Electric power operation risk discrimination method and system based on deep learning
Through deep learning technology, analyzing multi-source data and building a risk discrimination model, solving the problems of multi-source data fusion and insufficient behavior monitoring of operators in the existing technology, achieving more accurate risk prediction and evaluation of power operations, and increasing risks Reliability and accuracy of judgment.
Patent Information
- Application Number
- CN202410982720.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-22
- Publication Date
- 2024-11-08
AI Technical Summary
The existing power operation risk determination methods fail to effectively combine multi-source data for real-time deep learning analysis, and lack comprehensive monitoring of operator behavior and status, resulting in insufficient adaptability and prediction accuracy in complex and dynamic environments.
Using a deep learning-based power operation risk identification method, pre-processing and splicing by collecting multi-source data (operation data, personnel data, equipment data and environmental data), and using feature analysis layer, feature fusion layer and risk identification layer to construct risks Discriminant model to realize systematic evaluation and discrimination of power operation risks.
Provides more accurate risk prediction, improves the reliability and accuracy of risk judgment, can dynamically adjust attention allocation, optimize feature extraction and risk assessment, enhances the prediction ability of the model, and ensures accurate prediction of high-risk events. .
Smart Images

Figure CN118917667A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of electric power engineering technology, and in particular to a method and system for identifying risks in electric power operations based on deep learning. Background Art
[0002] Power operations refer to a series of activities such as equipment installation, commissioning, maintenance and operation in the power system. These activities usually include power generation, transmission and distribution, as well as maintenance and repair of power equipment. Power operations involve a wide range of technical and professional knowledge, and operators need to have a deep understanding of the functions and safe operation of the power system. The stability and efficiency of the power system are directly related to the normal operation of the national economy and social life. Therefore, it is extremely important to ensure the safety and efficiency of power operations.
[0003] The risks of power operations mainly include electric shock, equipment failure, operating errors and natural disasters. The main purpose of identifying the risks of power operations is to discover potential dangers in advance, so as to take preventive measures to avoid accidents. Through effective risk identification, the life safety of operators can be protected, property losses can be prevented, and the stability and reliability of power supply can be ensured.
[0004] The traditional method of identifying risks in power operations mainly relies on regular equipment inspections. Although this method can prevent accidents to a certain extent, it still has many shortcomings. For example, regular inspections cannot cover all operating moments of all equipment, making it difficult to achieve comprehensive monitoring. In addition, with the expansion of the scale of the power system and the complexity of equipment, traditional methods are even more inadequate and cannot meet the needs of efficient and precise management.
[0005] Existing methods for identifying risks in power operations mostly rely on technical monitoring and data analysis. Monitoring systems are widely used to monitor equipment status and operating environment in real time. By collecting data such as current, voltage, temperature and other parameters, anomaly detection is performed using preset thresholds. Data analysis methods, especially machine learning-based models, can learn and predict potential risks from historical accident data and achieve early identification of risks. In addition, simulation and digital twin technologies are also used to simulate power system operations, predict risks and test different response strategies in a virtual environment. These methods improve the accuracy and efficiency of risk management and provide solid technical support for power operation safety.
[0006] For example, the Chinese patent with the prior art application number CN202011508995.4 discloses a risk management system for on-site operations in the power industry. First, by establishing a comprehensive on-site operation risk project ledger, major risk factors are systematically identified and monitored in real time, and risk management is further strengthened by formulating and implementing targeted preventive maintenance measures. This method effectively integrates risk warning, early risk education, on-site preventive maintenance, and post-construction summary to ensure that every operation link is strictly monitored and managed. Through this method, managers can obtain important risk information in a timely manner, take measures in advance to prevent accidents, and ensure the safety of power operations. In addition, the Chinese patent with the application number CN202310655491.2 discloses an early warning system for power operation risks, including a power operation order risk analysis module and an early warning analysis module to achieve comprehensive risk management of power operation orders. The power operation order acquisition module is responsible for collecting operation orders and passing them to the risk analysis module, which deeply analyzes these operation orders, identifies potential risk factors and related historical accident information, and forms a risk collection table. This table is then sent to the early warning analysis module, which uses historical data to conduct risk screening and issue corresponding warnings. This system not only supports immediate risk analysis and early warning of power operation orders, but also allows for long-term systematic risk assessment, so that staff can effectively manage operational risks from both short-term and long-term perspectives, greatly improving the safety and efficiency of power operations. However, existing power operation risk identification methods fail to consider real-time deep learning analysis combined with multi-source data fusion, and lack comprehensive monitoring of operator behavior and status, which leads to deficiencies in adaptability and prediction accuracy in dealing with complex and dynamic environments.
[0007] To this end, a method and system for identifying risks in power operations based on deep learning are proposed. Summary of the invention
[0008] The purpose of the present invention is to provide a method and system for identifying risks of power operations based on deep learning. First, multi-source data is collected, and the multi-source data includes first data, second data, third data and fourth data; then, the multi-source data is pre-processed to obtain first processed data, second processed data, third processed data and fourth processed data; further, the first processed data and the second processed data are spliced to obtain first spliced data, and the third processed data and the fourth processed data are spliced to obtain second spliced data; finally, the first spliced data and the second spliced data are input into the risk identification model, and successively pass through the feature analysis layer, the feature fusion layer and the risk identification layer to obtain the final risk identification result. This process aims to systematically evaluate and identify potential risks of power operations to ensure the safety of operations.
[0009] To achieve the above object, the present invention provides the following technical solutions:
[0010] A method for identifying risks of power operations based on deep learning, comprising:
[0011] Collecting multi-source data; the multi-source data includes first data, second data, third data and fourth data;
[0012] Preprocessing the multi-source data to obtain first processed data, second processed data, third processed data and fourth processed data;
[0013] splicing the first processed data and the second processed data to obtain first spliced data;
[0014] splicing the third processed data and the fourth processed data to obtain second spliced data;
[0015] Constructing a risk discrimination model, inputting the first spliced data and the second spliced data into the risk discrimination model, and sequentially passing through a feature analysis layer, a feature fusion layer, and a risk discrimination layer to obtain a risk discrimination result;
[0016] The feature analysis layer is used to extract and train features of the first spliced data and the second spliced data respectively to obtain a first analysis feature vector and a second analysis feature vector;
[0017] The feature fusion layer fuses the first analysis feature and the second analysis feature to obtain a comprehensive analysis feature; the comprehensive analysis feature is used to provide a basis for the final risk judgment;
[0018] The risk identification layer identifies the risk of the power operation according to the comprehensive analysis characteristics and outputs the risk identification result; the risk identification result is the probability value of the power operation risk being low risk, medium risk and high risk;
[0019] Collect historical power operation risk data and construct a power operation risk data set; the power operation risk data set is used to train the risk discrimination model.
[0020] The first data is operation data, including video data of operators on duty, which is obtained through a camera and is used to analyze the operator's wearing of safety equipment; the second data is personnel data, including the operator's heart rate data and whether he or she is drinking; the second data is obtained through a wearable device; the third data is equipment data, including the voltage, current and temperature of the equipment; the third data is obtained in real time through sensors; the fourth data is environmental data, including whether it is raining, temperature data and humidity data.
[0021] The specific steps of preprocessing the multi-source data include:
[0022] Extracting a key frame from the first data, the key frame being a frame of the operator's working area; resizing and graying the key frame; applying edge detection image processing technology to the key frame to identify and mark the wearing of safety equipment to obtain the first processed data; the safety equipment includes a safety helmet, insulating gloves and a safety rope;
[0023] Performing data cleaning on the second data to check and correct outliers or missing data;
[0024] Normalizing the heart rate data in the second data to between 0 and 1 to obtain preprocessed heart rate data;
[0025] The drinking state in the second data is converted into a binary feature, represented by 0 and 1, to obtain a pre-processed drinking state; 0 and 1 represent not drinking and drinking, respectively;
[0026] splicing the pre-processed heart rate data and the pre-processed drinking status to obtain the second processed data;
[0027] Performing data cleaning on the third data to check and correct abnormal values or missing data;
[0028] Use Butterworth filter to remove high frequency noise from voltage and current data;
[0029] performing normalization processing on the third data;
[0030] splicing the voltage, current and temperature of the device to obtain the third processed data;
[0031] Whether it is raining is converted into a binary feature, represented by 0 and 1, where 0 and 1 represent no rain and rain respectively;
[0032] Performing the data cleaning and normalization processing on the temperature data and humidity;
[0033] The whether it is raining, the temperature data and the humidity data are concatenated to obtain the fourth processed data.
[0034] The feature analysis layer includes a first network structure, a second network structure and a multi-risk interactive attention mechanism; the first network structure and the second network structure are both LSTM network structures; wherein the first network structure is used to analyze the first spliced data to obtain a first analysis feature; the second network structure is used to analyze the second spliced data to obtain a second analysis feature; the first analysis feature and the second analysis feature are processed by the multi-risk interactive attention mechanism to obtain the first analysis feature vector and the second analysis feature vector;
[0035] The processing process of the multi-risk interactive attention mechanism includes:
[0036] First, generate the interaction matrix. The specific formula is:
[0037]
[0038] Then, the attention score is calculated. The specific formula is:
[0039] a ij =v T σ(I ij +u i +u j );
[0040] Furthermore, the attention weight is normalized, and the specific formula is:
[0041]
[0042] Finally, feature fusion is performed, and the specific formula is:
[0043]
[0044] in, W represents feature concatenation, which is used to concatenate feature vectors from different sources; ij is a learnable weight matrix, customized for each pair of feature combinations; b ij is the bias term, used to adjust the output of the interaction matrix; F i and F j are the features processed by different network structures; v is the global weight vector used to convert the interaction matrix into a scalar attention score; σ is the activation function; u i and u j is with the feature F i and F j The associated bias vector; F i ·F j Represents the element-wise product of two features.
[0045] The specific process of training the risk discrimination model includes:
[0046] Collecting data: collecting 5,000 pieces of historical power operation risk data from the historical power operation risk data; the historical power operation risk data includes the first data, the second data, the third data and the fourth data;
[0047] Data marking: marking the risk identification result corresponding to the historical power operation risk data; the risk level is determined according to the risk identification result; the risk levels include 1, 2 and 3, corresponding to low risk, medium risk and high risk respectively;
[0048] Data preprocessing: performing preprocessing operations on the first data, the second data, the third data, and the fourth data respectively;
[0049] Divide the 5000 pieces of historical risk data into a training set and a validation set in a ratio of 7:3;
[0050] The risk discrimination model is trained by using the training set; and the risk discrimination model is verified and optimized by using the verification set.
[0051] The model training includes optimizing the loss function, and the formula of the optimized loss function is:
[0052]
[0053] Where C is the total number of categories, including no risk, low risk, medium risk and high risk in risk identification; w c is the weight of category c; y c is a one-hot encoded label. When the sample actually belongs to category c, y c is 1, otherwise y c is 0; p c The probability that the model predicts that the sample belongs to category c; y c+1 is the category label converted to calculate the sequential loss. If the actual category of the sample is greater than or equal to c+1, it is 1, otherwise it is 0; λ is the adjustment parameter.
[0054] A power operation risk identification system based on deep learning, comprising:
[0055] A data acquisition unit, used to acquire first data, second data, third data and fourth data;
[0056] a data processing unit, configured to perform preprocessing operations on the first data, the second data, the third data, and the fourth data, respectively, to obtain first processed data, second processed data, third processed data, and fourth processed data;
[0057] A data splicing unit, configured to splice the first processed data and the second processed data to obtain first spliced data; and splice the third processed data and the fourth processed data to obtain second spliced data;
[0058] The data analysis unit constructs a risk identification model for analyzing the first spliced data and the second spliced data to obtain a risk identification result.
[0059] The first data is operation data, including video data of operators on duty, which is obtained through a camera and is used to analyze the operator's wearing of safety equipment; the second data is personnel data, including the operator's heart rate data and whether he or she is drinking; the second data is obtained through a wearable device; the third data is equipment data, including the voltage, current and temperature of the equipment; the third data is obtained in real time through a sensor; the fourth data is environmental data, including whether it is raining, temperature value and humidity value.
[0060] The specific process of training the risk discrimination model includes:
[0061] Collecting data: collecting 5,000 pieces of historical power operation risk data from historical power operation risk data; the historical power operation risk data includes the first data, the second data, the third data and the fourth data;
[0062] Data marking: marking the risk identification results corresponding to the historical power operation risk data; the risk identification results include 1, 2 and 3, corresponding to low risk, medium risk and high risk respectively;
[0063] Data preprocessing: performing preprocessing operations on the first data, the second data, the third data, and the fourth data respectively;
[0064] Divide the 5000 pieces of historical risk data into a training set and a validation set in a ratio of 7:3;
[0065] The risk discrimination model is trained by using the training set; and the risk discrimination model is verified and optimized by using the verification set.
[0066] The model training includes optimizing the loss function, and the formula of the optimized loss function is:
[0067]
[0068] Where C is the total number of categories, including low risk, medium risk and high risk in risk identification; w c is the weight of category c; y c is a one-hot encoded label. When the sample actually belongs to category c, y cis 1, otherwise y c is 0; p c The probability that the model predicts that the sample belongs to category c; y c+1 is the category label converted to calculate the sequential loss. If the actual category of the sample is greater than or equal to c+1, it is 1, otherwise it is 0; λ is the adjustment parameter.
[0069] Compared with the prior art, the present invention has the following beneficial effects:
[0070] 1. The present invention analyzes the risks of power operations by collecting multi-source data, providing a comprehensive analysis perspective for the identification of power operation risks. By integrating operation data, personnel data, equipment data and environmental data, the method can comprehensively evaluate risk factors, thereby providing more accurate risk predictions. This comprehensive analysis method makes risk assessment not limited to a single perspective, but through multi-dimensional cross-validation, it greatly improves the reliability and accuracy of risk judgment.
[0071] 2. In the feature analysis stage, the present invention uses two LSTM network structures to perform detailed analysis on the first spliced data and the second spliced data respectively, effectively capturing the time-dependent characteristics in the data. In addition, combined with the multi-risk interactive attention mechanism, this method can dynamically adjust the attention allocation between different risk factors, further enhancing the prediction ability and accuracy of the model. This composite model structure not only optimizes the feature extraction process, but also provides solid data support for risk assessment of power operations by accurately identifying key risk drivers.
[0072] 3. In the process of model training, the present invention adopts an optimized loss function to better deal with the problem of class imbalance, and emphasizes the importance of minority classes by introducing weights, thereby effectively improving the risk discrimination model's ability to discriminate high-risk events. This method not only optimizes the learning process of the model, but also ensures that the potential high risks in power operations can be accurately predicted in practical applications, thereby providing a time window for taking preventive measures. BRIEF DESCRIPTION OF THE DRAWINGS
[0073] Figure 1 A schematic flow chart of a method for identifying risks of power operations based on deep learning provided by an embodiment of the present invention;
[0074] Figure 2 A schematic diagram of the structure of a power operation risk identification system based on deep learning provided by an embodiment of the present invention;
[0075] Figure 3 An example diagram of a key frame provided by an embodiment of the present invention;
[0076] Figure 4An example diagram of a key frame after preprocessing provided by an embodiment of the present invention;
[0077] Figure 5 A framework diagram of the risk identification model provided in an embodiment of the present invention. DETAILED DESCRIPTION
[0078] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. 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.
[0079] Power operation risks cover all kinds of safety and operational risks that may be encountered in the entire process from power generation, transmission to distribution and maintenance. These risks may come from equipment failure, operating errors, natural environmental factors and human factors. For example, high-voltage electrical equipment may fail due to improper maintenance or aging, and accidents may also occur if workers do not use safety equipment correctly or follow operating procedures. In addition, severe weather conditions such as heavy rain and high temperature may also affect the safe operation of power facilities. Therefore, power operation risks are a multi-faceted issue that requires comprehensive consideration.
[0080] The main purpose of identifying the risks of power operations is to prevent possible accidents, ensure personnel safety and the stable operation of the power system. By identifying and evaluating potential risks, timely measures can be taken to reduce or eliminate risk factors and avoid serious safety accidents and economic losses. In addition, systematic risk identification can also help improve the safety management level of the power industry and enhance the safety assurance capabilities of the entire industry by improving operating procedures and safety measures.
[0081] Electricity is the foundation of modern society, and its safety is directly related to social stability and people's quality of life. First, effective risk management and identification can reduce the occurrence of power outages, ensure the continuity and reliability of power supply, and support the normal operation of economic activities. Secondly, by reducing the occurrence of accidents, the lives of power industry workers and the public can be effectively protected and property losses can be avoided. Finally, systematic risk identification can also help power companies establish a good public image, enhance social responsibility, and promote sustainable development.
[0082] The purpose of the present invention is to provide a method for identifying risks in power operations based on deep learning. This method is applied to a system for identifying risks in power operations based on deep learning. For specific method flow charts and system structure diagrams, please refer to Figure 1 and Figure 2 .
[0083] Embodiment 1
[0084] This embodiment analyzes the maintenance scenario of the power system line workers in Nanjing A area on the high-voltage transmission line.
[0085] In this embodiment, an advanced power operation risk identification system based on deep learning is provided, which greatly improves the ability of power operation risk management by efficiently integrating and analyzing information from different data sources. The system first collects multi-source data including operating environment, employee behavior, equipment status and environmental conditions, and then obtains first spliced data and second spliced data by performing detailed preprocessing and effective splicing on these data. Subsequently, deep learning technology is used to conduct in-depth analysis of these data, and accurate risk identification results are finally generated through layer-by-layer processing of feature analysis layer, feature fusion layer and risk identification layer. The application of this system not only improves the safety of power operations, but also optimizes resource allocation, reduces potential costs caused by risk events, and brings substantial safety guarantees and economic benefits to the power industry.
[0086] Reference Figure 1 S10 in which S10 is applied to a data acquisition unit of a power operation risk identification system based on deep learning, and is used to collect multi-source data; the multi-source data includes first data, second data, third data and fourth data;
[0087] Furthermore, the first data is operation data, including video data of operators taking up their posts, which is acquired through a camera and used to analyze the wearing of safety equipment by the operators;
[0088] The second data is personnel data, including heart rate data and drinking status of the operator; the second data is obtained through a wearable device;
[0089] The third data is device data, including voltage, current and temperature of the device; the third data is acquired in real time through sensors; the fourth data is environmental data, including whether it is raining, temperature data and humidity data.
[0090] By integrating and analyzing multi-source data, the power operation risk identification system can provide a comprehensive safety assessment, thereby improving the safety of power industry operations. The first data includes video data of operators obtained through cameras, which can be used to monitor and ensure that employees wear safety equipment correctly, including helmets, insulating gloves and safety ropes, which is crucial to preventing accidents. The second data involves the heart rate and drinking status of personnel. This information helps to assess the health and suitability of employees to determine whether they are safe to work. The third data includes the voltage, current and temperature of the equipment. Through real-time monitoring of these data, equipment abnormalities can be detected in time to prevent power failures. The fourth data involves environmental factors such as weather conditions and temperature and humidity, which are also factors that cannot be ignored in safe operations, especially when working outdoors. Through this comprehensive data analysis, the system can not only improve the personal safety of employees, but also ensure the operating efficiency and safety of the entire power system.
[0091] Reference Figure 1 S20 in which S20 is applied to a data processing unit of a power operation risk identification system based on deep learning, and is used to perform preprocessing operations on the first data, the second data, the third data, and the fourth data, respectively, to obtain first processed data, second processed data, third processed data, and fourth processed data;
[0092] Furthermore, the specific steps of preprocessing the multi-source data include:
[0093] Extracting a key frame from the first data, the key frame being a frame of the operator's working area; resizing and graying the key frame; applying edge detection image processing technology to the key frame to identify and mark the wearing of safety equipment to obtain the first processed data; the safety equipment includes a safety helmet, safety gloves and a safety rope;
[0094] Reference Figure 3 An example diagram of a key frame provided in this embodiment, refer to Figure 4 This is an example diagram of a key frame after preprocessing provided in this embodiment.
[0095] like Figure 4 As shown, the operator is wearing all safety equipment.
[0096] Performing data cleaning on the second data to check and correct outliers or missing data;
[0097] Normalizing the heart rate data in the second data to between 0 and 1 to obtain preprocessed heart rate data;
[0098] The drinking state in the second data is converted into a binary feature, represented by 0 and 1, to obtain a pre-processed drinking state; 0 and 1 represent not drinking and drinking, respectively;
[0099] splicing the pre-processed heart rate data and the pre-processed drinking status to obtain the second processed data;
[0100] Performing data cleaning on the third data to check and correct abnormal values or missing data;
[0101] Use Butterworth filter to remove high frequency noise from voltage and current data;
[0102] performing normalization processing on the third data;
[0103] splicing the voltage, current and temperature of the device to obtain the third processed data;
[0104] Whether it is raining is converted into a binary feature, represented by 0 and 1, where 0 and 1 represent no rain and rain respectively;
[0105] Performing the data cleaning and normalization processing on the temperature data and humidity;
[0106] The whether it is raining, the temperature data and the humidity data are concatenated to obtain the fourth processed data.
[0107] In this embodiment, the specific values of the collected multi-source data are shown in Table 1:
[0108] Table 1 Specific numerical values of multi-source data collected in Nanjing A area
[0109]
[0110] By performing different preprocessing steps on multi-source data, the power operation risk identification system significantly improves the quality of data and the accuracy of analysis. From extracting and processing key video frames of operators to ensure that safety equipment is worn correctly, to standardizing heart rate and drinking status data, each step is designed to ensure the consistency and reliability of input data. These steps provide deep learning models with an easy-to-process format by removing noise, standardizing data ranges, and transforming key variables, laying a solid foundation for subsequent feature fusion and risk assessment. Through this refined data preprocessing step, the model can more accurately analyze and predict potential risks in power operations, thereby improving safety operation standards and reducing accident rates. This approach not only enhances the availability and quality of data, but also ensures that the results of risk identification are more reliable, thereby effectively supporting safety management in the power industry.
[0111] Reference Figure 1S30 in which S30 is applied to a data splicing unit of a power operation risk identification system based on deep learning, and is used to splice the first processed data and the second processed data to obtain first spliced data; and splice the third processed data and the fourth processed data to obtain second spliced data;
[0112] The first spliced data and the second spliced data are obtained by the splicing operation, creating a multi-dimensional data input for the risk identification model. This splicing enhances the expressiveness of the data and makes the risk identification more accurate and comprehensive.
[0113] Reference Figure 1 S40 in the embodiment is applied to a data analysis unit of a power operation risk identification system based on deep learning. By constructing a risk identification model, the first spliced data and the second spliced data are input into the risk identification model, and the risk identification result is obtained in sequence through the feature analysis layer, the feature fusion layer and the risk identification layer; Figure 5 This is the framework diagram of the risk discrimination model.
[0114] Furthermore, the feature analysis layer is used to perform feature extraction and training on the first spliced data and the second spliced data respectively to obtain a first analysis feature vector and a second analysis feature vector;
[0115] Furthermore, the feature analysis layer includes a first network structure, a second network structure and a multi-risk interactive attention mechanism; the first network structure and the second network structure are both LSTM network structures; wherein the first network structure is used to analyze the first spliced data to obtain a first analysis feature; and the second network structure is used to analyze the second spliced data to obtain a second analysis feature;
[0116] The first analysis feature and the second analysis feature are processed by the multi-risk interactive attention mechanism to obtain the first analysis feature vector and the second analysis feature vector;
[0117] The processing process of the multi-risk interactive attention mechanism includes:
[0118] First, generate the interaction matrix. The specific formula is:
[0119]
[0120] Then, the attention score is calculated. The specific formula is:
[0121] a ij =v T σ(I ij +u i +u j );
[0122] Furthermore, the attention weight is normalized, and the specific formula is:
[0123]
[0124] Finally, feature fusion is performed, and the specific formula is:
[0125]
[0126] in, W represents feature concatenation, which is used to concatenate feature vectors from different sources; ij is a learnable weight matrix, customized for each pair of feature combinations; b ij is the bias term, used to adjust the output of the interaction matrix; F i and F j are the features processed by different network structures; v is the global weight vector used to convert the interaction matrix into a scalar attention score; σ is the activation function; u i and u j is with the feature F i and F j The associated bias vector; F i ·F j Represents the element-wise product of two features.
[0127] In the power operation risk identification system based on deep learning, the feature analysis layer design greatly improves the depth and effect of data processing. By using two LSTM network structures, the system can efficiently extract time-dependent features from the first spliced data and the second spliced data, which represent the complex combination of operator behavior and environment and equipment status. Furthermore, the application of multi-risk interactive attention mechanism enables the system to not only process these data in a single way, but also evaluate and optimize the mutual influence and importance between different data sources. This mechanism generates an interaction matrix, calculates and normalizes the attention weights, and finally achieves highly refined feature integration in the feature fusion stage. This enables the risk identification model to understand and respond to the potential complex risks in power operations at a deeper level, thereby greatly enhancing the accuracy of risk identification and the response speed of the system.
[0128] Furthermore, the feature fusion layer fuses the first analysis feature and the second analysis feature to obtain a comprehensive analysis feature; the comprehensive analysis feature is used to provide a basis for the final risk judgment;
[0129] The risk identification layer identifies the risk of the power operation according to the comprehensive analysis characteristics and outputs the risk identification result; the risk identification result is the probability value of the power operation risk being low risk, medium risk and high risk;
[0130] Furthermore, historical power operation risk data is collected to construct a power operation risk data set; the power operation risk data set is used to perform model training on the risk discrimination model.
[0131] Furthermore, the specific process of training the risk identification model includes:
[0132] Collecting data: collecting 5,000 pieces of historical power operation risk data from the historical power operation risk data; the historical power operation risk data includes the first data, the second data, the third data and the fourth data;
[0133] Data marking: marking the risk identification result corresponding to the historical power operation risk data; the risk level is determined according to the risk identification result; the risk levels include 1, 2 and 3, corresponding to low risk, medium risk and high risk respectively;
[0134] Data preprocessing: performing preprocessing operations on the first data, the second data, the third data, and the fourth data respectively;
[0135] Divide the 5000 pieces of historical risk data into a training set and a validation set in a ratio of 7:3;
[0136] The risk discrimination model is trained by using the training set; and the risk discrimination model is verified and optimized by using the verification set.
[0137] In this embodiment, the risk discrimination model is systematically trained to enhance the discrimination accuracy and practicality of the risk discrimination model. By marking each piece of historical power operation risk data, the risk level from low to high is reflected, providing a practical training target for the model. By dividing the data into training sets and validation sets, the quality and consistency of the data are ensured, so that the model can effectively discriminate and predict risks in actual scenarios. This comprehensive training and validation process not only improves the model's sensitivity to risk, but also optimizes its application effect in different operating environments.
[0138] Furthermore, the model training includes optimizing a loss function, and the formula for optimizing the loss function is:
[0139]
[0140] Where C is the total number of categories, including no risk, low risk, medium risk and high risk in risk identification; w c is the weight of category c; y c is a one-hot encoded label. When the sample actually belongs to category c, y c is 1, otherwise yc is 0; p c The probability that the model predicts that the sample belongs to category c; y c+1 is the category label converted to calculate the sequential loss. If the actual category of the sample is greater than or equal to c+1, it is 1, otherwise it is 0; λ is the adjustment parameter.
[0141] The optimization loss function proposed in this embodiment ensures that the model can effectively learn and distinguish different levels of risk. The optimization loss function is specially designed to deal with class imbalance and orderliness issues. By introducing class weights and sequential loss calculations, it not only improves the performance of the model on frequently occurring categories, but also strengthens the model's ability to distinguish rare but critical risk levels. This method significantly improves the efficiency and accuracy of the model in practical applications by accurately adjusting the importance between different risk levels.
[0142] Embodiment 2
[0143] This embodiment analyzes the maintenance scenario of the power system line workers in Nanjing B area on the high-voltage transmission line.
[0144] Reference Figure 1 S10 in which S10 is applied to a data acquisition unit of a power operation risk identification system based on deep learning, and is used to collect multi-source data; the multi-source data includes first data, second data, third data and fourth data;
[0145] Furthermore, the first data is operation data, including video data of operators on duty, which is acquired through a camera and used to analyze the operator's wearing of safety equipment; the second data is personnel data, including the operator's heart rate data and drinking status; the second data is acquired through a wearable device; the third data is equipment data, including the voltage, current and temperature of the equipment; the third data is acquired in real time through a sensor; the fourth data is environmental data, including whether it is raining, temperature data and humidity data.
[0146] By integrating and analyzing multi-source data, the power operation risk identification system can provide a comprehensive safety assessment, thereby improving the safety of power industry operations. The first data includes video data of operators obtained through cameras, which can be used to monitor and ensure that employees wear safety equipment correctly, including helmets, insulating gloves and safety ropes, which is crucial to preventing accidents. The second data involves the heart rate and drinking status of personnel. This information helps to assess the health and suitability of employees to determine whether they are safe to work. The third data includes the voltage, current and temperature of the equipment. Through real-time monitoring of these data, equipment abnormalities can be detected in time to prevent power failures. The fourth data involves environmental factors such as weather conditions and temperature and humidity, which are also factors that cannot be ignored in safe operations, especially when working outdoors. Through this comprehensive data analysis, the system can not only improve the personal safety of employees, but also ensure the operating efficiency and safety of the entire power system.
[0147] Reference Figure 1 S20 in which S20 is applied to a data processing unit of a power operation risk identification system based on deep learning, and is used to perform preprocessing operations on the first data, the second data, the third data, and the fourth data, respectively, to obtain first processed data, second processed data, third processed data, and fourth processed data;
[0148] Furthermore, the specific steps of preprocessing the multi-source data include:
[0149] Extracting a key frame from the first data, the key frame being a frame of the operator's working area; resizing and graying the key frame; applying edge detection image processing technology to the key frame to identify and mark the wearing of safety equipment to obtain the first processed data; the safety equipment includes a safety helmet, safety gloves and a safety rope;
[0150] Performing data cleaning on the second data to check and correct outliers or missing data;
[0151] Normalizing the heart rate data in the second data to between 0 and 1 to obtain preprocessed heart rate data;
[0152] The drinking state in the second data is converted into a binary feature, represented by 0 and 1, to obtain a pre-processed drinking state; 0 and 1 represent not drinking and drinking, respectively;
[0153] splicing the pre-processed heart rate data and the pre-processed drinking status to obtain the second processed data;
[0154] Performing data cleaning on the third data to check and correct abnormal values or missing data;
[0155] Use Butterworth filter to remove high frequency noise from voltage and current data;
[0156] performing normalization processing on the third data;
[0157] splicing the voltage, current and temperature of the device to obtain the third processed data;
[0158] Whether it is raining is converted into a binary feature, represented by 0 and 1, where 0 and 1 represent no rain and rain respectively;
[0159] Performing the data cleaning and normalization processing on the temperature data and humidity;
[0160] The whether it is raining, the temperature data and the humidity data are concatenated to obtain the fourth processed data.
[0161] In this embodiment, the specific values of the collected multi-source data are shown in Table 2:
[0162] Table 2 Specific numerical values of multi-source data collected in Nanjing B area
[0163]
[0164] Reference Figure 1 S30 in which S30 is applied to a data splicing unit of a power operation risk identification system based on deep learning, and is used to splice the first processed data and the second processed data to obtain first spliced data; and splice the third processed data and the fourth processed data to obtain second spliced data;
[0165] The first spliced data and the second spliced data are obtained by the splicing operation, creating a multi-dimensional data input for the risk identification model. This splicing enhances the expressiveness of the data and makes the risk identification more accurate and comprehensive.
[0166] Reference Figure 1 S40 in the embodiment is applied to a data analysis unit of a power operation risk identification system based on deep learning. By constructing a risk identification model, the first spliced data and the second spliced data are input into the risk identification model, and the risk identification results are obtained in sequence through a feature analysis layer, a feature fusion layer and a risk identification layer;
[0167] Furthermore, the feature analysis layer is used to perform feature extraction and training on the first spliced data and the second spliced data respectively to obtain a first analysis feature vector and a second analysis feature vector;
[0168] Furthermore, the feature analysis layer includes a first network structure, a second network structure and a multi-risk interactive attention mechanism; the first network structure and the second network structure are both LSTM network structures; wherein the first network structure is used to analyze the first spliced data to obtain a first analysis feature; the second network structure is used to analyze the second spliced data to obtain a second analysis feature; the first analysis feature and the second analysis feature are processed by the multi-risk interactive attention mechanism to obtain the first analysis feature vector and the second analysis feature vector;
[0169] The processing process of the multi-risk interactive attention mechanism includes:
[0170] First, generate the interaction matrix. The specific formula is:
[0171]
[0172] Then, the attention score is calculated. The specific formula is:
[0173] a ij =v T σ(I ij +u i +u j );
[0174] Furthermore, the attention weight is normalized, and the specific formula is:
[0175]
[0176] Finally, feature fusion is performed, and the specific formula is:
[0177]
[0178] in, W represents feature concatenation, which is used to concatenate feature vectors from different sources; ij is a learnable weight matrix, customized for each pair of feature combinations; b ij is the bias term, used to adjust the output of the interaction matrix; F i and F j are the features processed by different network structures; v is the global weight vector used to convert the interaction matrix into a scalar attention score; σ is the activation function; u i and u j is with the feature F i and F j The associated bias vector; F i ·F j Represents the element-wise product of two features.
[0179] Furthermore, the feature fusion layer fuses the first analysis feature and the second analysis feature to obtain a comprehensive analysis feature; the comprehensive analysis feature is used to provide a basis for the final risk judgment;
[0180] The risk identification layer identifies the risk of the power operation according to the comprehensive analysis characteristics and outputs the risk identification result; the risk identification result is the probability value of the power operation risk being low risk, medium risk and high risk;
[0181] Furthermore, historical power operation risk data is collected to construct a power operation risk data set; the power operation risk data set is used to perform model training on the risk discrimination model.
[0182] Furthermore, the specific process of training the risk identification model includes:
[0183] Collecting data: collecting 5,000 pieces of historical power operation risk data from the historical power operation risk data; the historical power operation risk data includes the first data, the second data, the third data and the fourth data;
[0184] Data marking: marking the risk identification result corresponding to the historical power operation risk data; the risk level is determined according to the risk identification result; the risk levels include 1, 2 and 3, corresponding to low risk, medium risk and high risk respectively;
[0185] Data preprocessing: performing preprocessing operations on the first data, the second data, the third data, and the fourth data respectively;
[0186] Divide the 5000 pieces of historical risk data into a training set and a validation set in a ratio of 7:3;
[0187] The risk discrimination model is trained by using the training set; and the risk discrimination model is verified and optimized by using the verification set.
[0188] Furthermore, the model training includes optimizing a loss function, and the formula for optimizing the loss function is:
[0189]
[0190] Where C is the total number of categories, including no risk, low risk, medium risk and high risk in risk identification; w c is the weight of category c; y c is a one-hot encoded label. When the sample actually belongs to category c, y c is 1, otherwise y c is 0; p cThe probability that the model predicts that the sample belongs to category c; y c+1 is the category label converted to calculate the sequential loss. If the actual category of the sample is greater than or equal to c+1, it is 1, otherwise it is 0; λ is the adjustment parameter.
[0191] 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 method for identifying power operation risks based on deep learning, characterized in that: include: Collect data from multiple sources; The multi-source data includes first data, second data, third data and fourth data; Preprocessing the multi-source data to obtain first processed data, second processed data, third processed data and fourth processed data; splicing the first processed data and the second processed data to obtain first spliced data; splicing the third processed data and the fourth processed data to obtain second spliced data; Constructing a risk discrimination model, inputting the first spliced data and the second spliced data into the risk discrimination model, and sequentially passing through a feature analysis layer, a feature fusion layer, and a risk discrimination layer to obtain a risk discrimination result; The feature analysis layer is used to extract and train features of the first spliced data and the second spliced data respectively to obtain a first analysis feature vector and a second analysis feature vector; The feature fusion layer fuses the first analysis feature and the second analysis feature to obtain a comprehensive analysis feature; the comprehensive analysis feature is used to provide a basis for the final risk judgment; The risk identification layer identifies the risk of power operation according to the comprehensive analysis characteristics and outputs the risk identification result; The risk identification result is the probability value of the power operation risk being low risk, medium risk and high risk; Collect historical power operation risk data and build a power operation risk data set; The electric power operation risk data set is used to perform model training on the risk discrimination model.
2. The method for identifying power operation risks based on deep learning according to claim 1 is characterized in that: The first data is operation data, including video data of operators on duty, which is obtained through a camera and is used to analyze the operator's wearing of safety equipment; the second data is personnel data, including the operator's heart rate data and whether he or she is drinking; the second data is obtained through a wearable device; the third data is equipment data, including the voltage, current and temperature of the equipment; the third data is obtained in real time through sensors; the fourth data is environmental data, including whether it is raining, temperature data and humidity data.
3. The method for identifying power operation risks based on deep learning according to claim 1 is characterized in that: The specific steps of preprocessing the multi-source data include: Extracting a key frame from the first data, the key frame being a frame of the operator's working area; resizing and graying the key frame; applying edge detection image processing technology to the key frame to identify and mark the wearing of safety equipment to obtain the first processed data; the safety equipment includes a safety helmet, insulating gloves and a safety rope; Performing data cleaning on the second data to check and correct outliers or missing data; Normalizing the heart rate data in the second data to between 0 and 1 to obtain preprocessed heart rate data; The drinking state in the second data is converted into a binary feature, represented by 0 and 1, to obtain a pre-processed drinking state; 0 and 1 represent not drinking and drinking, respectively; splicing the pre-processed heart rate data and the pre-processed drinking status to obtain the second processed data; Performing data cleaning on the third data to check and correct abnormal values or missing data; Use Butterworth filter to remove high frequency noise from voltage and current data; performing normalization processing on the third data; splicing the voltage, current and temperature of the device to obtain the third processed data; Whether it is raining is converted into a binary feature, represented by 0 and 1, where 0 and 1 represent no rain and rain respectively; Performing the data cleaning and normalization processing on the temperature data and humidity; The whether it is raining, the temperature data and the humidity data are concatenated to obtain the fourth processed data.
4. The method for identifying power operation risks based on deep learning according to claim 1 is characterized in that: The feature analysis layer includes a first network structure, a second network structure and a multi-risk interactive attention mechanism; the first network structure and the second network structure are both LSTM network structures; wherein the first network structure is used to analyze the first spliced data to obtain a first analysis feature; the second network structure is used to analyze the second spliced data to obtain a second analysis feature; the first analysis feature and the second analysis feature are processed by the multi-risk interactive attention mechanism to obtain the first analysis feature vector and the second analysis feature vector; The processing process of the multi-risk interactive attention mechanism includes: First, generate the interaction matrix. The specific formula is: Then, the attention score is calculated. The specific formula is: a ij =v T σ(I ij +u i +u j ): Furthermore, the attention weight is normalized, and the specific formula is: Finally, feature fusion is performed, and the specific formula is: in, W represents feature concatenation, which is used to concatenate feature vectors from different sources; ij is a learnable weight matrix, customized for each pair of feature combinations; b ij is the bias term, used to adjust the output of the interaction matrix; F i and F j are the features processed by different network structures; v is the global weight vector used to convert the interaction matrix into a scalar attention score; σ is the activation function; u i and u j is with the feature F i and F j The associated bias vector; F i ·F j Represents the element-wise product of two features.
5. The method for identifying power operation risks based on deep learning according to claim 1, characterized in that: The specific process of training the risk discrimination model includes: Collecting data: collecting 5,000 pieces of historical power operation risk data from the historical power operation risk data; the historical power operation risk data includes the first data, the second data, the third data and the fourth data; Data marking: marking the risk identification result corresponding to the historical power operation risk data; the risk level is determined according to the risk identification result; the risk levels include 1, 2 and 3, corresponding to low risk, medium risk and high risk respectively; Data preprocessing: performing preprocessing operations on the first data, the second data, the third data, and the fourth data respectively; Divide the 5000 pieces of historical risk data into a training set and a validation set in a ratio of 7:3; The risk discrimination model is trained by using the training set; and the risk discrimination model is verified and optimized by using the verification set.
6. A method for identifying power operation risks based on deep learning according to claim 5, characterized in that: The model training includes optimizing the loss function, and the formula of the optimized loss function is: Where C is the total number of categories, including no risk, low risk, medium risk and high risk in risk identification; w c is the weight of category c; y c is a one-hot encoded label. When the sample actually belongs to category c, y c is 1, otherwise y c is 0; p c The probability that the model predicts that the sample belongs to category c; y c+1 is the category label converted to calculate the sequential loss. If the actual category of the sample is greater than or equal to c+1, it is 1, otherwise it is 0; λ is the adjustment parameter.
7. A power operation risk identification system based on deep learning, characterized in that: include: A data acquisition unit, used to acquire first data, second data, third data and fourth data; a data processing unit, configured to perform preprocessing operations on the first data, the second data, the third data, and the fourth data, respectively, to obtain first processed data, second processed data, third processed data, and fourth processed data; A data splicing unit, configured to splice the first processed data and the second processed data to obtain first spliced data; and splice the third processed data and the fourth processed data to obtain second spliced data; The data analysis unit constructs a risk identification model for analyzing the first spliced data and the second spliced data to obtain a risk identification result.
8. The power operation risk identification system based on deep learning according to claim 7 is characterized in that: The first data is operation data, including video data of operators on duty, which is obtained through a camera and is used to analyze the operator's wearing of safety equipment; the second data is personnel data, including the operator's heart rate data and whether he or she is drinking; the second data is obtained through a wearable device; the third data is equipment data, including the voltage, current and temperature of the equipment; the third data is obtained in real time through a sensor; the fourth data is environmental data, including whether it is raining, temperature value and humidity value.
9. The power operation risk identification system based on deep learning according to claim 7 is characterized in that: The specific process of training the risk identification model includes: Collecting data: collecting 5,000 pieces of historical power operation risk data from historical power operation risk data; the historical power operation risk data includes the first data, the second data, the third data and the fourth data; Data marking: marking the risk identification results corresponding to the historical power operation risk data; the risk identification results include 1, 2 and 3, corresponding to low risk, medium risk and high risk respectively; Data preprocessing: performing preprocessing operations on the first data, the second data, the third data, and the fourth data respectively; Divide the 5000 pieces of historical risk data into a training set and a validation set in a ratio of 7:3; The risk discrimination model is trained by using the training set; and the risk discrimination model is verified and optimized by using the verification set.
10. The power operation risk identification system based on deep learning according to claim 9, characterized in that: The model training includes optimizing the loss function, and the formula of the optimized loss function is: Where C is the total number of categories, including low risk, medium risk and high risk in risk identification; w c is the weight of category c; y c is a one-hot encoded label. When the sample actually belongs to category c, y c is 1, otherwise y c is 0; p c The probability that the model predicts that the sample belongs to category c; y c+1 is the category label converted to calculate the sequential loss. If the actual category of the sample is greater than or equal to c+1, it is 1, otherwise it is 0; λ is the adjustment parameter.
Citation Information
Cited By
Construction safety gridding management optimization method and system based on artificial intelligence
CN119784157A
Construction safety grid management optimization method and system based on artificial intelligence
CN119784157B
