Power frequency electromagnetic field intelligent identification method and system based on deep learning

Abstract

The invention relates to the technical field of power frequency electromagnetic field intelligent identification, and discloses a power frequency electromagnetic field intelligent identification method and system based on deep learning. The method comprises the following steps: inputting a time-frequency spectrogram into a multi-branch convolutional network to extract a comprehensive feature vector, calculating and identifying uncertainty indexes, and marking a low-confidence sample; calculating a difference vector of the low-confidence-coefficient sample comprehensive feature vector and the high-confidence-coefficient feature center vector, positioning a deviation dimension set, and determining a deviation time-frequency spectrogram; and carrying out local resampling on the deviation time-frequency spectrogram corresponding to the original signal time period to generate a local time-frequency spectrogram, extracting a local deviation correction feature vector, calculating a fusion weight according to a difference vector, carrying out weighted fusion to generate a corrected feature vector, and executing secondary identification. According to the invention, the identification accuracy, stability and real-time response capability of the power frequency electromagnetic field intelligent identification method based on deep learning in a complex electromagnetic environment are improved.

Description

Technical Field

[0001] This application relates to the field of intelligent identification technology of power frequency electromagnetic fields, and in particular to an intelligent identification method and system for power frequency electromagnetic fields based on deep learning. Background Technology

[0002] In existing technologies, deep learning-based intelligent identification methods for power frequency electromagnetic fields automatically analyze and recognize patterns in power frequency electromagnetic field signals by constructing deep neural network models. A typical implementation involves acquiring time-domain or frequency-domain signal data of power frequency electromagnetic fields using electromagnetic field sensors. After preprocessing, the data is input into deep learning models such as convolutional neural networks, recurrent neural networks, or long short-term memory networks for feature extraction and classification. These methods utilize the multi-layer nonlinear transformation capabilities of deep neural networks to automatically learn deep feature representations of electromagnetic field signals. Compared to traditional signal processing methods, they exhibit stronger recognition capabilities when processing complex electromagnetic field data and have been widely applied in scenarios such as power equipment fault diagnosis and transmission line condition monitoring.

[0003] However, existing deep learning-based recognition methods have significant shortcomings. First, current methods typically employ time-frequency analysis at a single time scale, failing to comprehensively capture the transient changes, short-term fluctuations, and long-term trends of power frequency electromagnetic fields across different time dimensions. This results in insufficient adaptability of the models to complex electromagnetic environments. Second, existing methods lack a quantitative evaluation mechanism for the model's recognition confidence. When the model's feature extraction for certain samples is biased, it cannot actively identify these low-confidence samples and can only passively accept potentially erroneous recognition results. Third, when recognition accuracy is unsatisfactory, existing methods typically improve the model by retraining the entire model or increasing its complexity. However, these methods are computationally expensive, time-consuming, and cannot specifically address feature extraction deficiencies in particular samples. Summary of the Invention

[0004] This application provides a deep learning-based intelligent identification method and system for power frequency electromagnetic fields. It solves the problems of insufficient model generalization ability, inability to actively detect feature extraction defects, lack of targeted correction ability, and inability to reuse correction experience in the prior art by constructing a progressive feature correction mechanism that includes multi-scale time-frequency feature extraction, uncertainty quantification screening, accurate deviation positioning, adaptive weight fusion, and accumulation of correction experience. It improves the identification accuracy, stability, and real-time response capability of the deep learning-based intelligent identification method for power frequency electromagnetic fields in complex electromagnetic environments.

[0005] Firstly, this application provides a deep learning-based intelligent identification method for power frequency electromagnetic fields, the deep learning-based intelligent identification method for power frequency electromagnetic fields comprising:

[0006] Step S1: Collect power frequency electromagnetic field signals, apply short time windows, medium time windows and long time windows to the power frequency electromagnetic field signals to perform time-frequency transformation, and generate three sets of time-frequency spectrum diagrams at different time scales;

[0007] Step S2: Input the time-spectrum image into a multi-branch convolutional network to extract a comprehensive feature vector, calculate the recognition uncertainty index of the comprehensive feature vector, and mark it as a low-confidence sample when the uncertainty index is greater than a preset threshold;

[0008] Step S3: Calculate the difference vector between the comprehensive feature vector of the low-confidence sample and the feature center vector of the high-confidence sample, locate the deviation dimension set based on the difference vector, and determine the corresponding deviation time spectrum.

[0009] Step S4: Locally resample the original signal time period corresponding to the deviation time spectrum to generate a local time spectrum, extract the local correction feature vector, calculate the fusion weight of each dimension according to the difference vector, and perform weighted fusion of the local correction feature vector and the comprehensive feature vector according to the fusion weight to generate the corrected feature vector and perform secondary recognition.

[0010] Secondly, this application provides a deep learning-based intelligent identification system for power frequency electromagnetic fields, the deep learning-based intelligent identification system for power frequency electromagnetic fields comprising:

[0011] The acquisition module is used to acquire power frequency electromagnetic field signals, and to apply short time windows, medium time windows and long time windows to the power frequency electromagnetic field signals to perform time-frequency transformation, generating three sets of time-frequency spectrum diagrams at different time scales;

[0012] The input module is used to input the time-spectrum image into a multi-branch convolutional network to extract a comprehensive feature vector, calculate the recognition uncertainty index of the comprehensive feature vector, and mark it as a low-confidence sample when the uncertainty index is greater than a preset threshold.

[0013] The calculation module is used to calculate the difference vector between the comprehensive feature vector of the low-confidence sample and the feature center vector of the high-confidence sample, locate the deviation dimension set based on the difference vector, and determine the corresponding deviation time spectrum.

[0014] The fusion module is used to locally resample the original signal time period corresponding to the deviation time spectrum to generate a local time spectrum, extract the local correction feature vector, calculate the fusion weight of each dimension according to the difference vector, and perform weighted fusion of the local correction feature vector and the comprehensive feature vector according to the fusion weight to generate the corrected feature vector and perform secondary recognition.

[0015] In the technical solution provided in this application, three sets of time-frequency spectra are generated at different time scales by applying short, medium, and long time windows to the power frequency electromagnetic field signal in step S1. This allows subsequent feature extraction to simultaneously acquire transient change information, harmonic variation characteristics, and long-term trend patterns of the electromagnetic field. Compared with the existing technology that uses a single time scale analysis, the multi-scale time-frequency spectra can more comprehensively depict the multi-dimensional characteristics of the power frequency electromagnetic field in complex electromagnetic environments, providing a richer feature representation basis for the model and enhancing the model's adaptability to different operating conditions and abnormal modes. In step S2, by calculating the identification uncertainty index of the comprehensive feature vector and marking low-confidence samples, this application realizes the model's proactive evaluation of its own feature extraction quality. Compared with the existing technology that passively accepts identification results, this self-diagnosis mechanism based on uncertainty quantification can promptly identify samples with potential defects in feature extraction, providing triggering conditions for subsequent targeted correction and avoiding adverse consequences caused by erroneous identification results in practical applications.

[0016] In step S3, by calculating the difference vector between the comprehensive feature vector of low-confidence samples and the center vector of high-confidence features and locating the set of deviation dimensions, this application achieves precise spatial localization of feature extraction deviations. It can clearly identify which dimensions in the comprehensive feature vector have deviated and which time scale feature extraction branch the deviation corresponds to. Compared with the existing technology that can only make overall adjustments, this precise localization capability allows the correction operation to focus on the feature dimensions and time-frequency regions where problems truly exist, avoiding unnecessary interference with correct features. In step S4, by locally resampling the original signal time period corresponding to the deviation time-frequency spectrum to generate a local time-frequency spectrum and extracting the local correction feature vector, this application performs targeted feature re-extraction for the located deviation region. Compared with the existing technology's overall reprocessing method, local resampling can capture the detailed features of the deviation region with higher time-frequency resolution while reducing computational overhead. More importantly, this application employs a mechanism that calculates the fusion weights for each dimension based on the difference vector and performs weighted fusion according to these weights. This mechanism enables differentiated feature correction based on the severity of the deviation. Dimensions with severe deviations adopt more corrective features for correction, while dimensions with smaller deviations mainly retain the original features. Compared to the simple replacement or uniform fusion methods of existing technologies, this adaptive weighted fusion strategy can effectively correct the weak links in feature extraction while retaining the existing effective feature information to the greatest extent. It improves the accuracy of recognition while maintaining the integrity and stability of feature representation, thereby significantly improving the reliability and generalization ability of the deep learning-based intelligent recognition method for power frequency electromagnetic fields in practical applications. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of an embodiment of the deep learning-based intelligent identification method for power frequency electromagnetic fields in this application.

[0019] Figure 2 This is a schematic diagram comparing the uncertainty indicators before and after correction in the embodiments of this application;

[0020] Figure 3 This is a schematic diagram illustrating the accumulation of the corrective knowledge base and the optimization of response time in an embodiment of this application. Detailed Implementation

[0021] This application provides a method and system for intelligent identification of power frequency electromagnetic fields based on deep learning. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0022] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the deep learning-based intelligent identification method for power frequency electromagnetic fields in this application includes:

[0023] Step S1: Collect power frequency electromagnetic field signals, apply short time windows, medium time windows and long time windows to the power frequency electromagnetic field signals to perform time-frequency transformation, and generate three sets of time spectrum diagrams at different time scales;

[0024] Specifically, applying three different time windows to the acquired power frequency electromagnetic field signal for time-frequency transformation aims to capture the characteristics of electromagnetic field anomalies at different time scales. The short time window captures transient abrupt changes such as electromagnetic pulses caused by switching operations; the medium time window captures short-term changes in the power frequency and its harmonic components; and the long time window extracts long-term trends and periodic features. These three time-spectrum diagrams are stacked along the channel dimension to form multi-channel input data, enabling the subsequent network to simultaneously acquire short-term abrupt changes, medium-term fluctuation characteristics, and long-term trend patterns. Compared to existing technologies that only use a single time scale, multi-scale time-spectrum diagrams can more comprehensively characterize the energy distribution characteristics of the power frequency electromagnetic field signal across different time dimensions.

[0025] Step S2: Input the time-frequency spectrum into a multi-branch convolutional network to extract the comprehensive feature vector, calculate the recognition uncertainty index of the comprehensive feature vector, and mark it as a low-confidence sample when the uncertainty index is greater than a preset threshold;

[0026] Specifically, the three parallel branches of the multi-branch convolutional network process spectrograms at different time scales, with each branch automatically extracting deep feature representations at its corresponding scale through multi-layer convolution operations. The feature vectors extracted by the three branches are concatenated to form a comprehensive feature vector, which integrates feature information from multiple scales. The uncertainty index is quantified by performing multiple forward propagations on the same sample and calculating the standard deviation of the output probability distribution. A larger standard deviation indicates that the model's recognition of the sample is less stable and has lower confidence. Compared to existing technologies that directly output recognition results, this invention, through an uncertainty quantification mechanism, can proactively identify samples where the model's feature extraction may have defects, providing triggering conditions for subsequent correction—a crucial step in achieving self-diagnosis.

[0027] Step S3: Calculate the difference vector between the comprehensive feature vector of the low-confidence sample and the feature center vector of the high-confidence sample. Based on the difference vector, locate the set of deviation dimensions and determine the corresponding deviation time spectrum.

[0028] The high-confidence feature center vector is obtained by selecting samples with low uncertainty and correct identification from the training set, extracting the comprehensive feature vector of these samples, and calculating the average of samples of the same category. The difference vector is formed by calculating the absolute value of the difference between the comprehensive feature vector of low-confidence samples and the high-confidence feature center vector dimension by dimension. The difference value of each dimension reflects the degree of deviation of the feature extraction in that dimension. The deviation threshold is dynamically determined based on the statistical characteristics of the difference vector, and dimensions with difference values ​​exceeding the threshold are identified as deviation dimensions. Since the comprehensive feature vector is composed of features from three branches concatenated in sequence, the position index of the deviation dimension in the vector can be used to infer whether the deviation mainly occurs in the branch processing short-time, medium-time, or long-time spectrograms, thereby accurately locating the target time-time spectrogram that needs to be reprocessed. Compared with the existing technology of overall re-identification, this positioning mechanism can find the specific weak links in feature extraction.

[0029] Step S4: Locally resample the original signal time period corresponding to the deviation time spectrum to generate a local time spectrum, extract the local correction feature vector, calculate the fusion weight of each dimension according to the difference vector, and perform weighted fusion of the local correction feature vector and the comprehensive feature vector according to the fusion weight to generate the corrected feature vector and perform secondary recognition.

[0030] Specifically, local resampling of the original signal time period corresponding to the deviation time spectrum improves time-frequency resolution by shortening the time window length and increasing the window overlap rate, generating a more refined local time spectrum and extracting local correction feature vectors. The calculation of fusion weights employs an adaptive strategy based on the degree of deviation. For the deviation dimension, a deviation coefficient is calculated based on the degree of deviation of its difference value relative to the mean and standard deviation of the difference vector; the more severe the deviation, the larger the coefficient. Then, the sigmoid function maps the coefficient to a fusion weight between 0.5 and 1, ensuring that dimensions with severe deviations adopt more correction features during fusion, while non-deviation dimensions are given lower fixed weights to retain the original features. The expanded correction feature vector and the original comprehensive feature vector are weighted and summed dimension by dimension according to these weights, generating a corrected feature vector that corrects both the deviation and correct dimensions. Compared to existing techniques that simply replace or uniformly fuse, this differentiated correction strategy can specifically enhance weak links in feature extraction while retaining existing effective information, thereby improving the model's recognition accuracy and generalization ability in complex electromagnetic environments.

[0031] In one specific embodiment, step S1 includes:

[0032] The original power frequency electromagnetic field signal is obtained by acquiring the original time-domain signals in the X-axis, Y-axis and Z-axis directions using a triaxial electromagnetic field sensor.

[0033] Short-time, medium-time, and long-time windows are applied to the original power frequency electromagnetic field signal, and short-time, medium-time, and long-time spectrum diagrams are generated by short-time Fourier transform, respectively.

[0034] The short-time spectrogram, medium-time spectrogram, and long-time spectrogram are normalized to grayscale image format respectively;

[0035] The short-time, medium-time, and long-time spectrograms are stacked according to the channel dimension to form a three-dimensional tensor data as the time spectrogram.

[0036] Specifically, when applying three different time windows to the original power frequency electromagnetic field signal for short-time Fourier transform, the short time window is set to 50 milliseconds, mainly to capture the transient change characteristics of the electromagnetic field, with a time resolution of 2 milliseconds and a frequency resolution of 20 Hz, which can accurately locate the occurrence time of sudden electromagnetic pulses; the medium time window is set to 200 milliseconds to capture the short-term changes of the 50 Hz power frequency and its harmonic components, with a time resolution of 8 milliseconds and a frequency resolution of 5 Hz, which can clearly show the dynamic evolution process of harmonic energy; the long time window is set to 1 second to extract the long-term trend and periodicity of the electromagnetic field, with a time resolution of 40 milliseconds and a frequency resolution of 1 Hz, which can reflect the overall distribution characteristics of electromagnetic field energy. The setting of the three time windows follows the basic principle of the mutual constraint between time resolution and frequency resolution in time-frequency analysis. The short time window provides high time resolution but low frequency resolution, while the long time window provides high frequency resolution but low time resolution. Through multi-scale combination, the characteristics of the power frequency electromagnetic field signal can be comprehensively characterized in both the time and frequency domains.

[0037] After normalizing the three time-frequency spectrograms to grayscale image format, they are stacked according to the channel dimension. The resulting 3D tensor data contains 9 channels: the first 3 channels correspond to short-time spectrograms along the X, Y, and Z axes; the middle 3 channels correspond to medium-time spectrograms along the three axes; and the last 3 channels correspond to long-time spectrograms along the three axes. This stacking method preserves the time-frequency distribution information in the spatial dimension of the 3D tensor data and integrates multi-scale and multi-directional features in the channel dimension, providing structured input data for feature extraction in subsequent multi-branch convolutional networks. Normalizing the grayscale value range to 0 to 255 ensures that the spectrograms at different time scales have a uniform numerical scale, avoiding over-enhancement or suppression of features at a certain scale in subsequent processing due to differences in energy amplitude. This ensures that features at all three time scales can be learned and extracted equally during network training.

[0038] In one specific embodiment, step S2 includes:

[0039] The spectrograms corresponding to the short time window, the medium time window, and the long time window in the time-spectrum map are respectively input into three parallel convolutional branch networks. Each branch network contains four convolutional layers. Each convolutional layer is followed by a batch normalization layer and a max pooling layer to extract short-time feature vectors, medium-time feature vectors, and long-time feature vectors, respectively.

[0040] The short-time feature vector, medium-time feature vector, and long-time feature vector are concatenated to form a comprehensive feature vector.

[0041] The comprehensive feature vector is input into the main recognition network for forward propagation. The main recognition network consists of two fully connected layers and an output layer. The output layer uses the Softmax activation function to output the probability distribution of each anomaly category.

[0042] The same sample is propagated forward multiple times, with different neurons being randomly discarded each time. The standard deviation of the prediction results of multiple probability distributions is calculated as an uncertainty index. When the uncertainty index is greater than a preset threshold, the sample is marked as a low-confidence sample.

[0043] Specifically, the three parallel convolutional branch networks are designed to extract features from spectrograms at different time scales. The parameters of the four convolutional layers in each branch are configured as follows: the first layer has a 7x7 kernel size, 32 kernels, and a stride of 2; the second layer has a 5x5 kernel size, 64 kernels, and a stride of 2; and the third and fourth layers both have a 3x3 kernel size, 128 and 256 kernels respectively, and a stride of 1 for both. Each convolutional layer is followed by a batch normalization layer to accelerate network convergence and prevent gradient vanishing. A max-pooling layer uses a 2x2 pooling kernel to downsample the feature map. After four convolutional layers, the input 256x256 pixel spectrogram is progressively reduced to an 8x8 feature map, expanding the number of channels from 3 to 256. Then, a global average pooling layer compresses the 8x8x256 feature map into a 256-dimensional feature vector. The short-term feature vector, medium-term feature vector, and long-term feature vector output by the three branches each contain 256 feature dimensions. They are concatenated in sequence to form a 768-dimensional comprehensive feature vector, which integrates short-term transient features, medium-term harmonic features, and long-term trend features.

[0044] The main recognition network employs a fully connected layer structure to map the comprehensive feature vector to the anomaly category space. The first fully connected layer contains 512 neurons and uses the ReLU activation function and Dropout regularization. The Dropout rate is set to 0.5 to prevent overfitting. The second fully connected layer contains 4 neurons corresponding to four anomaly categories: normal state, partial discharge, equipment failure, and external interference. The output layer uses the Softmax activation function to convert the output values ​​of the 4 neurons into a probability distribution. The sum of the probabilities of the four categories is 1, and the category with the highest probability is the recognition result. During network training, the cross-entropy loss function is used to measure the difference between the predicted probability distribution and the true label. The network weight parameters are updated through the backpropagation algorithm and the Adam optimizer. The learning rate is set to 0.001, the batch size is set to 32, and the training epochs are set to 100. The model performance is monitored on the validation set, and the model parameters with the highest validation accuracy are saved.

[0045] The uncertainty index is quantified using the Monte Carlo Dropout method. During the testing phase, the Dropout layer remains active. The same sample undergoes 20 forward passes, with 50% of the neurons in the first fully connected layer randomly discarded each time, resulting in 20 different 4D probability distributions. The standard deviation of these 20 probability distributions across each category dimension is calculated, and the average of the four category standard deviations is taken as the uncertainty index for that sample. A larger standard deviation indicates greater discrepancies in prediction results obtained from different forward passes, suggesting instability in the model's feature representation of the sample and low recognition confidence. A preset threshold of 0.15 is set; samples with an uncertainty index greater than 0.15 are classified as low-confidence samples, triggering subsequent feature correction mechanisms. This sample selection method based on uncertainty quantification, compared to existing techniques that rely solely on output probability values, can more accurately identify samples with defects in model feature extraction.

[0046] Figure 2 This is a schematic diagram comparing uncertainty indicators before and after correction in an embodiment of this application. Figure 2As shown, samples from four anomaly categories—normal state, partial discharge, equipment failure, and external interference—were initially identified and then re-identified after correction. The identification uncertainty indices for each category were calculated and compared. During the initial identification, the uncertainty indices for the four anomaly categories were 0.142, 0.185, 0.203, and 0.178, respectively. The equipment failure category had the highest uncertainty index at 0.203, exceeding the preset confidence threshold of 0.15, indicating a significant bias in the model's feature extraction for this category. After processing by the feature correction mechanism of this invention, the uncertainty indices for the four categories decreased to 0.089, 0.112, 0.125, and 0.108, respectively, representing reductions of 37.3%, 39.5%, 38.4%, and 39.3%. All uncertainty indices were below the confidence threshold, verifying that this invention, through bias dimension localization, local resampling, and adaptive weight fusion mechanisms, can effectively correct feature extraction defects and significantly improve the stability and confidence of the model's identification.

[0047] In one specific embodiment, step S3 includes:

[0048] Samples with uncertainty indices below the confidence threshold and correct identification are selected from the training set. The comprehensive feature vectors of the samples are extracted and stored according to the anomaly category to construct a high-confidence feature library for each category.

[0049] Based on the probability distribution of low-confidence samples, determine the most likely anomaly category to which they belong, extract multiple reference feature vectors from the high-confidence feature library of the corresponding category, and calculate the average value of the reference feature vectors to obtain the high-confidence feature center vector.

[0050] Calculate the difference vectors between the comprehensive feature vector of low-confidence samples and the feature center vector of high-confidence samples in each feature dimension. The difference value of each dimension represents the degree of deviation in the corresponding dimension.

[0051] The deviation threshold is calculated based on the mean and standard deviation of the difference vector. Dimensions with a difference value greater than the deviation threshold are recorded as the deviation dimension set. The corresponding time scale branch is deduced based on the index position of the deviation dimension set in the comprehensive feature vector. The time spectrum diagram corresponding to the corresponding time scale branch is then determined as the deviation time spectrum diagram.

[0052] Specifically, the high-confidence feature library is constructed by traversing 10,000 samples in the training set. For each sample, the recognition uncertainty index in step S2 is calculated and compared with the true label for verification. Samples with a recognition uncertainty index less than 0.10 and recognition results consistent with the true label are selected, totaling 6,800 samples. The 768-dimensional comprehensive feature vectors of these 6,800 high-confidence samples are extracted and stored according to four anomaly categories, forming four sub-libraries: normal state feature library, partial discharge feature library, equipment fault feature library, and external interference feature library. Each sub-library contains the comprehensive feature vectors of all high-confidence samples in that category. These vectors represent the feature space distribution that the model should present when accurately extracting features from samples of that category, providing a reference standard for subsequently judging the feature deviation of low-confidence samples.

[0053] For low-confidence samples, their probability distribution is a 4-dimensional vector output by the main recognition network in step S2, containing the probability values ​​of the sample belonging to four categories: normal state, partial discharge, equipment failure, and external interference. The category with the highest probability value is selected as the most likely abnormal category to which the sample belongs, and 100 reference feature vectors are randomly extracted from the high-confidence feature library of the corresponding category. The average value of these 100 reference feature vectors across 768 feature dimensions is calculated to obtain a 768-dimensional high-confidence feature center vector, which represents the center position of the high-confidence sample of this category in the feature space. Then, the absolute value of the difference between the 768-dimensional comprehensive feature vector of the low-confidence sample and the high-confidence feature center vector is calculated in each dimension to form a 768-dimensional difference vector. The value of the i-th element in the difference vector reflects the degree to which the sample deviates from the normal feature extraction result in the i-th feature dimension.

[0054] The deviation threshold is calculated based on the statistical properties of the difference vector. First, the mean and standard deviation of the 768 elements of the difference vector are calculated. The deviation threshold is set as the mean plus 1.5 times the standard deviation. This setting is based on the outlier detection principle in statistics, which can identify feature dimensions that significantly deviate from the normal range. The 768 dimensions of the difference vector are traversed, and dimensions with difference values ​​greater than the deviation threshold are indexed and recorded in the deviation dimension set. Since the 768-dimensional comprehensive feature vector is composed of three 256-dimensional branch feature vectors concatenated sequentially, index positions 0 to 255 correspond to features extracted by the short-term branch, index positions 256 to 511 correspond to features extracted by the medium-term branch, and index positions 512 to 767 correspond to features extracted by the long-term branch. The number of deviation dimensions in each index interval of the statistical deviation dimension set is counted. The time scale branch corresponding to the interval with the most deviation dimensions is selected as the main deviation branch. This branch has the most significant defects in the feature extraction process. The time spectrum corresponding to the input of this branch, namely the short-time spectrum, medium-time spectrum, or long-time spectrum, is determined as the deviation time spectrum, which specifies the target for subsequent local resampling and feature correction.

[0055] In one specific embodiment, step S4 includes:

[0056] For each deviation dimension in the deviation dimension set, the activation weight matrix corresponding to each deviation dimension is calculated by the class activation mapping technique. The activation weight matrix is ​​then upsampled and binarized to obtain the time-frequency region mask for each dimension.

[0057] Perform a logical OR operation on the time-frequency region masks of each dimension to obtain the comprehensive deviation region mask. The comprehensive deviation region mask identifies the time-frequency positions in the deviation time-frequency spectrum that cause feature extraction deviation.

[0058] Based on the comprehensive deviation region mask, the corresponding original signal time period is extracted. The original signal time period is then locally resampled using the sliding window method. The window length is set to half of the original window length and the window overlap rate is set to a preset overlap ratio to generate a local time spectrum.

[0059] The local temporal spectrogram is input into a lightweight convolutional network to extract local bias correction feature vectors.

[0060] Specifically, activation mapping is used to establish the correspondence between the deviation feature dimensions and the spatial locations of the original time-frequency spectrogram. The specific implementation involves tracing the activation source of each deviation dimension in the multi-branch convolutional network. Taking the short-time branch as an example, the feature map output after the fourth convolutional layer and before global average pooling is 8x8x256, with 256 channels corresponding to 256 feature dimensions. For a given deviation dimension, an 8x8 activation feature map of its corresponding channel is extracted. The value at each position in this feature map represents the contribution of the corresponding region in the original time-frequency spectrogram to that feature dimension. The 8x8 activation feature map is upsampled to 256x256 using bilinear interpolation to match the spatial resolution of the original time-frequency spectrogram. A binarization threshold is set to 0.6 times the maximum value of the activation feature map. Positions with upsampled activation values ​​greater than the threshold are marked as 1, and positions with values ​​less than the threshold are marked as 0, resulting in the time-frequency region mask corresponding to that deviation dimension. Repeat the above operation for all deviation dimensions in the deviation dimension set to generate multiple time-frequency region masks. Then perform a pixel-by-pixel logical OR operation on these masks. That is, if a pixel position is marked as 1 in any mask, then that position is marked as 1 in the comprehensive deviation region mask. The final comprehensive deviation region mask identifies all the key time-frequency positions in the deviation time-frequency spectrum that cause feature extraction deviation.

[0061] Based on the pixel positions marked as 1 in the composite deviation region mask, the corresponding time periods in the original power frequency electromagnetic field signal are deduced. Since the horizontal axis of the time-spectrum graph represents time and the vertical axis represents frequency, the horizontal coordinates in the mask can be directly mapped to the time index of the original signal. The horizontal coordinate range of all pixels marked as 1 in the mask is extracted to determine the original signal time periods that need to be reprocessed. Local resampling is performed on these time periods using a sliding window method. Assuming the main deviation occurs in the 50-millisecond window corresponding to the short-time branch, the window length is adjusted to 25 milliseconds, half the original window length, and the window overlap rate is set to 75%, meaning there is an 18.75-millisecond overlap between adjacent windows. By reducing the window length and increasing the overlap rate, higher temporal resolution can be obtained, allowing the generated local time-spectrum graph to more precisely depict the transient changes in the electromagnetic field within the deviation time periods. The local temporal spectrogram has a size of 128 x 128 x 3 and is input into a lightweight convolutional network containing three convolutional layers. The first layer has a kernel size of 5 x 5 and 32 kernels, the second layer has a kernel size of 3 x 3 and 64 kernels, and the third layer has a kernel size of 3 x 3 and 128 kernels. After convolution and global average pooling, a 128-dimensional local bias correction feature vector is output. This vector is specifically designed to provide supplementary feature representations for weak temporal and frequency regions in the original feature extraction.

[0062] In one specific embodiment, calculating the fusion weights for each dimension based on the difference vector includes:

[0063] The local correction feature vector is extended to the same dimension as the comprehensive feature vector by a mapping network;

[0064] For each deviation dimension in the deviation dimension set, the deviation severity coefficient is calculated based on the difference value of the deviation dimension and the mean and standard deviation of the difference vector. The deviation severity coefficient indicates the severity of the deviation dimension.

[0065] For each biased dimension, the bias degree coefficient is mapped to the corresponding dimension's fusion weight using the Sigmoid function, while a fixed low fusion weight is set for the non-biased dimensions.

[0066] The expanded local correction feature vector and the comprehensive feature vector are weighted and fused dimension by dimension according to the fusion weight to generate the corrected feature vector.

[0067] Specifically, the mapping network employs a single fully connected layer structure to extend the 128-dimensional local bias correction feature vector to 768 dimensions. The weight matrix of this fully connected layer has a size of 128 x 768 and contains 98,304 trainable parameters. The weight matrix is ​​initialized using the Xavier initialization method, and the Tanh activation function is used to constrain the output value to between -1 and +1. The training process of the mapping network is synchronized with the main recognition network. During the model training phase, the backpropagation algorithm is used to learn how to effectively map the 128 dimensions of the local bias correction feature vector to the 768-dimensional space of the comprehensive feature vector, ensuring that the extended bias correction feature vector maintains a dimensional correspondence with the original comprehensive feature vector in the feature space. The extended local bias correction feature vector and the original comprehensive feature vector have the same 768-dimensional structure, laying the foundation for subsequent dimension-wise weighted fusion.

[0068] The deviation coefficient is calculated based on the difference vector and its statistical properties obtained in step S3. For the i-th deviation dimension in the deviation dimension set, the difference value of that dimension in the difference vector is first extracted. This difference value represents the absolute distance of the low-confidence sample from the center of the high-confidence feature in that dimension. The mean and standard deviation of all 768 dimensions of the difference vector are calculated. The deviation coefficient is equal to the difference value of that dimension minus the mean of the difference vector, and then divided by the standard deviation of the difference vector. This calculation process is essentially a standardization process for the difference values, converting the difference values ​​of different dimensions into a dimensionless standardized deviation. A positive deviation coefficient, and a larger value, indicates a more severe deviation in that dimension, requiring more corrective features to be adopted for correction.

[0069] The calculation of the fusion weights uses the Sigmoid function to non-linearly map the bias coefficient. The mathematical form of the Sigmoid function is that the output value equals 1 divided by 1 plus the negative power of the bias coefficient (e). Since the bias coefficient is positive, the Sigmoid function maps it to fusion weights between 0.5 and 1. The larger the bias coefficient, the closer the mapped fusion weight is to 1, meaning that this dimension adopts more corrective features during fusion. For non-biased dimensions not in the biased dimension set, their fusion weights are uniformly set to a fixed low weight of 0.2, indicating that the original feature extraction for these dimensions is relatively accurate, requiring only a slight introduction of corrective information. The calculation method for the dimension-wise weighted fusion is as follows: for the i-th dimension of the comprehensive feature vector, the value of the corrected feature vector for that dimension is equal to the fusion weight multiplied by the value of the expanded locally corrected feature vector for that dimension, plus 1 minus the fusion weight, and then multiplied by the value of the original comprehensive feature vector for that dimension. This weighted fusion mechanism ensures that dimensions with severe biases primarily adopt the corrective information from the bias-correcting features, while dimensions that were originally extracted correctly primarily retain their original features. This achieves differentiated feature correction based on the degree of bias, which, compared to existing technologies that simply replace or uniformly fuse features, can retain effective feature information to the greatest extent while correcting feature defects.

[0070] In one specific embodiment, performing secondary identification includes:

[0071] The corrected feature vector is input into the secondary recognition network for classification, and the secondary recognition network outputs the probability distribution of the secondary recognition categories.

[0072] Uncertainty index for secondary identification calculated by performing multiple forward propagations on the same sample;

[0073] When the uncertainty index of the secondary identification is less than that of the initial identification and the secondary identification result is correct, the abnormal category label, deviation dimension set, deviation degree coefficient set, local resampling parameters and fusion weight vector of the corresponding sample are stored as correction experience entries in the correction knowledge base.

[0074] When processing new low-confidence samples, the similarity between the deviation dimension set of the current sample and the deviation dimension set of each entry in the correction knowledge base is calculated. When the similarity is greater than the preset similarity threshold, the local resampling parameters and fusion weight vector recorded by the correction experience entries are directly reused for fast correction.

[0075] Specifically, the secondary recognition network employs a three-layer fully connected structure to reclassify the corrected feature vectors. The first fully connected layer contains 512 neurons and uses the PReLU activation function. PReLU, or Parametrically Modified Linear Unit activation function, provides a non-zero gradient for negative inputs compared to ReLU, thus enhancing the network's expressive power. The second fully connected layer contains 256 neurons, also using the PReLU activation function. Through the nonlinear transformation of the two fully connected layers, the 768-dimensional corrected feature vectors are mapped to a more compact feature space. The third layer, the output layer, contains four neurons corresponding to four anomaly categories and uses the Softmax activation function to output the probability distribution of the secondary recognition categories. The training process of the secondary recognition network uses the same loss function and optimizer configuration as the main recognition network, with a learning rate of 0.0005 and a batch size of 16. During training, an augmented training set containing corrected samples is used for parameter updates. The uncertainty index of secondary identification is calculated in the same way as in step S2. The same corrected sample is forward-propagated 20 times while maintaining the Dropout activation of the first fully connected layer. The average standard deviation of the output of the 20 probability distributions is calculated as the uncertainty index of secondary identification.

[0076] The error correction knowledge base employs a key-value pair storage structure. Each error correction experience entry contains five parts: anomaly category label, a set of deviation dimensions, a set of deviation severity coefficients, local resampling parameters, and a fusion weight vector. The anomaly category label identifies the sample category to which the error correction experience applies. The set of deviation dimensions records which of the 768 feature dimensions have deviated. The set of deviation severity coefficients records the standardized deviation severity values ​​corresponding to each deviation dimension. The local resampling parameters include the adjusted window length and window overlap rate. The fusion weight vector records the fusion weight values ​​for each of the 768 dimensions. When the uncertainty index of the secondary identification is less than that of the initial identification, it indicates that the error correction operation has effectively reduced the uncertainty of identification. The secondary identification result is then compared with the true label. If the identification is correct, it means that the error correction strategy has successfully corrected the feature extraction defects. At this point, the complete error correction experience for that sample is stored as an entry in the knowledge base. The knowledge base is indexed using a composite key composed of the anomaly category label and the main deviation branch number, facilitating rapid retrieval of similar error correction experiences.

[0077] When processing new low-confidence samples, the deviation dimension set calculated in step S3 is first extracted. Then, the Jaccard similarity between the deviation dimension sets is calculated by traversing all entries in the correction knowledge base. The Jaccard similarity is calculated by dividing the number of elements in the intersection of the current sample's deviation dimension set and the knowledge base entry's deviation dimension set by the number of elements in the union of the two sets. The similarity value ranges from 0 to 1, with a higher value indicating a more similar feature deviation pattern between the two samples. A preset similarity threshold is set to 0.6. When the similarity of a knowledge base entry is greater than 0.6, the correction experience recorded in that entry is deemed applicable to the current sample. The local resampling parameters and fusion weight vector recorded in that entry are directly reused, skipping the time-consuming operations in step S4 such as class activation mapping, mask generation, local re-extraction, and weight calculation. Only local resampling and weighted fusion are needed to generate the corrected feature vector. This knowledge base-based fast correction mechanism reduces the correction time for a single sample from an average of 120 milliseconds to 15 milliseconds, significantly improving the system's response speed in real-time monitoring scenarios. The knowledge base is optimized after running 1,000 samples. The number of times each entry is reused and the success rate of correction are counted. Inefficient entries that are reused less than 5 times or have a success rate of less than 0.7 are deleted, while high-frequency and efficient correction patterns are retained. This allows the knowledge base to gradually converge to the set of correction strategies that are most suitable for the current application scenario.

[0078] Figure 3 This is a schematic diagram illustrating the accumulation of corrective knowledge base and optimization of response time in an embodiment of this application. For example... Figure 3 As shown, as the number of low-confidence samples processed by the system increases from 0 to 10,000, the number of valid correction experience entries stored in the correction knowledge base gradually accumulates, growing from the initial 0 to 478. After processing approximately 4,000 samples, the growth rate of the number of knowledge base entries slows down and approaches saturation, indicating that the system has learned the main characteristic deviation patterns in this application scenario. Simultaneously, the average correction time per sample decreases significantly from the initial 120 milliseconds to 15 milliseconds, a reduction of 87.5%. This is because the correction experience entries accumulated in the knowledge base can be directly reused by subsequent samples with similar deviation patterns, skipping time-consuming operations such as class activation mapping, mask generation, and weight calculation. This figure verifies that the correction knowledge base mechanism of this invention can not only continuously accumulate effective correction experience but also significantly improve the system's response speed in real-time monitoring scenarios through experience reuse, enabling the system to possess both high accuracy and high real-time performance.

[0079] The above describes the deep learning-based intelligent identification method for power frequency electromagnetic fields in the embodiments of this application. The following describes the deep learning-based intelligent identification system for power frequency electromagnetic fields in the embodiments of this application. One embodiment of the deep learning-based intelligent identification system for power frequency electromagnetic fields in the embodiments of this application includes:

[0080] The acquisition module is used to acquire power frequency electromagnetic field signals, and to apply short time windows, medium time windows and long time windows to the power frequency electromagnetic field signals to perform time-frequency transformation, generating three sets of time-frequency spectrum diagrams at different time scales;

[0081] The input module is used to input the time-spectrum image into the multi-branch convolutional network to extract the comprehensive feature vector, calculate the recognition uncertainty index of the comprehensive feature vector, and mark it as a low confidence sample when the uncertainty index is greater than a preset threshold.

[0082] The calculation module is used to calculate the difference vector between the comprehensive feature vector of the low-confidence sample and the feature center vector of the high-confidence sample, locate the deviation dimension set based on the difference vector, and determine the corresponding deviation time spectrum.

[0083] The fusion module is used to locally resample the original signal time period corresponding to the deviation time spectrum to generate a local time spectrum, extract the local correction feature vector, calculate the fusion weight of each dimension according to the difference vector, and perform weighted fusion of the local correction feature vector and the comprehensive feature vector according to the fusion weight to generate the corrected feature vector and perform secondary recognition.

[0084] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligent identification of power frequency electromagnetic fields based on deep learning, characterized in that, The method includes: Step S1: Collect power frequency electromagnetic field signals, apply short time windows, medium time windows and long time windows to the power frequency electromagnetic field signals to perform time-frequency transformation, and generate three sets of time-frequency spectrum diagrams at different time scales; Step S2: Input the time-spectrum image into a multi-branch convolutional network to extract a comprehensive feature vector, calculate the recognition uncertainty index of the comprehensive feature vector, and mark it as a low-confidence sample when the uncertainty index is greater than a preset threshold; Step S3: Calculate the difference vector between the comprehensive feature vector of the low-confidence sample and the feature center vector of the high-confidence sample, locate the deviation dimension set based on the difference vector, and determine the corresponding deviation time spectrum. Step S4: Locally resample the original signal time period corresponding to the deviation time spectrum to generate a local time spectrum, extract the local correction feature vector, calculate the fusion weight of each dimension according to the difference vector, and perform weighted fusion of the local correction feature vector and the comprehensive feature vector according to the fusion weight to generate the corrected feature vector and perform secondary recognition.

2. The intelligent identification method for power frequency electromagnetic fields based on deep learning according to claim 1, characterized in that, Step S1 includes: The original power frequency electromagnetic field signal is obtained by acquiring the original time-domain signals in the X-axis, Y-axis and Z-axis directions using a triaxial electromagnetic field sensor. Short-time, medium-time, and long-time windows are applied to the original power frequency electromagnetic field signal, and short-time, medium-time, and long-time spectrum diagrams are generated by short-time Fourier transform, respectively. The short-time spectrogram, medium-time spectrogram, and long-time spectrogram are normalized into grayscale image format respectively. The short-time spectrum, medium-time spectrum, and long-time spectrum are stacked according to the channel dimension to form three-dimensional tensor data as the time spectrum.

3. The intelligent identification method for power frequency electromagnetic fields based on deep learning according to claim 1, characterized in that, Step S2 includes: The spectrum maps corresponding to the short time window, the medium time window, and the long time window in the time spectrum map are respectively input into three parallel convolutional branch networks. Each branch network contains four convolutional layers. Each convolutional layer is connected to a batch normalization layer and a max pooling layer to extract short-time feature vectors, medium-time feature vectors, and long-time feature vectors, respectively. The short-time feature vector, medium-time feature vector, and long-time feature vector are concatenated to form a comprehensive feature vector. The comprehensive feature vector is input into the main recognition network for forward propagation. The main recognition network contains two fully connected layers and an output layer. The output layer uses the Softmax activation function to output the probability distribution of each anomaly category. The same sample is propagated forward multiple times, with different neurons being randomly discarded each time. The standard deviation of the prediction results of multiple probability distributions is calculated as an uncertainty index. When the uncertainty index is greater than a preset threshold, the sample is marked as a low-confidence sample.

4. The intelligent identification method for power frequency electromagnetic fields based on deep learning according to claim 1, characterized in that, Step S3 includes: Samples with uncertainty indices less than the confidence threshold and correct identification are selected from the training set. The comprehensive feature vectors of these samples are extracted and stored according to the anomaly category to construct a high-confidence feature library for each category. Based on the probability distribution of the low-confidence samples, determine the most likely anomaly category to which they belong, extract multiple reference feature vectors from the high-confidence feature library of the corresponding category, and calculate the average value of the reference feature vectors to obtain the high-confidence feature center vector. Calculate the difference vectors between the comprehensive feature vector of the low-confidence sample and the feature center vector of the high-confidence sample in each feature dimension. The difference value of each dimension represents the degree of deviation in the corresponding dimension. The deviation threshold is calculated based on the mean and standard deviation of the difference vector. Dimensions with a difference value greater than the deviation threshold are recorded as a deviation dimension set. The corresponding time scale branch is deduced based on the index position of the deviation dimension set in the comprehensive feature vector. The time spectrum diagram corresponding to the corresponding time scale branch is determined as the deviation time spectrum diagram.

5. The intelligent identification method for power frequency electromagnetic fields based on deep learning according to claim 1, characterized in that, Step S4 includes: For each deviation dimension in the deviation dimension set, the activation weight matrix corresponding to each deviation dimension is calculated by class activation mapping technology. The activation weight matrix is ​​then upsampled and binarized to obtain the time-frequency region mask for each dimension. A logical OR operation is performed on the time-frequency region masks of each dimension to obtain a comprehensive deviation region mask. The comprehensive deviation region mask identifies the time-frequency positions in the deviation time-frequency spectrum that cause feature extraction deviation. Based on the comprehensive deviation region mask, the corresponding original signal time period is extracted. The original signal time period is then locally resampled using the sliding window method. The window length is set to half of the original window length and the window overlap rate is set to a preset overlap ratio to generate a local time spectrum. The local temporal spectrogram is input into a lightweight convolutional network to extract local bias correction feature vectors.

6. The intelligent identification method for power frequency electromagnetic fields based on deep learning according to claim 5, characterized in that, The calculation of the fusion weights for each dimension based on the difference vector includes: The local correction feature vector is expanded to the same dimension as the comprehensive feature vector by a mapping network; For each deviation dimension in the deviation dimension set, a deviation severity coefficient is calculated based on the difference value of the deviation dimension and the mean and standard deviation of the difference vector. The deviation severity coefficient represents the severity of the deviation dimension. For each deviation dimension, the deviation degree coefficient is mapped to the corresponding dimension's fusion weight using the Sigmoid function, and a fixed low fusion weight is set for the non-deviation dimensions. The expanded local correction feature vector and the comprehensive feature vector are weighted and fused dimension by dimension according to the fusion weight to generate the corrected feature vector.

7. The intelligent identification method for power frequency electromagnetic fields based on deep learning according to claim 6, characterized in that, The secondary identification process includes: The corrected feature vector is input into a secondary recognition network for classification, and the secondary recognition network outputs the probability distribution of the secondary recognition categories. Uncertainty index for secondary identification calculated by performing multiple forward propagations on the same sample; When the uncertainty index of the secondary identification is less than that of the initial identification and the secondary identification result is correct, the abnormal category label, deviation dimension set, deviation degree coefficient set, local resampling parameter and fusion weight vector of the corresponding sample are stored as correction experience entries in the correction knowledge base. When processing new low-confidence samples, the similarity between the deviation dimension set of the current sample and the deviation dimension set of each entry in the correction knowledge base is calculated. When the similarity is greater than a preset similarity threshold, the local resampling parameters and fusion weight vector recorded in the correction experience entry are directly reused for fast correction.

8. A deep learning-based intelligent identification system for power frequency electromagnetic fields, characterized in that, For implementing the deep learning-based intelligent identification method for power frequency electromagnetic fields as described in any one of claims 1-7, the deep learning-based intelligent identification system for power frequency electromagnetic fields comprises: The acquisition module is used to acquire power frequency electromagnetic field signals, and to apply short time windows, medium time windows and long time windows to the power frequency electromagnetic field signals to perform time-frequency transformation, generating three sets of time-frequency spectrum diagrams at different time scales; The input module is used to input the time-spectrum image into a multi-branch convolutional network to extract a comprehensive feature vector, calculate the recognition uncertainty index of the comprehensive feature vector, and mark it as a low-confidence sample when the uncertainty index is greater than a preset threshold. The calculation module is used to calculate the difference vector between the comprehensive feature vector of the low-confidence sample and the feature center vector of the high-confidence sample, locate the deviation dimension set based on the difference vector, and determine the corresponding deviation time spectrum. The fusion module is used to locally resample the original signal time period corresponding to the deviation time spectrum to generate a local time spectrum, extract the local correction feature vector, calculate the fusion weight of each dimension according to the difference vector, and perform weighted fusion of the local correction feature vector and the comprehensive feature vector according to the fusion weight to generate the corrected feature vector and perform secondary recognition.

9. The system according to claim 8, characterized in that, Power frequency electromagnetic field signals are acquired, and time-frequency transformations are performed on the power frequency electromagnetic field signals by applying short, medium, and long time windows respectively, generating three sets of time-frequency spectrum diagrams at different time scales, including: The original power frequency electromagnetic field signal is obtained by acquiring the original time-domain signals in the X-axis, Y-axis and Z-axis directions using a triaxial electromagnetic field sensor. Short-time, medium-time, and long-time windows are applied to the original power frequency electromagnetic field signal, and short-time, medium-time, and long-time spectrum diagrams are generated by short-time Fourier transform, respectively. The short-time spectrogram, medium-time spectrogram, and long-time spectrogram are normalized into grayscale image format respectively. The short-time spectrum, medium-time spectrum, and long-time spectrum are stacked according to the channel dimension to form three-dimensional tensor data as the time spectrum.

10. The system according to claim 8, characterized in that, The time-spectral graph is input into a multi-branch convolutional network to extract a comprehensive feature vector. The recognition uncertainty index of the comprehensive feature vector is calculated. When the uncertainty index is greater than a preset threshold, the sample is marked as a low-confidence sample, including: The spectrum maps corresponding to the short time window, the medium time window, and the long time window in the time spectrum map are respectively input into three parallel convolutional branch networks. Each branch network contains four convolutional layers. Each convolutional layer is connected to a batch normalization layer and a max pooling layer to extract short-time feature vectors, medium-time feature vectors, and long-time feature vectors, respectively. The short-time feature vector, medium-time feature vector, and long-time feature vector are concatenated to form a comprehensive feature vector. The comprehensive feature vector is input into the main recognition network for forward propagation. The main recognition network contains two fully connected layers and an output layer. The output layer uses the Softmax activation function to output the probability distribution of each anomaly category. The same sample is propagated forward multiple times, with different neurons being randomly discarded each time. The standard deviation of the prediction results of multiple probability distributions is calculated as an uncertainty index. When the uncertainty index is greater than a preset threshold, the sample is marked as a low-confidence sample.

Images