An edge deployment-oriented power quality disturbance lightweight identification method

Abstract

The application discloses a power quality disturbance lightweight identification method for edge deployment. The method first converts a one-dimensional disturbance signal into a fusion angle domain image through a dual-domain angle mapping, and constructs a robust shrinkage teacher network to pre-train the fusion angle domain image to obtain noise-resistant discrimination ability; then an adaptive channel interaction lightweight network is constructed as an edge side identification model, local channel interaction attention, grouping feature generation and channel rearrangement mechanisms are introduced in the lightweight unit to enhance key disturbance feature expression and reduce model complexity; finally, through soft discrimination transfer training, the noise-resistant discrimination information learned by the teacher network is transferred to the edge side lightweight network. The method can maintain high identification accuracy under low computational complexity and small parameter size, significantly alleviate the dependence of edge monitoring devices on computing power, storage and real-time inference ability, and is suitable for power quality disturbance online identification in substation, power distribution terminal and other low-computing-power scenarios.

Description

Technical Field

[0001] This invention relates to the field of power quality monitoring and analysis in power systems, and specifically to a lightweight identification method for power quality disturbances in edge deployments. Background Technology

[0002] With the large-scale integration of new energy sources and the widespread use of power electronic equipment, power quality disturbances have evolved from simple disturbances to complex composite disturbances, exhibiting severe feature aliasing and significantly increasing the difficulty of identification. Traditional power quality disturbance classification methods typically consist of two stages: first, extracting the time-frequency features of the disturbance signal using time-frequency analysis methods such as Short-Time Fourier Transform (STFT), Wavelet Transform (WT), and S-Transform (ST); and then using classifiers such as Random Forest, Support Vector Machine (SVM), and Decision Tree to identify the disturbance type. However, with the increasing complexity of the power grid environment, traditional identification and classification methods, heavily reliant on human experience, can no longer meet the rapidly increasing demands for identifying power quality disturbance types in new power systems.

[0003] In recent years, deep learning methods have been widely adopted in the field of power quality disturbance identification due to their end-to-end feature learning capabilities and strong nonlinear fitting abilities. Existing research mainly focuses on optimizing architectures such as Convolutional Neural Networks (CNN), Long Short-Term Memory Networks (LSTM), and Deep Residual Networks (ResNet), significantly improving disturbance classification performance. However, two major bottlenecks remain. First, the original power quality disturbance signal is essentially a one-dimensional time series, and simply mining its temporal features is insufficient to fully capture deep feature correlations, leading to inadequate feature extraction, especially in complex disturbances and high-noise environments where recognition accuracy drops significantly. Second, existing deep learning models are complex in structure and have a large number of parameters, consuming excessive computational and storage resources, making them difficult to deploy in low-computing-power scenarios such as substation edge equipment. Some studies have attempted to utilize the feature extraction capabilities of image classification networks by transforming one-dimensional signals into two-dimensional feature maps (such as Markov transition fields, angular domain mapping, etc.), but the image transformation process may introduce additional computational overhead, and the model complexity is not significantly reduced.

[0004] To address the aforementioned issues, some improvements have attempted to strike a balance between recognition accuracy and model complexity. Examples include hybrid models integrating CNN and LSTM, CNN structures incorporating attention mechanisms, and classification methods based on continuous wavelet transform (CWT) and transfer learning. However, most of these studies are based on ideal simulation data and have not fully explored robustness under high-noise conditions, nor have they deeply considered the model's deployment adaptability and real-time inference performance on low-computing-power edge devices. Therefore, how to significantly reduce the computational complexity and number of parameters of the model while ensuring high recognition accuracy, making it suitable for resource-constrained scenarios such as substations, has become a pressing technical challenge in the field of power quality disturbance identification. Summary of the Invention

[0005] The technical problem to be solved in this application is: to address the key technical defects in existing power quality disturbance identification technologies, such as the difficulty in feature extraction and low identification accuracy caused by the aliasing of composite disturbance features, and the difficulty in deploying deep learning models with complex structures and large number of parameters on low-computing-power edge devices, a lightweight power quality disturbance identification method is proposed that combines high identification accuracy, low model complexity and edge deployment adaptability.

[0006] To address the aforementioned technical issues, this application proposes a lightweight power quality disturbance identification method for edge deployment. This method first employs a dual-domain angular mapping to convert a one-dimensional power quality disturbance time series into a fused angular domain image, enhancing the representation of local transient and global change features of the disturbance signal. Based on this, a robust shrinking teacher network is constructed as a high-precision noise-resistant model, and an adaptive channel interaction lightweight network is designed as the edge-side identification model. The edge-side identification model improves channel modeling capabilities through local channel interaction attention, reduces computational complexity, and enhances cross-channel information interaction through grouped feature generation and channel rearrangement mechanisms. Finally, soft discriminative transfer training is employed, with the pre-trained robust shrinking teacher network guiding the training of the edge-side lightweight network, achieving significantly reduced model complexity with minimal accuracy loss. This method includes the following steps:

[0007] S101, Acquires raw power quality disturbance signals And normalize it;

[0008] S102, the original power quality disturbance signal is converted into a fused angular domain image by using dual-domain angular mapping, the dual-domain angular mapping including global angular domain mapping and differential angular domain mapping;

[0009] S103, construct a robust shrinking teacher network, and pre-train the robust shrinking teacher network using the fused angular domain image to obtain a robust shrinking teacher network with noise-resistant discrimination capability;

[0010] S104, Construct an adaptive channel interaction benchmark lightweight network as an edge-side identification model. The adaptive channel interaction benchmark lightweight network includes local channel interaction attention, group feature generation, and channel rearrangement mechanism.

[0011] S105, The edge-side identification model is trained through soft discriminative transfer training, so that the edge-side identification model learns the soft discriminative information and true category label information output by the robust shrinking teacher network;

[0012] S106 uses the trained edge-side recognition model to classify and identify the fused angular domain image under test, and outputs the power quality disturbance type.

[0013] Optionally, the detailed steps of S102 include:

[0014] The overall trend of power quality signal changes is captured by global angular domain mapping; local fluctuations, transient changes and short-term changes of power quality signal are captured by differential angular domain mapping; and a fused angular domain image is obtained by combining them through spatial region-level stitching.

[0015] Optionally, the detailed steps of S103 include:

[0016] The robust shrinking teacher network embeds an adaptive shrinking layer on the basis of the residual feature extraction structure and is trained on the obtained image training set using a pre-training method to obtain a robust shrinking teacher network with noise-resistant discrimination ability.

[0017] Optionally, the detailed steps of S104 include:

[0018] In the bottleneck structure of the adaptive channel interaction lightweight network, the original channel attention structure is replaced by local channel interaction attention. The local channel interaction attention adaptively determines the cross-channel interaction range according to the number of input feature channels and does not perform dimensionality reduction. In the lightweight feature generation unit, a group feature generation mechanism is adopted, and a channel rearrangement operation is introduced after the group feature generation. The improved edge recognition model outputs feature maps, and the parameters of the entire model are initialized randomly.

[0019] Optionally, the detailed steps of S105 include:

[0020] The soft discriminative transfer training generates soft discriminative information by pre-training a robust shrinking teacher network, and combines it with real class labels to jointly guide the training of the edge-side identification model. During training, a preset transfer temperature, soft supervision weights, and learning rate scheduling strategy are adopted to enable the edge-side identification model to learn both the inter-class correlation information of the robust shrinking teacher network and the real label discriminative information.

[0021] The lightweight power quality disturbance identification method proposed in this application for edge deployment has the following advantages compared with traditional deep learning models and existing lightweight networks:

[0022] 1. High Model Lightweighting: This application employs a soft-discriminative transfer learning framework, using a high-precision robust shrinking teacher network to guide the training of an adaptive channel-interaction lightweight network. Compared to a highly complex robust shrinking teacher network, the edge-side identification model reduces floating-point computation by over 97%, parameter count by over 92%, and weight file size by over 91% in a 20dB noise environment, significantly improving the number of frames processed per second. It can be effectively deployed on resource-constrained hardware platforms such as substation edge monitoring equipment.

[0023] 2. Minimal Accuracy Loss: This application converts one-dimensional perturbation signals into fused angular domain images through dual-domain angular mapping, fully exploring local transient and global change features; at the same time, it adopts soft discriminative transfer training, enabling the lightweight model on the edge side to learn the inter-class correlation information contained in the robust shrinking teacher network output, maintaining high recognition accuracy while keeping low complexity.

[0024] 3. Excellent noise resistance: This application selects a robust shrinking teacher network as the teacher model. Its adaptive shrinking mechanism can effectively suppress noise and extract discriminative features in complex perturbation scenarios with high noise and severe feature aliasing. Through soft discriminative transfer training, this noise-resistant discriminative ability is transferred to the edge-side recognition model.

[0025] 4. Strong adaptability to edge deployment: This application enhances cross-channel information interaction while compressing the computational and parameter volume of the model through local channel interactive attention, group feature generation and channel rearrangement mechanism, avoiding the significant decrease in accuracy caused by simple pruning or channel reduction, and is more suitable for real-time online identification of low computing power terminals. Attached Figure Description

[0026] Figure 1 This is an overall flowchart of a lightweight identification method for power quality disturbances in edge deployment as described in an embodiment of this application.

[0027] Figure 2 This is a fused corner domain image composed of global corner domain mapping and differential corner domain mapping in an embodiment of this application. Wherein (a) is global corner domain mapping, (b) is differential corner domain mapping, and (c) is fused corner domain image.

[0028] Figure 3 It is a lightweight unit structure for adaptive channel interaction.

[0029] Figure 4 This is a schematic diagram of a soft-discriminative transfer training framework. Detailed Implementation

[0030] A lightweight identification method for power quality disturbances in edge deployments, specifically including the following implementation methods:

[0031] S101, Acquires raw power quality disturbance signals And normalize it;

[0032] S102, the original power quality disturbance signal is converted into a fused angular domain image by using dual-domain angular mapping, the dual-domain angular mapping including global angular domain mapping and differential angular domain mapping;

[0033] S103, construct a robust shrinking teacher network, and pre-train the robust shrinking teacher network using the fused angular domain image to obtain a robust shrinking teacher network with noise-resistant discrimination capability;

[0034] S104, Construct an adaptive channel interaction benchmark lightweight network as an edge-side identification model. The adaptive channel interaction benchmark lightweight network includes local channel interaction attention, group feature generation, and channel rearrangement mechanism.

[0035] S105, The edge-side identification model is trained through soft discriminative transfer training, so that the edge-side identification model learns the soft discriminative information and true category label information output by the robust shrinking teacher network;

[0036] S106 uses the trained edge-side recognition model to classify and identify the fused angular domain image under test, and outputs the power quality disturbance type.

[0037] Figure 2 This diagram illustrates the process of stitching together a global angle domain map and a differential angle domain map to form a fused angle domain image. The upper portion of the global angle domain map and the lower portion of the differential angle domain map are selected for stitching in this diagram.

[0038] Figure 3 This paper presents an adaptive channel interaction lightweight unit structure. In the original lightweight feature generation structure, the core feature map and auxiliary feature map are usually fused by direct concatenation, resulting in insufficient cross-channel information interaction and easy underutilization of subsequent features. To solve this problem, this application introduces channel rearrangement operation, which promotes cross-group information flow through group reorganization, dimension transformation, and flattening merging, thereby enhancing the interaction between features.

[0039] Figure 4 The soft-discriminative transfer training process is demonstrated. Before training begins, the robust shrinking teacher network is pre-trained and its parameters are frozen to provide soft-discriminative information. During transfer training, the robust shrinking teacher network generates soft-discriminative information at a preset transfer temperature, which, together with the real class labels, constitutes joint supervision. The main steps include soft-discriminative information generation, definition of the combined loss function, and iterative updates of the edge-side identification model.

[0040] To verify the effectiveness of the method described in this application, simulation experiments were conducted. The experimental setup is as follows: fundamental frequency... sampling frequency =5120Hz. Sampling lasted for 10 cycles, with each signal containing 1024 data points. In addition to the noise-free signal, noisy signals with signal-to-noise ratios of 20 dB and 40 dB were also generated. The generated signals were converted into a fused angular domain image through dual-domain angular mapping, constructing an image with a resolution of [resolution missing]. The dataset consists of pixels. This dataset is then divided into training, validation, and test sets in a 7:2:1 ratio.

[0041] Data augmentation techniques, including random pruning, scaling, and horizontal flipping, are employed during training to enhance the model's generalization ability. The edge-side identification model learns the noise-resistant discriminative information and real label information of the robust shrinking teacher network through soft discriminative transfer training.

[0042] This application further validates the method from two dimensions: model recognition performance and model complexity. Model recognition performance is primarily evaluated using overall accuracy (OA), F1 score, and frames per second (FPS). Overall accuracy represents the proportion of correctly classified samples to the total number of test samples; F1 score comprehensively measures the precision and recall of various perturbations; and FPS measures the model's real-time processing capability during actual inference. Model complexity is primarily evaluated using the number of parameters, floating-point operations per second (FLOPs), and weight file size. The number of parameters reflects the model's storage and memory overhead, FLOPs reflects the theoretical computational complexity of the model, and the weight file size reflects the storage space required for model deployment. The specific mathematical definitions of each metric are as follows:

[0043]

[0044]

[0045]

[0046]

[0047]

[0048]

[0049]

[0050] Where C represents the number of power quality disturbance categories; This represents the number of samples correctly identified as belonging to class i; N represents the total number of samples. This represents the precision of the i-th class, reflecting the proportion of samples that the model predicts to be of the i-th class that actually belong to that class; This represents the recall rate of the i-th class, reflecting the proportion of real samples in the i-th class that are correctly identified by the model; Let F1 score represent the F1 score of class i; M represents the total number of image samples for which the model completed inference. This represents the total time consumed by the model to complete the inference for these samples; Indicates the number of input channels; K represents the number of output channels; H represents the kernel size; H and W represent the height and width of the output feature map, respectively.

[0051] To further verify the overall performance of the proposed method compared with other lightweight networks, this application compares MobileNetv2, MobileNetv3, ShuffleNetv1, ShuffleNetv2, benchmark lightweight networks, benchmark lightweight network v2, and edge identification models under different signal-to-noise ratio conditions. All comparison models were tested under the same dataset partitioning, training epochs, and transfer training settings. The results are shown in Table 1.

[0052] Table 1. Recognition results of different models under different signal-to-noise ratio conditions.

[0053] Model Noise-free training 40dB of training alone Individual training 20dB Noise-free transfer training Transfer training 40dB Transfer training 20dB MobileNetv2 99.50% 99.50% 95.04% 99.73% 99.77% 96.65% MobileNetv3 99.42% 99.38% 95.23% 99.77% 99.65% 96.15% ShuffleNetv1 99.42% 99.19% 94.88% 99.77% 99.46% 95.85% ShuffleNetv2 99.50% 99.27% 94.92% 99.81% 99.58% 96.38% Benchmark Lightweight Network 99.50% 99.19% 95.15% 99.54% 99.54% 96.88% Benchmark Lightweight Networks v2 99.54% 99.27% 94.58% 99.62% 99.81% 96.50% Edge-side identification model 99.42% 99.31% 95.15% 99.69% 99.58% 96.85%

[0054] As shown in Table 1, each lightweight network achieved varying degrees of accuracy improvement after soft-discriminative transfer training, indicating that the soft-discriminative information output by the robustly shrinking teacher network can provide richer category relevance information for the edge-side identification model than hard labels. Under 20dB strong noise conditions, the overall accuracy of the edge-side identification model in this application after soft-discriminative transfer training reached 96.85%, significantly higher than MobileNetv3, ShuffleNetv1, ShuffleNetv2, and the benchmark lightweight network v2. While maintaining high recognition accuracy, it has lower parameter count and smaller weight file, making it more suitable for deployment in computing-constrained scenarios such as substation edge equipment.

[0055] To verify the compression effect of the edge-side identification model of this application relative to the robust shrinking teacher network, this application further compares the robust shrinking teacher network and the edge-side identification model, and the results are shown in Table 2.

[0056] Table 2 Comparison between the Robust Shrinkage Teacher Model and the Edge-Side Identification Model

[0057] Model OA F1 FLPOs Parameters FPS Weight size Robust shrinking teacher network 97.04% 97.04% 3693.90M 26.36M 38.40 90.99MB Edge-side identification model 96.85% 96.85% 102.48M 1.98M 49.22 7.65MB

[0058] As shown in Table 2, under 20dB noise conditions, the overall accuracy of the edge-side identification model is 96.85%, only 0.19 percentage points lower than that of the robust shrinking teacher network. However, its FLOPs decreased from 3693.90M to 102.48M, the number of parameters decreased from 26.36M to 1.98M, the weight size decreased from 90.99MB to 7.65MB, and the FPS increased from 38.40 to 49.22. In other words, compared with the robust shrinking teacher network, the edge-side identification model reduces computational complexity by approximately 97.2%, reduces the number of parameters by approximately 92.5%, reduces weight file size by approximately 91.59%, and increases inference speed by approximately 28.18%. These results demonstrate that this application transfers the noise-resistant recognition capability of the robust shrinking teacher network to the lightweight model through soft discriminative transfer training, significantly reducing hardware resource requirements while maintaining high recognition accuracy.

[0059] In summary, this application proposes a lightweight power quality disturbance identification method for edge deployment. It converts a one-dimensional disturbance signal into a fused angular domain image through dual-domain angular mapping, extracts noise-resistant discriminative knowledge using a robust shrinking teacher network, and optimizes the edge-side identification model structure through local channel interactive attention, grouped feature generation, and channel rearrangement mechanisms. Ultimately, it achieves an effective balance between recognition accuracy, noise resistance, and model lightweighting within a soft discriminative transfer training framework. Multiple simulation examples demonstrate that the proposed method maintains high power quality disturbance identification accuracy even in 20dB high-noise environments, while significantly reducing model floating-point computation, parameter count, and weight file size, thus improving real-time inference capabilities. This provides an effective technical solution for online power quality disturbance identification in substations, distribution terminals, and other low-computing-power edge devices, demonstrating strong engineering application value.

Claims

1. A lightweight identification method for power quality disturbances in edge deployments, characterized in that, include: S101, Acquires raw power quality disturbance signals And normalize it; S102, the original power quality disturbance signal is converted into a fused angular domain image by using dual-domain angular mapping, the dual-domain angular mapping including global angular domain mapping and differential angular domain mapping; S103, construct a robust shrinking teacher network, and pre-train the robust shrinking teacher network using the fused angular domain image to obtain a robust shrinking teacher network with noise-resistant discrimination capability; S104, Construct an adaptive channel interaction benchmark lightweight network as an edge-side identification model. The adaptive channel interaction benchmark lightweight network includes local channel interaction attention, group feature generation, and channel rearrangement mechanism. S105, The edge-side identification model is trained through soft discriminative transfer training, so that the edge-side identification model learns the soft discriminative information and true category label information output by the robust shrinking teacher network; S106 uses the trained edge-side recognition model to classify and identify the fused angular domain image under test, and outputs the power quality disturbance type.

2. The lightweight identification method for power quality disturbances for edge deployment according to claim 1, characterized in that, In step S102, the global angular domain mapping is used to characterize the overall trend of power quality disturbance signals, and the differential angular domain mapping is used to characterize local fluctuations, transient changes and short-term disturbance characteristics, and a fused angular domain image is formed by region stitching or channel stitching.

3. The lightweight identification method for power quality disturbances in edge deployment according to claim 1, characterized in that, In step S103, the robust shrinking teacher network includes a residual feature extraction structure and an adaptive shrinking structure, which are used to suppress redundant perturbations and extract discriminative features in a noisy environment.

4. The lightweight identification method for power quality disturbances in edge deployment according to claim 1, characterized in that, In step S104, the local channel interaction attention adaptively determines the cross-channel interaction range based on the number of input feature channels, and completes channel weight recalibration without channel dimensionality reduction; the group feature generation mechanism sets the number of groups based on the network size, number of input channels or number of output channels of the edge-side identification model, and introduces a channel rearrangement mechanism after the group features are generated to enhance cross-group information interaction.

5. The lightweight identification method for power quality disturbances in edge deployment according to claim 1, characterized in that, In step S105, the soft-discriminative transfer training includes using the soft-discriminative information output by the robust shrinking teacher network and the true class labels together as supervision signals, and applying them according to the loss function. Update the edge-side recognition model parameters, where, For soft supervision weights, For migration temperature, To robustly shrink the soft-discrimination probabilities output by the teacher network at the migration temperature, This represents the soft discrimination probability output by the edge-side identification model at the migration temperature. The cross-entropy loss is the value corresponding to the true category label. This represents the divergence loss between soft-discrimination probabilities.

6. The lightweight identification method for power quality disturbances in edge deployment according to claim 1, characterized in that, In step S106, the trained edge recognition model is deployed to the edge monitoring device, and the model's recognition performance, real-time inference capability, and storage overhead are comprehensively evaluated based on overall accuracy, F1 score, number of frames processed per second, number of parameters, floating-point operation volume, and weight file size.

7. The lightweight identification method for power quality disturbances for edge deployment according to claim 1, characterized in that, The edge-side identification model is deployed to the edge monitoring device to receive the power quality disturbance signal under test in real time, and then performs normalization processing, dual-domain angle mapping, edge-side model inference and disturbance category output in sequence.

Images