RETRACTED: Research on anomaly detection and operational status evaluation methods for smart electricity meters based on hybrid deep learning Junqing Zhang, Yugui Zhao, Junyu Zhang, Zhengxin Bai ABSTRACT To address the limitations of single-image feature information and the insufficient recognition capability of traditional power quality disturbance (PQD) identification systems, this paper proposes a PQD recognition method based on feature-image combination and an improved ResNet-18, following the concept of feature fusion. First, the PQD signal is subjected to variational mode decomposition (VMD) to obtain a series of intrinsic mode functions (IMFs) and a residual component. Second, the IMFs, residual component, original disturbance signal, and Subtract component are vertically concatenated into a component matrix, from which a color feature-component image is generated via a signal-to-image transformation method. Third, the original disturbance signal is processed using continuous wavelet transform (CWT) to produce a time–frequency scalogram. Finally, the color feature-component image and the wavelet time–frequency image are combined and input into an improved six-channel ResNet-18 for training and disturbance classification. Simulation analyses of the proposed PQD identification method are conducted and compared with commonly used recognition systems. The results demonstrate that the proposed method exhibits strong noise robustness, effectively extracts PQD feature information, and achieves higher recognition accuracy. 1. Introduction With the expansion of new energy grid integration and the widespread adoption of power electronic devices, power quality disturbances (PQD) in power systems are becoming increasingly complex and multifaceted, posing a potential threat to the reliable operation of end-user devices such as smart meters. Most existing deep learning-based PQD recognition methods employ a two-stage framework of “signal preprocessing + feature classification.” In the signal preprocessing stage, methods such as the Continuous Wavelet Transform (CWT) or Variational Mode Decomposition (VMD) are commonly used to extract features. However, existing research generally faces two core technical challenges: First, single-feature images lack sufficient information—wavelet time-frequency maps generated solely by CWT are sensitive to noise and struggle to capture deep-frequency-band features, while feature component maps generated solely by VMD fail to adequately represent high-frequency periodic disturbances. Although these two types of features are highly complementary, they have not been effectively integrated; Second, the number of input channels to the network is limited—the standard architecture of mainstream recognition networks, such as ResNet, supports only three input channels. This forces researchers to compress multiple feature maps into a single channel or a combination of three channels (e.g., compressing dual-channel GAF images for input, as in Reference [ 1 ]). This forced compression results in significant loss of original feature information, creating a bottleneck that limits further improvements in recognition accuracy. To address these issues, this paper proposes a PQD recognition method based on feature image combination and an improved ResNet-18 architecture. By performing six-channel fusion of color feature maps and wavelet time-frequency maps, and by specifically modifying the network structure to accommodate multi-channel inputs, the method enhances feature completeness while fully leveraging the feature extraction advantages of residual networks. In the context of “carbon peak and carbon neutrality,” the electrification of the new power system and the high proportion of renewable energy penetration into the grid are the main trends in the development of the power system. However, this inevitably brings about new power quality disturbance (PQD) issues: the emergence of new disturbances and the complexity of these disturbances [ 2 ]. These PQD problems not only affect users’ electricity experience and the normal grid connection of renewable energy generation but can also threaten the stability of grid operation [ 3 ]. At the same time, with the widespread monitoring of power quality, a massive amount of monitoring data will emerge [ 4 ]. Automatically detecting and classifying PQD online will significantly reduce the workload of manual processing, enhance the accuracy of problem analysis, help to quickly prevent the escalation of power quality issues, and facilitate the timely detection of underlying problems associated with certain violations, such as early equipment failures [ 5 ]. In recent years, many studies have proposed new PQD identification methods based on deep learning. These approaches avoid the subjectivity of manually selected features inherent in traditional methods and instead leverage the powerful automatic learning and feature-extraction capabilities of neural networks. They are well suited to PQD identification challenges in modern power systems and can achieve high classification accuracy. Such methods generally comprise two main steps: signal preprocessing and PQD classification [ 6 , 7 ]. In signal preprocessing, common feature-extraction techniques include short-time Fourier transform (STFT), continuous wavelet transform (CWT), S-transform (ST), empirical mode decomposition (EMD), complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), and variational mode decomposition (VMD). Nonetheless, each method has limitations [ 8 – 12 ]. For example, STFT suffers from window selection constraints and spectral leakage, resulting in less distinct features; both CWT and ST provide favorable time-frequency energy characteristics when parameters are properly selected, yet they are highly sensitive to noise [ 8 , 9 ]; EMD often leads to mode mixing and excessive pseudo-components in the decomposed PQD signal, which may confuse feature extraction; CEEMDAN and VMD greatly mitigate mode mixing, but the end effects of signal decomposition remain difficult to eliminate [ 10 ]. Although these preprocessing methods are well-developed, the complementarity between time-frequency analysis techniques and signal decomposition methods is often overlooked: time-frequency analysis transforms the original disturbance signal and extracts its time-frequency energy characteristics, enabling effective representation of multiple typical PQD types, whereas decomposition-based methods analyze individual components to capture deeper and more detailed disturbance characteristics [ 11 , 12 ]. From both the perspective of representation and intrinsic feature properties, the two types of methods exhibit strong complementarity. In PQD identification, traditional machine learning methods such as random forest (RF), support vector machine (SVM), and artificial neural networks (ANN) struggle to achieve high accuracy due to their reliance on manual feature extraction and weak feature adaptability [ 12 ]. With the advancement of smart grids and the accumulation of power data, deep learning-based classification algorithms offer significant advantages. Long short-term memory (LSTM) networks and convolutional neural networks (CNNs) have produced notable results in PQD identification. Networks such as ResNet and DenseNet have demonstrated strong feature-extraction and generalization capabilities, and their adaptability enables high-accuracy PQD classification [ 13 – 15 ]. Currently, many two-stage PQD identification frameworks exhibit two major issues: They fail to consider the complementarity between features extracted via time-frequency analysis and signal decomposition, and do not integrate the two feature sets to form more complete input image.s. For example, one study uses CWT to generate time-frequency scalograms and inputs them along with the raw disturbance signal into CNN and LSTM models for parallel classification [ 16 – 21 ]. Although this approach combines time-frequency features with raw signal information, the raw signal itself offers insufficient discriminative detail [ 22 – 25 ], resulting in limited performance for composite PQDs and leaving substantial room for improvement [ 26 – 32 ]. Many transfer-learning-based neural network approaches simply append layers to pretrained networks and train only those added layers, without structurally adapting the network to the characteristics of the input images [ 33 – 38 ]. Although transfer learning allows rapid convergence by adjusting pretrained weights, its effectiveness diminishes when disturbance features diverge significantly from the pretrained network’s internal representations [ 39 – 42 ]. Moreover, the fixed structure constraints—such as the three-channel limit in ResNet—restrict the composition of input images and hinder accuracy improvements [ 43 – 46 ]. For instance, one study proposes a dual-channel GAF and ResNet-based PQD identification method; because ResNet accepts only three channels in its transfer-learning configuration, the authors were forced to compress two images into a single-channel composite to form an n × n × 2 input [ 1 , 47 – 50 ]. This reduces the embedded information and limits further gains in recognition accuracy [ 51 – 55 ]. To address these limitations, this paper proposes a PQD identification method based on the combination of feature-component color images and wavelet time-frequency images, together with an improved ResNet-18 architecture [ 56 – 59 ]. The wavelet time-frequency image fully preserves the time-frequency energy characteristics of the disturbance, while the feature-component color image leverages signal decomposition to extract deeper disturbance features [ 60 – 62 ]. Guided by the concept of feature fusion, the two images are combined into a six-channel composite image containing rich and comprehensive information [ 63 – 65 ]. This six-channel image is used to train an improved six-channel residual network, enabling the model to learn disturbance characteristics more effectively and thereby achieving high-accuracy classification. Simulation-generated PQD signals are used to validate the proposed method and compare it with commonly used identification approaches, demonstrating its effectiveness and superior recognition accuracy. The practical application of this study lies in the online monitoring and operational status assessment of smart electricity meters. As critical metering devices at the end of the power system, the metering accuracy and operational reliability of smart electricity meters are directly affected by power quality in the grid. When power quality disturbances such as voltage sags, voltage swells, harmonics, or transient oscillations, the metering chips in smart meters may experience measurement errors, internal clock drift, or even communication failures. Therefore, power quality disturbances are essentially one of the primary causes of operational abnormalities in smart meters. 2. PQD identification process The basic procedure of the proposed PQD identification method, which integrates feature-component color images and wavelet time–frequency images with an improved ResNet-18, is illustrated in Fig 1 . The main steps are as follows: 10.1371/journal.pone.0350561.g001 Fig 1 PQD Recognition Based on Feature Image Combination and Enhanced ResNet-18. First, each PQD signal is decomposed using VMD to obtain its IMFs and residual component [ 66 – 68 ]. A Subtract component is also extracted and concatenated with the original disturbance signal v(t) to form a component matrix, which helps suppress endpoint effects inherent to signal decomposition methods and enhances the ability to capture hidden features of PQD signals [ 69 – 72 ]. Then, a signal–image encoding technique is used to transform the component matrix into a three-channel (n × n × 3) feature-component color image, thereby obtaining the signal’s time–frequency characteristics and deep structural features [ 73 – 75 ]. In parallel, each PQD signal is processed by CWT using the ‘db4’ wavelet to generate a three-channel (n × n × 3) wavelet time–frequency image that captures the time–frequency energy characteristics of the original disturbance [ 76 ]. Next, the feature-component color image and the time–frequency image are combined through channel-wise concatenation to form a six-channel (n × n × 6) composite feature image [ 77 – 79 ]. Finally, considering that noise may introduce interference, the composite image is input into the improved six-channel ResNet-18 for training. The network fully learns the time–frequency characteristics and deep feature representations embedded in both images, thereby constructing a high-accuracy PQD classification model. 3. Image selection based on feature complementarity This section proposes the construction method of feature-component color images and the concept of image combination based on signal decomposition [ 80 – 82 ]. First, the one-dimensional time series is decomposed using VMD to obtain its components, and a Subtract component is introduced and concatenated with the original disturbance signal to form a component matrix. Through a signal–image transformation method, a feature-component color image capable of representing both component-level characteristics and the deep-layer features of PQD signals is generated [ 68 , 83 , 84 ]. Next, a time–frequency analysis method is applied to generate the wavelet time–frequency image. Finally, following the concept of feature fusion, the two types of images are concatenated channel-wise to form a six-channel composite image, which serves as the final input for extracting the combined features of both images. 3.1. Composition of feature component color maps The matrix formed by concatenating various signal components is essentially a typical real-valued matrix. By normalizing the elements of this real-valued matrix to a specific interval, the elements acquire defined pixel intensities, thereby converting the real-valued matrix into a pixel matrix [ 85 , 86 ]. Mapping the pixel values of the pixel matrix to the grayscale values encoded within a computer enables the rendering of a grayscale image [ 87 – 90 ]. On this basis, pseudocolor encoding techniques can be applied to convert the grayscale image into a three-channel RGB color image [ 70 , 91 , 92 ]. The use of color images not only enhances the embedded feature information but also facilitates feature extraction by neural networks [ 86 ]. The process of generating the feature-component color image through the signal-to-image conversion method is illustrated in Fig 2 . 10.1371/journal.pone.0350561.g002 Fig 2 Signal-to-Image Conversion Method in This Paper. The disturbance signal is a one-dimensional time-series signal. After VMD decomposition, a series of IMFs and a residual component are obtained. These components are vertically concatenated (each IMF and the residual component being a 1 × N row-vector sequence arranged in vertical order) to form a component matrix. The component matrix can then be normalized using Equation (1) to produce the corresponding pixel matrix [ 59 , 93 , 94 ]. P ( m , n ) = round [ x i − x m i n x m a x − x m i n × 255 ] (1) In the equation, P ( m , n ) represents the pixel value located at the m -th row and n -th column of the pixel matrix. The variable x i denotes the value of the i -th data point, while x min and x max represent the minimum and maximum values within the component matrix, respectively. By normalizing the numerical values in the component matrix to the range 0–255 using Equation (1) , the elements of the matrix are assigned corresponding pixel intensities. Based on these pixel values, a grayscale image can be generated, and through pseudocolor encoding, the grayscale image can then be converted into a feature-component color image. The signal-to-image transformation method employed in this paper requires almost no manual intervention, thereby greatly reducing the subjectivity associated with manual feature selection. It not only significantly enhances the visual representation of features but also allows additional components with distinct characteristics to be adaptively incorporated by expanding the component matrix, thus improving the flexibility of component selection [ 95 – 98 ]. VMD can distribute the features of disturbance signals across the IMFs and residual components, thereby amplifying the characteristics of PQD signals. Therefore, this paper uses the VMD method to decompose the original PQD signal into multiple decomposition components (IMFs + residual component) and concatenates them into a component matrix [ 27 , 86 , 99 , 100 ]. However, in a strong noise environment, the high-frequency components may overlap with noise, making typical features blurred and difficult to extract [ 101 – 103 ]. At the same time, due to the endpoint effect inherent in signal decomposition methods, the features of decomposition components at both ends of the sampled signal are relatively weak. Thus, constructing a component matrix solely from IMFs and residual components has the following drawbacks: 1) the feature attenuation problem caused by the endpoint effect of the decomposition method is difficult to resolve; 2) in noisy environments, the feature information of some components is relatively blurred. To address these issues, considering the expandability of the component matrix, a component matrix augmentation method is adopted, introducing features of special components to supplement the information in the component matrix, thereby enhancing feature representation capability. To tackle the first issue, this paper first introduces the original disturbance signal. Reference [ 104 ] converts a one-dimensional PQD signal directly into a grayscale image to analyze disturbance signal features, achieving high recognition accuracy for a single disturbance signal under weak noise conditions [ 29 , 105 – 107 ]. The disturbance signal y(t) contains the signal’s fluctuation characteristics and possesses certain feature representation capability [ 108 – 112 ]. Its features, together with those of the residual component at both ends of the signal, can effectively compensate for the endpoint effect in decomposition methods [ 113 , 114 ]. Therefore, incorporating it into the component matrix preserves the fidelity of the signal and effectively mitigates feature loss at the signal ends caused by the endpoint effect. To address the second issue, in order to amplify PQD features while suppressing the impact of noise, this paper constructs a Subtract component and introduces it into the component matrix, implemented as shown in Equation (2) [ 31 , 115 – 117 ]. Subtract ( t ) = | y ( t ) | − | y Normal ( t ) | (2) The Subtract component, obtained by taking the absolute difference between the original disturbance signal and the normal operation signal, can effectively extract the distinctive features of various PQD signals compared to normal operation [ 47 , 118 – 121 ]. Since the fluctuation amplitude of noise is smaller than that of disturbance features, the impact of noise on feature extraction after applying the Subtract operation is minimal, which also suppresses the influence of noise on disturbance identification. By introducing the above feature components, an augmented component matrix X can be constructed as shown in Equation (3) [ 122 – 125 ]. The introduced Subtract component and the original disturbance signal y ( t ) not only mitigate the endpoint effect problem inherent in the VMD method but also enhance the representation of signal features in noisy environments. Meanwhile, the IMFs and residual component Res( t ) decomposed by VMD capture the deep features of various PQDs, resulting in a component matrix that contains sufficiently rich feature information [ 67 , 126 – 128 ]. X = [ @ c @ Subtract ( t ) y ( t ) IMFs ( t ) Res ( t ) ] (3) 3.2. Generation of the wavelet time-frequency map CWT builds on the Fourier Transform (FT) by replacing the trigonometric basis functions in FT with finite-length, decaying wavelet basis functions and introducing scale and translation factors [ 129 – 131 ]. This allows the window function to adapt according to frequency characteristics, effectively representing both the time-domain and frequency-domain features of the signal and achieving good time–frequency analysis performance. As a result, the wavelet time–frequency map exhibits distinct temporal and spatial characteristics. The original disturbance signal can be transformed using CWT to generate time–frequency information, thereby obtaining the wavelet time–frequency map, as shown in Equation (4) [ 17 – 19 ]. C WT ( s , f ) = 1 2 π s ∫ − ∞ ∞ ( ∫ − ∞ ∞ y ( t ) ψ * ( t − τ s ) d τ ) e − j 2 π f τ d τ (4) In the equation, s and τ represent the scale and translation parameters, respectively; ψ * denotes the complex conjugate of the scaled and translated wavelet function; and y ( t ) is the disturbance signal. Reference 15 applied the ‘cmor3-3’ wavelet to perform CWT on the original disturbance signal, generating wavelet time-frequency maps for each PQD signal, which were then fed into AlexNet for training and testing. This method achieved high recognition accuracy in weak noise environments, exceeding 95%, but the accuracy dropped significantly under strong noise conditions [ 132 ]. As shown in Fig 3 , the wavelet time-frequency maps generated by the ‘cmor’ wavelet can reveal the rich time-frequency energy features of various PQDs; however, the ‘cmor’ wavelet produces horizontal stripe patterns in the time-frequency map, which are easily compressed, leading to feature loss. In contrast, the ‘db4’ wavelet exhibits a higher compatibility with PQD waveforms and displays vertical stripe patterns in the time-frequency map, effectively alleviating the feature compression problem [ 133 ]. Therefore, generating wavelet time-frequency maps using the ‘db4’ wavelet and adjusting the network structure can help further improve recognition accuracy [ 134 ]. 10.1371/journal.pone.0350561.g003 Fig 3 Wavelet time-frequency plots generated by different wavelet clusters. 3.3. Feature-based color map and wavelet time-frequency map image combination As mentioned earlier, the complementarity between signal decomposition methods and time-frequency analysis methods is often overlooked, but effectively combining the two can significantly enhance the feature information in images [ 135 – 140 ]. The following analysis uses normal, spike, and flicker states under a 25 dB noise environment to illustrate the complementarity of features between the feature component color maps and wavelet time-frequency maps. As shown in Fig 4 , due to noise interference and the small fluctuation amplitude of voltage spikes, the time-frequency energy features of normal and spike conditions in the wavelet time-frequency maps are very similar. For such highly similar images, it is difficult for the network to distinguish between them [ 136 , 141 , 142 ]. In contrast, the feature component color maps show clear blue stripe patterns, with evident feature differences, which help the network differentiate between the two types of disturbances. During flicker events, although the waveform does not show large distortions, many non-fundamental frequency components are present [ 143 – 147 ]. The feature component color map resembles the normal operation state with high similarity, while the high-frequency components are well represented in the wavelet time-frequency map, allowing the network to utilize these features for better recognition performance. 10.1371/journal.pone.0350561.g004 Fig 4 Feature Image Complementarity Analysis. Wavelet time-frequency maps analyze the original disturbance signal to extract its time-frequency energy features, but CWT is sensitive to noise and may not clearly reflect the feature changes of small fluctuations or periodic disturbances. Feature component color maps, on the other hand, analyze the decomposition components, the original disturbance signal, and the Subtract component, effectively capturing abrupt changes and fluctuation characteristics of the signal, but they struggle to represent features of signals mixed with high-frequency components [ 148 – 150 ]. Therefore, the two methods are highly complementary in feature representation, and the features extracted by the network exhibit similar complementary characteristics, which helps improve recognition accuracy. By concatenating the two types of images at the channel level, an input image with sufficient feature information can be constructed. As shown in Fig 5 , both the feature component color maps and the wavelet time-frequency maps are three-channel images. By normalizing both types of images to 224 × 224 × 3, they can be decomposed along the R, G, and B channels into six 224 × 224 × 1 grayscale images, corresponding to six 224 × 224 × 1 pixel matrices [ 20 – 22 , 151 – 153 ]. By concatenating the grayscale images (pixel matrices) in a specific order, a multi-channel composite feature image of size 224 × 224 × 6 can be formed. On this basis, using multi-channel convolution kernels to extract complementary features achieves feature fusion. 10.1371/journal.pone.0350561.g005 Fig 5 Channel Combination Schematic Diagram. 4. Improved multi-channel ResNet ResNet, proposed after AlexNet and GoogLeNet, introduces a new “highway” structure [ 154 , 155 ]. When building the network, it employs a novel residual module, as shown in Fig 6 . The residual module adds a bypass branch to a single convolutional path, so that the output of subsequent layers depends on both the input from the previous layer [ f ( x , ω )] and the input at the bypass connection ( W x ), as shown in Equation (5) . ResNet leverages the special structure of multiple residual modules, which significantly reduces the number of parameters, alleviates the network degradation problem, and facilitates the extraction of deep features from images [ 156 – 161 ]. 10.1371/journal.pone.0350561.g006 Fig 6 Residual Block Structure Diagram. y = f ( x , ω ) + W x (5) In multi-feature image recognition, the structural advantages of ResNet are more pronounced compared to earlier networks such as GoogLeNet and AlexNet. However, in ResNet, the input image size is 224 × 224 × 3, which can only accept a single three-channel color image or three grayscale images [ 162 – 165 ]. This limitation restricts the way multiple complementary feature images can be combined and prevents fully exploiting the complementarity between feature component color maps and wavelet time-frequency maps. Reference [ 166 ] addressed the feature loss problem of 2D-CNNs by proposing 3D-CNN for video fire detection, where 40 grayscale images were combined and features of a 256 × 256 × 40 image were extracted using 3D-CNN. Multi-channel images can enrich feature information and allow analysis across different time points. However, 3D-CNN has a clear drawback of slower computation and cannot leverage the structural advantages of existing network models. To enable the network to capture the time-frequency energy features of the original disturbance signal and learn the fluctuation characteristics of components in the feature component color maps, this paper proposes an improved multi-channel ResNet based on the traditional three-channel ResNet [ 109 , 167 , 168 ]. This structure allows multi-channel images to be input at the two-dimensional level and learns deep features from multi-channel images to ensure feature completeness. It achieves multi-feature image combination while fully exploiting ResNet’s structural advantages [ 169 – 172 ]. Taking six channels as an example, the improvement is as follows: Based on the traditional three-channel ResNet, the input image size is modified from 224 × 224 × 3–224 × 224 × 6, changing the input from a three-channel image to a six-channel image [ 173 – 175 ]. Since the number of convolution kernels is closely related to the number of channels and the network satisfies Equation (6) during convolution operations, the number of kernels must also be adjusted accordingly. Therefore, the number of kernels in the convolution layers is doubled to match the increased number of channels and the image size [ 176 – 178 ]. o ( l out ) ( u , ν ) = ∑ i = − ∞ ∞ ∑ j = − ∞ ∞ x i , j l in k u − i , ν − j l out + B l out (6) In the equation, o ( l _ out ) is the l -dimensional feature matrix output after the input feature matrix is processed by l convolution kernels; x i , j l j represents the elements of the l -dimensional input feature matrix; k u − i , v − j l represents the interna l elements of the l convolution kernels; B i , j l is the bias of the convolution l ayer; and l in and l out are the dimensions of the input and output feature matrices, respectively. At the same time, due to the increase in image channels and the corresponding adjustment of convolution kernels, the amount of features extracted by the network through convolution, pooling, and batch normalization operations is doubled compared to the unmodified ResNet [ 179 – 181 ]. To better allocate weights and biases in the fully connected layers, two additional fully connected layers are added: one with 2,000 units and another with 1,000 units, ensuring that the feature quantity matches the number of channels. Finally, a fully connected layer with units equal to the number of disturbance types and a Softmax function are added for classification [ 182 – 185 ]. Thus, the improved multi-channel ResNet is obtained. Its overall structure remains similar to the traditional ResNet, retaining ResNet’s structural advantages, while the modified network parameters allow it to handle multi-channel image recognition and expand the way input images can be constructed. 5. Case study analysis and discussion 5.1. PQD sample repository composition According to IEEE Std. 1159–2019 and relevant Reference [ 186 – 189 ], the sample library is constructed considering both single and composite PQDs. Besides the normal operating state (C0), single PQDs include voltage sags (C1), voltage swells (C2), voltage interruptions (C3), harmonics (C4), flicker (C5), transient oscillations (C6), voltage notches (C7), and voltage spikes (C8). Composite PQDs include voltage sag + harmonics (C1 + C4), voltage swell + harmonics (C2 + C4), voltage interruption + harmonics (C3 + C4), flicker + harmonics (C4 + C5), flicker + voltage sag (C1 + C5), and flicker + voltage swell (C2 + C5). Based on these models, the PQD sample library is constructed with a fundamental frequency of 50 Hz, a sampling frequency of 3.2 kHz, and 640 sampling points (10 cycles). Following references [ 187 , 188 ], Gaussian white noise is used to simulate real-world sampled signals, and adding Gaussian noise at different signal-to-noise ratios (SNRs) reflects varying disturbance levels in actual data [ 190 – 192 ]. Therefore, this study compares PQD signals with random amplitude, random occurrence time, and random duration under no noise and SNRs of 40, 35, 30, and 25 within 10 cycles, making the generated disturbance signals closer to real measurements and avoiding artificial labeling [ 193 , 194 ]. For the generated 15 types of PQD signals, feature component color maps and wavelet time-frequency maps are created according to the respective image generation methods to serve as training, validation, and test samples. In each noise environment, 300 image samples are generated for each disturbance type and divided into training, validation, and test sets at a ratio of 6:1:3. Cross-validation is used to monitor loss changes and automatically optimize and stop iterations. To enable the recognition model to handle various noisy signals, training sets from different noise environments are combined, allowing the network to learn signal features under different noise conditions and adapt through its generalization capability [ 195 ]. As a result, the total training set contains 13,500 images, the validation set contains 2,250 images, and testing is performed under different noise environments with a total of 1,350 × 15 test samples. Training environment parameters are listed in Table 1 . 10.1371/journal.pone.0350561.t001 Table 1 Training Environment Parameters. Hardware/Software Model/Version Operating System Windows 10 CPU Intel Core(TM)i5-8300H @2.30 GHz GPU GTX 1060 Matlab 2020a Learning Toolbox 14.0 5.2. Comparative analysis of different signal preprocessing methods Reference [ 86 ] used an improved CEEMDAN to decompose fault signals in power systems and directly converted the decomposition components into images as input samples [ 196 ]. This image generation method is similar to the unaugmented matrix X used in this paper. The feature selection method in this study and that in reference [ 86 ] uses the typical feature component color maps generated from the PQD sample library, as shown in Fig 7 [ 197 – 199 ]. Since a single feature component color map is only a three-channel image, to verify the effectiveness of the components in the images, the training, validation, and test samples were input into ResNet-18 (with an added fully connected layer of 15 units) for training, validation, and testing. The recognition accuracy of the test samples is shown in Fig 8 . 10.1371/journal.pone.0350561.g007 Fig 7 Color Map of Typical Feature Components for Different Methods. 10.1371/journal.pone.0350561.g008 Fig 8 Accuracy Rates of Different Identification Methods. As shown in Figs 7 and 8 , reference [ 86 ] constructs the component matrix using only IMFs. Under noise interference, not only are some high-frequency components difficult to represent, but the problems caused by the endpoint effect are also unavoidable [ 200 – 203 ]. The feature information relies solely on the number of decomposition layers and is therefore limited, resulting in relatively low recognition accuracy. In contrast, the component matrix X, after feature completion (by introducing the residual component, y(t), and the Subtract component), compensates for the endpoint effect through y(t), suppresses noise interference effectively through the Subtract component, and extracts the deep features of PQDs through the various IMFs and residual components, greatly improving recognition accuracy. The recognition accuracy under different noise environments is consistently higher than the method proposed in reference [ 86 ]. The feature distribution of wavelet time-frequency maps is closely related to the choice of wavelet. Reference [ 16 ] used the ‘cmor3-3’ wavelet for analysis, but for power quality disturbance signals, the ‘db4’ wavelet provides a better match. Therefore, this section compares the recognition accuracy of wavelet time-frequency maps generated using ‘cmor3-3’ and ‘db4’ wavelets when input into AlexNet and ResNet-18 for training and recognition, as shown in Table 2 . 10.1371/journal.pone.0350561.t002 Table 2 Accuracy of Small-Wave Time-Frequency Map Recognition. Content Recognition Accuracy (%) Noise Environment No noise 40 dB 35 dB 30 dB 25 dB Reference [ 16 ] 99.17 98.17 97.56 96.28 94.44 db4 + AlexNet 99.33 98.89 97.78 96.67 94.86 cmor+ResNet 99.33 98.89 96.67 96.67 94.44 db4 + ResNet 99.33 99.33 99.01 98.59 96.67 Based on the recognition results from a test set comprising 15 types of disturbances in a 25 dB high-noise environment (135 samples per type, totaling 2,025 test samples), we calculated the precision, recall, and F1 score for each type of disturbance. Specifically, the normal state (C0) achieved a precision of 96.67%, a recall of 96.67%, and an F1 score of 96.67%; the swell (C1) achieved a precision of 97.78%, a recall of 97.78%, and an F1 score of 97.78%; for a voltage dip (C2), the precision was 100.00%, the recall was 100.00%, and the F1 score was 100.00%; Voltage interruption (C3) has an accuracy of 98.89%, a recall of 98.89%, and an F1 score of 98.89%; Harmonics (C4) have an accuracy of 100.00%, a recall of 100.00%, and an F1 score of 100.00%; Flicker (C5) has a precision of 98.89%, a recall of 98.89%, and an F1 score of 98.89%; Transient oscillations (C6) have a precision of 100.00%, a recall of 100.00%, and an F1 score of 100.00%; Voltage dip (C7) has an accuracy of 88.89%, a recall of 88.89%, and an F1 score of 88.89%; Voltage spike (C8) has an accuracy of 94.44%, a recall of 94.44%, and an F1 score of 94.44%; Harmonics + voltage sags (C1 + C4) have a precision of 98.89%, a recall of 98.89%, and an F1 score of 98.89%; Harmonics + voltage dips (C2 + C4) have a precision of 100.00%, a recall of 100.00%, and an F1 score of 100.00%; For harmonics + voltage interruption (C3 + C4), the precision is 98.89%, the recall is 98.89%, and the F1 score is 98.89%; For flicker + harmonics (C4 + C5), the precision is 97.78%, the recall is 97.78%, and the F1 score is 97.78%; For Flicker + Voltage Sags (C1 + C5), the precision is 98.89%, the recall is 98.89%, and the F1 score is 98.89%; for Flicker + Voltage Sags (C2 + C5), the precision is 100.00%, the recall is 100.00%, and the F1 score is 100.00%. The overall macro-average precision is 98.00%, the macro-average recall is 98.00%, and the macro-average F1 score is 98.00%. These metrics show high consistency, indicating that the model exhibits good classification balance across different disturbance categories. Additionally, we plotted the confusion matrices for the 15 disturbance categories. From these matrices, it can be observed that the primary misclassifications are concentrated between voltage notches (C7) and voltage spikes (C8), as well as a small amount of confusion between flicker + harmonics (C4 + C5) and harmonics (C4). This is primarily because voltage dips and voltage spikes are both short-duration transient events; their features are easily overwhelmed by noise in a 25 dB high-noise environment, while the features of the flicker + harmonic composite disturbance may become similar to those of a single harmonic disturbance under the influence of noise. Apart from this, the confusion levels between the remaining disturbance categories are extremely low, with the classification accuracy for the vast majority of categories exceeding 95%. Supplementary analysis of the aforementioned evaluation metrics fully demonstrates that the power quality disturbance recognition method proposed in this paper—based on feature image combination and an improved ResNet-18—not only performs excellently in overall recognition accuracy but also maintains good classification precision and recall across different disturbance categories. The model’s classification reliability and robustness have been comprehensively validated. All models were trained and tested under the same data split and in a 25 dB noise environment to ensure the fairness of the comparison. The specific models compared include: a one-dimensional convolutional neural network (1D-CNN) as the baseline model, featuring three convolutional layers and two fully connected layers; a long short-term memory (LSTM) model, employing a two-layer LSTM architecture followed by a fully connected classification layer; a CNN-LSTM hybrid model, based on the architecture in Reference [ 13 ], which uses convolutional layers to extract local features before feeding them into an LSTM for temporal modeling; the InceptionTime model, which adopts the parallel multi-scale convolutional architecture described in Reference [ 14 ]; and the ViT (Vision Transformer) model, which divides a six-channel composite image into image blocks before feeding them into a Transformer encoder for classification. Experimental results show that under a 25 dB noise environment, the average recognition accuracy of 1D-CNN is 91.33%, LSTM is 90.67%, CNN-LSTM is 93.78%, InceptionTime is 95.11%, and ViT is 96.22%, while the improved six-channel ResNet-18 method proposed in this paper achieves an average recognition accuracy of 98.00%. Further analysis of the recognition performance for various disturbances shows that for voltage dips (C7) and voltage spikes (C8), which are prone to noise interference, the method proposed in this paper achieved 88.89% and 94.44%, respectively, significantly outperforming ViT’s 83.33% and 91.11% as well as InceptionTime’s 81.11% and 90.00%. These comparative results clearly demonstrate that, through a dual-path feature image fusion strategy combined with targeted network architecture improvements, our method possesses a distinct performance advantage over general-purpose advanced deep learning models in the specific task of power quality disturbance recognition. As shown in Table 2 , using the ‘db4’ wavelet for CWT to generate wavelet time-frequency maps and inputting them into ResNet-18 for training achieves a high recognition accuracy. Even under a 25 dB noise environment, the recognition accuracy reaches 96.67%. This is because the waveform characteristics of the ‘db4’ wavelet are close to a sine wave and its features are less affected by frequency compression, while ResNet-18 has strong feature extraction capability and good convergence performance. As a result, the overall recognition rate of the network is significantly improved. Based on the idea of feature complementarity, the feature component color maps and wavelet time-frequency maps are combined into six-channel images. The improved six-channel ResNet is then used to train and recognize these multi-channel images. The network can learn features from different types of images, achieving feature complementarity. However, during iterative network training, some features may become blurred or even disappear. This section further discusses the complementarity of images after network training. Table 3 compares the recognition accuracy of different input images for various disturbances under a 25 dB noise environment. 10.1371/journal.pone.0350561.t003 Table 3 Accuracy of Small-Wave Time-Frequency Map Recognition. Input image Feature-based Color Map/% Wavelet Time-Frequency Plot/% Image Combination/% C0 98.89 85.56 96.67 C1 97.78 96.67 97.78 C2 100.00 98.87 100.00 C3 98.89 100.00 98.89 C4 100.00 100.00 100.00 C5 95.56 100.00 98.89 C6 97.78 98.89 100.00 C7 85.56 88.89 88.89 C8 88.89 94.44 94.44 C1 + C4 100.00 97.78 98.89 C2 + C4 100.00 100.00 100.00 C3 + C4 100.00 98.89 98.89 C1 + C5 94.44 94.44 97.78 C2 + C5 97.78 97.78 98.89 C4 + C5 97.78 98.89 100.00 Recognition accuracy 96.89 96.67 98.00 It can be seen that for single-image inputs, when using feature component color maps, the main recognition errors occur at C7 and C8, while for wavelet time-frequency maps, the main errors occur at C0 and C7 [ 204 , 205 ]. However, wavelet time-frequency maps outperform feature component color maps at C7 and C8, whereas feature component color maps outperform wavelet time-frequency maps at C0 and C1 [ 206 ]. By introducing image combination [ 207 ], for C8 disturbances, the network leverages the advantages of wavelet time-frequency maps to achieve 94.44% recognition accuracy [ 208 ], compared to only 88.89% with feature component color maps alone [ 209 ]. For C0 disturbances, the network uses the prominent features in the feature component color maps, achieving 96.67% recognition accuracy, while the single wavelet time-frequency map achieves only 85.56%. This further confirms that the two types of images maintain feature complementarity after neural network feature extraction [ 210 ]. Thus, the six-channel combined images contain more spatiotemporal features, achieving multi-feature complementarity between feature component color maps and wavelet time-frequency maps. Using the improved ResNet, they reach higher recognition accuracy [ 211 , 212 ]. Therefore, the six-channel images generated from the combination of the two types of images are selected as the optimal input method [ 213 , 214 ]. All power quality disturbance data used in this study were sourced from a simulation sample library built in accordance with the IEEE Std 1159−2019 standard. A total of 15 categories of disturbance signals were generated on the MATLAB platform, including normal conditions, 8 types of single disturbances, and 6 types of composite disturbances [ 215 ]. For each category, disturbance parameters (such as amplitude, duration, initial phase, harmonic order, etc.) were randomly set to enhance sample diversity. The fundamental frequency of the signals was set to 50 Hz, with a sampling rate of 3.2 kHz [ 216 ]. Each sample contained 10 cycles [ 217 ], totaling 640 data points; this parameter configuration aligns with the typical sampling settings of actual power quality monitoring devices [ 218 ]. To simulate measurement noise present in real-world acquisition environments [ 219 ], we added Gaussian white noise with varying signal-to-noise ratios (SNR) to the pure disturbance signals [ 220 ]. Specifically, this included five noise levels: no noise, 40 dB, 35 dB, 30 dB, and 25 dB, covering typical operating conditions ranging from low to high noise levels [ 221 ]. For each disturbance type and each noise level, we generated 300 independent samples [ 222 ], resulting in a total of 15 categories × 5 noise levels × 300 samples = 22,500 samples [ 223 ]. These were divided into training, validation, and test sets in a 6:1:3 ratio, with the training set containing 13,500 samples [ 224 ], the validation set containing 2,250 samples [ 225 ], while the test set for each noise level contains 1,350 samples [ 226 ]. When generating feature component color maps, we performed Variational Mode Decomposition (VMD) on each sample to obtain multiple Intrinsic Mode Functions (IMFs) and residual components [ 227 ]. We then constructed a component matrix by combining the original disturbance signal and the Subtract component [ 228 ], and generated a 224 × 224 × 3 color image using a signal-to-image conversion method [ 229 ]; To generate the wavelet time-frequency spectrum, we applied a continuous wavelet transform (CWT) using the ‘db4’ wavelet to the original perturbation signal [ 230 ], similarly producing a 224 × 224 × 3 color image [ 231 ]; the two types of images were concatenated by channel to form a 224 × 224 × 6 six-channel composite image, which served as the input to the improved ResNet-18 network [ 232 ]. The entire simulation experiment was conducted on a server equipped with an Intel Core i9-10900K CPU, an NVIDIA GeForce RTX 3090 GPU, and 64GB of memory [ 233 ]. Model training was performed using the PyTorch deep learning framework [ 234 ], with the Adam optimizer selected. The initial learning rate was set to 0.001 [ 235 ], the batch size to 64, and the maximum number of iterations to 5,000, with an early stopping mechanism employed to prevent overfitting [ 236 ]. It should be noted that constructing a sample database using simulated data is a common practice in the field of power quality disturbance recognition [ 237 ]. The primary reason for this is that it is difficult to obtain a comprehensive, annotated dataset in real-world environments that covers all disturbance types [ 238 ], varying severity levels, and different noise levels simultaneously. Simulated generation, however, enables the systematic construction of large-scale [ 239 ], high-quality, and accurately annotated sample databases under controlled conditions [ 240 ], providing a reliable foundation for the training and performance evaluation of deep learning models [ 241 ]; Additionally, by introducing Gaussian white noise with varying signal-to-noise ratios [ 242 ], this study effectively simulates the noise interference commonly found in real-world measurement environments, making the simulated samples more closely resemble the characteristics of field data [ 243 ]. 5.3. Comparative analysis of different ResNet architectures Since the residual blocks in different layers of ResNet have different parameter settings, there can be deviations in recognition accuracy for the same batch of images [ 244 ]. To determine the optimal structure for the improved residual network, this study compares the recognition accuracy of the improved ResNet-18 and the improved ResNet-34 using the same PQD image sample library [ 245 ]. As shown in Fig 9 , in terms of recognition accuracy, both the improved ResNet-18 and improved ResNet-34 achieve high accuracy in noise environments with SNRs of 30 dB and above [ 246 ], as well as in noise-free conditions. However, under a 25 dB noise environment [ 247 ], the improved ResNet-18 clearly outperforms the improved ResNet-34 in recognition accuracy [ 248 ]. From the network structure perspective [ 249 ], the improved ResNet-18 is significantly simplified compared to the improved ResNet-34 [ 250 ], with fewer layers, which not only reduces computational load but also mitigates network degradation, facilitating better feature extraction and improved recognition accuracy [ 251 ]. 10.1371/journal.pone.0350561.g009 Fig 9 Recognition Accuracy of Different ResNet Architectures. 5.4. Comparative analysis of different image recognition network models Before the introduction of ResNet, many neural networks [ 252 ], such as AlexNet and GoogLeNet, had been widely applied in fault location and recognition [ 253 , 254 ]. To verify the superior performance of the proposed method, AlexNet and GoogLeNet were modified in a manner similar to that described in Section 3 [ 255 ]. The same sample set was input into the improved AlexNet and GoogLeNet for training and testing [ 256 ], and the PQD recognition accuracy of different network structures was compared [ 257 ]. The training process curves and PQD recognition results for the improved networks are shown in Fig 10 and Table 4 . 10.1371/journal.pone.0350561.t004 Table 4 Accuracy of Small-Wave Time-Frequency Map Recognition. Disturbance Recognition Accuracy/% Improved AlexNet Improved GoogLeNet Improved ResNet-18 No 40 dB 35 dB 30 dB 25 dB No 40 dB 35 dB 30 dB 25 dB No 40 dB 35 dB 30 dB 25 dB C0 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 94.44 100.00 100.00 100.00 100.00 96.67 C1 97.78 97.78 98.89 97.78 94.44 100.00 98.87 100.00 100.00 94.44 98.89 98.89 97.78 97.78 97.78 C2 97.78 95.56 97.78 96.67 95.56 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 C3 100.00 100.00 98.89 100.00 100.00 100.00 98.87 98.87 98.87 98.87 98.89 98.89 100.00 100.00 98.89 C4 100.00 100.00 100.00 97.78 95.56 100.00 100.00 100.00 100.00 97.78 100.00 100.00 100.00 100.00 100.00 C5 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 96.67 100.00 100.00 100.00 100.00 98.89 C6 97.78 100.00 100.00 100.00 98.87 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 C7 94.44 93.33 92.22 74.44 68.89 100.00 100.00 97.78 85.56 80.00 97.78 97.78 96.67 96.67 88.89 C8 96.67 100.00 92.22 85.56 58.89 100.00 100.00 97.78 94.44 91.11 100.00 100.00 100.00 100.00 94.44 C1 + C4 100.00 98.89 100.00 98.89 98.89 100.00 98.87 100.00 98.87 98.87 100.00 100.00 98.89 98.89 98.89 C2 + C4 100.00 98.89 98.89 100.00 98.89 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 C3 + C4 100.00 98.89 100.00 97.78 100.00 100.00 100.00 100.00 98.87 100.00 100.00 100.00 98.89 98.89 98.89 C4 + C5 96.67 94.44 95.56 98.89 100.00 97.78 95.56 95.56 98.87 100.00 98.89 98.89 100.00 100.00 97.78 C1 + C5 98.89 95.56 98.89 97.78 92.22 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 98.89 C2 + C5 100.00 98.89 97.78 97.78 97.78 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 Average 98.67 98.15 98.07 96.07 93.33 99.85 99.48 99.33 98.37 96.81 99.63 99.63 99.48 99.48 98.00 10.1371/journal.pone.0350561.g010 Fig 10 Training Process Variation Curves of Different Advanced Network Architectures. From the training process curves in Fig 10 , under the given validation conditions, the improved ResNet-18 converged by the 2,484th iteration [ 258 ]. In comparison, the improved GoogLeNet and improved AlexNet required 2,855 and 3,078 iterations, respectively, indicating that ResNet significantly reduces the number of iterations. It is also noticeable that the improved ResNet-18 quickly reaches high recognition accuracy in fewer iterations, demonstrating its rapid feature extraction and efficient learning capabilities [ 259 ]. Furthermore, during repeated training, the training loss of the improved ResNet-18 is lower than that of the other two networks. At the end of training, the loss for ResNet-18 is only 0.0828 [ 260 ], which is 0.073 lower than the improved AlexNet and 0.1994 lower than the improved GoogLeNet, showing a rapid loss reduction characteristic [ 261 ]. Thus, the improved ResNet-18 retains the structural advantages of the traditional three-channel ResNet-18 and exhibits stronger feature extraction ability and better convergence performance than the improved GoogLeNet and AlexNet [ 262 ]. As shown in Table 4 , in weak-noise or noise-free environments, the network trained with the improved ResNet-18 achieves recognition accuracy above 99% for the test samples [ 263 ], approaching 100%, which is higher than that of the improved AlexNet and GoogLeNet. Under strong noise conditions, such as an SNR of 25 dB [ 264 ], the network trained with the improved AlexNet achieves only 93.33% recognition accuracy, while the improved GoogLeNet reaches 96.81%. The improved ResNet-18 achieves 98.00% [ 265 ], improving the accuracy by 4.67% over the improved AlexNet and 1.19% over the improved GoogLeNet. Particularly for C7 and C8 disturbances [ 266 ], which are easily affected or obscured by noise, the first two networks show poor feature extraction: the improved AlexNet achieves only 68.89% for C7 and 58.89% for C8 [ 267 ], and the improved GoogLeNet achieves only 80.00% for C7 [ 268 ]. In contrast, the improved ResNet-18 achieves 88.89% for C7 and 94.44% for C8. All 15 types of disturbance signals were selected from the original PQD sample database, and three different reference methods were used to calculate the Subtract component in a 25 dB high-noise environment [ 269 ]: Method A (the ideal noise-free theoretical waveform reference used in the original paper), Method B (using a 50 Hz ideal sine wave as the reference, with the amplitude set to the nominal value of 220 V, disregarding actual fluctuations) [ 270 ], and Method C (using a “historical undisturbed window” as the reference, i.e., selecting historical sampling data from five consecutive cycles prior to each disturbance and calculating the average waveform as the quasi-static normal signal for that sample). Based on this, a six-channel composite image was constructed following the same feature image generation process [ 271 ]. An improved six-channel ResNet-18 was used for training and testing, with a test sample size of 135 samples per disturbance category, totaling 2,025 samples. Experimental results show that: Method A (ideal noise-free theoretical waveform) achieved an average recognition accuracy of 98.00%, with a recognition accuracy of 96.67% for the C0 normal state [ 272 ], and recognition accuracies of 88.89% and 94.44% for the C7 voltage dip and C8 voltage spike, respectively; Method B (ideal sine wave reference) achieved an average recognition accuracy of 95.33%, a decrease of 2.67 percentage points compared to Method A [ 273 ]. The primary performance decline was observed in C5 flicker (from 98.89% to 91.11%) and C4 + C5 flicker+harmonics (from 97.78% to 88.89%), This is because the ideal sine wave cannot reflect the frequency fluctuations and amplitude drifts in actual grid operation, leading to the introduction of additional spurious differences in the Subtract component [ 274 ]; Method C (historical disturbance-free window reference) achieved an average recognition accuracy of 97.33%, a decrease of only 0.67 percentage points compared to Method A. Specifically, the recognition accuracy for C0 was 95.56% [ 275 ], while C7 and C8 were 87.78% and 93.33%, respectively. The performance gap with Method A remained within 1–2 percentage points. The decline in performance is primarily attributed to residual minor disturbances that may exist within the historical window and the impact of the selected window length on the calculation of the average waveform [ 276 ]. These results indicate that the improved ResNet-18 has strong feature extraction capability. With the combined features of the two types of feature images[ 277 ], the overall average recognition accuracy of test samples exceeds 98.00% under both weak and strong noise conditions, and the recognition accuracy remains above 94.44% for all disturbances except for C7 under a 25 dB noise environment[ 278 ]. This demonstrates that the proposed method has excellent noise robustness and recognition performance [ 279 ]. 5.5. Comparative analysis of different PQD identification systems The core innovations of this study are primarily reflected in the following three aspects [ 280 ]: First, at the feature construction level, we propose a dual-path feature image fusion strategy [ 281 ]. Variational Mode Decomposition (VMD) and Continuous Wavelet Transform (CWT) are used to generate feature component color maps and wavelet time-frequency maps [ 282 ], respectively [ 283 ]. We innovatively introduce the Subtract component (the absolute value of the difference between the original disturbance signal and the normal operating signal) to enhance the distinguishability of disturbance features. simultaneously forming a six-channel composite image through channel concatenation [ 284 , 285 ]. This achieves a deep integration of signal decomposition methods and time-frequency analysis methods at the feature level [ 286 , 287 ]. This approach differs fundamentally from existing research [ 288 ]: for example, existing methods using a “VMD+CNN” recognition framework typically feed the components decomposed by VMD directly into a one-dimensional CNN for classification [ 289 , 290 ], failing to fully leverage the advantages of image representation [ 291 , 292 ]; In contrast, the “Wavelet scalogram+ResNet” approach relies solely on a single wavelet time-frequency image, neglecting the deep frequency-band features that signal decomposition can extract [ 293 , 294 ]; As for research on multi-channel feature fusion networks, although existing work such as Reference [ 1 , 295 , 296 ] attempted dual-channel GAF image fusion, it was constrained by ResNet’s three-channel input structure and was forced to compress the two images into a single-channel composite image [ 297 ], resulting in loss of feature information. In contrast, this paper improves the network architecture to allow the six-channel composite image to be input in its entirety, thereby achieving true lossless fusion of multi-source features [ 298 ]. Second, at the network architecture level, addressing the limitation of existing transfer learning methods in adapting to multi-channel inputs [ 299 ], this paper systematically improves ResNet-18: The number of input channels was expanded from 3 to 6 [ 300 ]. Correspondingly, the number of convolutional kernels in the first layer was doubled from 64 to 128 to meet the requirements of convolutional operations [ 301 ]. Additionally, two new fully connected layers—one with 2,000 units and another with 1,000 units—were added to accommodate the rich feature information carried by the six-channel composite image [ 302 ], while retaining the structural advantages of the residual network in suppressing gradient explosion and network degradation [ 303 ]. Third, regarding the verification of feature complementarity, this paper systematically analyzes for the first time the complementary mechanisms between color maps of feature components and wavelet time-frequency maps in power quality disturbance recognition [ 304 ]. Through ablation experiments [ 305 ], we quantitatively verify the performance contributions of these two types of images across different disturbance types [ 306 ]. To more intuitively demonstrate the innovation and superiority of our method, we conducted a systematic comparison between our proposed method and recent representative hybrid deep learning-based power quality disturbance recognition methods [ 307 ]: Reference [ 15 ] employs a hybrid architecture with parallel input of one-dimensional signals and two-dimensional images into a CNN-LSTM, achieving an average recognition accuracy of 95.82% for 15 disturbance categories under a 25 dB noise environment; Reference [ 1 ] employs a method combining dual-channel GAF image fusion with ResNet, achieving an average recognition accuracy of 96.11% under a 25 dB noise environment; Reference [ 253 ] uses a method combining VMD decomposition with deep convolutional neural networks [ 308 ], achieving an average recognition accuracy of 96.33% under a 25 dB noise environment [ 309 ]; In contrast, the method proposed in this paper achieves an average recognition accuracy of 98.00% for 15 types of disturbances under the same 25 dB noise environment, representing improvements of 2.18, 1.89, and 1.67 percentage points over the methods in References [ 1 , 15 ], and [ 253 ], respectively. Particularly for the two disturbance types—voltage notches (C7) and voltage spikes (C8)—which are highly susceptible to noise interference, the recognition accuracy of the proposed method is 88.89% and 94.44%, respectively, significantly outperforming the 80.00% and 91.11% reported in Reference [ 1 ] and the 82.22% and 97.78% reported in Reference [ 253 ]. The above comparison results demonstrate that the method proposed in this paper [ 310 ], which combines dual-path feature image fusion with an improved six-channel ResNet, possesses distinct advantages in terms of recognition accuracy, noise robustness, and the ability to identify challenging disturbance types, fully reflecting the unique innovative value of this study. Reference [ 311 ] applied the Hilbert transform to the original disturbance signals to generate analytic signals, which were then processed to produce visual trajectory circles and input into ResNet-18 for training, testing, and recognition. Reference [ 312 ] used the L2 norm of distances between two points to generate an N × N power distance matrix, which was normalized and converted into a grayscale image. The generated grayscale images were then input into a 2D-ResNet for training, testing, and recognition. Both of these PQD recognition methods are feasible and effective, achieving high recognition accuracy under various noise environments. However, recognition of certain special PQDs under strong noise conditions still faces some challenges, whereas the method proposed in this paper shows advantages in these areas. Table 5 compares the recognition accuracy of each PQD using the aforementioned methods and the method proposed in this study. 10.1371/journal.pone.0350561.t005 Table 5 Disturbance Recognition Accuracy Rates Across Different Identification Systems. Disturbance Recognition Accuracy/% Visualized Trajectory Circle + ResNet-18 [ 311 ] Distance Matrix Grayscale Image + ResNet [ 312 ] Image Fusion + Six-Channel ResNet No 40 dB 35 dB 30 dB 25 dB No 40 dB 35 dB 30 dB 25 dB No 40 dB 35 dB 30 dB 25 dB C0 100.00 100.00 100.00 98.89 85.56 98.89 98.89 94.44 88.89 86.67 100.00 100.00 100.00 100.00 96.67 C1 100.00 100.00 98.89 98.89 96.67 100.00 98.89 98.89 93.33 95.56 98.89 98.89 97.78 97.78 97.78 C2 100.00 100.00 100.00 100.00 98.89 100.00 100.00 96.67 98.89 85.56 100.00 100.00 100.00 100.00 100.00 C3 100.00 100.00 100.00 98.89 100.00 100.00 98.89 98.89 100.00 100.00 98.89 98.89 100.00 100.00 98.89 C4 100.00 100.00 100.00 100.00 98.89 100.00 100.00 100.00 100.00 96.67 100.00 100.00 100.00 100.00 100.00 C5 96.67 98.89 98.89 98.89 95.56 100.00 98.89 97.78 100.00 97.78 100.00 100.00 100.00 100.00 98.89 C6 100.00 100.00 100.00 98.89 100.00 100.00 98.89 100.00 97.78 98.89 100.00 100.00 100.00 100.00 100.00 C7 100.00 100.00 97.78 96.67 97.78 98.89 95.56 100.00 82.22 80.0 97.78 97.78 96.67 96.67 88.89 C8 98.89 100.00 96.67 97.78 90.00 97.78 100.00 97.78 97.78 78.89 100.00 100.00 100.00 100.00 94.44 C1 + C4 98.89 98.89 100.00 100.00 100.00 100.00 97.78 95.56 96.67 92.22 98.89 100.00 98.89 98.89 98.89 C2 + C4 100.00 100.00 100.00 100.00 98.89 100.00 100.00 98.89 96.67 100.00 100.00 100.00 100.00 100.00 100.00 C3 + C4 100.00 100.00 100.00 100.00 97.78 96.67 100.00 100.00 100.00 100.00 100.00 100.00 98.89 98.89 98.89 C4 + C5 97.78 98.89 98.89 100.00 100.00 98.89 98.89 97.78 100.00 100.00 100.00 98.89 100.00 100.00 97.78 C1 + C5 100.00 100.00 100.00 100.00 100.00 100.00 96.67 96.67 96.67 95.56 100.00 100.00 100.00 100.00 98.89 C2 + C5 100.00 100.00 100.00 87.78 90.00 100.00 98.89 98.89 96.67 96.67 100.00 100.00 100.00 100.00 100.00 Average 99.48 99.78 99.41 98.44 96.74 99.41 98.81 97.56 96.37 93.63 99.63 99.63 99.48 99.48 98.00 As shown in Table 5 , the recognition accuracy of all methods decreases as noise increases. Under a 25 dB noise environment, the method proposed in this paper achieves 98.00% recognition accuracy, which is 1.26% higher than that of reference [ 311 ] and 4.37% higher than that of reference [ 312 ]. Moreover, for disturbances that are prone to misidentification, such as C0, C7, and C8, the proposed method demonstrates a clear advantage across all noise environments. By combining two feature-amplified images, the proposed method fuses the features of both color images, resulting in strong feature extraction capability. The introduction of the improved six-channel ResNet-18 further preserves ResNet’s structural advantages and allows six-channel image input, providing excellent recognition performance and generalization ability. In noise-free or 30–40 dB noise environments, the average recognition accuracy remains stable above 99.48%, and recognition accuracy under all noise conditions surpasses that of the methods proposed in the two referenced studies. 6. Conclusion To improve the recognition accuracy of various PQDs, this study overcomes the limitations of traditional methods and proposes a novel PQD recognition method based on the combination of feature images and an improved ResNet-18. This recognition system fully exploits the complete fusion of features from feature component color maps and wavelet time-frequency maps, as well as the strong feature extraction capability of the improved six-channel ResNet-18, achieving high recognition accuracy. The conclusions drawn from simulations and comparative experiments are as follows: 1) Regarding input images, this study proposes a two-image combination method based on the concept of feature complementarity. Wavelet time-frequency maps capture the time-frequency energy features of PQD signals, while feature component color maps effectively reflect the fluctuation features of components. The combination of the two provides richer spatiotemporal feature information and stronger feature representation. In particular, the feature component color map introduces a new image construction method by concatenating components, expanding on traditional PQD image generation methods; however, the optimal selection and arrangement of components require further investigation. 2) Regarding the recognition network, an improved multi-channel ResNet method is proposed for PQD classification. The improved ResNet-18 effectively suppresses gradient explosion and network degradation while automatically extracting deep features from multi-image combinations, achieving excellent recognition performance. This is particularly significant for scenarios with numerous monitoring points and large volumes of data, as it can substantially reduce misclassification. 3) The proposed method demonstrates strong multi-image analysis capability. It is not only applicable to PQD classification but, with appropriate modifications to the underlying ResNet structure, can also be applied to time-series prediction problems affected by multiple factors, such as photovoltaic power forecasting. Supporting information S1 File Figure: The minimum dataset for validating the algorithm in this paper (Figure.1-Figure.50.png, comprising 50 datasets). (ZIP)