Deep learning-based multi-energy complementary power distribution system data analysis method and system

Abstract

According to the multi-energy complementary power distribution system data analysis method based on deep learning, firstly, preprocessing and feature extraction are carried out on multi-source heterogeneous sensor data, and key indexes reflecting the health state of equipment are obtained; and obtaining globally consistent equipment state representation through a spatio-temporal data fusion technology. And then a multi-fault mode knowledge base is constructed, and a multi-label classification algorithm of deep learning is adopted to identify fault types. And finally, establishing a mapping relation between an equipment topological structure and a fault propagation path by using an attention mechanism and a graph convolutional network, and reasoning a specific position of equipment damage through fault traceability and influence analysis. The robustness and generalization ability of the model are improved by adopting methods of data enhancement, transfer learning and the like, and real-time and reliable fault location is realized through edge calculation. According to the method, the equipment fault in the multi-energy complementary power distribution system can be accurately positioned, and the reliability and safety of the system are improved.

Description

Technical Field

[0001] The present invention relates to the field of electric power operation technology, and specifically to a data analysis method and system for a multi-energy complementary distribution system based on deep learning. Background Art

[0002] In the multi-energy complementary power distribution system, equipment failure and damage often occur due to the wide variety and distribution of equipment. Traditional manual inspection and fault location methods are inefficient and difficult to meet the needs of smart grids. How to use deep learning technology and combine multi-source heterogeneous sensor data to accurately locate the location of equipment damage is a technical problem that needs to be solved urgently. Summary of the invention

[0003] In order to make up for the shortcomings of the prior art, the present invention proposes a data analysis method and system for a multi-energy complementary distribution system based on deep learning.

[0004] The technical solution adopted by the present invention to solve its technical problem is: The data analysis method of multi-energy complementary distribution system based on deep learning includes: Step 1, acquiring multi-source heterogeneous sensor data in a multi-energy complementary power distribution system, and preprocessing the multi-source heterogeneous sensor data, wherein the preprocessing includes cleaning, alignment and standardization processing, eliminating noise and abnormal values ​​in the multi-source heterogeneous sensor data, and obtaining preprocessed sensor data; Step 2, for the preprocessed sensor data, using a feature extraction algorithm to extract health status indicators and fault characteristic indicators, based on the health status indicators, determine whether the device is in a healthy, sub-healthy or faulty state, if the device is in a sub-healthy or faulty state, based on the fault characteristic indicators, filter out a feature subset of the device damage position to obtain the device status feature; Step 3, based on the device state characteristics, the device state characteristics are synchronized and aligned by using spatiotemporal data fusion technology to obtain a globally consistent device state representation; Step 4: Build a knowledge base containing multiple fault modes, use a multi-label classification algorithm based on deep learning to identify and judge the globally consistent device state representation, and determine the device fault type; Step 5: After determining the type of equipment failure, establish the equipment topology structure, use the attention mechanism and graph convolutional network technology to obtain the mapping relationship of the fault propagation path, and determine the specific location of the equipment damage through fault tracing and impact analysis.

[0005] Preferably, step 6 is also included, using data enhancement and transfer learning methods to expand the knowledge base containing multiple fault modes, and through edge computing and distributed reasoning technology, a multi-label classification algorithm based on deep learning is used to identify and judge the globally consistent device state representation to obtain a globally consistent fault mode recognition result.

[0006] Preferably, in step 1, the multi-source heterogeneous sensor data is cleaned according to the data cleaning rules to obtain the cleaned sensor data, and the cleaning rules include removing missing values ​​and outliers; the cleaned sensor data is standardized by using the minimum-maximum normalization method, and the data is mapped to a unified scale space of [0,1] to obtain standardized data; the standardized data is aligned in the time dimension using the dynamic time warping algorithm to obtain aligned sensor data, and the aligned sensor data is denoised by wavelet transform, the number of decomposition layers is selected, and the high-frequency components of each layer are threshold processed, and then reconstructed to obtain the denoised sensor data; the reconstructing to obtain the denoised sensor data includes using a density-based spatial clustering algorithm to identify outliers in the aligned sensor data, marking the outlier points as noise points and eliminating them, so as to obtain the final cleaned, aligned, denoised and outlier-eliminated sensor data.

[0007] Preferably, in step 2, the preprocessed sensor data is subjected to frequency domain analysis by using fast Fourier transform to extract frequency domain features of spectrum energy and spectrum entropy; at the same time, the preprocessed sensor data is subjected to time-frequency domain analysis by using wavelet packet decomposition to extract wavelet coefficients of the preprocessed sensor data at different scales as time domain features; the extracted frequency domain features and time domain features of spectrum entropy reflect the health status and fault features of the equipment; Preset the threshold range of the health status of the device, compare the extracted frequency domain characteristics and time domain characteristics of the spectrum entropy with the preset threshold, and judge whether the device is currently in a healthy, sub-healthy or faulty state through the threshold judgment method; If the judgment result is that the equipment is in a sub-healthy or faulty state, then the Pearson correlation coefficient is used to calculate the correlation between each feature and the equipment damage location based on the extracted fault features, and the top N features with the highest correlation are selected as the equipment status feature subset. The screened equipment status feature subset is used as input, and the support vector machine algorithm is used to train the equipment status diagnosis model; the grid search and cross-validation methods are used to optimize the kernel function type and kernel function parameters of the support vector machine. When new equipment sensor data is input, after preprocessing and feature extraction, the extracted spectral entropy frequency domain features and time domain features are input into the trained equipment status diagnosis model. The trained equipment status diagnosis model performs classification prediction based on the input features to determine the health status of the equipment; If the equipment is diagnosed as sub-healthy or faulty, the fault feature subset screened out by the diagnosis result is used to train the support vector machine prediction model of the equipment damage location using the support vector regression algorithm; The support vector machine predicts the continuous value of the equipment damage location by finding the optimal regression hyperplane in the feature space; The fault feature subset is input into the trained support vector machine prediction model, and the trained support vector machine prediction model performs regression prediction based on the input features to obtain the equipment status features.

[0008] Preferably, in step 3, the device state characteristics are synchronized in time through time interpolation, and the device state characteristics are unified on the same time axis. The device state characteristics are spatially aligned through a coordinate transformation method, and the device state characteristics are unified in the same spatial coordinate system. A spatiotemporal data fusion algorithm is selected, and a Kalman filter is used to establish a data fusion model. The sensor data after time synchronization and spatial alignment are used as input, and the state parameters of the device are calculated through the spatiotemporal data fusion algorithm. In the data fusion process, a dynamic model of the device state is established, and the Kalman filter algorithm is used to predict and update the device state. The prediction result is corrected in combination with the multi-sensor observation data at the current moment to obtain an estimated value of the device state, and the device state estimated value obtained by data fusion is used as a globally consistent device state representation.

[0009] Preferably, determining the equipment failure type in step 4 includes: S401: Obtain historical operation data and fault records of the equipment, and build a knowledge base containing multiple equipment failure modes, where the knowledge base contains feature descriptions, impact scope, and severity of various failure modes; S402: For globally consistent device state representation, a convolutional neural network is used to extract local features and capture key patterns in the data; at the same time, a long short-term memory network is used to extract time series features and capture the dynamic change trend of the data; S403: Fusing the local features extracted by the convolutional neural network and the temporal features extracted by the long short-term memory network to obtain a comprehensive feature representation of the device state as input to the deep learning multi-label classification algorithm; S404: using a deep learning multi-label classification algorithm to classify and identify the comprehensive feature representation of the device status, outputting the probabilities of various fault modes through a softmax activation function layer, determining whether the device has a fault, and if so, outputting the top N possible fault modes with the highest probability; S405: Match the classification and identification results with the fault modes in the knowledge base of multiple equipment fault modes, obtain detailed information of the fault mode, including the cause of the fault, the scope of impact and the severity, and form an equipment fault diagnosis report; S406: According to the equipment fault diagnosis report, the corresponding fault handling process is automatically triggered, including alarm notification, emergency plan and maintenance work order, to determine the type of equipment fault.

[0010] Preferably, in step 5, the device topology information and fault information are obtained to construct a device topology map; for the device topology map, a graph attention network is used to perform representation learning on the nodes to obtain a low-dimensional vector representation of the nodes; based on the fault information, known fault nodes are marked on the device topology map; a graph convolutional network is used to perform end-to-end fault propagation paths on the device topology map to obtain potential fault propagation paths; based on the potential fault propagation paths, the root node of the fault is determined to determine the cause of the equipment failure; for the identified root node, a graph attention network is used to perform impact analysis on its neighborhood subgraph to obtain a set of affected nodes; the location information of the root node and the affected node in the device topology structure is fused to determine the specific location of the damaged equipment and output the equipment fault diagnosis result.

[0011] Preferably, in step 6, identifying and judging the globally consistent device state representation includes: S601: Acquire device status representation and perform preprocessing to obtain a standardized device status representation data set; S602: Using data enhancement technology, by rotating, translating, scaling, and adding noise to the standardized device state characterization data set, generating data samples with various changes, and expanding the training data set; S603: using a transfer learning method, selecting a convolutional neural network pre-training model related to the fault mode recognition task, and using the expanded data set to adjust the convolutional neural network pre-training model; S604: Build a multi-label classification model based on deep learning, design the network structure and loss function of the model, and initialize the model parameters using the adjusted pre-trained model; S605: Divide the device state representation data set preprocessed in S601 into a training set and a validation set, use the training set to train the multi-label classification model, evaluate the model performance on the validation set, and adjust the model hyperparameters according to the evaluation results; S606: Deploy the trained multi-label classification model to the edge device, use edge computing technology to implement local reasoning, and determine the failure mode of the device in real time based on the device status characterization data set; S607: Distributed reasoning technology is used to aggregate and fuse the reasoning results on multiple edge devices to obtain a globally consistent fault mode recognition result.

[0012] A multi-energy complementary power distribution system data analysis system based on deep learning, used to implement a multi-energy complementary power distribution system data analysis method based on deep learning, including: Data preprocessing module: used to obtain multi-source heterogeneous sensor data in the multi-energy complementary power distribution system, and preprocess the multi-source heterogeneous sensor data, wherein the preprocessing includes cleaning, alignment and standardization processing, eliminating noise and abnormal values ​​in the multi-source heterogeneous sensor data, and obtaining preprocessed sensor data; Feature extraction module: used for extracting health status indicators and fault characteristic indicators from the preprocessed sensor data by using a feature extraction algorithm, and judging whether the device is in a healthy, sub-healthy or faulty state based on the health status indicators. If the device is in a sub-healthy or faulty state, a feature subset of the device damage position is screened out based on the fault characteristic indicators to obtain the device status characteristics; Spatiotemporal data fusion module: used to synchronize and align the device status characteristics based on the device status characteristics through spatiotemporal data fusion technology to obtain a globally consistent device status representation; Fault identification module: used to build a knowledge base containing multiple fault modes, using a multi-label classification algorithm based on deep learning to identify and judge the globally consistent device status representation and determine the type of device fault; Analysis module: After determining the type of equipment fault, it is used to establish the equipment topology structure, use the attention mechanism and graph convolutional network technology to obtain the mapping relationship of the fault propagation path, and determine the specific location of the equipment damage through fault tracing and impact analysis.

[0013] It also includes: Data enhancement transfer learning module: used to establish the equipment topology structure after determining the type of equipment failure, use the attention mechanism and graph convolutional network technology to obtain the mapping relationship of the fault propagation path, and determine the specific location of equipment damage through fault tracing and impact analysis.

[0014] Compared with the prior art, the present invention has the following beneficial effects: The present invention discloses a data analysis method for a multi-energy complementary power distribution system based on deep learning. This method aims at the equipment failure problem in the multi-energy complementary power distribution system, and extracts health status and fault characteristic indicators through multi-source heterogeneous sensor data collection and preprocessing. The spatiotemporal data fusion technology is used to obtain a globally consistent representation of the equipment state, and a knowledge base containing multiple fault modes is constructed. A deep learning multi-label classification algorithm is used to identify the type of equipment failure, and the attention mechanism and graph convolution network technology are used in combination with the equipment topology to determine the specific location of the equipment damage. The present invention also introduces data enhancement and transfer learning methods to expand the fault mode knowledge base, and improves the fault identification efficiency through edge computing and distributed reasoning technology. This method can effectively improve the fault diagnosis accuracy and efficiency of the multi-energy complementary power distribution system, realize the precise positioning and impact analysis of equipment failures, and provide important support for the safe and stable operation of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A flowchart of a data analysis method for a multi-energy complementary distribution system based on deep learning; Figure 2 This is the structural block diagram of the data analysis system for multi-energy complementary distribution system based on deep learning.

[0016] Explanation of reference numerals: 201 - data processing module; 202 - feature extraction module; 203 - spatiotemporal data fusion module; 204 - fault identification module; 205 - analysis module. DETAILED DESCRIPTION

[0017] The technical solutions in the embodiments of the present invention are described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0018] Example 1 like Figure 1 As shown, the data analysis method of multi-energy complementary distribution system based on deep learning includes: Step 1, acquiring multi-source heterogeneous sensor data in a multi-energy complementary power distribution system, and preprocessing the multi-source heterogeneous sensor data, wherein the preprocessing includes cleaning, alignment and standardization processing, eliminating noise and abnormal values ​​in the multi-source heterogeneous sensor data, and obtaining preprocessed sensor data; Step 2, for the preprocessed sensor data, a feature extraction algorithm is used to extract health status indicators and fault characteristic indicators, and based on the health status indicators, it is determined whether the device is in a healthy, sub-healthy or faulty state. If the device is in a sub-healthy or faulty state, based on the fault characteristic indicators, a feature subset of the damaged position of the relevant device is screened out to obtain the device status feature; Step 3, based on the device state characteristics, the device state characteristics are synchronized and aligned by using spatiotemporal data fusion technology to obtain a globally consistent device state representation; Step 4: Build a knowledge base containing multiple fault modes, use a multi-label classification algorithm based on deep learning to identify and judge the globally consistent device state representation, and determine the device fault type; Step 5: After determining the type of equipment failure, establish the equipment topology structure, use the attention mechanism and graph convolutional network technology to obtain the mapping relationship of the fault propagation path, and determine the specific location of the equipment damage through fault tracing and impact analysis.

[0019] Preferably, step 6 is also included, using data enhancement and transfer learning methods to expand the knowledge base containing multiple fault modes, and through edge computing and distributed reasoning technology, a multi-label classification algorithm based on deep learning is used to identify and judge the globally consistent device state representation to obtain a globally consistent fault mode recognition result.

[0020] Preferably, in step 1, the multi-source heterogeneous sensor data is cleaned according to the data cleaning rules to obtain cleaned sensor data, and the cleaning rules include removing missing values ​​and outliers; the cleaned sensor data is standardized by using the minimum-maximum normalization method, and the data is mapped to a unified scale space of [0,1] to obtain standardized data; the standardized data is aligned in the time dimension using the dynamic time warping algorithm to obtain aligned sensor data; The data is normalized to the interval [0,1], the formula is as follows:

[0021] in, represents the normalized data; Represents the original data; Indicates the minimum value of the data; Indicates the maximum value of the data.

[0022] The aligned sensor data are denoised by wavelet transform, the appropriate number of decomposition layers is selected, the high-frequency components of each layer are thresholded, and then the denoised sensor data are reconstructed; the density-based spatial clustering algorithm is used to identify outliers in the aligned sensor data, mark the outlier points as noise points and eliminate them, and obtain the final cleaned, aligned, denoised and outlier-eliminated sensor data.

[0023] Sensor data cleaning is an important part of data processing, which aims to improve data quality and reliability. The minimum-maximum normalization method allows different types of sensor data to be compared and comprehensively analyzed. The dynamic time warping algorithm is used to solve the problem of asynchronous time series data. Wavelet transform denoising is an effective signal processing method. Through wavelet decomposition, the signal can be decomposed into sub-signals of different frequency bands. Threshold processing of the high-frequency part can effectively remove noise while retaining the main features of the signal. Density-based spatial clustering algorithm is used to identify outliers. In multidimensional sensor data, normal data tends to cluster, while outliers are sparsely distributed. By setting a suitable density threshold, isolated data points can be identified as outliers. The purpose of this series of processing steps is to improve data quality and lay the foundation for subsequent analysis. The cleaned data is more reliable, and the standardized processing makes data from different sources comparable. Time alignment ensures the synchronization of multi-source data and creates conditions for multi-sensor fusion analysis. Denoising improves the signal-to-noise ratio of the signal, which is conducive to extracting effective features. Outlier removal further improves the representativeness and credibility of the data. In quality control, sensor data after cleaning can more accurately reflect product characteristics, helping to improve production efficiency and product quality.

[0024] Preferably, in step 2, the preprocessed sensor data is subjected to frequency domain analysis by using fast Fourier transform to extract frequency domain features of spectrum energy and spectrum entropy; at the same time, the preprocessed sensor data is subjected to time-frequency domain analysis by using wavelet packet decomposition to extract wavelet coefficients of the preprocessed sensor data at different scales as time domain features; the extracted frequency domain features and time domain features of spectrum entropy reflect the health status and fault features of the equipment; Preset the threshold range of the health status of the device, compare the extracted frequency domain characteristics and time domain characteristics of the spectrum entropy with the preset threshold, and judge whether the device is currently in a healthy, sub-healthy or faulty state through the threshold judgment method; If the judgment result is that the equipment is in a sub-healthy or faulty state, then the Pearson correlation coefficient is used to calculate the correlation between each feature and the equipment damage location based on the extracted fault features, and the top N features with the highest correlation are selected as the equipment status feature subset. The screened equipment status feature subset is used as input, and the support vector machine algorithm is used to train the equipment status diagnosis model; the grid search and cross-validation methods are used to optimize the kernel function type and kernel function parameters of the support vector machine. When new equipment sensor data is input, after preprocessing and feature extraction, the extracted spectral entropy frequency domain features and time domain features are input into the trained equipment status diagnosis model. The trained equipment status diagnosis model performs classification prediction based on the input features to determine the health status of the equipment; If the equipment is diagnosed as sub-healthy or faulty, the fault feature subset screened out by the diagnosis result is used to train the support vector machine prediction model of the equipment damage location using the support vector regression algorithm; The support vector machine predicts the continuous value of the equipment damage location by finding the optimal regression hyperplane in the feature space; The fault feature subset is input into the trained support vector machine prediction model, and the trained support vector machine prediction model performs regression prediction based on the input features to obtain the equipment status features.

[0025] Fast Fourier transform is an efficient frequency domain analysis method that can convert time domain signals into frequency domain representation. In the multi-energy complementary power distribution system, frequency domain analysis of pre-processed sensor data can extract features such as spectral energy and spectral entropy. Spectral energy reflects the energy distribution of the signal on different frequency components, while spectral entropy measures the complexity of the spectrum. These features can effectively reflect the vibration characteristics and operating status of the equipment. Wavelet packet decomposition is a multi-scale analysis method that can provide local features of the signal in both the time domain and the frequency domain. By performing wavelet packet decomposition on sensor data, wavelet coefficients can be extracted as time domain features at different scales. These features can capture transient changes and local features during the operation of the equipment, which helps to identify early signs of faults. The judgment of the health status of the equipment is usually based on a preset threshold range. If the extracted spectral energy exceeds this range, it may indicate that the equipment is in a sub-healthy or faulty state. This threshold judgment method is simple and intuitive, but it is necessary to determine the appropriate threshold based on historical data and expert experience. The Pearson correlation coefficient is a statistic that measures the degree of linear correlation between two variables. In fault diagnosis, it can be used to calculate the correlation between each feature and the location of equipment damage. By selecting the most relevant features as input, the accuracy and efficiency of the fault diagnosis model can be improved. Support vector machine (SVM) is a powerful machine learning algorithm that is suitable for classification and regression problems. In equipment status diagnosis, SVM can build a classification model that can distinguish different health states by learning the feature patterns in historical data. The choice of kernel function has an important impact on the performance of SVM. Commonly used kernel functions include linear kernel, polynomial kernel and radial basis function (RBF) kernel. Through grid search and cross validation, the performance of different kernel functions and parameter combinations can be systematically evaluated to find the configuration that best suits the current problem. Support vector regression (SVR) is an extension of SVM for regression problems.

[0026] When predicting the location of equipment damage, SVR can establish a mapping relationship between features and the coordinates of the damage location. This data-driven fault diagnosis and location prediction method has stronger adaptability and generalization capabilities than traditional rule-based methods. It can learn complex patterns from a large amount of historical data and make accurate predictions on new data. This not only improves the efficiency of equipment maintenance, but also enables predictive maintenance and reduces economic losses caused by equipment failures.

[0027] Preferably, in step 3, the device state characteristics are synchronized in time through time interpolation, and the device state characteristics are unified on the same time axis. The device state characteristics realize spatial alignment of data through coordinate transformation method, and the device state characteristics are unified under the same spatial coordinate system. A spatiotemporal data fusion algorithm is selected, and a Kalman filter is used to establish a data fusion model. The sensor data after time synchronization and spatial alignment is used as input, and the state parameters of the device are calculated by the spatiotemporal data fusion algorithm. In the data fusion process, a dynamic model of the device state is established, and the Kalman filter algorithm is used to predict and update the device state. The prediction result is corrected in combination with the multi-sensor observation data at the current moment to obtain the device state estimate. The device state estimate obtained by data fusion is used as a globally consistent device state representation.

[0028] The time synchronization of equipment state characteristics is achieved through time interpolation. Through methods such as linear interpolation or spline interpolation, all sensor data can be unified on the same time axis to achieve time alignment of data. This can ensure the time consistency of each sensor data in subsequent analysis and improve the accuracy of data fusion. Spatial alignment is achieved through coordinate transformation methods. In order to unify the analysis, all sensor data need to be converted to the same global coordinate system. This can be achieved through a combination of rotation matrix and translation vector. The spatiotemporal data fusion algorithm uses Kalman filtering to establish a data fusion model. Kalman filtering is a recursive estimation algorithm that can effectively process multi-source data containing noise. In equipment state monitoring, a dynamic model of equipment state, such as position, velocity, acceleration, etc., can be established. In the data fusion process, prediction is first made based on the state estimate and dynamic model of the previous moment. Then, the prediction result is corrected in combination with the multi-sensor observation data at the current moment. Kalman filtering automatically adjusts the weights of different data sources according to the measurement error characteristics of each sensor to obtain the optimal state estimate. The advantage of this fusion method is that it can make full use of the complementarity of multi-source data to improve the accuracy and reliability of state estimation. By fusing these two types of data through Kalman filtering, we can get an accurate and real-time position estimate. The estimated value of the equipment state obtained by data fusion serves as a global consistent representation of the equipment state, providing a reliable basis for subsequent state monitoring, fault diagnosis and predictive maintenance. This globally consistent state representation can comprehensively reflect the overall operating status of the equipment, avoiding the one-sidedness that may be caused by a single sensor data.

[0029] Preferably, determining the equipment failure type in step 4 includes: S401: Obtain historical operation data and fault records of the equipment, and build a knowledge base containing multiple equipment failure modes, where the knowledge base contains feature descriptions, impact scope, and severity of various failure modes; S402: For globally consistent device state representation, a convolutional neural network is used to extract local features and capture key patterns in the data; at the same time, a long short-term memory network is used to extract time series features and capture the dynamic change trend of the data; S403: Fusing the local features extracted by the convolutional neural network and the temporal features extracted by the long short-term memory network to obtain a comprehensive feature representation of the device state as input to the deep learning multi-label classification algorithm; S404: using a deep learning multi-label classification algorithm to classify and identify the comprehensive feature representation of the device status, outputting the probabilities of various fault modes through a softmax activation function layer, determining whether the device has a fault, and if so, outputting the top N possible fault modes with the highest probability; S405: Match the classification and identification results with the fault modes in the knowledge base of multiple equipment fault modes, obtain detailed information of the fault mode, including the cause of the fault, the scope of impact and the severity, and form an equipment fault diagnosis report; S406: According to the equipment fault diagnosis report, the corresponding fault handling process is automatically triggered, including alarm notification, emergency plan and maintenance work order, to determine the type of equipment fault.

[0030] The core of the equipment fault diagnosis system is to build a comprehensive fault mode knowledge base. This knowledge base contains the feature descriptions, impact scope and severity of various faults, providing an important reference for subsequent fault identification and disposal. In the feature extraction stage, the system adopts a method combining convolutional neural networks and long short-term memory networks. Convolutional neural networks are good at capturing local patterns in data, such as abnormal peaks in equipment vibration signals. Long short-term memory networks focus on capturing time series features, such as abnormal temperature change trends over time. Feature fusion is a key step to improve the accuracy of fault diagnosis. By combining local features with time series features, the system can obtain a more comprehensive representation of the equipment status. The application of multi-label classification algorithms enables the system to identify multiple possible fault modes at the same time. Through the softmax activation function, the system outputs the probability distribution of various fault modes. This multi-label output helps maintenance personnel to comprehensively evaluate the equipment status and formulate more targeted maintenance strategies. The generation of fault diagnosis reports is a key step in converting machine learning results into actionable information. The report not only contains the fault type, but also provides detailed information such as the cause of the fault, the scope of impact and the severity. Automatically triggering the fault disposal process is an important part of improving maintenance efficiency. Depending on the severity and type of the fault, the system can automatically send an alarm notification, initiate an emergency plan, or generate a maintenance work order. This automated processing greatly shortens fault response time and reduces the risk of equipment damage and safety accidents.

[0031] Preferably, in step 5, the device topology information and fault information are obtained to construct a device topology map; for the device topology map, a graph attention network is used to perform representation learning on the nodes to obtain a low-dimensional vector representation of the nodes; The formula for constructing the graph attention network is as follows:

[0032] in, Indicates that node i is in +1 layer of feature representation; represents the activation function; represents the neighbor set of node i; represents the attention weight; Indicates The weight matrix of the layer; Indicates that node j is in Feature representation of the layer. This formula describes the update process of node features in the graph attention network.

[0033] According to the fault information, the known fault nodes are marked on the equipment topology map; the graph convolutional network is used to perform end-to-end fault propagation paths on the equipment topology map to obtain potential fault propagation paths; based on the potential fault propagation paths, the root node of the fault is determined to determine the cause of the equipment failure; for the identified root node, the graph attention network is used to perform impact analysis on its neighborhood subgraph to obtain the affected node set; the location information of the root node and the affected node in the equipment topology structure is fused to determine the specific location of the equipment damage and output the equipment fault diagnosis results.

[0034] The acquisition of equipment topology information and fault information is the basis for fault diagnosis. Fault information may be the tripping of a circuit breaker in a substation. When constructing a device topology graph, each device can be used as a node, and the physical or logical connections between devices can be used as edges to form a complete network structure. The graph attention network learns the representation of nodes and can capture the local structural features and global location information of nodes. This low-dimensional vector representation makes subsequent analysis more efficient.

[0035] Known fault nodes are marked on the equipment topology, such as the substation where the tripped circuit breaker is located. This provides a starting point for fault propagation analysis. On this basis, the graph convolutional network performs end-to-end fault propagation path analysis to simulate how the fault spreads in the network. By analyzing the fault propagation path, the root node that initially caused the problem can be traced back. In the power system, this may be a short-circuited transformer. After determining the cause of the fault, targeted repair measures can be taken to improve system reliability. The graph attention network performs impact analysis on the neighborhood subgraph of the root node, which can accurately locate the affected range. This analysis helps to assess the severity of the fault and formulate emergency plans. Finally, the location information of the root node and the affected node in the equipment topology is fused to intuitively display the impact range of the fault. In the power system, this may be manifested as power outages or instability in a certain area. In this way, the specific part of the equipment that is damaged can be quickly located, such as a specific section of a transmission line. This graph network-based fault diagnosis method has a global perspective and can effectively deal with fault propagation problems in complex systems. It can not only accurately locate the source of the fault, but also predict the chain reaction that the fault may cause, providing a basis for timely preventive measures. This is of great significance for ensuring the stable operation of large and complex systems.

[0036] Preferably, the identification and judgment of the globally consistent device state representation in step 6 includes: S601: acquiring the device state representation and performing preprocessing to obtain a standardized device state representation data set.

[0037] S602: Using data enhancement technology, by rotating, translating, scaling, and adding noise to the standardized device state characterization data set, generating data samples with various changes, and expanding the training data set; S603: using a transfer learning method, selecting a convolutional neural network pre-training model related to the fault mode recognition task, and using the expanded data set to adjust the convolutional neural network pre-training model; S604: Build a multi-label classification model based on deep learning, design the network structure and loss function of the model, and initialize the model parameters using the adjusted pre-trained model; S605: Divide the device state representation data set preprocessed in S601 into a training set and a validation set, use the training set to train the multi-label classification model, evaluate the model performance on the validation set, and adjust the model hyperparameters according to the evaluation results; S606: Deploy the trained multi-label classification model to the edge device, use edge computing technology to implement local reasoning, and determine the failure mode of the device in real time based on the device status characterization data set; S607: Distributed reasoning technology is used to aggregate and fuse the reasoning results on multiple edge devices to obtain a globally consistent fault mode recognition result.

[0038] Preprocessing of equipment status characterization data is the basis of fault diagnosis. First, clean, denoise and standardize the raw data to ensure data quality. Normalize the temperature data to make features of different dimensions comparable. Data augmentation technology can effectively expand the training samples. Generate new samples by rotating and translating the raw data. Add Gaussian noise to the temperature curve to enhance the robustness of the model to noise. This can improve the generalization ability of the model and reduce the risk of overfitting. Transfer learning uses pre-trained models to accelerate model training. ResNet or VGG networks pre-trained on large-scale image datasets can be selected and migrated to equipment fault diagnosis tasks. By fine-tuning network parameters, it can quickly adapt to new tasks and improve model performance. This method is particularly suitable for situations with limited training data. Multi-label classification models can identify multiple fault modes at the same time. A multi-task learning framework can be used to design a network structure that shares the underlying feature extractor and multiple classification heads. The loss function can be a weighted sum of binary cross entropy losses.

[0039] This structure can make full use of the correlation between fault modes and improve classification accuracy. Model training and evaluation is an iterative optimization process. The k-fold cross-validation method can be used to divide the data set into a training set and a validation set. The model performance can be continuously optimized by adjusting hyperparameters such as learning rate and batch size. Evaluation indicators can be selected such as precision, recall rate and F1 score to comprehensively measure the classification effect of the model. Edge computing technology realizes local real-time reasoning. The trained model is deployed on the edge device. At the same time, the model complexity is reduced through technologies such as model quantization and pruning to adapt to the computing resource limitations of the edge device. Distributed reasoning technology integrates the results of multiple edge devices. A federated learning framework can be used to aggregate the reasoning results of multiple devices while protecting data privacy. A more reliable global fault diagnosis result can be obtained through weighted averaging or voting mechanism. This method can make full use of the advantages of distributed systems and improve the accuracy and reliability of diagnosis.

[0040] like Figure 2 As shown, the data preprocessing module is used to obtain multi-source heterogeneous sensor data in a multi-energy complementary power distribution system, and preprocess the multi-source heterogeneous sensor data, wherein the preprocessing includes cleaning, alignment and standardization processing, eliminating noise and abnormal values ​​in the multi-source heterogeneous sensor data, and obtaining preprocessed sensor data; Feature extraction module: used for extracting health status indicators and fault characteristic indicators from the preprocessed sensor data by using a feature extraction algorithm, and judging whether the device is in a healthy, sub-healthy or faulty state based on the health status indicators. If the device is in a sub-healthy or faulty state, a feature subset of the device damage position is screened out based on the fault characteristic indicators to obtain the device status characteristics; Spatiotemporal data fusion module: used to synchronize and align the device status characteristics based on the device status characteristics through spatiotemporal data fusion technology to obtain a globally consistent device status representation; Fault identification module: used to build a knowledge base containing multiple fault modes, using a multi-label classification algorithm based on deep learning to identify and judge the globally consistent device status representation and determine the type of device fault; Analysis module: After determining the type of equipment fault, it establishes the equipment topology structure, uses the attention mechanism and graph convolutional network technology to obtain the mapping relationship of the fault propagation path, and determines the specific location of the equipment damage through fault tracing and impact analysis; It also includes a data enhancement and transfer learning module: it uses data enhancement and transfer learning methods to expand the knowledge base containing multiple fault modes, and through edge computing and distributed reasoning technology, uses a multi-label classification algorithm based on deep learning to identify and judge the globally consistent equipment status representation, and obtains a globally consistent fault mode recognition result.

[0041] The present invention discloses a data analysis method for a multi-energy complementary power distribution system based on deep learning. This method aims at the equipment failure problem in the multi-energy complementary power distribution system, and extracts health status and fault characteristic indicators through multi-source heterogeneous sensor data collection and preprocessing. The spatiotemporal data fusion technology is used to obtain a globally consistent representation of the equipment state, and a knowledge base containing multiple fault modes is constructed. A deep learning multi-label classification algorithm is used to identify the type of equipment failure, and the attention mechanism and graph convolution network technology are used in combination with the equipment topology to determine the specific location of the equipment damage. The present invention also introduces data enhancement and transfer learning methods to expand the fault mode knowledge base, and improves the fault identification efficiency through edge computing and distributed reasoning technology. This method can effectively improve the fault diagnosis accuracy and efficiency of the multi-energy complementary power distribution system, realize the precise positioning and impact analysis of equipment failures, and provide important support for the safe and stable operation of the system.

[0042] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A data analysis method for a multi-energy complementary power distribution system based on deep learning, characterized in that: include: Step 1, acquiring multi-source heterogeneous sensor data in a multi-energy complementary power distribution system, and preprocessing the multi-source heterogeneous sensor data, wherein the preprocessing includes cleaning, alignment and standardization processing, eliminating noise and abnormal values ​​in the multi-source heterogeneous sensor data, and obtaining preprocessed sensor data; Step 2, for the preprocessed sensor data, using a feature extraction algorithm to extract health status indicators and fault characteristic indicators, based on the health status indicators, determine whether the device is in a healthy, sub-healthy or faulty state, if the device is in a sub-healthy or faulty state, based on the fault characteristic indicators, filter out a feature subset of the device damage position to obtain the device status feature; Step 3, based on the device state characteristics, the device state characteristics are synchronized and aligned by using spatiotemporal data fusion technology to obtain a globally consistent device state representation; Step 4: Build a knowledge base containing multiple fault modes, use a multi-label classification algorithm based on deep learning to identify and judge the globally consistent device state representation, and determine the device fault type; Step 5: After determining the type of equipment failure, establish the equipment topology structure, use the attention mechanism and graph convolutional network technology to obtain the mapping relationship of the fault propagation path, and determine the specific location of the equipment damage through fault tracing and impact analysis.

2. The multi-energy complementary power distribution system data analysis method based on deep learning according to claim 1 is characterized in that: It also includes step 6, using data enhancement and transfer learning methods to expand the knowledge base containing multiple fault modes, and through edge computing and distributed reasoning technology, using the multi-label classification algorithm based on deep learning to identify and judge the globally consistent equipment state representation, and obtain a globally consistent fault mode recognition result.

3. The method for analyzing data of a multi-energy complementary power distribution system based on deep learning according to claim 1 is characterized in that: In step 1, the multi-source heterogeneous sensor data is cleaned according to the data cleaning rules to obtain the cleaned sensor data. The cleaning rules include removing missing values ​​and outliers. The cleaned sensor data is standardized by using the minimum-maximum normalization method, and the data is mapped to a unified scale space of [0,1] to obtain the standardized data. A dynamic time warping algorithm is used to align the standardized data in the time dimension to obtain aligned sensor data, and the aligned sensor data is denoised by wavelet transform. The number of decomposition layers is selected, and the high-frequency components of each layer are threshold processed to reconstruct the denoised sensor data. The denoised sensor data is reconstructed, and the denoised sensor data is obtained. The denoised sensor data is reconstructed, including using a density-based spatial clustering algorithm to identify outliers in the aligned sensor data, marking the outlier points as noise points and eliminating them, to obtain the final cleaned, aligned, denoised and outlier-eliminated sensor data.

4. The method for analyzing data of a multi-energy complementary power distribution system based on deep learning according to claim 1 is characterized in that: In step 2, the preprocessed sensor data is subjected to frequency domain analysis by using fast Fourier transform to extract frequency domain features of spectrum energy and spectrum entropy; at the same time, the preprocessed sensor data is subjected to time-frequency domain analysis by using wavelet packet decomposition to extract wavelet coefficients of the preprocessed sensor data at different scales as time domain features; The extracted spectrum entropy frequency domain features and time domain features reflect the health status and fault characteristics of the equipment; Preset the threshold range of the health status of the device, compare the extracted frequency domain characteristics and time domain characteristics of the spectrum entropy with the preset threshold, and judge whether the device is currently in a healthy, sub-healthy or faulty state through the threshold judgment method; If the judgment result is that the equipment is in a sub-healthy or faulty state, then the Pearson correlation coefficient is used to calculate the correlation between each feature and the equipment damage location based on the extracted fault features, and the top N features with the highest correlation are selected as the equipment status feature subset. The screened equipment status feature subset is used as input, and the support vector machine algorithm is used to train the equipment status diagnosis model; the grid search and cross-validation methods are used to optimize the kernel function type and kernel function parameters of the support vector machine. When new equipment sensor data is input, after preprocessing and feature extraction, the extracted spectral entropy frequency domain features and time domain features are input into the trained equipment status diagnosis model. The trained equipment status diagnosis model performs classification prediction based on the input features to determine the health status of the equipment; If the equipment is diagnosed as sub-healthy or faulty, the fault feature subset screened out by the diagnosis result is used to train the support vector machine prediction model of the equipment damage location using the support vector regression algorithm; The support vector machine predicts the continuous value of the equipment damage location by finding the optimal regression hyperplane in the feature space; The fault feature subset is input into the trained support vector machine prediction model, and the trained support vector machine prediction model performs regression prediction based on the input features to obtain the equipment status features.

5. The method for analyzing data of a multi-energy complementary power distribution system based on deep learning according to claim 1 is characterized in that: In step 3, the device state characteristics are synchronized by time interpolation, and the device state characteristics are unified on the same time axis. The device state characteristics are spatially aligned by coordinate transformation method, and the device state characteristics are unified in the same spatial coordinate system. The spatiotemporal data fusion algorithm is selected, and the Kalman filter is used to establish a data fusion model. The sensor data after time synchronization and spatial alignment is used as input, and the state parameters of the device are calculated by the spatiotemporal data fusion algorithm. In the data fusion process, a dynamic model of the device state is established, and the Kalman filter algorithm is used to predict and update the device state. The prediction result is corrected in combination with the multi-sensor observation data at the current moment to obtain the device state estimate, and the device state estimate obtained by data fusion is used as a globally consistent device state representation.

6. The method for analyzing data of a multi-energy complementary power distribution system based on deep learning according to claim 1 is characterized in that: In step 4, determine the type of equipment failure, including: S401: Obtain historical operation data and fault records of the equipment, and build a knowledge base containing multiple equipment failure modes, where the knowledge base contains feature descriptions, impact scope, and severity of various failure modes; S402: For globally consistent device state representation, a convolutional neural network is used to extract local features and capture key patterns in the data; at the same time, a long short-term memory network is used to extract time series features and capture the dynamic change trend of the data; S403: Fusing the local features extracted by the convolutional neural network and the temporal features extracted by the long short-term memory network to obtain a comprehensive feature representation of the device state as input to the deep learning multi-label classification algorithm; S404: using a deep learning multi-label classification algorithm to classify and identify the comprehensive feature representation of the device status, outputting the probabilities of various fault modes through a softmax activation function layer, determining whether the device has a fault, and if so, outputting the top N possible fault modes with the highest probability; S405: Match the classification and identification results with the fault modes in the knowledge base of multiple equipment fault modes, obtain detailed information of the fault mode, including the cause of the fault, the scope of impact and the severity, and form an equipment fault diagnosis report; S406: According to the equipment fault diagnosis report, the corresponding fault handling process is automatically triggered, including alarm notification, emergency plan and maintenance work order, to determine the type of equipment fault.

7. The method for analyzing data of a multi-energy complementary power distribution system based on deep learning according to claim 1 is characterized in that: In step 5, the device topology information and fault information are obtained to construct a device topology map; For the equipment topology map, a graph attention network is used to learn the representation of the nodes and obtain the low-dimensional vector representation of the nodes; according to the fault information, the known fault nodes are marked on the equipment topology map; the graph convolutional network is used to perform end-to-end fault propagation paths on the equipment topology map to obtain potential fault propagation paths; according to the potential fault propagation paths, the root node of the fault is judged to determine the cause of the equipment failure; for the identified root node, a graph attention network is used to perform impact analysis on its neighborhood subgraph to obtain the affected node set; the location information of the root node and the affected node in the equipment topology structure is fused to determine the specific location of the equipment damage and output the equipment fault diagnosis results.

8. The method for analyzing data of a multi-energy complementary power distribution system based on deep learning according to claim 2 is characterized in that: In step 6, the globally consistent device status representation is identified and judged, including: S601: Acquire device status representation and perform preprocessing to obtain a standardized device status representation data set; S602: Using data enhancement technology, by rotating, translating, scaling, and adding noise to the standardized device state characterization data set, generating data samples with various changes, and expanding the training data set; S603: using a transfer learning method, selecting a convolutional neural network pre-training model related to the fault mode recognition task, and using the expanded data set to adjust the convolutional neural network pre-training model; S604: Build a multi-label classification model based on deep learning, design the network structure and loss function of the model, and initialize the model parameters using the adjusted pre-trained model; S605: Divide the device state representation data set preprocessed in S601 into a training set and a validation set, use the training set to train the multi-label classification model, evaluate the model performance on the validation set, and adjust the model hyperparameters according to the evaluation results; S606: Deploy the trained multi-label classification model to the edge device, use edge computing technology to implement local reasoning, and determine the failure mode of the device in real time based on the device status characterization data set; S607: Distributed reasoning technology is used to aggregate and fuse the reasoning results on multiple edge devices to obtain a globally consistent fault mode recognition result.

9. A multi-energy complementary power distribution system data analysis system based on deep learning, used to implement the multi-energy complementary power distribution system data analysis method based on deep learning according to any one of claims 1 to 8, characterized in that: include: Data preprocessing module: used to obtain multi-source heterogeneous sensor data in the multi-energy complementary power distribution system, and preprocess the multi-source heterogeneous sensor data, wherein the preprocessing includes cleaning, alignment and standardization processing, eliminating noise and abnormal values ​​in the multi-source heterogeneous sensor data, and obtaining preprocessed sensor data; Feature extraction module: used for extracting health status indicators and fault characteristic indicators from the preprocessed sensor data by using a feature extraction algorithm, and judging whether the device is in a healthy, sub-healthy or faulty state based on the health status indicators. If the device is in a sub-healthy or faulty state, a feature subset of the device damage position is screened out based on the fault characteristic indicators to obtain the device status characteristics; Spatiotemporal data fusion module: used to synchronize and align the device status characteristics based on the device status characteristics through spatiotemporal data fusion technology to obtain a globally consistent device status representation; Fault identification module: used to build a knowledge base containing multiple fault modes, using a multi-label classification algorithm based on deep learning to identify and judge the globally consistent device status representation and determine the type of device fault; Analysis module: After determining the type of equipment fault, it is used to establish the equipment topology structure, use the attention mechanism and graph convolutional network technology to obtain the mapping relationship of the fault propagation path, and determine the specific location of the equipment damage through fault tracing and impact analysis.

10. The multi-energy complementary power distribution system data analysis system based on deep learning according to claim 9 is characterized in that: Also includes: Data-enhanced transfer learning module: used to establish the equipment topology after determining the type of equipment fault, use the attention mechanism and graph convolutional network technology to obtain the mapping relationship of the fault propagation path, and determine the specific location of the equipment damage through fault tracing and impact analysis.

Images