Multi-scene visual algorithm evaluation method based on electric power artificial intelligence platform

Abstract

The invention discloses a multi-scene visual algorithm evaluation method based on an electric power artificial intelligence platform, and the method comprises the following steps: S1, integrating data in an electric power scene, and generating a multi-modal data set; s2, storing and preprocessing the data set, and extracting key features; s3, based on a multi-stage active learning mechanism, selectively sampling data and optimizing a visual algorithm; s4, the optimized algorithm is migrated to different environments, parameters are adjusted through a self-adaptive strategy, and models adapting to different conditions are generated; s5, in combination with deep learning and expert knowledge, optimizing an evaluation process; s6, performing real-time fault injection and elasticity evaluation by using a fault mode library and a simulation technology; s7, constructing a double-layer digital twinning virtual evaluation environment, and verifying the validity of the algorithm under different conditions; and S8, recording and tracking the evaluation data and generating an evaluation report by using a distributed data management system. According to the invention, through an intelligent tuning and evaluation mechanism, the efficiency and reliability of power equipment monitoring and fault prediction are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power system monitoring and fault detection, and in particular to a multi-scenario visual algorithm evaluation method based on an electric power artificial intelligence platform. Background Art

[0002] With the rapid development and improvement of the intelligence level of power systems, the monitoring and fault detection technology of power equipment has become the key to ensure the stable operation of power systems. However, the existing power equipment monitoring and fault detection methods have many shortcomings in practical applications and are difficult to meet the needs of efficient, safe and reliable operation of modern power systems. Specifically, the existing technology has obvious defects in the following aspects:

[0003] 1. Single data and insufficient information: Traditional power equipment monitoring methods mainly rely on a single type of sensor data, such as current, voltage, and temperature. Although this data can provide some operating status information of the equipment, it cannot fully reflect the health status of the equipment. The limitations of a single data source make it difficult to identify complex fault types, resulting in low accuracy and reliability of fault detection.

[0004] 2. Low efficiency of manual inspection: The inspection of existing power equipment mainly relies on manual operation, and the inspectors regularly inspect the equipment. This method has problems such as high labor intensity, low efficiency, low inspection frequency, and inability to monitor the equipment status in real time. At the same time, the professional level and experience of the inspectors will also affect the accuracy of the test results, and there will be subjective judgment errors, making it difficult to detect equipment failures in a timely manner.

[0005] 3. Static monitoring systems lack dynamic adaptability: Many existing monitoring systems use fixed rules or preset thresholds for monitoring, which makes it difficult to cope with the complexity of power systems and changes in environmental conditions. Static monitoring systems are difficult to adapt to changing operating environments and cannot adjust monitoring strategies in a timely manner, resulting in frequent false alarms and missed alarms. For example, when environmental conditions change, fixed thresholds and rules cannot effectively detect faults, thus affecting the accuracy of monitoring.

[0006] 4. Lack of data fusion and intelligent analysis capabilities: Existing monitoring systems usually only use a single data source and lack the ability to comprehensively utilize and intelligently analyze multi-source data. Traditional monitoring equipment mainly relies on single sensor data and lacks multimodal data fusion and adaptive feature extraction technology. It is difficult to extract effective information from massive data and conduct in-depth fault analysis and prediction. This limits the adaptability and detection accuracy of the monitoring system when facing complex faults and changing environments.

[0007] 5. Imperfect evaluation mechanism: Traditional fault detection evaluation mainly relies on offline evaluation and static testing, which cannot simulate the real operating environment and fault scenarios. The lack of a complete fault injection and elasticity evaluation mechanism makes it impossible to fully evaluate the robustness and extreme performance of the algorithm, resulting in poor results in actual applications. For example, in offline evaluation, it is impossible to consider the dynamic changes and complex fault modes of the equipment in actual operation, and the evaluation results often deviate from the actual situation.

[0008] 6. System isolation and lack of coordination: The existing monitoring system lacks effective integration and coordination with other systems, resulting in insufficient information sharing and comprehensive analysis capabilities, making it difficult to form comprehensive monitoring and fault diagnosis of power equipment. The isolation of the monitoring system makes it difficult for data between different systems to communicate, and the information island phenomenon is serious, affecting the efficiency of overall monitoring and fault detection.

[0009] Therefore, how to provide a multi-scenario visual algorithm evaluation method based on an electric power artificial intelligence platform is an issue that technical personnel in this field urgently need to solve. Summary of the invention

[0010] One object of the present invention is to propose a multi-scenario visual algorithm evaluation method based on an electric power artificial intelligence platform. The present invention makes full use of multimodal data fusion, adaptive feature extraction, active learning, transfer learning, intelligent evaluation and digital twin technology, and describes in detail the intelligent implementation process of power equipment monitoring and fault detection, and has the advantages of high efficiency, accuracy and reliability.

[0011] A multi-scenario visual algorithm evaluation method based on an electric power artificial intelligence platform according to an embodiment of the present invention comprises the following steps:

[0012] S1. Integrate visible light, infrared, lidar and environmental sensor data in power scenarios and use adaptive fusion algorithms to generate multimodal data sets;

[0013] S2, store and preprocess multimodal datasets and extract key features using adaptive feature extraction techniques;

[0014] S3, based on a multi-stage active learning mechanism, uses the extracted multimodal feature data for selective data sampling, and optimizes the visual algorithm in stages through an active learning algorithm;

[0015] S4. Apply the automated tuning technology of transfer learning to migrate the optimized visual algorithm to the target domain, automatically adjust key parameters through adaptive strategies, and generate models that adapt to different environmental conditions;

[0016] S5. Based on the model after automatic tuning, combined with multimodal data sets, a human-machine collaborative intelligent evaluation method based on deep learning and expert knowledge fusion is adopted to optimize the evaluation process through real-time expert feedback and intelligent annotation models;

[0017] S6. Combined with the optimized model, using real-time fault injection and resilience evaluation mechanism, the simulation technology based on the power system fault mode library is used to generate fault scenarios. By dynamically adjusting the complexity of the scenarios, the robustness and limit performance of the algorithm are comprehensively evaluated.

[0018] S7. Build a two-layer digital twin virtual evaluation environment, integrate the optimization model and multimodal data, and verify the effectiveness of the algorithm under different operating conditions through two-way twin simulation technology;

[0019] S8. A distributed data management system using multi-layer encryption and blockchain technology can record and track evaluation data and results, support cross-regional multi-party collaboration, and generate evaluation reports.

[0020] Optionally, the S2 specifically includes:

[0021] S21, storing the generated multimodal data set in a distributed database, and performing data preprocessing on the stored data set;

[0022] S22, extract key features from the preprocessed multimodal data set using adaptive feature extraction technology based on power equipment features;

[0023] S23. Through nonlinear dynamic system feature extraction technology, the multi-dimensional nonlinear mapping function of complex dynamic systems is used to extract features from data:

[0024]

[0025] Among them, x n is the value of the nth iteration, r and γ are system parameters, k i is the frequency parameter, m is the number of frequency parameters;

[0026] S24, feature extraction result x n Input into the complex domain feature transformation step, perform complex domain transformation on the image data, and map the data from the real domain to the complex domain:

[0027]

[0028] Among them, Z is the complex domain feature, X and Y are the data in the real domain, θ is the phase angle, δ and γ are the adjustment parameters, ω j and are the frequency and phase parameters respectively, and n is the number of parameters;

[0029] S25, input the complex domain feature transformation result Z into the high-dimensional space mapping step, use the high-dimensional space mapping technology to map the multimodal data to a higher-dimensional feature space, and realize feature extraction in the high-dimensional space through an adaptive high-dimensional kernel function:

[0030]

[0031] Among them, K(Z1,Z2) is the kernel function, and is the result after complex domain feature mapping, σ, α, β, p and q are adaptive parameters;

[0032] S26. Store the extracted key feature data back into the distributed database.

[0033] Optionally, the S3 specifically includes:

[0034] S31, using the multimodal data set after the key features are extracted as input data, initializing the multimodal training data set for model training, and marking the samples in the data set, the markings including the normal state and fault state of the power equipment;

[0035] S32, performing preliminary data sampling, selecting a portion of samples from the multimodal training data set using a random sampling method, and constructing an initial training set;

[0036] S33, train the initial visual algorithm model, input the initial training set data, and use the generalized Gaussian mixture network structure to extract features from the image data:

[0037]

[0038] Among them, O(i,j) represents the value of the (i,j)th position of the output feature map, I(i+m,j+n) represents the value of the (i+m,j+n)th position of the input image, μ and ν are the mean parameters of the Gaussian distribution, σ is the standard deviation of the Gaussian distribution, b is the bias value, and k is the size of the convolution kernel;

[0039] S34, performing scene change detection based on the output result of the initial visual algorithm model, and calculating the detection error by comparing the detection result with the true mark;

[0040] S35, according to the detection error result, using the extracted multimodal feature data for selective data sampling, using the extracted complex domain feature transformation result Z and the nonlinear dynamic system feature extraction result x n To determine the sample selection strategy, give priority to samples with larger errors for further training;

[0041] S36. Add the selectively sampled samples to the training set, retrain the visual algorithm model, and optimize the visual algorithm in stages through a multi-stage active learning algorithm:

[0042]

[0043] Among them, θ t is the model parameter of the tth iteration, η is the learning rate, J(θ) is the loss function, is the gradient of the loss function, γ is the adjustment parameter, K(Z k ,Z k+1 ) is a high-dimensional kernel function, L is the number of feature pairs, and Z is the result of feature transformation in the complex domain;

[0044] S37, in each iteration, use the complex domain feature transformation result Z and the nonlinear dynamic system feature extraction result x n As part of the input data;

[0045] S38, repeating steps S34 to S37 until the detection error converges to a preset threshold, forming a final optimized visual algorithm;

[0046] S39. Apply the optimized visual algorithm to multi-scenario visual inspection of power equipment, verify the effectiveness of the algorithm in actual application, and record and store the verification results.

[0047] Optionally, the S4 specifically includes:

[0048] S41, migrate the optimized visual algorithm model from the source domain to the target domain, and continue to collect multimodal data sets in the target domain;

[0049] S42, preprocessing the multimodal dataset of the target domain to obtain a preprocessed multimodal dataset;

[0050] S43, applying an adaptive strategy to extract features from the multimodal data set, and generating a feature vector using an adaptive feature extraction algorithm;

[0051] S44. Based on the data features of the target domain, transfer learning is performed on the initial visual algorithm model, model parameters are adjusted through a multi-layer neural network, and model weights and bias parameters are adjusted in combination with the results of adaptive feature extraction;

[0052] S45. Automatically tune the visual algorithm model after transfer learning, use the automatic parameter tuning algorithm to optimize the model's hyperparameters, and combine the adaptive feature extraction results and high-dimensional kernel functions;

[0053] S46. Use the improved co-evolution algorithm to evaluate the tuned model. Combined with the multi-objective optimization strategy, the co-evolution evaluation is performed considering the accuracy, robustness and computational efficiency of the model:

[0054]

[0055] Among them, F(θ) represents the fitness function of the co-evolution, α, β and γ are adaptive weight parameters, Accuracy(θ, Z, x n ) represents the combination of complex domain feature transformation result Z and nonlinear dynamic system feature extraction result x n The model accuracy after θ is θ, Z, x n ) represents the combination of Z and x n The robustness of the model after Efficiency(θ,Z,x n ) represents the combination of Z and x n The model calculation efficiency after ∈ is a very small positive number;

[0056] S47. According to the optimization results of the improved co-evolutionary algorithm, the model parameter combination with the best performance is selected, and the tuned visual algorithm model is applied to the visual inspection of power equipment in the target field.

[0057] Optionally, the S6 specifically includes:

[0058] S61. Construct a simulation platform based on a power system fault mode library, where the fault mode library includes common power equipment fault types and their corresponding fault feature data;

[0059] S62. Perform preliminary testing on the optimized visual algorithm model on the simulation platform, and record the performance parameters of the model under normal working conditions;

[0060] S63, selecting multiple fault types from the fault mode library, and injecting these fault types into the simulation platform in sequence according to the preset injection frequency and time interval, so as to simulate the fault scenarios encountered by the actual power equipment;

[0061] S64. Real-time monitoring of the visual algorithm model's response to fault detection during the fault injection process, and recording the model's detection accuracy, false alarm rate, and missed alarm rate for each fault type:

[0062]

[0063] Among them, TP is the number of true positives, TN is the number of true negatives, FP is the number of false positives, and FN is the number of false negatives; θ is the model parameter, Z is the result of complex domain feature transformation, and x n is the feature extraction result of nonlinear dynamic system, α i and β i is the weight parameter, and ψ(Z i ,x n) is the characteristic function;

[0064] S65. Compare and analyze the performance changes of the visual algorithm model before and after fault injection, and use a flexible evaluation mechanism to dynamically evaluate the robustness and adaptability of the model by adjusting the frequency and intensity of fault injection;

[0065] S66. Evaluate the comprehensive detection capability of the model in complex fault scenarios by injecting a combination of multiple fault types;

[0066] S67. Based on the elasticity evaluation results, identify the performance bottlenecks of the model and make optimization suggestions to further improve the visual algorithm model.

[0067] Optionally, the S7 specifically includes:

[0068] S71. Create a digital twin virtual environment based on the updated multimodal dataset and optimized visual algorithm model;

[0069] S72. Generate a digital twin model of the power equipment in a virtual environment and use simulation technology to simulate the operating status of the power equipment under different operating conditions:

[0070]

[0071] Among them, M(t) represents the state of the digital twin model at time t, α i and β i is the weight parameter, and ψ(Z i ,x n ) are the complex domain feature transformation results Z i And the nonlinear dynamic system feature extraction result x n The characteristic function of

[0072] S73, synchronizing the data in the virtual environment and the actual environment, and comparing the operating status of the virtual model and the actual equipment in real time through a data synchronization algorithm;

[0073] S74. Apply the bidirectional twin simulation technology based on time series and power equipment behavior to perform synchronous simulation of power equipment under different operating conditions and verify the effectiveness of the virtual model under different conditions;

[0074] S75, analyzing the simulation results, and calculating the errors between the virtual model and the actual equipment under different operating conditions through an error analysis algorithm;

[0075] S76. Based on the error analysis results, optimize the digital twin model and adjust the weight parameter α i and β i :

[0076]

[0077] in, and is the optimized weight parameter, η is the learning rate, λ is the regularization parameter, and is the partial derivative of the error with respect to the weight parameter, and Represents a composite function For the weight parameter α i and β i The sum of the partial derivatives of , m is the number of composite functions;

[0078] S77. Apply the optimized digital twin model to the operation monitoring and fault prediction of actual power equipment to verify the effectiveness and reliability of the model under different operating conditions.

[0079] The beneficial effects of the present invention are:

[0080] (1) The present invention significantly improves the efficiency and accuracy of power equipment monitoring and fault detection by combining multimodal data fusion, adaptive feature extraction technology and active learning mechanism. In particular, in terms of comprehensive utilization and intelligent analysis of multi-source data, the present invention overcomes the limitations of traditional methods and can extract effective information from multiple data sources for in-depth fault analysis and prediction.

[0081] (2) The present invention optimizes the adaptability and performance of the visual algorithm in different environments and improves the versatility and stability of the model by introducing the automatic tuning technology and intelligent evaluation method of transfer learning. At the same time, the human-machine collaborative intelligent evaluation method based on the fusion of deep learning and expert knowledge is adopted to realize the intelligence and real-time feedback of the evaluation process and enhance the accuracy of fault detection.

[0082] (3) The present invention utilizes real-time fault injection and resilience evaluation mechanisms, combined with digital twin technology, to provide an efficient and accurate evaluation environment, enabling the system to simulate real operating conditions and conduct extreme performance testing, thereby comprehensively evaluating the robustness and reliability of the algorithm.

[0083] (4) The present invention achieves efficient monitoring and fault prediction of power equipment by comprehensively applying multiple data processing and analysis technologies, thereby improving the safety and reliability of the power system. The system can adapt to complex operating environments, automatically process multimodal data, and optimize fault detection and monitoring processes, thereby ensuring the stable operation of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0085] Figure 1 This is a flow chart of a multi-scenario visual algorithm evaluation method based on an electric power artificial intelligence platform proposed by the present invention;

[0086] Figure 2 This is a flow chart of multimodal data fusion and adaptive feature extraction proposed by the present invention;

[0087] Figure 3 A flow chart for constructing a double-layer digital twin virtual evaluation environment proposed in the present invention. DETAILED DESCRIPTION

[0088] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0089] refer to Figure 1-3 , a multi-scenario visual algorithm evaluation method based on an electric power artificial intelligence platform includes the following steps:

[0090] S1. Integrate visible light, infrared, lidar and environmental sensor data in power scenarios and use adaptive fusion algorithms to generate multimodal data sets;

[0091] S2, store and preprocess multimodal datasets and extract key features using adaptive feature extraction techniques;

[0092] S3, based on a multi-stage active learning mechanism, uses the extracted multimodal feature data for selective data sampling, and optimizes the visual algorithm in stages through an active learning algorithm;

[0093] S4. Apply the automated tuning technology of transfer learning to migrate the optimized visual algorithm to the target domain, automatically adjust key parameters through adaptive strategies, and generate models that adapt to different environmental conditions;

[0094] S5. Based on the model after automatic tuning, combined with multimodal data sets, a human-machine collaborative intelligent evaluation method based on deep learning and expert knowledge fusion is adopted to optimize the evaluation process through real-time expert feedback and intelligent annotation models;

[0095] S6. Combined with the optimized model, using real-time fault injection and resilience evaluation mechanism, the simulation technology based on the power system fault mode library is used to generate fault scenarios. By dynamically adjusting the complexity of the scenarios, the robustness and limit performance of the algorithm are comprehensively evaluated.

[0096] S7. Build a two-layer digital twin virtual evaluation environment, integrate the optimization model and multimodal data, and verify the effectiveness of the algorithm under different operating conditions through two-way twin simulation technology;

[0097] S8. A distributed data management system using multi-layer encryption and blockchain technology can record and track evaluation data and results, support cross-regional multi-party collaboration, and generate evaluation reports.

[0098] In this implementation, S2 specifically includes:

[0099] S21, storing the generated multimodal data set in a distributed database, and performing data preprocessing on the stored data set;

[0100] S22, extract key features from the preprocessed multimodal data set using adaptive feature extraction technology based on power equipment features;

[0101] S23. Through nonlinear dynamic system feature extraction technology, the multi-dimensional nonlinear mapping function of complex dynamic systems is used to extract features from data:

[0102]

[0103] Among them, x n is the value of the nth iteration, r and γ are system parameters, k i is the frequency parameter, m is the number of frequency parameters;

[0104] S24, feature extraction result x n Input into the complex domain feature transformation step, perform complex domain transformation on the image data, and map the data from the real domain to the complex domain:

[0105]

[0106] Among them, Z is the complex domain feature, X and Y are the data in the real domain, θ is the phase angle, δ and γ are the adjustment parameters, ω j and are the frequency and phase parameters respectively, and n is the number of parameters;

[0107] S25, input the complex domain feature transformation result Z into the high-dimensional space mapping step, use the high-dimensional space mapping technology to map the multimodal data to a higher-dimensional feature space, and realize feature extraction in the high-dimensional space through an adaptive high-dimensional kernel function:

[0108]

[0109] Among them, K(Z1,Z2) is the kernel function, and is the result after complex domain feature mapping, σ, α, β, p and q are adaptive parameters;

[0110] S26. Store the extracted key feature data back into the distributed database.

[0111] In this implementation, S3 specifically includes:

[0112] S31, using the multimodal data set after the key features are extracted as input data, initializing the multimodal training data set for model training, and marking the samples in the data set, the markings including the normal state and fault state of the power equipment;

[0113] S32, performing preliminary data sampling, selecting a portion of samples from the multimodal training data set using a random sampling method, and constructing an initial training set;

[0114] S33, train the initial visual algorithm model, input the initial training set data, and use the generalized Gaussian mixture network structure to extract features from the image data:

[0115]

[0116] Among them, O(i,j) represents the value of the (i,j)th position of the output feature map, I(i+m,j+n) represents the value of the (i+m,j+n)th position of the input image, μ and ν are the mean parameters of the Gaussian distribution, σ is the standard deviation of the Gaussian distribution, b is the bias value, and k is the size of the convolution kernel;

[0117] S34, performing scene change detection based on the output result of the initial visual algorithm model, and calculating the detection error by comparing the detection result with the true mark;

[0118] S35, according to the detection error result, using the extracted multimodal feature data for selective data sampling, using the extracted complex domain feature transformation result Z and the nonlinear dynamic system feature extraction result x n To determine the sample selection strategy, give priority to samples with larger errors for further training;

[0119] S36. Add the selectively sampled samples to the training set, retrain the visual algorithm model, and optimize the visual algorithm in stages through a multi-stage active learning algorithm:

[0120]

[0121] Among them, θ t is the model parameter of the tth iteration, η is the learning rate, J(θ) is the loss function, is the gradient of the loss function, γ is the adjustment parameter, K(Z k ,Z k+1) is a high-dimensional kernel function, L is the number of feature pairs, and Z is the result of feature transformation in the complex domain;

[0122] S37, in each iteration, use the complex domain feature transformation result Z and the nonlinear dynamic system feature extraction result x n As part of the input data;

[0123] S38, repeating steps S34 to S37 until the detection error converges to a preset threshold, forming a final optimized visual algorithm;

[0124] S39. Apply the optimized visual algorithm to multi-scenario visual inspection of power equipment, verify the effectiveness of the algorithm in actual application, and record and store the verification results.

[0125] In this implementation manner, the S4 specifically includes:

[0126] S41, migrate the optimized visual algorithm model from the source domain to the target domain, and continue to collect multimodal data sets in the target domain;

[0127] S42, preprocessing the multimodal dataset of the target domain to obtain a preprocessed multimodal dataset;

[0128] S43, applying an adaptive strategy to extract features from the multimodal data set, and generating a feature vector using an adaptive feature extraction algorithm;

[0129] S44. Based on the data features of the target domain, transfer learning is performed on the initial visual algorithm model, model parameters are adjusted through a multi-layer neural network, and model weights and bias parameters are adjusted in combination with the results of adaptive feature extraction;

[0130] S45. Automatically tune the visual algorithm model after transfer learning, use the automatic parameter tuning algorithm to optimize the model's hyperparameters, and combine the adaptive feature extraction results and high-dimensional kernel functions;

[0131] S46. Use the improved co-evolution algorithm to evaluate the tuned model. Combined with the multi-objective optimization strategy, the co-evolution evaluation is performed considering the accuracy, robustness and computational efficiency of the model:

[0132]

[0133] Among them, F(θ) represents the fitness function of the co-evolution, α, β and γ are adaptive weight parameters, Accuracy(θ, Z, x n ) represents the combination of complex domain feature transformation result Z and nonlinear dynamic system feature extraction result x n The model accuracy after θ is θ, Z, x n) represents the combination of Z and x n The robustness of the model after Efficiency(θ,Z,x n ) represents the combination of Z and x n The model calculation efficiency after ∈ is a very small positive number;

[0134] S47. According to the optimization results of the improved co-evolutionary algorithm, the model parameter combination with the best performance is selected, and the tuned visual algorithm model is applied to the visual inspection of power equipment in the target field.

[0135] In this implementation manner, S6 specifically includes:

[0136] S61. Construct a simulation platform based on a power system fault mode library, where the fault mode library includes common power equipment fault types and their corresponding fault feature data;

[0137] S62. Perform preliminary testing on the optimized visual algorithm model on the simulation platform, and record the performance parameters of the model under normal working conditions;

[0138] S63, selecting multiple fault types from the fault mode library, and injecting these fault types into the simulation platform in sequence according to the preset injection frequency and time interval, so as to simulate the fault scenarios encountered by the actual power equipment;

[0139] S64. Real-time monitoring of the visual algorithm model's response to fault detection during the fault injection process, and recording the model's detection accuracy, false alarm rate, and missed alarm rate for each fault type:

[0140]

[0141] Among them, TP is the number of true positives, TN is the number of true negatives, FP is the number of false positives, and FN is the number of false negatives; θ is the model parameter, Z is the result of complex domain feature transformation, and x n is the feature extraction result of nonlinear dynamic system, α i and β i is the weight parameter, and ψ(Z i ,x n ) is the characteristic function;

[0142] S65. Compare and analyze the performance changes of the visual algorithm model before and after fault injection, and use a flexible evaluation mechanism to dynamically evaluate the robustness and adaptability of the model by adjusting the frequency and intensity of fault injection;

[0143] S66. Evaluate the comprehensive detection capability of the model in complex fault scenarios by injecting a combination of multiple fault types;

[0144] S67. Based on the elasticity evaluation results, identify the performance bottlenecks of the model and make optimization suggestions to further improve the visual algorithm model.

[0145] In this implementation manner, the S7 specifically includes:

[0146] S71. Create a digital twin virtual environment based on the updated multimodal dataset and optimized visual algorithm model;

[0147] S72. Generate a digital twin model of the power equipment in a virtual environment and use simulation technology to simulate the operating status of the power equipment under different operating conditions:

[0148]

[0149] Among them, M(t) represents the state of the digital twin model at time t, α i and β i is the weight parameter, and ψ(Z i ,x n ) are the complex domain feature transformation results Z i And the nonlinear dynamic system feature extraction result x n The characteristic function of

[0150] S73, synchronizing the data in the virtual environment and the actual environment, and comparing the operating status of the virtual model and the actual equipment in real time through a data synchronization algorithm;

[0151] S74. Apply the bidirectional twin simulation technology based on time series and power equipment behavior to perform synchronous simulation of power equipment under different operating conditions and verify the effectiveness of the virtual model under different conditions;

[0152] S75, analyzing the simulation results, and calculating the errors between the virtual model and the actual equipment under different operating conditions through an error analysis algorithm;

[0153] S76. Based on the error analysis results, optimize the digital twin model and adjust the weight parameter α i and β i :

[0154]

[0155] in, and is the optimized weight parameter, η is the learning rate, λ is the regularization parameter, and is the partial derivative of the error with respect to the weight parameter, and Represents a composite function For the weight parameter α iand β i The sum of the partial derivatives of , m is the number of composite functions;

[0156] S77. Apply the optimized digital twin model to the operation monitoring and fault prediction of actual power equipment to verify the effectiveness and reliability of the model under different operating conditions.

[0157] Embodiment 1:

[0158] In order to verify the feasibility of the present invention, the present invention is applied to the equipment fault detection project of East China Electric Power Company. The power company has a wide variety of equipment, which is widely distributed and has a complex and changeable operating environment. Traditional fault detection methods are inefficient and rely on manual detection, which is not only time-consuming and error-prone, but also unable to meet the needs of timely and accurate fault detection. In order to improve the accuracy and efficiency of fault detection, we introduced a multi-scenario visual algorithm evaluation method based on the power artificial intelligence platform.

[0159] In the daily operation of the power company, multimodal sensors are installed at each site, including visible light cameras, infrared cameras, lidars, and environmental sensors. These sensors collect images and environmental data of the equipment in real time, and generate multimodal data sets through adaptive fusion algorithms. All collected data are transmitted to the distributed database in the central control room via a wireless network for storage and preprocessing. During the data preprocessing process, the system uses adaptive feature extraction technology to extract key features of the multimodal data set. For example, infrared cameras can detect the temperature distribution on the surface of the equipment, lidars can capture the three-dimensional structural changes of the equipment, and environmental sensors can provide temperature and humidity information around the equipment. After these feature data are processed, they are input into the active learning algorithm.

[0160] The active learning algorithm uses selective data sampling to prioritize samples with large detection errors for further training. During the training process, the system uses a generalized Gaussian mixture network structure to extract features from image data and gradually optimizes the visual algorithm model through a multi-stage active learning algorithm. At each stage of model training and optimization, the system dynamically adjusts model parameters based on real-time feedback from experts and the results of the intelligent annotation model to ensure the accuracy and robustness of the model.

[0161] When the initial optimization of the model is completed, the system migrates the model to the target field for application. During the migration process, the system continues to collect and preprocess the multimodal data set of the target field, and uses transfer learning and automatic tuning technology to further optimize the model parameters and generate a fault detection model that adapts to different environmental conditions. In actual applications, the system uses real-time fault injection and elasticity evaluation mechanisms to comprehensively evaluate the optimized visual algorithm model. Through a simulation platform based on the power system fault mode library, the system simulates various fault scenarios encountered by actual power equipment, and dynamically adjusts the complexity of the scenario to comprehensively evaluate the robustness and extreme performance of the algorithm. At the same time, the system constructs a two-layer digital twin virtual evaluation environment, and verifies the effectiveness of the algorithm under different operating conditions through bidirectional twin simulation technology.

[0162] In order to demonstrate the effectiveness and advantages of the method of the present invention, we conducted actual tests at multiple sites and central control rooms of East China Electric Power Company from May to October 2023. During the test, the system showed good robustness in different seasons and weather conditions, and adapted to various complex environments. One day in July 2023, an equipment anomaly occurred at a site of East China Electric Power Company. The traditional method requires dispatching maintenance personnel to the site for inspection, which takes 1 hour. However, with the new system, the staff in the central control room discovered the anomaly through multimodal sensor data within 3 minutes, the system automatically generated a detailed fault report, and simulated the fault scenario through the digital twin model to confirm the problem. In the end, the staff only needed to adjust the equipment parameters remotely to solve the fault, avoid long-term downtime, and ensure the continuity of power supply.

[0163] Table 1: Comparison between traditional method and the method of the present invention

[0164] index Traditional methods Method of the present invention Increase Total query time per day (minutes) 300 60 Save 80% Query error rate (%) 15％ 2％ 86.7% reduction User satisfaction (%) 60 95 35% improvement Stability in the face of complex input Unstable, high error rate Stable, low error rate - Ability to handle unusual issues Low high - Fault detection accuracy (%) 85 95 Improve by 10% Time required for each test (minutes) 10 3 3 times improvement False alarm rate (%) 10 3 70% reduction Missing rate (%) 5 1 80% reduction Number of expert interventions per month (times) 20 5 75% reduction Cost per test (yuan) 200 50 Save 75% Monthly testing cost savings (yuan) - - 15000 Annual cost savings (yuan) - - 180000

[0165] Table 2: Specific detection example data

[0166] date Detection Type Detection time Fault location Processing results 2023-07-01 Device abnormality 3 minutes Accurately locate faults Remotely adjust device parameters and solve faults 2023-07-15 Device abnormality 3 minutes Accurately locate faults Remotely adjust device parameters and solve faults 2023-08-05 Device abnormality 3 minutes Accurately locate faults Remotely adjust device parameters and solve faults 2023-08-20 Device abnormality 3 minutes Accurately locate faults Remotely adjust device parameters and solve faults

[0167] Through the above data and examples, it can be seen that the new system is significantly superior to traditional methods in terms of fault detection accuracy, speed, and economic benefits. It not only solves the problems of high false alarm and missed alarm rates in traditional methods, but also greatly reduces the detection cost and improves the reliability and efficiency of power equipment operation. These beneficial effects not only improve the operating efficiency of power companies, but also provide reference for other similar industries.

[0168] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A multi-scenario visual algorithm evaluation method based on an electric power artificial intelligence platform, characterized in that: The steps include: S1. Integrate visible light, infrared, lidar and environmental sensor data in power scenarios and use adaptive fusion algorithms to generate multimodal data sets; S2, store and preprocess multimodal datasets and extract key features using adaptive feature extraction techniques; S3, based on a multi-stage active learning mechanism, uses the extracted multimodal feature data for selective data sampling, and optimizes the visual algorithm in stages through an active learning algorithm; S4. Apply the automated tuning technology of transfer learning to migrate the optimized visual algorithm to the target domain, automatically adjust key parameters through adaptive strategies, and generate models that adapt to different environmental conditions; S5. Based on the model after automatic tuning, combined with multimodal data sets, a human-machine collaborative intelligent evaluation method based on deep learning and expert knowledge fusion is adopted to optimize the evaluation process through real-time expert feedback and intelligent annotation models; S6. Combined with the optimized model, using real-time fault injection and resilience evaluation mechanism, the simulation technology based on the power system fault mode library is used to generate fault scenarios. By dynamically adjusting the complexity of the scenarios, the robustness and limit performance of the algorithm are comprehensively evaluated. S7. Build a two-layer digital twin virtual evaluation environment, integrate the optimization model and multimodal data, and verify the effectiveness of the algorithm under different operating conditions through two-way twin simulation technology; S8. A distributed data management system using multi-layer encryption and blockchain technology can record and track evaluation data and results, support cross-regional multi-party collaboration, and generate evaluation reports.

2. According to claim 1, a multi-scenario visual algorithm evaluation method based on an electric power artificial intelligence platform is characterized in that: The S2 specifically includes: S21, storing the generated multimodal data set in a distributed database, and performing data preprocessing on the stored data set; S22, extract key features from the preprocessed multimodal data set using adaptive feature extraction technology based on power equipment features; S23. Through nonlinear dynamic system feature extraction technology, the multi-dimensional nonlinear mapping function of complex dynamic systems is used to extract features from data: Among them, x n is the value of the nth iteration, r and γ are system parameters, k i is the frequency parameter, m is the number of frequency parameters; S24, feature extraction result x n Input into the complex domain feature transformation step, perform complex domain transformation on the image data, and map the data from the real domain to the complex domain: Among them, Z is the complex domain feature, X and Y are the data in the real domain, θ is the phase angle, δ and γ are the adjustment parameters, ω j and are the frequency and phase parameters respectively, and n is the number of parameters; S25, input the complex domain feature transformation result Z into the high-dimensional space mapping step, use the high-dimensional space mapping technology to map the multimodal data to a higher-dimensional feature space, and realize feature extraction in the high-dimensional space through an adaptive high-dimensional kernel function: Among them, K(Z1,Z2) is the kernel function, and is the result after complex domain feature mapping, σ, α, β, p and q are adaptive parameters; S26. Store the extracted key feature data back into the distributed database.

3. According to claim 1, a multi-scenario visual algorithm evaluation method based on an electric power artificial intelligence platform is characterized in that: The S3 specifically includes: S31, using the multimodal data set after the key features are extracted as input data, initializing the multimodal training data set for model training, and marking the samples in the data set, the markings including the normal state and fault state of the power equipment; S32, performing preliminary data sampling, selecting a portion of samples from the multimodal training data set using a random sampling method, and constructing an initial training set; S33, train the initial visual algorithm model, input the initial training set data, and use the generalized Gaussian mixture network structure to extract features from the image data: Among them, O(i,j) represents the value of the (i,j)th position of the output feature map, I(i+m,j+n) represents the value of the (i+m,j+n)th position of the input image, μ and ν are the mean parameters of the Gaussian distribution, σ is the standard deviation of the Gaussian distribution, b is the bias value, and k is the size of the convolution kernel; S34, performing scene change detection based on the output result of the initial visual algorithm model, and calculating the detection error by comparing the detection result with the true mark; S35, according to the detection error result, using the extracted multimodal feature data for selective data sampling, using the extracted complex domain feature transformation result Z and the nonlinear dynamic system feature extraction result x n To determine the sample selection strategy, give priority to samples with larger errors for further training; S36. Add the selectively sampled samples to the training set, retrain the visual algorithm model, and optimize the visual algorithm in stages through a multi-stage active learning algorithm: Among them, θ t is the model parameter of the tth iteration, η is the learning rate, J(θ) is the loss function, is the gradient of the loss function, γ is the adjustment parameter, K(Z k ,Z k+1 ) is a high-dimensional kernel function, L is the number of feature pairs, and Z is the result of feature transformation in the complex domain; S37, in each iteration, use the complex domain feature transformation result Z and the nonlinear dynamic system feature extraction result x n As part of the input data; S38, repeating steps S34 to S37 until the detection error converges to a preset threshold, forming a final optimized visual algorithm; S39. Apply the optimized visual algorithm to multi-scenario visual inspection of power equipment, verify the effectiveness of the algorithm in actual application, and record and store the verification results.

4. According to claim 1, a multi-scenario visual algorithm evaluation method based on an electric power artificial intelligence platform is characterized in that: The S4 specifically includes: S41, migrate the optimized visual algorithm model from the source domain to the target domain, and continue to collect multimodal data sets in the target domain; S42, preprocessing the multimodal dataset of the target domain to obtain a preprocessed multimodal dataset; S43, applying an adaptive strategy to extract features from the multimodal data set, and generating a feature vector using an adaptive feature extraction algorithm; S44. Based on the data features of the target domain, transfer learning is performed on the initial visual algorithm model, model parameters are adjusted through a multi-layer neural network, and model weights and bias parameters are adjusted in combination with the results of adaptive feature extraction; S45. Automatically tune the visual algorithm model after transfer learning, use the automatic parameter tuning algorithm to optimize the model's hyperparameters, and combine the adaptive feature extraction results and high-dimensional kernel functions; S46. Use the improved co-evolution algorithm to evaluate the tuned model. Combined with the multi-objective optimization strategy, the co-evolution evaluation is performed considering the accuracy, robustness and computational efficiency of the model: Among them, F(θ) represents the fitness function of the co-evolution, α, β and γ are adaptive weight parameters, Accuracy(θ, Z, x n ) represents the combination of complex domain feature transformation result Z and nonlinear dynamic system feature extraction result x n The model accuracy after θ is θ, Z, x n ) represents the combination of Z and x n The robustness of the model after Efficiency(θ,Z,x n ) represents the combination of Z and x n The model calculation efficiency after ∈ is a very small positive number; S47. According to the optimization results of the improved co-evolutionary algorithm, the model parameter combination with the best performance is selected, and the tuned visual algorithm model is applied to the visual inspection of power equipment in the target field.

5. According to claim 1, a multi-scenario visual algorithm evaluation method based on an electric power artificial intelligence platform is characterized in that: The S6 specifically includes: S61. Construct a simulation platform based on a power system fault mode library, where the fault mode library includes common power equipment fault types and their corresponding fault feature data; S62. Perform preliminary testing on the optimized visual algorithm model on the simulation platform, and record the performance parameters of the model under normal working conditions; S63, selecting multiple fault types from the fault mode library, and injecting these fault types into the simulation platform in sequence according to the preset injection frequency and time interval, so as to simulate the fault scenarios encountered by the actual power equipment; S64. Real-time monitoring of the visual algorithm model's response to fault detection during the fault injection process, and recording the model's detection accuracy, false alarm rate, and missed alarm rate for each fault type: Among them, TP is the number of true positives, TN is the number of true negatives, FP is the number of false positives, and FN is the number of false negatives; θ is the model parameter, Z is the result of complex domain feature transformation, and x n is the feature extraction result of nonlinear dynamic system, α i and β i is the weight parameter, and ψ(Z i ,x n ) is the characteristic function; S65. Compare and analyze the performance changes of the visual algorithm model before and after fault injection, and use a flexible evaluation mechanism to dynamically evaluate the robustness and adaptability of the model by adjusting the frequency and intensity of fault injection; S66. Evaluate the comprehensive detection capability of the model in complex fault scenarios by injecting a combination of multiple fault types; S67. Based on the elasticity evaluation results, identify the performance bottlenecks of the model and make optimization suggestions to further improve the visual algorithm model.

6. The multi-scenario visual algorithm evaluation method based on the power artificial intelligence platform according to claim 1 is characterized in that: The S7 specifically includes: S71. Create a digital twin virtual environment based on the updated multimodal dataset and optimized visual algorithm model; S72. Generate a digital twin model of the power equipment in a virtual environment and use simulation technology to simulate the operating status of the power equipment under different operating conditions: Among them, M(t) represents the state of the digital twin model at time t, α i and β i is the weight parameter, and ψ(Z i ,x n ) are the complex domain feature transformation results Z i And the nonlinear dynamic system feature extraction result x n The characteristic function of S73, synchronizing the data in the virtual environment and the actual environment, and comparing the operating status of the virtual model and the actual equipment in real time through a data synchronization algorithm; S74. Apply the bidirectional twin simulation technology based on time series and power equipment behavior to perform synchronous simulation of power equipment under different operating conditions and verify the effectiveness of the virtual model under different conditions; S75, analyzing the simulation results, and calculating the errors between the virtual model and the actual equipment under different operating conditions through an error analysis algorithm; S76. Based on the error analysis results, optimize the digital twin model and adjust the weight parameter α i and β i : in, and is the optimized weight parameter, η is the learning rate, λ is the regularization parameter, and is the partial derivative of the error with respect to the weight parameter, and Represents a composite function For the weight parameter α i and β i The sum of the partial derivatives of , m is the number of composite functions; S77. Apply the optimized digital twin model to the operation monitoring and fault prediction of actual power equipment to verify the effectiveness and reliability of the model under different operating conditions.

Images