentiable neural architecture search network. During actual inference flow, this mechanism dynamically activates target convolutional kernel weight branches that match the current background based on the identified environmental features, and temporarily shuts down other unnecessary data paths. Combined with lightweight operations such as spatially depth-separable convolution, this architecture rationally allocates computational resources. This adaptive computational scheduling method enables deep learning networks to run with lower storage consumption and computational overhead, thus adapting to edge-side physical measurement devices with relatively limited computing power.

[0020] 3. This invention proposes a dynamic margin adaptive adjustment strategy based on offline historical statistical distribution. By extracting the variance of the loss distribution of the target branch on the validation set to construct a preset local smoothing constraint term, and using it as a negative bias to subtract from the anomaly state category logic output of the fully connected layer, the model can adaptively reduce the anomaly confidence probability during the inference phase based on its "historical uncertainty" when facing different types of noise. This compensation mechanism, which introduces a one-way prior penalty, improves the model's comprehensive anti-maloperation capability when dealing with complex and imbalanced data environments.

[0021] 4. This invention constructs an abnormal waveform analysis architecture with feature decoupling, computing power scheduling, and adaptive discrimination. This scheme reduces the protection malfunction rate by stripping away complex electromagnetic noise through environmental perception; it controls the computational overhead of the inference stage by allocating edge computing power on demand using a soft routing gating mechanism; and it improves the model's comprehensive adaptability under data imbalance conditions by employing a dynamic margin compensation strategy. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the overall anomaly analysis and discrimination process of the present invention. Figure 2 This is a flowchart illustrating the construction process of the multimodal load addressing matrix of the present invention. Figure 3 This is a flowchart of the soft routing gating mechanism based on neural architecture search of the present invention; Figure 4 This is a flowchart of the spatiotemporal decoupling calculation and abnormal distortion feature extraction of the present invention. Detailed Implementation

[0023] 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.

[0024] Please see Figures 1 to 4 This invention provides a method for analyzing and identifying anomalies in current waveform features based on deep learning. The technical solution is as follows: acquire discrete sampling signals of power distribution lines, split and extract the power frequency fundamental wave data stream and high-frequency transient discrete sequence; fuse the steady-state slow-varying envelope features extracted based on the power frequency fundamental wave data stream and the power frequency starting step envelope features, and perform unsupervised density clustering analysis to construct a multimodal load addressing matrix that characterizes the current power grid load operation background features.

[0025] The multimodal load addressing matrix is ​​used as a conditional trigger. A preset differentiable neural architecture search network is input to calculate and generate an architecture probability weight matrix. Based on the architecture probability weight matrix, the target convolutional kernel weight branch in the preset target depth network that matches the background noise under the current load operation background features is dynamically activated through a soft routing gating mechanism.

[0026] The high-frequency transient discrete sequence is mapped to a two-dimensional phase space trajectory matrix and fed into the target depth network. The spatiotemporal decoupling calculation is performed by the activated target convolutional kernel weight branch to filter out the switching transient background noise and extract the dimension-reduced abnormal distortion feature vector.

[0027] Obtain the preset local smoothing constraint term corresponding to the weight branch of the target convolutional kernel, adjust the Softmax activation function bias term in the classification fully connected layer at the end of the target depth network, and generate a dynamic margin discrimination bound corresponding to the current power grid background noise intensity; input the abnormal distortion feature vector into the adjusted classification fully connected layer, and output the distribution line anomaly confidence probability through the dynamic margin discrimination bound mapping.

[0028] Example 1: Further, the step of splitting and extracting the power frequency fundamental wave data stream and the high-frequency transient discrete sequence includes: applying a discrete wavelet packet decomposition algorithm to divide the frequency band boundary of the discrete sampling signal of the power distribution line, separating the current sampling points whose frequency range corresponds to the rated fundamental frequency of the power distribution system and the harmonics that are integer multiples of the rated frequency within a preset number, and recombining them into a power frequency fundamental wave data stream according to the time series; extracting and saving the edge detail current components whose frequency is within a preset high-frequency characteristic frequency band, and recombining them into a high-frequency transient discrete sequence according to the time series; extracting the effective value of the fundamental current and the current crest coefficient of the power frequency fundamental wave data stream within a preset number of continuous sliding time windows, and combining them to form a steady-state slowly varying envelope feature; extracting the first-order difference data of the current amplitude within a preset number of complete cycles after the zero-point level change point of the power frequency fundamental wave data stream, and combining them to form a power frequency start-up step envelope feature.

[0029] Specifically, discrete sampled signals of power distribution lines with a sampling rate of not less than 256kHz are acquired. The Daubechies wavelet family (e.g., db8) is selected as the mother wavelet, and the number of wavelet packet decomposition levels is set to [value missing]. (For example To accommodate the limited memory capacity of the microcontroller unit in the edge monitoring terminal, a directional frequency band pruning decomposition strategy is applied: abandoning the conventional full binary tree traversal calculation. During the layer-by-layer recursion of wavelet packet decomposition, the actual physical frequency span range corresponding to each node is calculated based on the sampling rate and the current layer number. Only nodes whose frequency range covers the target low-frequency band (including the rated fundamental frequency and preset harmonics, such as 0 to 1250Hz) or the target high-frequency characteristic band (such as 10kHz to 100kHz) continue to execute the decomposition operation of the next-level child nodes. For nodes whose frequency range falls completely outside the above two target frequency bands, their branch expansion is directly terminated and the memory cache occupied by that node is immediately released. Through directional pruning node reconstruction, current sampling points whose frequency range corresponds to the rated fundamental frequency and integer multiples of the rated frequency harmonics of the power distribution system are separated and reassembled into a power frequency fundamental frequency data stream according to the time series. Edge detail current components whose frequencies are within the preset high-frequency characteristic band range are extracted and saved, and reassembled into a high-frequency transient discrete sequence according to the time series.

[0030] By reconstructing wavelet packet nodes, frequency band nodes are separated that include the rated fundamental frequency of the power distribution system (50Hz or 60Hz) and its integer multiples of harmonics (typically up to the 21st to 25th harmonics, i.e., a frequency upper limit of 1050Hz to 1250Hz). Although it contains harmonic components, in the technical context of this invention, this low-frequency band generally characterizes the macroscopic operating mechanism of the load under the power frequency cycle. The above node components are superimposed and reconstructed in a time series, defined as the power frequency fundamental frequency data stream.

[0031] Extract and save edge detail current components whose frequencies fall within a preset high-frequency characteristic band (typically set to 10kHz to 100kHz). This band avoids low-frequency harmonic interference and high-frequency radio frequency noise, and is specifically designed to capture transient oscillations caused by load switching, arc discharge, etc. Reconstruct the node signals of this band and save them as a high-frequency transient discrete sequence in time series.

[0032] To characterize the steady-state properties of the load operation, a sliding time window feature extraction is performed on the power frequency fundamental wave data stream. The time window length is set to the length of a single power frequency cycle. (e.g., at a power frequency of 50Hz) =20ms). Set the preset quantity to (For example =50 represents the continuous extraction of envelope changes within 1 second. The sliding step size of adjacent time windows is set to one discrete sampling period to achieve data overlap and smoothing.

[0033] In the Within a time window, calculate the effective value of the fundamental current. The calculation formula is: in, The number of sampling points within a single time window. This represents the discrete current amplitude within the time window.

[0034] Extract the absolute value of the peak current within the same time window. And calculate the current crest coefficient. : Continuous A sequence of valid values ​​within a time window (length is...) One-dimensional sequence) and peak coefficient sequence (length is A one-dimensional sequence is assembled column-wise to form a structure of size [missing information]. The steady-state slowly varying envelope characteristic matrix of 2.

[0035] To extract the transient envelope containing the physical attributes at the moment of device switching, a first-order difference calculation is performed on the waveforms before and after the zero-level abrupt change point. This is to eliminate the random influence of drastic fluctuations in the starting current amplitude caused by closing the device at different voltage phase angles (such as voltage zero-crossing points and voltage peak points). The zero-level abrupt change point is specifically defined as: the starting moment when the amplitude change rate of the power frequency fundamental wave data stream exceeds the limit within five consecutive sampling points, and this moment corresponds to a specific delayed phase angle (such as a recorded value between 0 and 180 degrees) after the zero-crossing of the distribution network phase voltage. During extraction, this phase angle is recorded as an additional constraint dimension.

[0036] Starting from the abrupt change point, a preset number of periods M (e.g., M=3, i.e., 3 complete power frequency cycles) of the fundamental frequency data stream are extracted. To overcome the defect that the micro-differences between adjacent sampling points are submerged by quantization noise at high sampling rates, a macro-difference time step constant s is introduced (its value is set according to the system's sampling rate, typically corresponding to a physical time span of 1ms to 5ms; for example, at a sampling rate of 256kHz, s=256). Macro-first-order difference operations are then performed on the extracted discrete current sequence: Where n is the discrete sampling time number after the mutation point.

[0037] The absolute value of the envelope within a sliding window of half a power frequency cycle is extracted from the calculated difference sequence. This macroscopic difference envelope data not only effectively filters out steady-state bias constants and microscopic high-frequency quantization noise, but also accurately reproduces the macroscopic current rise rate characteristics during the initial stage of equipment startup. These are combined to form the power frequency startup step envelope characteristics. If the truncation period is insufficient due to data loss... For each cycle, zero-padding at the end of the data is used to maintain the consistency of the feature vector dimensions.

[0038] This invention addresses the problem of coupling between steady-state load and transient characteristics in distribution network signals. It decouples macroscopic steady-state features from microscopic transient details in the frequency domain through discrete wavelet packet decomposition. By combining a zero-point mutation alignment mechanism to extract the start-up step envelope features, it improves the feature fluctuation problem caused by the randomness of the initial phase angle of different devices closing, providing a stable data foundation for subsequent cluster analysis and enhancing the targeting and anti-interference capability of feature extraction.

[0039] Furthermore, the step of constructing a multimodal load addressing matrix by performing unsupervised density clustering analysis includes: performing a tensor concatenation operation along the data channel dimension on the steady-state slow-varying envelope feature and the power frequency starting step envelope feature, and inputting them into a density-based noise application space clustering algorithm model; calculating the Euclidean distance distribution within the neighborhood radius of the cluster core points, dividing sample points with Euclidean distance differences within a preset constant range into load equipment clusters of the same type, and calculating the spatial coordinates of the cluster centers of all load equipment clusters; extracting the spatial coordinates of all the cluster centers and the percentage of samples contained in each load equipment cluster, performing a combination operation according to a standard row and column format, and generating a multimodal load addressing matrix for characterizing the distribution status of equipment types within the distribution network.

[0040] Specifically, for the previously output steady-state slowly varying envelope features and the power frequency start-up step envelope features, since their physical dimensions are different (the former includes RMS values ​​and dimensionless peak coefficients, while the latter is the differential current amplitude), Z-score normalization is first performed on both sets of features. To address the significant misalignment in time series length between the two sets of features (the steady-state feature length is fixed to the preset number of time windows),... The transient feature length is determined by both the high-frequency sampling points and the truncation period. Therefore, a feature length alignment mechanism is introduced before tensor splicing: a one-dimensional adaptive max-pooling algorithm is used to uniformly compress and align the time axis length of the power frequency start-up step envelope feature to a specified value. This algorithm can preserve the most obvious current surge peak features during transient startup to the greatest extent while aligning dimensions, preventing key high-frequency distortion information from being diluted during significant compression. Subsequently, the standardized features are subjected to tensor concatenation along the data channel dimension to construct a multi-dimensional feature vector space.

[0041] Data from the multidimensional feature vector space is input into a density-based noise-based spatial clustering algorithm model. To avoid the subjective bias of manually preset constants and their incompatibility with different distribution zones, the neighborhood radius and density threshold of the cluster core points adopt a dynamic value selection logic: Get the total number of spliced ​​feature samples in the current calculation period. The density threshold is dynamically set based on the total sample size within the time window. To reduce the memory allocation overhead of the edge measurement terminal and accurately adapt to the local density distribution of data, a dynamic k-distance elbow method is used for optimization: calculating the distance from each sample point in the multidimensional feature vector space to its k-th elbow. The Euclidean distance between the nearest neighbors (where is set) The calculation results are arranged in ascending order to construct a k-distance sequence. The distance value corresponding to the maximum curvature in the distribution curve of this k-distance sequence is extracted and set as the initial elbow distance. Then, a preset margin decay constant is introduced. To balance inter-class edge bridging caused by high-frequency drift noise, the margin attenuation constant... The value range is limited to 0.10 to 0.25 (preferably 0.18, corresponding to retaining 75% to 90% of the initial elbow distance). The neighborhood radius for the current calculation cycle is calculated using the following formula. : Clustering core point determination and cluster partitioning: Calculate the Euclidean distance between any two sample points in the multidimensional feature vector space. When a sample point is in... The number of samples contained within the radius neighborhood is greater than or equal to When the time is right, mark it as a cluster core point (in a physical sense, a core point represents a specific type of load equipment, such as a variable frequency air conditioner or an induction motor, which is in a typical running or starting state within the current time window).

[0042] The Euclidean distance difference is in The core points within the area and their density-reachable boundary points are divided into load equipment clusters of the same type. For each of these load equipment clusters... Calculate the spatial coordinates of its cluster centers. The calculation formula is: in, Indicates the first The total number of sample points within each cluster is the multidimensional feature vector within the cluster.

[0043] Extract the spatial coordinates of the cluster centers of all clusters And calculate the percentage of samples contained in each load device cluster out of the total number of samples. The coordinate vectors and their corresponding percentage values ​​are combined using a standard row and column format to generate a multimodal load addressing matrix that characterizes the distribution of equipment types within the distribution network. It should be noted that there is a physical mapping relationship between the distribution of equipment types within the distribution network and the distribution of high-frequency background noise. This invention characterizes the background features of load operation through the multimodal load addressing matrix, essentially extracting the source-end topology prior information of high-frequency background noise. This implicit correlation between the load operation background features and specific high-frequency noise categories is pre-established by statistical analysis of historical samples during the subsequent offline joint training phase, thereby guiding the network to perform subsequent matching and routing.

[0044] If the power grid experiences a lightning strike or a severe short-circuit transient, the background noise will be extremely high. This will cause the noise spatial clustering algorithm to determine that all data points are outliers (i.e., the number of effective clusters is 0), and the multimodal load addressing matrix will face an empty set anomaly. At this time, bypass protection is triggered, and the "historical typical daily load addressing average distribution matrix" pre-stored in local non-volatile memory is called as the default matrix output to ensure that the subsequent neural architecture search network has a valid input tensor.

[0045] Subsequent searches for differentiable neural architecture networks (fully connected layers) require inputs of fixed dimensions. The upper limit of the preset matrix dimension is... The specific settings are based on: extracting the statistical distribution of the normal operation clustering results of the target distribution area over the past 30 days, and taking the 90th quantile of the number of equipment clusters with the largest historical effective load as the threshold. (Typical values ​​are set in this embodiment) =10). Each row of the multimodal load addressing matrix is ​​determined by the cluster center. The matrix is ​​composed of 3D feature space coordinates and the numerical value representing the proportion of each sample, i.e., the fixed number of columns is 1. +1.

[0046] If the number of device clusters generated by the current clustering Less than the preset upper limit of matrix dimension Then, zero-filling operation is used to fill in the remaining values. - Okay; if > Then sort by sample percentage from largest to smallest, and only the top few are selected. Data from the main load device clusters is used to discard the state of extremely small, weak edge devices. The final generated size is... The two-dimensional multimodal load addressing matrix is ​​flattened into a fixed-length one-dimensional feature vector before being fed into the continuous operating space of the differentiable neural architecture search network.

[0047] This invention employs an unsupervised density clustering algorithm to divide spliced ​​features into clusters. By calculating cluster centers and sample proportions, a multimodal load addressing matrix is ​​constructed, which can reflect the current distribution pattern of equipment in the distribution area. This mechanism reduces the model's dependence on prior human experience and improves the addressing matrix's adaptability to complex background noise.

[0048] Furthermore, the step of dynamically activating the target convolutional kernel weight branch through the soft routing gating mechanism includes: the preset differentiable neural architecture search network contains a continuous operation space; the multimodal load addressing matrix is ​​input into the continuous operation space; the normalized probability values ​​of the corresponding candidate branch nodes are calculated through a fully connected layer and Softmax normalization to generate the architecture probability weight matrix; based on the numerical distribution sequence in the architecture probability weight matrix, the Gumbel distribution sampling algorithm is applied to output discrete one-hot encoded vectors; the one-hot encoded vectors are multiplied by a matrix multiplication operation with the candidate feature processing path set inside the preset target depth network; independent data paths with non-zero multiplication results are defined and activated as the target convolutional kernel weight branch, and other data paths with zero multiplication results are shut down.

[0049] Specifically, the preset differentiable neural architecture search network is a lightweight routing module deployed on an edge computing terminal. This module has completed convergence through an offline joint training phase, and its training dataset covers typical load switching waveforms of power distribution lines in all four seasons, as well as background noise from various types of lightning strikes and switch oscillations.

[0050] To ensure the parameter determinism and engineering feasibility of the preset differentiable neural architecture search network and the preset target depth network, offline end-to-end joint initialization training of the network parameters of both is required before the power distribution line monitoring terminal is put into online operation. This joint training process covers sample set construction, loss function setting, and alternating gradient optimization strategy. Considering the limited computing power of edge computing terminals, this invention adopts a "teacher-student" joint distillation architecture on a cloud server. The specific steps are as follows: First, iterative training and soft label sequence generation of the cloud teacher network model are performed. A teacher network model with a relatively large initial parameter base is deployed on an independent server (cloud computing platform). Specifically, the teacher network model adopts a deep one-dimensional residual network (e.g., 1D-ResNet50) architecture, whose input is adapted to a long sequence one-dimensional feature tensor, contains multiple cascaded residual blocks, and connects a global average pooling layer and a fully connected layer at the end. An offline historical sample set with one-dimensional current waveform fault states and corresponding finely labeled environmental noise is retrieved, processed into a fixed-length feature vector, and then input into the teacher network model for multiple rounds of supervised iterative training. After convergence, a refining temperature coefficient greater than 1 is introduced. (For example, setting) =3), which is mapped through the Softmax layer to output a classification soft label sequence rich in globally distributed dark knowledge. Secondly, a joint edge network architecture is constructed. In a simulated server environment, a "differentiable neural architecture search network" to be deployed and a lightweight "student network model" (i.e., a pre-defined target depth network) are constructed in parallel. The differentiable neural architecture search network maps the multimodal load addressing matrix of the input through an internal fully connected layer; the student network model contains the aforementioned sets of candidate feature processing paths. To achieve spatiotemporal decoupling and satisfy the complete topology of network forward propagation, the overall architecture of the pre-defined target depth network consists of an input layer, a set of candidate feature processing paths, a global average pooling layer, and a classification fully connected layer cascaded sequentially. The input layer is used to receive a two-dimensional phase space trajectory matrix; the set of candidate feature processing paths contains C=3 parallel target convolutional kernel weight branches, each branch employing cascaded spatial feature extraction and temporal feature extraction operations. Specifically, branch one consists of 3×1 spatially separable convolutions (composed of a 3×1 spatial channel-wise convolution with a kernel size of 3×1, 64 output channels, a stride of 1, symmetric zero-filling at the edges, a Mish activation function, and a 1×1 pointwise convolution) cascaded with 3 layers of temporal causal convolutions (kernel length 7, output channels 64); branch two consists of 5×1 spatially separable convolutions (5×1 kernel) cascaded with the same 3 layers of temporal causal convolutions; branch three consists of dilated convolutions (3×1 spatial convolutions with a dilation rate of 2) cascaded with the same 3 layers of temporal causal convolutions. The spatial size and number of channels of the output feature maps of all branches are kept consistent to ensure dimensionality compatibility between the one-hot encoded vector and the dot product operation of the path set matrix. The outputs of each branch are ultimately compressed into a one-dimensional structure by a global average pooling layer and fed into the final classification fully connected layer.

[0051] Finally, joint backpropagation optimization of the architecture search and node weights is performed. During the offline joint training phase, features are extracted based on the aforementioned discrete sampling rate and time window length, constructing an offline historical sample set with a total sample size of no less than 50,000, and independently partitioning the training set and the test / validation set. The ratio of samples containing abnormal faults such as physical arcs to normal background samples in the original sample set is no less than 1:4. To prevent data leakage during the evaluation phase, during model iteration, SMOTE (Synthetic Minority Oversampling) is used only in the feature space of the training set to augment the feature space of minority class samples of fault arcs, overcoming the data imbalance caused by the scarcity of real physical arc waveform samples.

[0052] The offline historical sample set is input into the student network model, and the corresponding multimodal load addressing matrix is ​​input into the differentiable neural architecture search network.

[0053] Forward propagation: The differentiable neural architecture search network outputs architecture probability weights based on the addressing matrix, which are then sampled using a Gumbel distribution to activate the target candidate branches in the student network model. The student network model then outputs a predicted probability sequence after activating these branches. (This sequence has been fitted with a temperature coefficient aligned with the teacher network) (Smoothing is performed). Loss calculation: Calculate the classification soft label sequence. With the predicted probability sequence Relative entropy loss between The calculation formula is: Subsequently, the relative entropy loss Standard cross-entropy loss of student network model against real physical hard labels Perform weighted summation to construct the total distillation loss function. : Among them, the weighting coefficient The default value range is between 0.3 and 0.5.

[0054] Joint gradient update: The gradient is calculated based on the total distillation loss function using the backpropagation algorithm. This embodiment employs an alternating optimization strategy: independent Adam optimizers are configured for updating the differentiable neural architecture search network and the student network model, respectively. Based on the calculated gradient, the node weight parameters of the activated target candidate branches within the student network model are updated, with their initial learning rate set to... And apply a cosine annealing strategy to smoothly decay to a certain value within a preset training epoch (e.g., 200 epochs). On the other hand, the gradient of the total distillation loss with respect to the architecture probability distribution is calculated, and the weight parameters of the multilayer perceptron within the differentiable neural architecture search network are updated synchronously, with the learning rate of its architecture parameters independently set to a constant (e.g., ...). Through an end-to-end joint backpropagation mechanism, the total distillation loss function serves as an implicit evaluation metric for route optimization. When the multimodal load addressing matrix is ​​input, the system does not require manually provided preset route labels; instead, it autonomously tends to select the target convolutional branch that minimizes the current total distillation loss based directly on the gradient direction of backpropagation. This data-driven implicit mapping relationship enables the search network to adaptively select the optimal processing path that matches the background noise characteristics of the load operation context. This alternating independent update mechanism effectively prevents parameter collapse during the search for the optimal routing architecture and the convergence of local weights.

[0055] The multimodal load addressing matrix is ​​input into the routing module. The module internally constructs a continuous operation space, and performs feature dimensionality reduction and nonlinear mapping on the flattened one-dimensional multimodal load addressing matrix via a multilayer perceptron (MLP, i.e., a fully connected layer). Specifically, the MLP consists of an input layer, two hidden layers, and an output layer. The number of nodes in the input layer is strictly aligned with the dimensions of the flattened matrix. The first hidden layer has 64 nodes, and the second hidden layer has 32 nodes. Both layers use the LeakyReLU activation function to prevent gradient vanishing during feature mapping. To prevent severe routing bias (i.e., model collapse to a single branch) in the early stages of architecture search, the node weights of the MLP are initialized using a He normal distribution. Finally, the network passes through... The classification layer of each output node calculates the corresponding preset target depth network. Normalized fit probability values ​​of each candidate branch node (in ), by all This constitutes the architecture probability weight matrix.

[0056] To achieve the mapping from continuous probability space to discrete branch selection and to ensure the differentiability of the model during updates, the Gumbel-Softmax distribution sampling algorithm is applied.

[0057] Based on the numerical distribution sequence in the architecture probability weight matrix Inject noise terms, independently sampled from the standard Gumbel (0,1) distribution, into the logarithmic probability. Calculate the activation intensity of each candidate branch. : in, This refers to the degradation temperature coefficient. During the inference phase of real-time online monitoring of power distribution lines, to avoid numerical overflow or division-by-zero calculation errors when the underlying hardware executes exponential functions, the degradation temperature coefficient is... Set as a specific finite small positive number (e.g., set the temperature coefficient during the inference period). (The typical engineering value range is set between 0.001 and 0.1). Within this range, the above formula degenerates into an argmax discrete selection operation, such that the branch with the highest probability corresponds to... The output is 1, and the other branches output 0, thus outputting a value of length . The discrete one-hot encoded vector.

[0058] The generated one-hot encoded vector is multiplied by a matrix multiplication operation with the set of candidate feature processing paths within the predefined target depth network. Independent data paths whose multiplication result equals 1 are defined and activated as target convolutional kernel weight branches that match the background noise under the load's operating background characteristics.

[0059] The data flow path for any other dot product operation that results in 0 is shut down. In the underlying tensor computation framework, the shut-down branch does not participate in forward propagation memory allocation and floating-point operations, thereby reducing the static power consumption and inference latency of the edge monitoring device.

[0060] If the architecture probability weight matrix contains the maximum probability value Still below the preset confidence level (e.g., below) This indicates that the current power grid noise features are not covered by the offline training set, and the model is stuck in a state of ambiguity. At this point, Gumbel sampling is abandoned, and a preset safe default one-hot encoding is generated (usually pointing to the basic 3×1 spatially separable convolutional branch with the most stable parameters and the central receptive field) to prevent the model from choosing an extreme branch that would cause the extraction of distorted features to fail.

[0061] If multiple non-zero entries appear in the one-hot encoded vector due to hardware floating-point calculation precision issues, the non-zero bit with the smallest index number is truncated to 1 by the priority encoder, and the rest are set to 0 to ensure the physical uniqueness constraint of "single branch activation".

[0062] This invention introduces a differentiable neural architecture search and soft routing gating mechanism. One-hot encoding is generated through Gumbel distribution sampling, dynamically activating only the target convolutional branch that matches the current background noise, while physically shutting down other idle paths. This mechanism maintains the network's adaptability to various scenarios while reducing the dynamic power consumption and inference latency of edge computing terminals, thus improving resource utilization.

[0063] Furthermore, the step of mapping the high-frequency transient discrete sequence to a two-dimensional phase space trajectory matrix includes: applying a mutual information algorithm to calculate the optimal delay time constant of the high-frequency transient discrete sequence in the phase space reconstruction process; applying a false nearest neighbor algorithm to calculate the optimal embedding dimension constant of the high-frequency transient discrete sequence in the phase space reconstruction process; and performing a phase space delay coordinate reconstruction operation on the high-frequency transient discrete sequence in one-dimensional sequence format based on the optimal delay time constant and the optimal embedding dimension constant to generate the two-dimensional phase space trajectory matrix.

[0064] Specifically, the physical essence of the specific implementation steps for mapping a high-frequency transient discrete sequence into a two-dimensional phase space trajectory matrix is ​​based on Takens' embedding theorem. This maps the hidden nonlinear dynamic evolution laws in a one-dimensional time series to a high-dimensional topological space, thereby revealing the concealed transient characteristics caused by switching actions or arc discharges in the geometric form of attractors. The specific implementation process is as follows: Obtain the reconstructed one-dimensional high-frequency transient discrete sequence. To ensure that the reconstructed coordinate systems in phase space are both relatively independent and do not completely lose their correlation, the mutual information algorithm is used to calculate the optimal delay time constant. For a given delay step size Calculate the original sequence and the delayed sequence Shannon mutual information between : in, To estimate the probability density function of the corresponding sequence, the equal-width histogram method is used: the number of histogram bins is set. ,in Let be the total length of the high-frequency transient discrete sequence. Calculate the marginal probability density of the one-dimensional sequence based on this number of bins. and Simultaneously, a two-dimensional equal-width histogram is used to calculate the joint probability density. Gradually increase The value of will determine the mutual information. The step size corresponding to the first local minimum is set as the optimal delay time constant. .

[0065] If the currently captured high-frequency sequence is pure white noise or an extremely chaotic random transient, the mutual information function will be used within a preset maximum search step size (e.g., ...). Within a sample point (n), the correlation coefficient may exhibit a monotonically decreasing trend without a clear local minimum. In this case, the non-convergence truncation mechanism is triggered, switching to the autocorrelation function method to reduce the autocorrelation coefficient to its initial value. The corresponding step size is used as the default optimal delay time constant. This prevents the sequence from getting stuck in an infinite loop or from having excessively large values ​​that cause gaps in the sequence information.

[0066] The phase space must be large enough to ensure that attractor trajectories can fully unfold without self-crossing. Set the initial embedding dimension. Find any point in phase space nearest neighbor When the dimension changes from Increase to When the distance between two points changes, calculate the rate of change. Specifically, a joint determination is made using two threshold indicators: the first indicator (relative distance increase rate) calculates the change in distance between the two points. Euclidean distance in 3D reconstruction space and in The difference in distances in 3D space, divided by its distance in 4D space. The distance between two points in 3D space. When this ratio is greater than a preset first threshold (usually limited to 10 to 15), it is determined to be a false nearest neighbor; the second criterion (system geometric boundary constraint): calculate the distance between two points in 3D space. The ratio of the Euclidean distance in the 1D space to the standard deviation of the original 1D high-frequency transient discrete sequence. When this ratio is greater than a preset second threshold (usually limited to 2), it is also judged as a false nearest neighbor.

[0067] If either of the two criteria is satisfied, the node is designated as a false nearest neighbor. This varies with the dimension. The percentage of false nearest neighbors is calculated based on the increasing percentage of the total distance. Considering the interference from the actual background noise of the distribution network, a preset minimum threshold is set. (Limited to the range of 0.01 to 0.05, preferably 5%). When this percentage decreases to near or below... When the attractor is fully expanded, the corresponding dimension is... Set as the optimal embedding dimension constant .

[0068] If the high-frequency transient sequence contains discharge noise with extremely high degrees of freedom, the proportion of spurious nearest neighbors may continue to oscillate and fail to converge to the threshold. The following logic is activated at this point: The adaptive convergence protection logic is then initiated: The distribution curve of the proportion of false nearest neighbors is monitored in real time, and the dimension corresponding to the inflection point where the slope of this curve first flattens out is directly used as the optimal embedding dimension constant. Output. To prevent the computational dimension from expanding infinitely and causing the tensor size of subsequent 2D convolutional networks to overflow, a preset empirical upper limit is set (e.g., ...). =20). If iterates to The threshold was not reached. If no obvious slope inflection point is detected, the hardware computing power truncation mechanism is triggered, and the output is directly interrupted. As the optimal embedding dimension constant .

[0069] The optimal delay time constant obtained based on the above calculations With the optimal embedding dimension constant For a total length of One-dimensional high-frequency transient discrete sequence Perform phase space delay coordinate reconstruction operation.

[0070] Constructed two-dimensional phase space trajectory matrix For a size of The matrix, where Let be the total number of phase points after reconstruction. The mathematical expression of the matrix is: The two-dimensional phase space trajectory matrix The geometric edge features of the high-frequency transient signal in multidimensional space are preserved and used as a standard input tensor. These are then fed into the weight branches of the activated target convolutional kernel to perform subsequent spatiotemporal decoupling calculations.

[0071] Considering the requirement of fixed-dimensional input tensors for the underlying hardware operators of edge computing devices, when generating an initial size of... After obtaining the phase space trajectory matrix, interpolation or truncation operations are performed along the phase point temporal direction and the embedding dimension direction. To prevent conventional smooth interpolation algorithms from erasing the subtle geometric glitches at the attractor edges, nearest neighbor interpolation is used to discretize and map the two-dimensional phase space trajectory matrix, uniformly scaling and aligning it to a preset fixed-size matrix. (For example, setting) , The row directions of this fixed-size two-dimensional phase space trajectory matrix correspond to the geometric unfolding dimension of the attractor in phase space. There is a temporal correlation between adjacent rows with the optimal delay time constant. This correlation satisfies the semantic assumption of the local receptive field and provides a basis for extracting the geometric distortion features of the phase map edge through subsequent spatial depth separable convolution.

[0072] Further, the step of the activated target convolutional kernel weight branch performing spatiotemporal decoupling calculation and extracting the anomalous distortion feature vector includes: the calculation process of the target convolutional kernel weight branch includes cascaded spatial depth separable convolution operations and temporal causal convolution operations; performing the spatial depth separable convolution operation on the two-dimensional phase space trajectory matrix, using a one-dimensional spatial convolution kernel to perform local feature mapping calculation on the two-dimensional phase space trajectory matrix along the embedding dimension, generating a first feature tensor characterizing the geometric distortion of the phase space trajectory edge; performing the temporal causal convolution operation on the first feature tensor, performing causal convolution operation according to the time series unidirectional dependency mask, suppressing isolated jump data interference, compressing the spatial dimension parameters through a global average pooling layer, and outputting the one-dimensional anomalous distortion feature vector.

[0073] Specifically, the two-dimensional phase space trajectory matrix The topological tensor, possessing both time and state attributes, is input to the target convolutional branch. To accommodate the heterogeneous physical meaning of the row and column dimensions of the phase space matrix, a spatially depthwise separable convolution is first performed, using a size of [size missing]. An asymmetric one-dimensional spatial convolution kernel (e.g., 3×1 or 5×1) is used. This kernel performs local feature mapping computation only along the column vectors of the matrix (i.e., the embedding dimension direction), capturing the nonlinear topological correlation between different delayed coordinates at the same time, thereby generating a first feature tensor representing the transient geometry of the phase space trajectory, while avoiding premature smoothing interference in the time dimension.

[0074] The geometric topology of the phase space trajectory (such as the diffusion range of attractors and edge spikes) directly reflects the strength of the system's transient nonlinearity. This step focuses on extracting the geometric distortion features of the trajectory edges without performing cross-channel feature mixing, thus greatly reducing the number of network parameters and avoiding excessive smoothing of spatial information. If the optimal embedding dimension constant calculated in the previous step is... Extremely small (e.g.) <3) This results in the spatial dimension of the input matrix being smaller than the convolution kernel size, triggering an automatic padding mechanism. Symmetric zero-value padding is used to expand the dimensions of the upper and lower edges of the matrix, ensuring the validity of the convolution operation and the integrity of the feature receptive field.

[0075] After convolution, the first feature tensor representing the geometric distortion of the phase space trajectory edge is generated by mapping through a nonlinear activation function (such as ReLU or Mish).

[0076] The first feature tensor is fed into a temporal causal convolutional layer. To adhere to the temporal unidirectionality of the physical transient evolution of the power system (i.e., current distortion at the current moment can only be caused by the physical state at the current and past moments, and cannot be affected by future moments), a temporal series unidirectional dependency mask is introduced. This mask causes the convolutional kernel to be positioned along the time axis... The weight distribution on the data satisfies the following condition: for data points at future times, their connection weights are reset to 0.

[0077] Based on this one-way dependency mask, a one-dimensional convolution operation is performed along the time dimension. Its calculation formula can be expressed as the output at the current time step. Depends only on the input sequence A historic moment: in, For causal convolution kernel weights, This represents the temporal receptive field length of the convolution kernel.

[0078] By combining the convergent convolutional kernel weights fitted by the target deep network during offline training, this operation effectively suppresses the influence of isolated jump data points caused by random errors of measuring instruments or external high-frequency radio frequency interference that do not conform to the historical continuous transient evolution pattern on subsequent feature vectors (such noise often lacks causal continuity on the time axis), thereby purifying the persistent transient features caused by real physical faults in the power grid.

[0079] To avoid the time series length from shrinking drastically with increasing network depth due to causal convolution operations, asymmetric causal zero-value padding is used at the beginning (left side) of the time series in the time dimension. The padding length is... This ensures that the feature tensor output by the convolution is aligned with the input in the time dimension, preventing the loss of weak distortion features in the early transient stages during the convolution process.

[0080] After temporal causal convolution, the feature tensor contains high-dimensional spatiotemporal information from multiple channels. To eliminate the difference in temporal dimension length caused by different truncation durations and to prevent overfitting in fully connected layers, this tensor is input into a global average pooling layer.

[0081] The global average pooling layer calculates the arithmetic mean of each channel of the feature tensor over the entire two-dimensional space (spatial embedding dimension and temporal evolution dimension), thus transforming the three-dimensional feature tensor... (Number of channels) Spatial dimension (Time step) directly compressed and dimensionality reduced to a dimension of A one-dimensional structure vector. This one-dimensional structure vector has high translation invariance and temporal robustness. It is defined and output as the aforementioned abnormal distortion feature vector, providing standardized input data for subsequent classifiers.

[0082] This invention employs cascaded spatially depthwise separable convolutions and temporally causal convolutions to achieve spatiotemporal decoupling. Spatial computation independently extracts phase diagram edge deformations, while temporally causal masks ensure that features depend only on historical states and filter out random, discontinuous jump noise. Finally, pooling dimensionality reduction outputs anomaly distortion feature vectors with translation invariance, improving the purity of fault features.

[0083] Further, the output of the anomaly confidence probability includes: retrieving the statistical variance of the cross-entropy loss distribution calculated by the target convolutional kernel weight branch based on a preset offline historical validation set; extracting the statistical variance and constructing the preset local smoothing constraint term by combining it with a preset penalty scaling factor; subtracting the preset local smoothing constraint term as a negative compensation parameter into the standard bias parameter of the abnormal state category in the classification fully connected layer to calculate the corrected classification bias, and setting the corrected classification bias as the dynamic margin discriminant bound; performing a tensor inner product operation on the abnormal distortion feature vector and the node weights of the classification fully connected layer, fusing the dynamic margin discriminant bound, performing an exponential normalization mapping via the Softmax activation function, and outputting the anomaly confidence probability.

[0084] Specifically, a pre-stored offline historical verification set is retrieved from the computing terminal's memory. This verification set is independent of the training set of the pre-defined target deep network, and its data comes from historical waveform archives of the physical power distribution network induced by real switch operations, lightning interference, and typical faults (such as single-phase ground faults), and has been manually verified and labeled by experts. For the activated target convolutional kernel weight branch, the branch-specific statistical variance constant pre-stored in the edge device's memory is retrieved. The specific calculation mechanism is as follows: In the offline verification stage, the pre-defined offline historical verification set is input into the trained differentiable neural architecture search network, and the verification set is automatically divided into independent sample subsets corresponding to different branches according to the output one-hot encoding; for each candidate branch in the pre-defined target deep network, the cross-entropy loss sequence generated during its forward inference process in its corresponding sample subset is extracted, and a static scalar statistical variance is calculated. This variance characterizes the uncertainty and volatility of the current branch when dealing with a specific type of grid noise.

[0085] To break the translation invariance of the activation function and achieve conservative decision-making, the statistical variance is extracted. And combined with the preset penalty scaling factor Construct a pre-defined local smoothing constraint term : Among them, the penalty scaling factor This is used to map the dimensions from the cross-entropy variance to the logistic scale of the classification fully connected layer. This negative bias prevents overscaling that could lead to "false negatives" (i.e., missed detections) in anomaly detection. The specific values ​​are determined during the offline network calibration phase: within the preset value range of [0.1, 1.0], a grid search algorithm is applied to optimize the parameters of the validation set, selecting the optimal solution that minimizes the false positive rate of normal background samples and ensures that the false negative rate of real anomalous samples does not exceed a preset safety limit (e.g., 5%), and this optimal solution is then fixed as a deterministic hyperparameter in the target deep network. Subsequently, the bias term of the activation function is adjusted: this adjustment only applies to the "abnormal state category" predicted by the network. The corresponding value in the classification fully connected layer is... The standard bias parameter for each abnormal state category is denoted as . The preset local smoothing constraint term As a negative compensation parameter, it is unidirectionally subtracted from the standard bias parameter of the abnormal state category to generate the dynamic margin discrimination boundary corresponding to the current power grid background noise intensity. : For categories characterizing normal power grid conditions, the bias parameters remain constant. However, when the power grid background noise is highly complex, leading to variance... When increasing, a local smoothing constraint term is preset. The dynamic margin of the abnormal state category increases. The corresponding reduction (i.e., the abnormal bias is suppressed downwards) means that before the network outputs a high-confidence fault alarm, it needs to extract strong fault features from the abnormal distortion feature vector to overcome the negative bias, thus achieving a safety redundancy mechanism where the greater the noise, the more conservative the decision.

[0086] Obtain the anomalous distortion feature vector of a one-dimensional structure This is compared with the node weight matrix of the fully connected classification layer. Perform tensor inner product operations and fuse the dynamic margin discriminant bound for the abnormal state category. Standard bias parameters for normal state categories Its classification logical value vector The corresponding number in the middle Components of each category The specific formula for generating this formula is: in, For classifying the corresponding fully connected layer, the first The node weight vector of the nth category; when the nth category... When a category is defined as an abnormal state category, the corresponding bias term When the first When a category is defined as a normal state category, the corresponding bias term Components of all categories Combine to generate a complete classification logical value vector .

[0087] Classification logical value vector Exponential normalization mapping is performed via the Softmax activation function. To prevent numerical overflow caused by hardware computation of the discrete exponential function, which could result in a NaN (Not a Number) probability output, a maximum value cancellation mechanism is introduced. For the... The probability of certainty for each state category The mapping calculation formula is as follows: in, For vectors The largest logical value in the array.

[0088] The probability values ​​output by the above formula are constrained to the range of [0, 1] and sum to 1. The calculated probability of power line anomaly is then output to the monitoring terminal or protection execution unit. If this probability exceeds the preset dead zone (e.g., 85%), a trip protection command for the physical disconnect switch of the power distribution network is directly triggered.

[0089] This invention addresses the problem of overconfidence and misclassification in existing classification networks under complex noise interference. It utilizes the cross-entropy loss variance of the offline validation set to construct a local smoothing constraint term, generating a dynamic margin discriminant bound and superimposing it onto the bias of the fully connected layer. By introducing decision penalty compensation, this method achieves an adaptive shrinkage mechanism of the classification threshold with the intensity of background noise, enabling the model to make conservative decisions in strong noise environments and enhancing the reliability of anomaly classification.

[0090] This invention proposes an anomaly waveform analysis architecture that aligns with the nonlinear physical mechanisms of power distribution networks, addressing the limitations of traditional deep learning models in terms of scenario adaptability and computational overhead at the edge of the power Internet of Things (IoT). First, by decoupling time-frequency domain signals and reconstructing phase space delays, one-dimensional complex electrical parameters are mapped to physically meaningful multi-dimensional topological features, laying the foundation for data analysis. Second, unsupervised multimodal load addressing and differentiable neural architecture search are introduced, endowing the algorithm with the ability to adaptively and dynamically reconstruct data paths in heterogeneous power grid environments, thus optimizing the dynamic allocation of hardware computing power at the edge. Finally, a classification and discrimination system based on spatiotemporal causal decoupling and variance penalty margin is established to filter out broadband background noise and mitigate the risk of model misjudgment. The overall solution, while maintaining existing hardware sensing configurations, helps improve the anti-interference capability and computational efficiency in power distribution network state perception and anomaly detection.

[0091] Example 2: Furthermore, the step of dynamically activating the target convolutional kernel weight branch through the soft routing gating mechanism also includes introducing a hysteresis switching debouncing mechanism: after obtaining the architecture probability weight matrix of the current time window, the historical routing state mask of the adjacent previous consecutive time window is retrieved; when the optimal candidate branch corresponding to the maximum probability value in the architecture probability weight matrix of the current time window changes compared to the previous time window, and the difference between the maximum probability value and the second maximum probability value in the current architecture probability weight matrix is ​​less than a preset hysteresis penalty threshold, the branch switching instruction is blocked, and the target convolutional kernel weight branch of the previous time window is maintained as the currently activated branch state; when the difference is greater than or equal to the preset hysteresis penalty threshold, switching and activation are performed according to the current optimal candidate branch.

[0092] Specifically, during the inference phase of real-time online monitoring of power distribution lines, the edge device's memory cache contains a queue of a preset time window size to store historical routing state masks. Within a time window, the differential neural architecture search network outputs an architecture probability weight matrix. Maximum fit probability value and its corresponding candidate branch index Simultaneously extract the second largest fit probability value. .

[0093] Retrieve The historical routing state mask at any given time is used to obtain the index of the branch that was actually active in the previous time window. If determined Calculate the probability difference .like If the value is less than the preset hysteresis penalty threshold (e.g., set to 0.15), it indicates that the current environmental noise interference is causing the model's decision on the optimal branch to be in a critical oscillation state of ambiguity. At this point, the one-hot encoding switching instruction generated based on Gumbel distribution sampling is disabled, and the model is activated. The corresponding convolution branch is closed, and the remaining pathways are shut down; if If this indicates a significant and substantial shift in the background noise characteristics of the power grid, then branch switching is permitted, activating... The corresponding target convolutional kernel weight branch. This hysteresis determination logic is implemented by the underlying conditional branch instructions of the edge microcontroller, resulting in extremely low computational complexity.

[0094] This invention introduces a hysteresis switching debouncing mechanism based on probability difference in soft routing gating. Under conditions where strong background noise from the power grid causes high-frequency fluctuations in the probability distribution, this mechanism alleviates the high-frequency oscillation switching between different convolutional branches of edge computing devices, reduces the additional computational overhead caused by memory context switching, and maintains the smooth continuity of feature extraction on the time axis, thereby improving the online inference stability and computational efficiency of deep networks.

[0095] Furthermore, the step of generating the multimodal load addressing matrix includes empty set bypass protection logic: real-time monitoring of the number of load device clusters output by clustering calculation; if it is determined that the number of load device clusters generated in the current time window is zero, the empty set bypass protection logic is triggered, and the pre-fixed historical stable background reference addressing parameters in the edge data processing terminal are retrieved as alternative coordinate outputs to generate a multimodal load addressing matrix with a preset default dimension, so as to ensure that the forward inference tensor dimension received by the subsequent neural architecture search network is valid.

[0096] Specifically, after the density clustering calculation process is completed, the total number of currently divided effective load device clusters is counted. When encountering extreme electromagnetic interference transient conditions such as lightning strikes, the steady-state and transient envelope feature distributions exhibit a random walk state, resulting in all sample points having substandard neighborhood densities and being marked as outlier noise points by the algorithm. At this point, the following conclusions are drawn. .

[0097] Once this boundary condition is detected, the main control program immediately interrupts the regular combination operation of the current addressing matrix and triggers the empty set bypass protection. It reads the "historical stable background baseline addressing parameters" burned into the edge computing terminal's internal read-only memory (these parameters are the long-term statistical average of the cluster center coordinates of the target distribution area under historical normal operation and without significant abnormal noise). The baseline addressing parameters, along with the evenly distributed sample proportions, are combined according to a standard row-column format to generate the default state multimodal load addressing matrix. This ensures that regardless of how drastically the physical power grid waveform is distorted, the multimodal load addressing matrix fed into the subsequent differentiable neural architecture search network always maintains strict consistency in the tensor dimension, fundamentally preventing operational anomalies caused by dimensional mismatches in tensor computation frameworks (such as TensorFlow Lite / ONNX Runtime).

[0098] This invention addresses the problem of clustering algorithms easily outputting empty sets due to feature distribution wandering under extreme transient conditions (such as lightning strikes or severe short circuits). It designs an empty set bypass protection logic. When all feature points are judged as noise due to lack of density clustering, it can automatically switch to preset baseline addressing parameters, preventing subsequent network crashes due to receiving zero-dimensional tensors. This enhances operational stability and continuous monitoring capabilities in the early stages of sudden abnormal physical faults.

[0099] Furthermore, the step of calculating the optimal embedding dimension constant using the false nearest neighbor algorithm includes a dimension hard truncation constraint mechanism based on hardware computing power: during the iterative calculation of the embedding dimension, the number of iteration steps of the algorithm is monitored simultaneously; if the iteratively calculated embedding dimension reaches the preset hardware computing power truncation limit, and the currently calculated proportion of false nearest neighbors has not yet decreased to the preset minimum threshold, then the iterative loop of the false nearest neighbor algorithm is terminated, and the hardware computing power truncation limit value is directly output as the optimal embedding dimension constant.

[0100] Specifically, before performing the spurious nearest neighbor algorithm on the high-frequency transient discrete sequence, a preset hardware computing power truncation limit must be pre-calculated to ensure that subsequent two-dimensional tensor forward propagation operations do not experience memory overflow. This upper limit is based on the available random access memory capacity of the edge device's microcontroller unit. (Unit: bytes) This is derived from the deterministic mapping formula: in, To reserve a safety margin factor for the intermediate cache between the underlying operating system kernel and the convolution operator (the typical engineering value range is set to 0.4 to 0.6, and this example uses 0.5). For the fixed phase point sequence length of subsequent truncation or alignment, This is the number of system bytes occupied by a single element of a floating-point tensor (typically 4 bytes). The system rounds down according to this formula, outputting a hardware computing power truncation limit with defined physical computational boundaries (e.g., based on memory capacity). Take 122880 bytes and convert them to the desired value. ).

[0101] In the FNN algorithm, starting from the initial dimension In the iterative optimization loop that begins with progressively increasing levels, the proportion of spurious nearest neighbors at the current dimension is calculated in real time. In conventional signal analysis, when When the value is less than a preset minimum threshold (e.g., 5%), output the value at that time. This is the optimal embedding dimension constant. However, when the distribution network input signal contains purely random high-frequency white noise (whose theoretical phase space degrees of freedom tend to infinity), It will continue to oscillate within a high numerical range without converging. At this point, the algorithm logic simultaneously evaluates the loop variable. Once the conditions are determined... and Once established, the main control logic executes a jump instruction, exiting the iterative loop code block of the current dimension, and... The value is assigned to the output parameter. At this point, the output dimension parameter does not physically represent the actual topological unfolded dimension of the system, but is defined as the maximum observable projective dimension that the edge network can handle. Subsequently, the values ​​are directly truncated according to this to generate a two-dimensional phase space trajectory matrix.

[0102] This invention introduces a dimension hard truncation constraint mechanism based on hardware computing power during phase space reconstruction. Addressing the shortcomings of high-frequency pure random white noise, which can lead to non-convergence of the proportion of spurious nearest neighbors and the algorithm falling into infinite loops, this mechanism prevents memory overflows in the edge computing terminal during the subsequent construction of the two-dimensional phase space trajectory matrix by setting a physical computation boundary for the embedded dimension. This ensures finite expansion of nonlinear dynamic characteristics while improving the timeliness of edge-side engineering calculations and the security of hardware operation.

[0103] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A deep learning-based current waveform feature anomaly analysis and discrimination method, characterized in that, include: Acquire discrete sampling signals from power distribution lines, and split and extract the power frequency fundamental wave data stream and high-frequency transient discrete sequence; By integrating the steady-state slow-varying envelope features extracted from the power frequency fundamental wave data stream with the power frequency start-up step envelope features, and performing unsupervised density clustering analysis, a multimodal load addressing matrix characterizing the current power grid load operation background features is constructed. Specifically, the steady-state slow-varying envelope features and the power frequency start-up step envelope features are subjected to tensor splicing along the data channel dimension and input into the density-based noise application space clustering algorithm model. Calculate the Euclidean distance distribution within the neighborhood radius of the cluster core point, divide the sample points whose Euclidean distance difference is within a preset constant range into load equipment clusters of the same type, and calculate the spatial coordinates of the cluster center of all load equipment clusters; extract the spatial coordinates of all the cluster center and the percentage of the number of samples contained in each load equipment cluster, perform combination operations according to the standard row and column format, and generate a multimodal load addressing matrix to characterize the distribution status of equipment types within the distribution network. The multimodal load addressing matrix is ​​used as a conditional trigger. A preset differentiable neural architecture search network is input to calculate and generate an architecture probability weight matrix. Based on the architecture probability weight matrix, the target convolutional kernel weight branch in the preset target depth network that matches the background noise under the current load operation background features is dynamically activated through a soft routing gating mechanism. High-frequency transient discrete sequences are mapped to two-dimensional phase space trajectory matrices and fed into the target depth network. The spatiotemporal decoupling calculation is performed by the activated target convolutional kernel weight branch to filter out switching transient background noise and extract the dimension-reduced abnormal distortion feature vector. Obtain the preset local smoothing constraint term corresponding to the weight branch of the target convolutional kernel, adjust the bias term of the Softmax activation function in the fully connected layer of the target depth network, and generate the dynamic margin discrimination bound corresponding to the current background noise intensity of the power grid. The abnormal distortion feature vector is input into the adjusted classification fully connected layer, and the probability of power line anomaly confidence is output through dynamic margin discriminant boundary mapping.

2. The deep learning-based current waveform feature anomaly analysis and discrimination method according to claim 1, characterized in that: The steps of splitting and extracting the power frequency fundamental wave data stream and the high-frequency transient discrete sequence include: applying a discrete wavelet packet decomposition algorithm to divide the frequency band boundaries of the discrete sampling signal of the power distribution line, separating the current sampling points whose frequency range corresponds to the rated fundamental frequency of the power distribution system and harmonics that are integer multiples of the rated frequency within a preset number, and recombining them into a power frequency fundamental wave data stream according to the time series; extracting and saving the edge detail current components whose frequency is within a preset high-frequency characteristic frequency band, and recombining them into a high-frequency transient discrete sequence according to the time series; extracting the effective value of the fundamental current and the current crest coefficient of the power frequency fundamental wave data stream within a preset number of continuous sliding time windows, and combining them to form a steady-state slowly varying envelope feature; extracting the first-order difference data of the current amplitude within a preset number of complete cycles after the zero-point level change point of the power frequency fundamental wave data stream, and combining them to form a power frequency start-up step envelope feature.

3. The method for analyzing and identifying anomalies in current waveform features based on deep learning according to claim 1, characterized in that: The step of dynamically activating the target convolutional kernel weight branch through a soft routing gating mechanism includes: the preset differentiable neural architecture search network contains a continuous operation space; the multimodal load addressing matrix is ​​input into the continuous operation space; the normalized probability values ​​of the corresponding candidate branch nodes are calculated through a fully connected layer and Softmax normalization to generate the architecture probability weight matrix; based on the numerical distribution sequence in the architecture probability weight matrix, the Gumbel distribution sampling algorithm is applied to output discrete one-hot encoded vectors; the one-hot encoded vectors are multiplied by a matrix multiplication operation with the candidate feature processing path set inside the preset target depth network; independent data paths with non-zero multiplication results are defined and activated as the target convolutional kernel weight branch, and other data paths with zero multiplication results are shut down.

4. The method for analyzing and identifying anomalies in current waveform features based on deep learning according to claim 1, characterized in that: The step of mapping a high-frequency transient discrete sequence into a two-dimensional phase space trajectory matrix includes: applying a mutual information algorithm to calculate the optimal delay time constant of the high-frequency transient discrete sequence in the phase space reconstruction process; applying a false nearest neighbor algorithm to calculate the optimal embedding dimension constant of the high-frequency transient discrete sequence in the phase space reconstruction process; and performing a phase space delay coordinate reconstruction operation on the high-frequency transient discrete sequence in one-dimensional sequence format based on the optimal delay time constant and the optimal embedding dimension constant to generate the two-dimensional phase space trajectory matrix.

5. The method for analyzing and identifying anomalies in current waveform features based on deep learning according to claim 1, characterized in that: The step of performing spatiotemporal decoupling calculation and extracting anomalous distortion feature vectors by the activated target convolutional kernel weight branch includes: the calculation process of the target convolutional kernel weight branch includes cascaded spatial depth separable convolution operations and temporal causal convolution operations; performing the spatial depth separable convolution operation on the two-dimensional phase space trajectory matrix, using a one-dimensional spatial convolution kernel to perform local feature mapping calculation on the two-dimensional phase space trajectory matrix along the embedding dimension, generating a first feature tensor characterizing the geometric distortion of the phase space trajectory edge; performing the temporal causal convolution operation on the first feature tensor, performing causal convolution operation according to the time series unidirectional dependency mask, suppressing isolated jump data interference, compressing the spatial dimension parameters through a global average pooling layer, and outputting the anomalous distortion feature vector with a one-dimensional structure.

6. The method for analyzing and identifying anomalies in current waveform features based on deep learning according to claim 1, characterized in that: The output of the anomaly confidence probability includes: retrieving the statistical variance of the cross-entropy loss distribution calculated by the target convolutional kernel weight branch based on a preset offline historical validation set; extracting the statistical variance and constructing the preset local smoothing constraint term by combining it with a preset penalty scaling factor; subtracting the preset local smoothing constraint term as a negative compensation parameter into the standard bias parameter of the abnormal state category in the classification fully connected layer to calculate the corrected classification bias, and setting the corrected classification bias as the dynamic margin discriminant bound; performing a tensor inner product operation on the abnormal distortion feature vector and the node weights of the classification fully connected layer, fusing the dynamic margin discriminant bound, performing an exponential normalization mapping via the Softmax activation function, and outputting the anomaly confidence probability.

Images