Electric power big data intelligent monitoring method and system based on HPLC (High Performance Liquid Chromatography)

Abstract

The invention discloses an electric power big data intelligent monitoring method and system based on HPLC. The method comprises the following steps: training based on an electrician identification sample data set to obtain an electrician face identification model; training based on the electrician dressing identification sample data set to obtain an electrician dressing identification model; and training based on the user electricity consumption information sample data set to obtain an electricity consumption detection model. Inputting the face image of the electrician, the dressing image of the electrician and the power consumption condition of the user, which are acquired in real time, into three models of a cloud side server to obtain an identity ID recognition result of the electrician, a dressing standard recognition result of the electrician and a power consumption condition recognition result of the user; the edge server sends an identification result to the II-type collector, the II-type collector transmits the identification result to the I-type concentrator through HPLC, and the I-type concentrator uploads the identification result to the cloud server. According to the invention, the processing speed of the whole data can be improved, the transmission time delay is reduced, and the intelligent real-time monitoring of the video monitoring data and the power consumption data is realized.

Description

Technical Field

[0001] The present invention belongs to the field of electric power monitoring, and in particular relates to an HPLC-based electric power big data intelligent monitoring method and system. Background Art

[0002] Traditional power information transmission does not have high requirements for real-time performance, and most of it belongs to non-real-time data transmission. However, with the in-depth application of power Internet of Things technology in the power system, the power industry pays more attention to the real-time processing and transmission of data. Existing power Internet of Things data transmission technologies include optical fiber, 5G, TD-LTE230, high-speed power line communication (HPLC), etc. Since HPLC is a communication technology that transmits data through power lines and does not require the re-laying of new communication lines, the construction cost is greatly reduced. It is currently a commonly used communication technology on the user side of the power system.

[0003] Driven by the power Internet of Things technology, the power industry is facing new challenges in real-time data transmission and processing. At the same time, power safety has become an important issue that needs urgent attention. In order to ensure the stable operation of the power system and the safety of power users, the power industry has begun to explore and apply various advanced technologies. Among them, the introduction of the video recognition system aims to improve the safety management level of the power user area. However, due to the bandwidth limitation of HPLC technology, there may be a delay when transmitting big data. When using HPLC technology to transmit video data, the existing technology is to compress the video in blocks to achieve its transmission on the power line, resulting in the maximum transmission distance is limited and susceptible to interference, and the quality of the video received by the server cannot be guaranteed; or by building a wireless communication system to achieve video transmission, which undoubtedly greatly increases the complexity and cost of the system.

[0004] In addition, the frequent occurrence of electricity theft also seriously threatens the security of electricity data and the stability of the power system. With the growth of energy demand, the economic pressure on power suppliers is increasing. Electricity theft not only causes economic losses, but also undermines fair competition in the market. Therefore, strengthening the security of electricity data and cracking down on electricity theft are crucial to ensuring the overall safety of the power system. Existing technologies can detect potential electricity theft by monitoring users' electricity usage behavior in real time through power monitoring systems. However, normal electricity usage patterns are sometimes misjudged as electricity theft, resulting in false alarms and customer dissatisfaction, and low data quality may lead to inaccurate analysis results, which in turn affects the effectiveness of electricity theft detection. Summary of the invention

[0005] In view of the problems existing in the prior art, the present invention provides a method and system for intelligent monitoring of electric power big data based on HPLC. By setting an edge computing system on the terminal side of the power substation, local video data and power consumption data are read in real time. Under the condition that data integrity, security and stability are guaranteed, the big data files are processed in the substation, the overall data processing speed is improved, the transmission delay is reduced, and intelligent real-time monitoring of the substation video surveillance data and power consumption data is realized.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions: a method for intelligent monitoring of power big data based on HPLC, performing the following steps:

[0007] S1. Construct an electrician recognition sample data set, including several electrician face images and corresponding identity IDs; construct an electrician face recognition neural network, and use the electrician recognition data set to train the electrician face recognition neural network to obtain an electrician face recognition model;

[0008] S2. Construct a sample data set for electrician clothing recognition, including several images of electricians’ clothing, and corresponding standard and irregular clothing labels; construct an electrician clothing recognition neural network, and use the sample data set for electrician clothing recognition to train the electrician clothing recognition neural network to obtain an electrician clothing recognition model;

[0009] S3, constructing a user electricity consumption information sample data set, including electricity consumption information of several users and corresponding normal or abnormal labels, constructing an electricity consumption detection neural network, and using the user electricity consumption information sample data set to train the electricity consumption recognition neural network to obtain an electricity consumption detection model;

[0010] S4. Deploy the electrician face recognition model, electrician clothing recognition model, and electricity usage detection model to the edge server, and input the electrician's face image, electrician's clothing image, and user's electricity usage collected in real time into the electrician face recognition model, electrician clothing recognition model, and electricity usage detection model in the cloud edge server to obtain the electrician's identity ID recognition result, the electrician's dress code recognition result, and the user's normal or abnormal electricity usage recognition result; the edge server sends the recognition result to the Type II collector, the Type II collector transmits the recognition result to the Type I concentrator through HPLC, and the Type I concentrator uploads the recognition result to the cloud server.

[0011] Furthermore, in the aforementioned step S1, the electrician recognition data set is used to train the electrician face recognition neural network specifically as follows: an Inception-ResNet module is used to extract the electrician face image, the Inception-ResNet module includes multiple MobileNet modules, each module is composed of a depth-separable convolution kernel, the depth-separable convolution kernel performs multi-scale feature extraction and splicing on the electrician face image, and then directly adds the spliced ​​features through residual connection, outputs the face feature vector, and finally compares the face feature vector with the template feature vector, calculates the feature distance to obtain the feature comparison result.

[0012] Furthermore, the aforementioned depth-wise separable convolution configures the standard convolution kernel to perform the following actions: first, each channel is convolved separately using depth-wise convolution, and then point-wise convolution is used to integrate features between channels.

[0013] Furthermore, the aforementioned training of the electrician face recognition neural network uses a combination function of ArcFace Loss and CenterLoss to optimize feature distribution;

[0014] ArcFace Loss increases the inter-class interval in the angle space. By setting the angle offset margin, the feature vectors of different categories are distributed in a larger angle interval to enhance the distinguishability.

[0015] Center Loss concentrates the feature vectors of samples of the same class in space and reduces the distance within the class.

[0016] The ArcFace Loss function is as follows:

[0017] Where: s is the scaling factor, which is used to amplify the distribution of feature vectors; m is the angle offset margin, which is used to increase the angle interval between categories. Indicates the correct category angle corresponding to the i-th sample;

[0018] The Center Loss function is as follows: Among them, f(x i ) is the feature vector of the i-th sample, c yi is the feature center of its corresponding category;

[0019] The total loss function is a weighted combination of the ArcFace Loss function and the Center Loss function:

[0020] Furthermore, in the aforementioned step S2, the electrician clothing recognition neural network includes: an input layer, a convolution layer, a global average pooling layer, a fully connected layer, and an output layer connected in sequence; the convolution layer includes: a first convolution unit, a first residual block, a second residual block, a third residual block, and a fourth residual block connected in sequence;

[0021] The input layer outputs the electrician's clothing image to the convolution layer. The first convolution unit extracts the initial image features of the electrician's clothing image and obtains the basic edge and texture information of the image.

[0022] The first residual block includes three residual units, which respectively extract shallow features from the initial image features output by the first convolution unit, capture local edge information in the image, and improve feature expression capabilities;

[0023] The second residual block includes four residual units, and the four residual units respectively extract deep features from the shallow features output by the first residual block;

[0024] The third residual block includes six residual units, and the six residual units further extract high-level semantic features from the deep features output by the second residual block;

[0025] The fourth residual block includes three residual units, which respectively extract global features of the image from the high-level semantic features output by the third residual block, capturing details and complex global structural information;

[0026] The global average pooling layer spatially averages each feature channel and outputs a one-dimensional feature vector;

[0027] The fully connected layer uses ReLU activation to output the one-dimensional feature vector as a 1024-dimensional activation vector;

[0028] The output layer outputs a probability vector through a number of neurons and a softmax activation function, and the probability vector corresponds to the probability of each clothing category.

[0029] Furthermore, in the aforementioned step S2, during the training of the electrician clothing recognition neural network, the cross entropy loss function is used to measure the gap between the probability distribution of the network output and the true label, as shown in the following formula:

[0030]

[0031] in, is the probability of each category predicted by the model, y i is the one-hot encoding of the true label; the network is trained using back-propagation and the Adam optimizer;

[0032] The Adam optimizer dynamically adjusts the learning rate to accelerate training by calculating the weighted average of the first-order moment and the second-order moment, as shown in the following formula:

[0033] in, and are the first-order and second-order moment estimates of the gradient, α is the learning rate, and ∈ is a constant to prevent division by zero errors.

[0034] Furthermore, the aforementioned power consumption detection neural network includes an input layer, a convolutional layer, an LSTM layer, an attention layer, a fully connected layer, an optimization layer, and an output layer connected in sequence;

[0035] The convolution layer includes a first convolution unit, a second convolution unit, a pooling unit, and a third convolution unit connected in sequence; the LSTM layer includes a first LSTM unit and a second LSTM unit connected in sequence; the fully connected layer includes a first fully connected unit and a second fully connected unit connected in sequence;

[0036] The input layer receives the user's electricity consumption information, including: the user's electricity consumption, electricity consumption time, voltage, current, and power time series window. The window size is T, and each time point has N features. The input is (T, N);

[0037] The first convolutional unit receives the input (T, N) and extracts local temporal features to capture small fluctuations in the power usage pattern;

[0038] The second convolution unit receives the output of the first convolution unit and further extracts deep features to capture abnormal fluctuations at different time scales;

[0039] The pooling unit reduces the data dimension of deep features and retains important features;

[0040] The third convolution unit captures high-level power usage features from the deep features output by the pooling layer;

[0041] The high-level electricity usage features output by the convolutional layer of the first LSTM unit capture the long-term dependencies of the time series, identifying cyclical and trend changes in electricity consumption, as well as long-term abnormal fluctuations;

[0042] The second LSTM unit further extracts the time-dependent features in the sequence based on the features output by the first LSTM unit;

[0043] The attention layer calculates the attention weights of each time step, allowing the network to focus on important time intervals and features;

[0044] The first fully connected unit further performs nonlinear transformation on the features output by the attention layer, integrating all the features extracted from the convolutional layer and LSTM. The second fully connected unit further refines the features and strengthens the discrimination ability of electricity theft detection.

[0045] The optimization layer is used to improve the stability of model training and accelerate the convergence process; the binary cross entropy loss function is used to distinguish electricity theft from normal electricity use:

[0046] in, is the predicted probability of the model, indicating that the sample is a positive class, that is, the probability of y = 1, the value range is [0,1], y is the true label, y = 1 indicates a positive class, y = 0 indicates a negative class, when the true label y = 1, the loss function only calculates Item; when the true label y = 0, the loss function only calculates item.

[0047] The output layer outputs the prediction result: 1 indicates electricity theft, and 0 indicates normal electricity use.

[0048] Furthermore, the aforementioned intelligent monitoring method for power big data based on HPLC, when extracting features, uses wavelet transform to process sudden changes and low-frequency changes in power signals, and then calculates the energy of wavelet coefficients at different scales to identify high-frequency mutations caused by power theft, and after decomposition through wavelet transform, extracts the low-frequency part of the signal to represent normal power consumption and the high-frequency part to represent power theft;

[0049] The discrete wavelet transform DWT is as follows:

[0050]

[0051] Among them, ψ(t) is the mother wavelet, a and b are scale and translation parameters respectively,

[0052] The energy calculation formula of wavelet coefficients is as follows:

[0053] Another aspect of the present invention provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of any method described in the present invention when executing the computer program.

[0054] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program implements the steps of any one of the methods described in the present invention when executed by a processor.

[0055] Compared with the prior art, the beneficial technical effects of the present invention using the above technical solution are as follows:

[0056] 1. Based on edge computing technology, the present invention places user-side big data files (video and electricity consumption data) on the edge server for processing, and transmits the results in the form of text to the cloud through HPLC, which effectively reduces resource consumption and transmission delay, and ensures the real-time nature of the data.

[0057] 2. The present invention sets up an edge computing system on the terminal side of the power substation, integrates the wireless / fiber optic communication system originally used to transmit big data into the HPLC system, and transmits it on the existing power lines through HPLC technology, effectively reducing the system complexity and reducing the system cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 It is a flow chart of the present invention.

[0059] Figure 2 This is the architecture diagram of the power Internet of Things system.

[0060] Figure 3 It is the algorithm flow chart of the present invention.

[0061] Figure 4 It is a flow chart of the face recognition algorithm in the present invention.

[0062] Figure 5 It is a flow chart of the clothing recognition algorithm of the present invention.

[0063] Figure 6 It is a flow chart of the anti-electricity theft monitoring algorithm of the present invention. DETAILED DESCRIPTION

[0064] In order to better understand the technical content of the present invention, specific embodiments are given and described as follows in conjunction with the accompanying drawings.

[0065] Various aspects of the invention are described herein with reference to the accompanying drawings, in which many illustrative embodiments are shown. The embodiments of the invention are not limited to those described in the accompanying drawings. It should be understood that the invention is implemented by any of the various concepts and embodiments described above, as well as the concepts and embodiments described in detail below, because the concepts and embodiments disclosed in the invention are not limited to any implementation. In addition, some aspects disclosed in the invention may be used alone or in any appropriate combination with other aspects disclosed in the invention.

[0066] refer to Figure 1 and Figure 2 , Figure 3 The present invention provides a method for intelligent monitoring of power big data based on HPLC, comprising the following steps:

[0067] S1. Construct an electrician recognition sample data set, including several electrician face images and corresponding identity IDs; construct an electrician face recognition neural network, and use the electrician recognition data set to train the electrician face recognition neural network to obtain an electrician face recognition model;

[0068] S2. Construct a sample data set for electrician clothing recognition, including several images of electricians’ clothing, and corresponding standard and irregular clothing labels; construct an electrician clothing recognition neural network, and use the sample data set for electrician clothing recognition to train the electrician clothing recognition neural network to obtain an electrician clothing recognition model;

[0069] S3, constructing a user electricity consumption information sample data set, including electricity consumption information of several users and corresponding normal or abnormal labels, constructing an electricity consumption detection neural network, and using the user electricity consumption information sample data set to train the electricity consumption recognition neural network to obtain an electricity consumption detection model;

[0070] S4. Deploy the electrician face recognition model, electrician clothing recognition model, and electricity usage detection model to the edge server, and input the electrician's face image, electrician's clothing image, and user's electricity usage collected in real time into the electrician face recognition model, electrician clothing recognition model, and electricity usage detection model in the cloud edge server to obtain the electrician's identity ID recognition result, the electrician's dress code recognition result, and the user's normal or abnormal electricity usage recognition result; the edge server sends the recognition result to the Type II collector, the Type II collector transmits the recognition result to the Type I concentrator through HPLC, and the Type I concentrator uploads the recognition result to the cloud server.

[0071] When implementing it, Figure 4 As shown, in step S1, an electrician recognition sample data set is constructed. A camera can be set in the electricity usage scene, the camera captures video information, the edge server receives the video stream shot by the camera through RTSP, intercepts the image in the video stream, uses the face detection algorithm to detect the face area from the image, intercepts the detected face area, unifies the image size, and normalizes the image so that its pixel value is between 0 and 1.

[0072] Common models used in the electrician face recognition neural network include VGGFace, FaceNet, ResNet, etc. The specific training of the electrician face recognition neural network using the electrician recognition data set is as follows: the Inception-ResNet module is used to extract the electrician face image. The Inception-ResNet module includes multiple MobileNet modules, each of which is composed of a deep separable convolution kernel. The deep separable convolution kernel performs multi-scale feature extraction and splicing on the electrician face image, and then directly adds the spliced ​​features through residual connection to output the face feature vector. Finally, by comparing the face feature vector and the template feature vector, the feature distance is calculated to obtain the feature comparison result. Specifically, the distance threshold is set according to the application scenario. If the distance is less than the threshold, it is considered to be the same person, otherwise it is considered to be different people. According to the distance calculation result, the recognition result is determined and the corresponding face ID is returned.

[0073] The Inception-ResNet module is used at the beginning of the electrician face recognition neural network to extract multi-scale features of face images. This module contains multiple parallel convolution paths, using convolution kernels of different sizes (such as 1×1, 3×3), and is equipped with a maximum pooling layer to extract detailed features of different scales. After the output of each path is spliced, the input is directly added through the residual connection to form the final output, avoiding gradient vanishing and improving the effectiveness of deep feature extraction.

[0074] Multiple MobileNet modules are embedded in the main network structure, and each module consists of a depthwise separable convolution. Depthwise separable convolution divides the standard convolution into two steps: first, each channel is convolved separately using depthwise convolution, and then pointwise convolution (1×1 convolution) is used to integrate features between channels. This design significantly reduces the number of parameters and computation, making the network more lightweight.

[0075] The combination of ArcFace Loss and Center Loss is used in model training to optimize feature distribution. ArcFace Loss increases the inter-class interval in the angle space. By setting the angle offset (margin), the feature vectors of different categories are distributed in a larger angle interval to enhance the distinguishability. Center Loss ensures that the feature vectors of samples of the same category are more concentrated in space and reduces the intra-class distance. The combination of the two improves the recognition accuracy of the model.

[0076] ArcFace Loss function:

[0077] Where: s is the scaling factor, which is used to amplify the distribution of feature vectors, m is the angular offset (margin), which is used to increase the angular interval between categories, Indicates the correct category angle corresponding to the i-th sample.

[0078] Center Loss function:

[0079] Among them, f(x i ) is the feature vector of the i-th sample. is the feature center of its corresponding category.

[0080] The total loss function is a weighted combination of the ArcFace Loss function and the Center Loss function:

[0081] The output after the Inception-ResNet module and the MobileNet module will enter the global average pooling layer, and finally generate a 128-dimensional or 512-dimensional feature vector to represent the face of the input image. This feature vector is very compact and discriminative, and can be used to compare the similarity between different faces.

[0082] When implementing it, Figure 5 As shown in the figure, when training the neural network for electrician clothing recognition, a sample library containing clothing recognition is constructed, and these samples are divided into training sets and test sets in a certain proportion to ensure that the model can effectively evaluate its performance in the training and verification stages. The sample images are preprocessed, including image cropping, resizing, normalization and other steps to ensure the consistency and quality of the input images; corresponding labels (standard and non-standard) are added to each image to identify the category of the image so that the model can perform supervised learning; training parameters such as learning rate, batch size, optimizer selection, etc. are adjusted to optimize the model training process. During the training process, the model is repeatedly tested and adjusted. After each training cycle, the model is evaluated using the test set, and the model parameters and structure are adjusted according to the evaluation results. After multiple iterations, a model with the best performance on the test set is finally obtained.

[0083] The electrician clothing recognition neural network in step S2 includes: an input layer, a convolution layer, a global average pooling layer, a fully connected layer, and an output layer connected in sequence; the convolution layer includes: a first convolution unit, a first residual block, a second residual block, a third residual block, and a fourth residual block connected in sequence;

[0084] The input layer outputs the electrician's clothing image to the convolution layer. The first convolution unit extracts the initial image features of the electrician's clothing image and obtains the basic edge and texture information of the image. The input data format is: the input is RGB image data. Assume that the input shape of each sample is 224×224×3, that is, 224 pixels high, 224 pixels wide, and 3 color channels. The convolution kernel size is 7×7, the number of convolutions is 64, the stride is 2, and the activation function is: ReLU.

[0085] The first residual block includes three residual units, which respectively extract shallow features from the initial image features output by the first convolution unit, capture local edge information in the image, and improve feature expression capabilities; input shape: (112, 112, 64), each residual unit contains the following convolution layers: convolution kernel sizes are: 1×1, 3×3, 1×1, number of convolutions: 64, 64, 256, respectively, activation function: ReLU

[0086] The second residual block includes four residual units, each of which contains the following convolutional layers: the convolution kernel sizes are 1×1, 3×3, and 1×1, and the number of convolutions is 128, 128, and 512 respectively. The four residual units extract deep features from the shallow features output by the first residual block to capture a wider range of patterns and local structures; input shape: (56, 56, 256), activation function: ReLU.

[0087] The third residual block includes six residual units, with an input shape of (28, 28, 512). Each unit contains the following convolutional layers: kernel size: 1×1, 3×3, 1×1, number of convolutions: 256, 256, 1024 respectively. The six residual units further extract high-level semantic features from the deep features output by the second residual block to enhance the model's understanding of complex image content. Activation function: ReLU.

[0088] The fourth residual block includes three residual units, input shape: (14, 14, 1024), including 3 residual units, each unit contains the following convolutional layers: convolution kernel size: 1×1, 3×3, 1×1, number of products: 512, 512, 2048 respectively, activation function: ReLU. The three residual units extract the global features of the image from the high-level semantic features output by the third residual block, capturing details and more complex global structural information.

[0089] The input shape of the global average pooling layer is: (7, 7, 2048), which spatially averages each feature channel and outputs a one-dimensional feature vector of length 2048.

[0090] The fully connected layer uses ReLU activation to output the one-dimensional feature vector as a 1024-dimensional activation vector;

[0091] The output layer outputs a probability vector through a number of neurons and a softmax activation function, and the probability vector corresponds to the probability of each clothing category.

[0092] During the training process, the cross entropy loss function is used to measure the gap between the probability distribution of the model output and the true label. The specific formula is:

[0093] in, is the probability of each category predicted by the model, y i is the one-hot encoding of the true label. The network is trained through back propagation and optimization algorithms (such as Adam). The Adam optimizer dynamically adjusts the learning rate to accelerate training by calculating the weighted average of the first-order moment and the second-order moment. The specific update rule is:

[0094] in, and are the first-order and second-order moment estimates of the gradient, α is the learning rate, and ∈ is a small constant to prevent division by zero errors.

[0095] Through this structure, ResNet50 first acts as a powerful feature extractor, and outputs category predictions after global average pooling and fully connected layers. During the training process, the cross entropy loss function is used for optimization, and the Adam optimizer helps adjust the weights to effectively achieve the clothing recognition task.

[0096] like Figure 6 As shown in the figure, when training the power consumption detection neural network, step 1: collect a large amount of user power consumption data. The data should include the user's power consumption, power consumption time, voltage, current and other information;

[0097] Step 2: Clean and preprocess the collected data, including removing outliers, processing missing data, and performing time series analysis to ensure data quality and consistency.

[0098] Step 3: Select features related to the user's electricity usage behavior, such as electricity usage pattern, peak and valley distribution of electricity usage, daily / weekly / monthly electricity usage pattern, energy-saving behavior, etc.

[0099] Step 4: Label the data according to the specific task to identify normal and abnormal power usage patterns;

[0100] Step 5: For the anomaly detection problem, a convolutional neural network model is used. The selected model is trained using a labeled dataset so that the model can learn the patterns of user electricity usage characteristics.

[0101] Step 6: Use cross-validation or reserve a portion of the data as a test set to evaluate the performance of the model. According to the performance in the actual application, optimize and adjust the model to improve the prediction accuracy.

[0102] The power consumption detection neural network includes a sequentially connected input layer, a convolutional layer, an LSTM layer, an attention layer, a fully connected layer, an optimization layer, and an output layer;

[0103] The convolution layer includes the first convolution unit, the second convolution unit, the pooling unit, and the third convolution unit connected sequentially; the input data of the convolution layer is usually time series data, such as current, voltage, power, etc. Assuming that the input of each sample is a time series window with a size of T, and each time point has N features, the input shape is (T, N).

[0104] The LSTM layer includes a first LSTM unit and a second LSTM unit connected sequentially; the fully connected layer includes a first fully connected unit and a second fully connected unit connected sequentially;

[0105] The input layer receives the user's electricity consumption information, including: the user's electricity consumption, electricity consumption time, voltage, current, and power time series window. The window size is T, and each time point has N features. The input is (T, N);

[0106] The first convolution unit receives input (T, N), convolution kernel size: (3, N) or (5, N), number of convolutions: 64, activation function: ReLU. The first convolution unit is used to extract local time features and capture small fluctuations in power consumption patterns.

[0107] The second convolution unit receives the output of the first convolution unit, with input shape: (T, 64) (from the output of the previous convolution layer), convolution kernel size: (3, 64), number of convolutions: 128, activation function: ReLU. The second convolution unit further extracts deep features and captures abnormal fluctuations at different time scales;

[0108] The pooling unit reduces the data dimension for deep features, and the pooling size is: (2, 1) The pooling layer is used to reduce the data dimension, improve computational efficiency, and retain important features.

[0109] The input shape of the third convolution unit is: (T / 2, 128), convolution kernel size: (3, 128), number of convolutions: 256, activation function: ReLU

[0110] The deep features output by the pooling layer capture high-level electricity usage characteristics and further improve the detection accuracy of electricity theft.

[0111] The first LSTM unit input shape is: (T / 2, 256), the number of LSTM units: 128, and the activation function: tanh. The high-level electricity usage features output by the first LSTM unit on the convolutional layer capture the long-term dependencies of the time series, identify periodic and trend changes in electricity consumption, and long-term abnormal fluctuations.

[0112] The number of LSTM units in the second LSTM unit is 64, and the activation function is tanh. The second LSTM unit further extracts the time-dependent features in the sequence based on the features output by the first LSTM unit.

[0113] The input shape of the attention layer is (T / 2, 64). The attention layer calculates the attention weight of each time step to make the model pay attention to more important time intervals and features. For example, the model can automatically learn which time period of power consumption fluctuation is more likely to be a manifestation of power theft. The structure of the attention layer is: using self-attention or weighted average strategy to assign different weights to different time steps.

[0114] The first fully connected unit input shape: (T / 2, 64), number of neurons: 128, activation function: ReLU. The first fully connected unit further nonlinearly transforms the features output by the attention layer and integrates all features extracted from the convolutional layer and LSTM. The second fully connected unit input shape: 128, number of neurons: 64, activation function: ReLU. The second fully connected unit is used to further refine features and enhance the discrimination ability of electricity theft detection.

[0115] The optimization layer is used to improve the stability of model training. BatchNormalization is added after the convolutional layer and LSTM layer to accelerate the convergence process. The binary cross entropy loss function is used to distinguish electricity theft from normal electricity use:

[0116] in, is the predicted probability of the model, indicating that the sample is a positive class, that is, the probability of y = 1, the value range is [0,1], y is the true label, y = 1 indicates a positive class, y = 0 indicates a negative class, when the true label y = 1, the loss function only calculates Item; when the true label y = 0, the loss function only calculates item.

[0117] The output layer outputs the prediction result: 1 indicates electricity theft, and 0 indicates normal electricity use.

[0118] Wavelet transform has good localization characteristics in time domain and frequency domain in time-frequency analysis, and is especially suitable for sudden changes and low-frequency changes in power signals. Wavelet transform can provide finer-grained time-frequency information than Fourier transform.

[0119] When extracting features, wavelet transform is used to process sudden changes and low-frequency changes in power signals, as shown in the following formula:

[0120] Discrete Wavelet Transform (DWT):

[0121] Among them, ψ(t) is the mother wavelet, a and b are scale and translation parameters respectively. By decomposing the signal into low-frequency and high-frequency components through wavelet transform, the instantaneous fluctuation caused by electricity theft can be effectively captured.

[0122] Energy characteristics: After decomposition through wavelet transform, the low-frequency part of the signal (usually representing normal power consumption) and the high-frequency part (possibly representing power theft) are extracted. The energy of the wavelet coefficients of each layer can be calculated by the following formula:

[0123]

[0124] By calculating the energy of wavelet coefficients at different scales, high-frequency mutations caused by electricity theft can be identified.

[0125] The trained model is used to directly identify whether the user's electricity consumption is normal or abnormal, as follows: the terminal device sends the electricity consumption information collection data to the edge server; the edge server determines whether the user's electricity consumption is abnormal through the convolutional neural network model. If there is an abnormality, the edge server sends the alarm information and electricity consumption information to the type II collector through RS-485; if there is no abnormality, the edge server sends the electricity consumption information to the type II collector through RS-485; the type II collector transmits the data to the type I concentrator through HPLC, and the type I concentrator uploads the data to the cloud; in the embodiment, the type II collector is a DCZL13-KLGC01 collector, and the type I concentrator is a DJGZ23-DXC concentrator.

[0126] Although the present invention has been described above with preferred embodiments, it is not intended to limit the present invention. A person skilled in the art of the present invention may make various modifications and improvements without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention shall be determined by the claims.

[0127] Another aspect of the present invention provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of any one of the methods described in the present invention when executing the computer program.

[0128] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program implements the steps of any one of the methods described in the present invention when executed by a processor.

Claims

1. A method for intelligent monitoring of power big data based on HPLC, characterized in that: Perform the following steps: S1. Construct an electrician recognition sample data set, including several electrician face images and corresponding identity IDs; construct an electrician face recognition neural network, and use the electrician recognition data set to train the electrician face recognition neural network to obtain an electrician face recognition model; S2. Construct a sample data set for electrician clothing recognition, including several images of electricians’ clothing, and corresponding standard and irregular clothing labels; construct an electrician clothing recognition neural network, and use the sample data set for electrician clothing recognition to train the electrician clothing recognition neural network to obtain an electrician clothing recognition model; S3, constructing a user electricity consumption information sample data set, including electricity consumption information of several users and corresponding normal or abnormal labels, constructing an electricity consumption detection neural network, and using the user electricity consumption information sample data set to train the electricity consumption recognition neural network to obtain an electricity consumption detection model; S4. Deploy the electrician face recognition model, electrician clothing recognition model, and electricity usage detection model to the edge server, input the electrician's face image, electrician clothing image, and user electricity usage collected in real time into the electrician face recognition model, electrician clothing recognition model, and electricity usage detection model in the cloud edge server, and obtain the electrician ID recognition result, the electrician clothing standard recognition result, and the user's normal or abnormal electricity usage recognition result; The edge server sends the recognition results to the Type II collector, which transmits the recognition results to the Type I concentrator through HPLC, and the Type I concentrator uploads the recognition results to the cloud server.

2. The HPLC-based power big data intelligent monitoring method according to claim 1 is characterized in that: In step S1, the electrician recognition data set is used to train the electrician face recognition neural network. Specifically, the Inception-ResNet module is used to extract the electrician face image. The Inception-ResNet module includes multiple MobileNet modules, each of which is composed of a depth-separable convolution kernel. The depth-separable convolution kernel performs multi-scale feature extraction and splicing on the electrician face image, and then directly adds the spliced ​​features through residual connection, outputs the face feature vector, and finally compares the face feature vector with the template feature vector to calculate the feature distance to obtain the feature comparison result.

3. The HPLC-based power big data intelligent monitoring method according to claim 2 is characterized in that: Depthwise separable convolution configures the standard convolution kernel to perform the following actions: first, use depthwise convolution to convolve each channel separately, and then use point-by-point convolution to integrate features between channels.

4. The HPLC-based power big data intelligent monitoring method according to claim 2 is characterized in that: When training the electrician face recognition neural network, a combination of ArcFace Loss and Center Loss is used to optimize feature distribution; ArcFace Loss increases the inter-class interval in the angle space and allows feature vectors of different categories to be distributed over a larger angle interval by setting the angle offset margin, thereby enhancing the distinguishability. Center Loss concentrates the feature vectors of samples of the same class in space and reduces the distance within the class. The ArcFace Loss function is as follows: Where: s is the scaling factor, which is used to amplify the distribution of feature vectors; m is the angle offset margin, which is used to increase the angle interval between categories. Indicates the correct category angle corresponding to the i-th sample; The Center Loss function is as follows: Among them, f(x i ) is the feature vector of the i-th sample, is the feature center of its corresponding category; The total loss function is a weighted combination of the ArcFace Loss function and the Center Loss function:

5. The HPLC-based power big data intelligent monitoring method according to claim 1 is characterized in that: In step S2, the electrician clothing recognition neural network includes: an input layer, a convolution layer, a global average pooling layer, a fully connected layer, and an output layer connected in sequence; the convolution layer includes: a first convolution unit, a first residual block, a second residual block, a third residual block, and a fourth residual block connected in sequence; The input layer outputs the electrician's clothing image to the convolution layer. The first convolution unit extracts the initial image features of the electrician's clothing image and obtains the basic edge and texture information of the image. The first residual block includes three residual units, which respectively extract shallow features from the initial image features output by the first convolution unit, capture local edge information in the image, and improve feature expression capabilities; The second residual block includes four residual units, and the four residual units respectively extract deep features from the shallow features output by the first residual block; The third residual block includes six residual units, and the six residual units further extract high-level semantic features from the deep features output by the second residual block; The fourth residual block includes three residual units, which respectively extract global features of the image from the high-level semantic features output by the third residual block, capturing details and complex global structural information; The global average pooling layer spatially averages each feature channel and outputs a one-dimensional feature vector; The fully connected layer uses ReLU activation to output the one-dimensional feature vector as a 1024-dimensional activation vector; The output layer outputs a probability vector through a number of neurons and a softmax activation function, and the probability vector corresponds to the probability of each clothing category.

6. The HPLC-based power big data intelligent monitoring method according to claim 5 is characterized in that: In step S2, during the training of the electrician clothing recognition neural network, the cross entropy loss function is used to measure the gap between the probability distribution of the network output and the true label, as shown in the following formula: in, is the probability of each category predicted by the model, y i is the one-hot encoding of the true label; the network is trained using back-propagation and the Adam optimizer; The Adam optimizer dynamically adjusts the learning rate to accelerate training by calculating the weighted average of the first-order moment and the second-order moment, as shown in the following formula: in, and are the first-order and second-order moment estimates of the gradient, α is the learning rate, and ∈ is a constant to prevent division by zero errors.

7. The HPLC-based power big data intelligent monitoring method according to claim 1, characterized in that: The power consumption detection neural network includes a sequentially connected input layer, a convolutional layer, an LSTM layer, an attention layer, a fully connected layer, an optimization layer, and an output layer; The convolution layer includes a first convolution unit, a second convolution unit, a pooling unit, and a third convolution unit connected in sequence; the LSTM layer includes a first LSTM unit and a second LSTM unit connected in sequence; the fully connected layer includes a first fully connected unit and a second fully connected unit connected in sequence; The input layer receives the user's electricity consumption information, including: the user's electricity consumption, electricity consumption time, voltage, current, and power time series window. The window size is T, and each time point has N features. The input is (T, N); The first convolutional unit receives the input (T, N) and extracts local temporal features to capture small fluctuations in the power usage pattern; The second convolution unit receives the output of the first convolution unit and further extracts deep features to capture abnormal fluctuations at different time scales; The pooling unit reduces the data dimension of deep features and retains important features; The third convolution unit captures high-level power usage features from the deep features output by the pooling layer; The high-level electricity usage features output by the convolutional layer of the first LSTM unit capture the long-term dependencies of the time series, identifying cyclical and trend changes in electricity consumption, as well as long-term abnormal fluctuations; The second LSTM unit further extracts the time-dependent features in the sequence based on the features output by the first LSTM unit; The attention layer calculates the attention weights of each time step, allowing the network to focus on important time intervals and features; The first fully connected unit further performs nonlinear transformation on the features output by the attention layer, integrating all the features extracted from the convolutional layer and LSTM. The second fully connected unit further refines the features and strengthens the discrimination ability of electricity theft detection. The optimization layer is used to improve the stability of model training and accelerate the convergence process; the binary cross entropy loss function is used to distinguish electricity theft from normal electricity use: in, is the predicted probability of the model, indicating that the sample is a positive class, that is, the probability of y = 1, the value range is [0,1], y is the true label, y = 1 indicates a positive class, y = 0 indicates a negative class, when the true label y = 1, the loss function only calculates Item; when the true label y = 0, the loss function only calculates item; The output layer outputs the prediction result: 1 indicates electricity theft, and 0 indicates normal electricity use.

8. The HPLC-based power big data intelligent monitoring method according to claim 7 is characterized in that: When extracting features, wavelet transform is used to process sudden changes and low-frequency changes in power signals, and then the energy of wavelet coefficients at different scales is calculated to identify high-frequency mutations caused by power theft. After decomposition through wavelet transform, the low-frequency part of the extracted signal represents normal power consumption and the high-frequency part represents power theft; The discrete wavelet transform DWT is as follows: Among them, ψ(t) is the mother wavelet, a and b are scale and translation parameters respectively, The energy calculation formula of wavelet coefficients is as follows:

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images